Planet Georezo » Séries temporelles (CESBIO)

Un Planet est un site Web dynamique qui agrÃ¨ge le plus souvent sur une seule page, le contenu de notes, d'articles ou de billets publiÃ©s sur des blogs ou sites Web afin d'accentuer leur visibilitÃ© et de faire ressortir des contenus pertinents aux multiples formats (texte, audio, vidÃ©o, Podcast). C'est un agrÃ©gateur de flux RSS. Il s'apparente Ã un portail web.
Vous pouvez lire le billet sur le blog La Minute pour plus d'informations sur les RSS !

Feeds

10000 items (0 unread) in 53 feeds

Dans la presse

Du côté des éditeurs

Toile géomatique francophone

Géomatique anglophone

Blogs GeoRezo.net

Abonnez vous à cette page (RSS) !

Séries temporelles (CESBIO)

Skip to page: 1 2 3

09 October 2020
15:18
A phase 0 study for Sentinel-HR
sur Séries temporelles (CESBIO)

=>

In 2018, Olivier Hagolle presented in this article the Sentinel-HR concept : an optical satellite in the metric spatial resolution class providing global, repetitive, systematic observations with an open access policy. By doing so, he continued the work of Gérard Dedieu who had already advocated for this concept in the previous years. Thanks to this fruitful lobbying, CNES started a phase 0 study to explore the concept.
The specificities of Sentinel-HR
A preview of what Sentinel-HR would look like : Pléiades images, acquired in Hérault, France, in 02.2019 (top) and 05.2019 (bottom), at 2m spatial resolution (2.8m resampled at 2m). This spatial resolution allows to identify hedges, tree and crop ranks, road networks and individual buildings. The constant viewing angle of Sentinel-HR would greatly facilitate change detection applications. The mission revisit frequency would allow to acquire more than a clear view per trimester on over most emerged lands (©CNES 2018-2019 - Airbus distribution - commercial use forbidden).
At first sight, there is nothing really new in the foreseen specifications of Sentinel-HR: it would provide a metric spatial resolution (between 1m and 2m), acquired in the classical spectral bands (red, green, blue, near infra-red). VHR commercial constellations provide better spatial resolutions, while Sentinel2 and its next generation provide a much richer spectral content. If Sentinel-HR aims at achieving similar geometric quality as VHR sensors and similar radiometric quality as Sentinel2, its originality lies not in its specifications, but in its acquisition plan: all emerged lands, acquired every 20 to 30 days, with constant (and close to Nadir) viewing angles. Those angles would allow to avoid geometric deregistration of above-ground objects and radiometric directional effects of non-lamberian surfaces, known to impair applications based on change detection for instance. Data would also be organized by tiles, similar to what is done for Sentinel2, so as to facilitate the building of spatial temporal cubes of data, with a free access policy and wide distribution. There is no doubt that Sentinel-HR would be a very welcome addition to Sentinel2 NG and VHR missions.
Complimentary 3D DTM generated from a single Venµs acquisition.
To complete change detection capabilities of Sentinel-HR, Digital Surface Model (DSM) acquisition capabilities are also considered. For some applications, such as monitoring the full set of glaciers in the world, DSM temporal series with an intra-annual revisit could be very useful. In order to assess the interest for such time series in this phas 0 study, we will rely on Venµs mission, which has a redundant row of detectors in the focal plane allowing to estimate a DSM for each acquisition, as presented by Amandine Rolland in this blog post.
Toward multi-sensors hybrid products
No need to recall that cloud coverage drastically impacts the actual frequency of clear pixels. From the EO Compass website, which provides stastitics on the full Sentinel2 archive, we can learn that the 30UXU tile located in Brittany, France, the average cloud cover is almost 60%, and the average time between two acquisitions with less that 10% of cloud cover is 33 days. In French Guyana, this delay increases to 45 days.

Many methods exist for the temporal resampling gap interpolation of Sentinel2 time series. It is for instance one of the steps implemented in the Iota2 processing chain, where images from different tiles are resampled on a common temporal grid, which allow to solve both the cloud coverage issue and the acquisition date heterogeneity from using different orbits together. With a foreseen revisit of 20 to 30 days, Sentinel-HR may provide only a few clear dates a year for a given location, which seems insufficient for a precise monthly interpolation of missing data for instance. However, we can try to guide this interpolation by using the more frequent, less spatially resolved corresponding Sentinel2 time series. This is an ongoing research topic here at CESBIO, and we will report our progress in further blog posts.
Aim of the phase 0 study Delineation (in red) of tree elements in Brittany, made by the Kermap company, from High Resolution optical images.
The Sentinel-HR phase 0 study has began this summer, for a duration of one year. Its main purpose is to demonstrate the usefulness of Sentinel-HR data, and the difficulty to collect such data with current sensors, through different applications identified by the mission group. The collected evidence may allow to influence the specifications of Sentinel2 NG, or convince the European agencies (Copernicus, EEA) that Sentinel-HR would be a welcome addition to Sentinel2 NG. The study will also establish a rough estimation of the cost and design. Last, the study aims at providing a prototype for hybrid products that leverage complementary missions, with methods that could be used in other contexts.
15:18
Une étude de phase 0 pour Sentinel-HR
sur Séries temporelles (CESBIO)

=>

En 2018, Olivier Hagolle vous avait parlé dans cet article du concept Sentinel-HR : une mission de résolution métrique fournissant des observations globales, répétitives, systématiques et libres d'accès. Il reprenait ainsi le flambeau de Gérard Dedieu qui portait déjà ce concept dans les années précédentes. Fruit de ces années de sensibilisation, une étude de phase 0 est en cours au CNES pour explorer ce concept et démontrer sa pertinence.
L'originalité du concept Sentinel-HR
Un aperçu de ce que serait Sentinel-HR: images Pléiades, Hérault 02.2019 (haut) et 05.2019 (bas), à 2m de résolution (2.8m ré-échantillonnés à 2m). A cette résolution, on peut distinguer les haies, les rangées d'arbres, les rangs de cultures, le réseau routier, et les bâtiments individuels. Les angles d'acquisition constants sur Sentinel-HR faciliteraient grandement les applications de détection de changement notamment. La revisite de la mission permettrait d'observer toutes les terres sans nuages tous les trimestres (©CNES 2018-2019 - Distribution Airbus - Usage commercial interdit).
De prime abord, il n'y a rien d'original dans les caractéristiques prévues pour Sentinel-HR : une résolution métrique (entre 1m et 2m), dans les bandes spectrales classiques (rouge, vert, bleu et proche infra-rouge). Certaines constellations commerciales THR font bien mieux en terme de résolution, tandis que Sentinel2 et ses successeurs fournissent une richesse spectrale bien plus importante. Si l'on vise effectivement une qualité géométrique proche de la THR et une qualité radiométrique proche de Sentinel2, la pertinence de Sentinel-HR ne se situe pas du coté des caractéristiques de l'imagerie, mais bien dans la philosophie de son plan de programmation : acquérir l'ensemble des terres émergées, tous les 20 à 30 jours, avec des angles d'acquisition constants (et proches du Nadir). Ces derniers permettraient de s'affranchir des effets géométriques de mouvement apparent du sur-sol, ou des effets radiométriques directionnels des surfaces non-lambertiennes, qui gênent certaines applications, notamment en détection de changement. Avec une mise en forme de la donnée similaire à Sentinel2, organisée en tuiles pour faciliter la constitution de séries temporelles par zone géographique, une licence libre et une diffusion large, nul doute que Sentinel-HR serait un complément très utile à l'offre Sentinel2 NG et aux missions THR.
La 3D en plus Un modèles numérique de terrain généré à partir d’une seule acquisition VENµS
Pour compléter la capacité à détecter les changements de Sentinel-HR, il est envisagé de doter la mission des capacités de restitution de Modèle Numérique de Surface (MNS). En effet, pour certaines applications, comme le suivi de l'ensemble des glaciers du globe par exemple, disposer d'une série temporelle de MNS avec une revisite intra-annuelle pourrait être précieux. Afin d'en étudier l'intérêt, nous allons au cours de l'étude de phase 0 nous appuyer sur les capacités de la mission Venµs, qui dispose d'une barrette redondante dans le plan focal permettant la restitution du relief pour chaque acquisition, comme présenté par Amandine Rolland dans cet article.

Vers des produits hybrides multi-capteurs
Nul besoin de rappeler que la nébulosité subie par les capteurs optiques diminuent fortement la fréquence d'observation. Ainsi le site EO Compass, qui fournit des statistiques de l'archive Sentinel2, nous apprend que pour la tuile 30UXU située en Bretagne, la couverture nuageuse moyenne des images Sentinel2 est de quasiment 60%, et le délai moyen entre deux acquisitions avec moins de 10% de nuages est de 33 jours. En Guyane (tuile 22NCL) ce délai monte à 45 jours.

Il existe déjà des techniques de ré-échantillonnage et d'interpolation temporelle pour les données Sentinel2 par exemple. C'est notamment l'objet d'une des étapes de la chaîne de production de carte d'occupation du sol Iota2, où l'on ré-échantillonne l'ensemble des images sur une grille temporelle commune, s'abstrayant ainsi de la nébulosité et de l'hétérogénéité des dates d'acquisition liée à l'utilisation d'orbites différentes. Avec une revisite de 20 à 30 jours, Sentinel-HR pourrait cependant ne fournir au final que quelques dates claires dans l'année pour un pixel donné, ce qui semble insuffisant pour une interpolation précise de la donnée manquante à une fréquence mensuelle par exemple. En revanche, on peut tenter de guider cette interpolation en s'appuyant sur la série temporelle Sentinel2 acquise durant la même période, qui observe ce pixel certes à moindre résolution spatiale, mais beaucoup plus fréquemment. Des travaux sont en cours au CESBIO pour explorer cette piste méthodologique.
Une étude de phase 0, pour quoi faire ? Extraction (en rouge) des éléments arborés du paysage, dont le bocage, en Bretagne par la société Kermap, à partir d'images à haute résolution optique.
La phase 0 Sentinel-HR a débuté cet été pour une durée de un an. Il s'agit avant tout de démontrer l'utilité de la donnée qui serait produite, et l'impossibilité de la collecter avec les missions actuelles, au travers des différentes applications portées par le groupe mission. Les éléments collectés permettront peut être d'influencer les caractéristiques de Sentinel2 NG, ou de convaincre les instances Européennes (Copernicus, EEA) du bénéfice qu'aurait une mission de ce type en complément à Sentinel2 NG. L'étude apportera également des éléments de dimensionnement et de coût. Enfin, l'étude vise à démontrer la possibilité de constituer des produits hybrides jouant sur la complémentarité des missions, avec des méthodologies qui pourront être ré-employées dans d'autre cadre.
05 October 2020
14:52
MAJA 4.2 est disponible et devient un logiciel libre.
sur Séries temporelles (CESBIO)
=>

MAJA est un logiciel de détection de nuage et de correction atmosphérique de qualité, développé par le CNES et le CESBIO, avec des contributions du DLR. Il est codé par CS GROUP. MAJA est utilisé pour traiter les données Sentinel-2 par le Pôle Theia, par les projets Sen2Agriet Sen4cap de l’agence spatiale Européenne, par le DLR, par le projet Neige&glace de l'agence Européenne de l'environnement, par la société KERMAP. et la plate-forme CODE-DE.

Nous souhaitions depuis longtemps faire de MAJA un logiciel libre, et nous y sommes enfin arrivés ! Après des années de discussions internes au CNES, la décision avait été prise cet hiver de libérer le code avec la version 4.2. Il nous a fallu 6 mois pour rendre cette décision effective, en partie pour terminer la validation et corriger les derniers bugs, et en partie pour mener un audit du code afin de vérifier que MAJA n'inclue pas de code avec des licences incompatibles avec la licence Apache version 2.0.

En conséquence, vous pouvez maintenant lire, télécharger, modifier et compiler le code à partir de cette adresse:
[https:]]
Vous vous rendrez compte que MAJA n'est pas un logiciel simple et qu'il s'appuie sur des caractéristiques avancées du langage C++. Cependant, si moi-même, un pythoniste amateur, j'ai pu y modifier quelques lignes, les vrais pros de l'informatique devraient y arriver.

Bien sûr, nous continuons à distribuer des exécutables MAJA pour linux (Ubuntu, Redhat),et la version 4.2 peut déjà être téléchargée depuis:
[https:]]
Le développement de MAJA a commencé il y a 12 ans, et depuis lors, la bibliothèque Orfeo Tool Box, sur laquelle MAJA s'appuie, a considérablement évolué. Il nous a donc semblé nécessaire de mettre MAJA à jour pour s'appuyer sur le mécanismes des applications OTB, et pour recoder l'orchestrateur des modules en python. La version 4 n'est donc qu'une simple mais importante mise à jour du code, avec peu d'améliorations concernant les méthodes, sauf une importante : lisez ce billet jusqu'au bout !

Voici la liste de nouveautés de la version 4.2 :
- C'est la première version libre de MAJA !
- L'orchestration a été réécrite en python, mais nous avons préservé les interface pour facilité la transition vers la version 4. Les seuls changements, légers, concernent les paramètres de traitement (GIPP). Les formats de MUSCATE sont les formats nominaux, mais les formats natifs ont été conservés provisoirement.
- Nous avons inclus un nouveau script pour aller chercher les données CAMS (Copernicus Atmosphere Monitoring service). Les données CAMS permette de contraindre le type d'aérosols dans les estimations d'épaisseurs optiques d'aérosols. Le nouveau (depuis juillet 2019) et l'ancien format de CAMS sont supportés. Nous devrions donc enfin pouvoir utiliser CAMS dans la production opérationnelle de Theia.
- Notre interface en ligne de commande pour lancer des traitements sur des séries temporelles entières, start-maja, a été inclus dans MAJA. Il a été complètement réécrit par Peter Kettig, et a gagné en clarté et modularité. Et Peter en a profité pour y ajouter de nouvelles possibilités bien pratiques :
  - Les MNT sont générés automatiquement. Il suffit de spécifier la tuile à traiter, et start_maaja préparera le MNT correspondant, s'il n'existe pas déjà.
  - Les paramètres de traitement (GIPP) et les LUTS sont aussi téléchargés automatiquement
- Nous avons implémenté un masque de neige pour VENµS (qui n'a pas de bande dans le moyen infra-rouge)
- Une dernière évolution a un fort impact positif sur la qualité de nos produits: il est possible d'améliorer la résolution des masques de nuages grâce à l'amélioration de la rapidité du module qui recherche les ombres de nuages. MAJA V4.2 peut maintenant détecter les nuages et les aérosols à une résolution de 120m au lieu de 240m avec une perte de temps de seulement 10% par rapport à la version 3.5.
Masque de nuage à 240m de résolution, exemple sur la Guyane Française Masque de nuage à 120m de résolution. Les petits nuages et leurs ombres sont bien mieux détectés

Un grand merci à Manuel Grizonnet et Pierre Lassalle, qui ont démarré la transformation de MAJA, et à Peter Kettig, qui a suivi le développement et réécrit et amélioré start-maja.Le codage a été rélisé par CS-SI, notamment par Benjamin Esquis et Julie Brossard. La participation de Jérôme Colin du CESBIO (CNRS) a été essentielle pour la validation. Pour ma part, j'ai seulement beaucoup insisté pour que l'on corrige le bug qui ralentissait les traitements à 120 m de résolution..

Merci aussi à ceux qui nous ont aidé à libérer le code de MAJA: cela nécessitait un changement culturel, et dans ce registre, le travail de pionniers de l'Orfeo ToolBox (OTB) au sein du CNES a été essentiel.
01 October 2020
18:24
MAJA 4.2 is available and open source
sur Séries temporelles (CESBIO)
=>

MAJA is the high quality cloud detection and atmospheric correction code developped by CNES and CESBIO, with contributions from DLR, and coded by CS GROUP. It is used within Pôle Theia, Sen2Agri, Sen4cap, DLR, EEA Snow&Ice, KERMAP... to process #Sentinel2 time series.

It's been a long time since we wanted to make MAJA available as an open source code, and we finally made it ! After years of internal discussions within CNES, the decision had been taken this winter to release MAJA as open source with version 4.2. But it took us 6 months from the decision to the release, half for the final validation of the results and correction of the last bugs, and half to perform an audit of the code. The audit was necessary to check that our developpment did not include code protected by other licences.

As a result, you may look, download and modify MAJA's code from here.
[https:]]
You will see the size of MAJA project: MAJA is certainly not a simple code, it uses very advanced features of C++. But anyway, as an amateur pythonist, I still have been able to make slight changes within MAJA, in a period I had some spare time. As a result, real computer professionals should still manage to contribute.

Of course, we still distribute compiled versions of MAJA for linux systems (Ubuntu, Redhat), and MAJA 4.2 has been added to the list of available versions. The executable code can be downloaded from here :
[https:]]

MAJA 's development had been started more than twelve years ago, and the Orfeo Tool Box, on which MAJA relies, has considerably evolved since then. As a result, we felt it was necessary to update MAJA to better use the framework of OTB Applications, and to replace the main C++ program which was hard to maintain, by a new orchestrator written in Python. The version 4 so far is mainly a transformation of the code with little improvement regarding the content of the algorithmic methods, but there is one which is important, so please read until the end of the list below.

Here is a list of what's new with MAJA V4.2, compared to MAJA V3 series :
- It is the first open source release of Maja !
- The orchestration has been rewritten in python, facilitating the use of Maja while keeping the legacy interfaces in place.Swapping to MAJA 4.2 should be very easy.
- Regarding the interfaces, only the processing parameters have been slightly changed in MAJA 4.2. MAJA nominal output formats are the same as those of MUSCATE, the Theia processing center. The formats that we callled "native" might be abandonned in the future versions
- A new download script for CAMS data (Copernicus Atmosphere Monitoring service) has been included
- The new and old CAMS formats are supported, the limitation after July 2019 is not active anymore. We will finally be able to swith MAJA production using CAMS in theia production center. CAMS data are used to specify the aerosol type before estimating the aerosol optical thickness.
- Our launcher to process time series, Start_maja, is now natively included within Maja. It was completely rewritten by Peter Kettig himself, and it is good to see the difference of a professional code compared to my version . Start Maja includes a lot of new features that make the use of MAJA much much easier.
  - There is an automatic DTM creation tool within Start_maja. You just need specify the tile to process and Start-MAJA fetches the data and generates the DTM files, if they do not exist yet.
  - The nominal processing parameters, and the LUTS (GIPP) are also downlaoded automatically
- We also implemented a Snow mask for VENµS (which does not have a SWIR band)
- One evolution has a large impact on the quality of the products generated by MAJA V4.2. It is the possibility to increase the resolution of cloud masks. This is possible thanks to an improvement of the speed of the module to find cloud shadows. MAJA V4.2 is able to process cloud masks and aerosol detection at 120 m resolution, with an increment of duration of about 10% compared to cloud masks at 240 m resolution in MAJA V3.x.
Cloud mask generated at 240m resolution, example in French Guyana Cloud mask generated at 120 m resolution. More small clouds and more shadows are detected.

A big "thank you" to Manuel Grizonnet and Pierre Lassalle, who started the refactoring, and Peter Kettig who handled most of the project and rewrote and improved start-maja. The coding was made by the CS-SI company, mainly Benjamin Esquis et Julie Brossard. The participation of Jerome Colin from CESBIO was essential for the validation, and I just insisted a lot, so that we solve the bug that prevented us from working at 120m resolution ... Many thanks also to those at CNES who helped us making MAJA's source available to everyone: it required a cultural change, and the pioneering work of the Orfeo ToolBox (OTB) team within CNES was essential.
28 September 2020
13:24
September snow captured by Sentinel-2
sur Séries temporelles (CESBIO)

A sunny Sunday just after an exceptional late-September snowfall... That's already a stroke of luck. But the passage of Sentinel-2 on the same day makes this weekend even more special!

The snow cover map was immediately processed and distributed by the new Copernicus Snow & Ice monitoring service powered by Cesbio's algorithms! You can explore it in your web browser here.
Sentinel-2 fractional snow cover map (white = 100% snow, black = 0% snow, grey = cloud) (Source)
Ground validation was performed on Sunday in the beautiful Vallée du Marcadau by students from ENS geosciences master program who were visiting the region as part of their field trip with Pierre René (association Moraine).

Aujourd’hui près du refuge Wallon 1866 m d’altitude : Moyennne de *26 cm de neige* mesurée par les étudiants @Geosciences_ENS @ENS_ULM @florencehabets @Meteo_Pyrenees pic.twitter.com/ii95WYqgwT

— Simon Gascoin (@sgascoin) September 27, 2020
22 September 2020
14:45
Estimation à la parcelle du rendement des grandes cultures sur la France métropolitaine avec des séries temporelles d'images Sentinel-2
sur Séries temporelles (CESBIO)

L'estimation du rendement
La connaissance du rendement agricole, que ce soit en prévision (avant la fin de la saison) ou en rétrospective (après la récolte) a beaucoup d'intérêt applicatif. On peut par exemple devenir riche en spéculant sur les marchés, et on peut prévoir des crises alimentaires (et aussi devenir riche en spéculant sur les marchés). Il existe d'autres usages de cette information comme par exemple l'étude de l'impact du changement climatique sur les cultures ou l'effet de certaines pratiques agricoles.

En France, le Service de la statistique et de la prospective (SSP) du Ministère chargé de l'agriculture produit chaque année la statistique agricole annuelle (SAA) et les données associées sont publiées sur le site Agreste. Pour ce qui concerne le rendement agricole, des estimations départementales pour chaque culture sont disponibles après chaque campagne agricole.

Ces estimations sont le résultat d'enquêtes auprès des exploitants, combinées par des experts à d'autres sources d'informations. Il s'agit d'un travail long et complexe qui pourrait bénéficier de l'observation par télédétection satellitaire.

Au CESBIO, nous avons eu l'opportunité d'avoir accès à des données de rendement issues de l'enquête TERLAB (TERres LABourables) menée par le SSP chaque année pour la production de la SAA. Nous avons utilisé ces données pour évaluer la possibilité d'estimer le rendement à l'échelle de la parcelle agricole à partir d'imagerie de télédétection.

Nous avons obtenu des résultats intéressants que nous vous présentons dans ce billet. Le rapport complet de l'étude est disponible en ligne.
L'enquête TERLAB
Les données que nous avons reçu correspondent au recueil par enquête téléphonique auprès des exploitants en 2017. Pour chaque exploitation, on dispose du rendement (en quintaux par hectare) pour chaque culture. Ce rendement n'est pas spatialisé, c'est à dire que chaque exploitant peut avoir plusieurs parcelles de la même culture avec des rendements différents. En croisant ces informations avec le Registre parcellaire graphique (RPG) de l'année correspondante, on peut associer à chaque parcelle d'une même exploitation le rendement obtenu par enquête, mais il sera différent du vrai rendement de la parcelle.

C'est avec ce type de données que nous travaillé : la géométrie de 225.367 parcelles distribuées sur 70 départements, chacune contenant l'information de type de culture ainsi que le rendement de cette culture pour l'exploitation.

Avec ces données, nous avons construit des modèles d'estimation de rendement à partir d'images Sentinel-2.
L'approche algorithmique
Nous avons choisi de travailler par apprentissage statistique et construire ainsi un modèle de régression non-linéaire qui génère une estimation de rendement par parcelle à partir d'observations satellite acquises pendant la campagne agricole.

Pour chaque parcelle agricole, nous avons construit un ensemble de descripteurs phénologiques simples à partir des séries d'images Sentinel-2 (des statistiques par trimestre de différents indices de végétation, vous trouverez les détails dans le rapport). Nous avons ensuite utilisé un algorithme d'apprentissage pour relier ces descripteurs à la valeur de rendement correspondante.

On pourrait se demander si le fait d'apprendre sur des données «fausses» (souvenez-vous que nous ne connaissons pas le rendement des parcelles, mais celui de l'ensemble de l'exploitation) permet d'apprendre un modèle fiable. En fait, si on choisit correctement l'algorithme d'apprentissage, on peut réussir à lisser un peu ce bruit dans les données. On fait aussi l'hypothèse que le rendement est très corrélé au climat, au type de sol et aux pratiques agricoles et que ces facteurs sont plutôt homogènes au sein d'une exploitation. En plus, le grand nombre d'échantillons disponibles permettent d'être robuste aux échantillons fortement aberrants qui sont minoritaires.
Les résultats
Dans cette étude, nous avons comparé deux approches. D'un côté, nous avons testé la possibilité de calibrer des modèles spatialement localisés pouvant travailler sur différentes cultures. Cette approche a montré des limites. Nous nous sommes ensuite concentrés sur un modèle mono-culture (nous avons choisi le blé tendre d'hiver, BTH pour ceux qui parlent couramment le RPG), mais applicable à l'ensemble du territoire. Nous avons donc utilisé 95.550 parcelles de blé pour apprendre (2/3 des parcelles) et tester (le 1/3 restant) le modèle de régression.

La figure ci-dessous montre les résultats de la validation. Les erreurs mesurées sont de 9.72 q/ha en RMSE. La figure montre une bonne corrélation entre les données TERLAB et les estimations issues de la donnée satellite. Des écarts importants existent pour les faibles rendements (inférieurs à 40 q/ha) pour lesquels la régression surestime. Il s'agit cependant de très peu de parcelles par rapport à l'ensemble de la validation (3.4 % pour un rendement TERLAB inférieur à 40 q/h).

Cette validation reste limitée, car, comme expliqué plus haut, nous n'avons pas accès au vrai rendement de chaque parcelle. Nous avons donc décidé de comparer à la SAA et avons pris toutes les parcelles, leurs surfaces et les rendements estimés et nous avons calculé la production totale et le rendement moyen pour chaque département.

La figure suivante montre le résultat de la comparaison des estimations satellite et les données Agreste. La colonne de gauche présente les départements pour lesquels l'enquête TERLAB fournit des données pour le BTH, et la colonne de droite contient les départements pour lesquels aucune exploitation de BTH n'a été enquêtée.

Nous observons que pour les départements TERLAB (colonne de gauche) les rendements sont proches de ceux de la SAA avec une légère sous-estimation pour les rendements élevés et une sur-estimation pour les faibles. C'est le même type de comportement observé pour le rendement à la parcelle. Les productions montrent la même sous-estimation pour les valeurs élevées, mais il ne semble pas y avoir de problème pour les faibles productions, ce qui est logique, car la pondération par les surfaces faibles fait chuter les erreurs.

La colonne de droite montre les résultats pour les départements sans donnée TERLAB pour le BTH. On constate une très bonne cohérence avec la SAA pour les productions. Il est intéressant de noter que la Moselle (57) constitue un cas atypique (production très élevée) dont l'estimation est très bonne. Le problème de la surestimation des faibles rendements est ici accentuée par le fait que beaucoup de départements non enquêtés ont de faibles rendements.
Conclusion
Cette étude a montré qu'il est possible d'estimer le rendement des grandes cultures à l'échelle de la parcelle à partir de séries d'images satellite optiques de type Sentinel-2.

Il reste encore du travail à faire, car il faudrait valider que d'autres cultures, comme le maïs ou le tournesol, peuvent être traités avec la même approche (nous avons des résultats partiaux qui vont dans le bon sens).

L'utilisation du satellite pourrait simplifier le travail d'enquête, car l'apprentissage des modèles peut être fait avec moins d'exploitations que celles utilisées actuellement. Il serait intéressant d'évaluer l'applicabilité du modèle appris sur des données d'une année aux images acquises sur l'année suivante, ce qui permettrait aussi de réduire la fréquence de l'enquête.

Enfin, l'application de l'approche au cours de saison, pourrait permettre d'avoir des prévisions de rendement, et donc de devenir riche en spéculant sur les marchés.
16 September 2020
13:21
Protégé : The Indian Water Rush
sur Séries temporelles (CESBIO)
XX^th century water reservoirs
(credits: A. Selles, S. Ferrant) XXI^st century water reservoirs
(credits: A. Selles, S. Ferrant)
The South-Indian states suffer from insufficient -yet abundant- monsoon rainfalls that do not fulfill any more the current urban and agricultural water needs.

During the last decades, fast urban population growth has disrupted existing tap-water supply systems in big cities, challenging communities with regular tap-water shortages (last big one in Chennai).
People stand in queues to fill vessels with drinking water from a water tanker in Chennai. June 2019 (AP Photo)
In the same time, Indian food production system has been boosted by irrigation. First with collective canals bringing surface water to crops in irrigated command area, and then with millions of individual bore wells pumping collective groundwater resource.

The water supply is permanently challenged by a tremendous population increase, but also by emerging food demand behaviors resulting in an increasing demand of meat or wine and beer, productions with a higher water footprint than the traditional vegetarian Indian food.

This led to a drastic drop of groundwater levels and many perennial rivers drying.

But how to further increase water supply?
That's the MBillion dollar question
The Kaleshwaram project: world's largest multi-stage lift irrigation project ...
... before the Colorado one!! India's adaptation strategy
to global changes
The case of Telangana, the youngest Indian state created in 2014 after Andhra Pradesh partition, reveals the Indian economy thirst for water. The urban area of its capital, Hyderabad, has grown from 6 to 12 million inhabitants in the last 10 years, as a result of the fast development of software industries (Google, Amazon, IBM etc.).
// push deferred map creation function window.WPLeafletMapPlugin = window.WPLeafletMapPlugin || []; window.WPLeafletMapPlugin.push(function () { var baseUrl = 'https://{s}.tile.openstreetmap.org/{z}/{x}/{y}.png'; var base = (!baseUrl && window.MQ) ? MQ.mapLayer() : L.tileLayer(baseUrl, { subdomains: 'abc', detectRetina: 0, minZoom: 0, maxZoom: 20, }); var options = L.Util.extend({}, { maxZoom: 20, minZoom: 0, layers: [base], zoomControl: 1, scrollWheelZoom: 0, doubleClickZoom: 0, attributionControl: false }, {"fitBounds":false,"maxBounds":[["17.9"," 78.8"],["18.9"," 79.2"]],"attribution":"Leaflet; \\r\\n\u00a9 OpenStreetMap contributeurs"}, {}); window.WPLeafletMapPlugin.createMap(options) .setView([44.67,-63.61],6); }); window.WPLeafletMapPlugin = window.WPLeafletMapPlugin || []; window.WPLeafletMapPlugin.push(function () { var previous_map = window.WPLeafletMapPlugin.getCurrentMap(), fitbounds = 0, is_image = previous_map.is_image_map, lat = 18.2, lng = 79, radius = 90000, group = window.WPLeafletMapPlugin.getCurrentGroup(); // update lat lng to previous map's center if (!lat && !lng && !is_image) { var map_center = previous_map.getCenter(); lat = map_center.lat; lng = map_center.lng; } var circle = L.circle([lat, lng], {radius: radius}); circle.setStyle({"radius":90000}); circle.addTo( group ); window.WPLeafletMapPlugin.circles.push( circle ); if (fitbounds) { // zoom the map to the polyline previous_map.fitBounds( circle.getBounds() ); } });
Located in the North of Telangana state, this mega project, initiated by the Telangana government, is a regional pride. It aims pumping the remaining Godaveri River base flow released from upstream dams, before entering the downstream Andhra Pradesh state. It then lifts water 200 meters higher with the world's most powerful pumping system to a network of big reservoirs linked with huge canals.

I let Kalvakuntla Taraka Rama Rao (alias KTR), the son of Kalvakuntla Chandrashekar Rao, the current Chief Minister of Telangana, speak about this enchanting achievement.

Kaleshwaram Lift Irrigation Project is world's largest lift irrigation project with 3 Barrages, 1531 km of Gravity Canals, 203 km of Tunnels, 20 Lifts, 19 Pump houses and 20 Reservoirs with total capacity of 147 TMC water irrigating 37 lakh acres#KaleshwaramProject pic.twitter.com/z5Tt08HQF4
— KTR (@KTRTRS) June 20, 2019

The figure below shows the regional footprint of the project: water pumped in the downstream north, lifted through a network of tunnels and canals to the upstream southern region in big reservoirs, up to Hyderabad water supply system.
Telangana relief and locations of pumps downstream north) and big reservoirs upstream south toward Hyderabad city area (170 to 400 m Above the sea level). Credit: A. Selles
The total cost of the project is estimated at 80 000 Crores Indian Rupees (9.7
Billions euros, funded by Telangana only). This is the costliest irrigation project undertaken by any state in India so far.

This mega lift program on Godaveri river could fuel the existing water">[https:] war with the downstream Andhra Pradesh state (created after the partition in 2014) that regularly breaks out, caused by a water quota release (or not) from the southern Krishna river dams.
Monitoring Adaptation from space
I have used two collections of Sentinel-1 images using Google Earth Engine, in September 2019 (blue) and 2020 (red) to extract surface water. Some of new big reservoirs appear in red and have been flooded during the last monsoon (July to September 2020).

The construction of those gigantic structures along the diagram (figure below) has been done in a short time as you can see with Sentinel-2 satellites. I have spotted 4 reservoirs among the 20 that are planned, the last monsoon period has filled up 3 of them. I have built the animated GIF (click on them to enlarge it) with the friendly smartphone app SnapPlanet.
Anicut reservoir Konda Pochamma reservoir Kaleshwaram reservoirs construction phase as seen with Sentinel-2
January 2018 to August 2020
(Click to enlarge figures) Anantagiri reservoir Srikomaravelly reservoir
Several villages - which represent 1000 houses for both Konda Pochamma and Srikomaravelly reservoirs only - have been submerged.

Gajwel town area, where Konda Pochamma has been constructed, was a study site of the ANR SHIVA research project (from 2009 to 2013) - Socio-economic Assessment of the rural Vulnerability of water users under stressors of global changes in the Hard rock area of South India - in the framework of which I started working on the Indian Agro-hydrological context in 2010.
At this time, we, scientists, could not imagine such an adaptation strategy!

We have been visiting the Konda Pochamma construction site in October 2018 during a land use survey field trip and took those pictures of the structure (located here [https:]] ).
Granite block on the inside side of the Konda Pochamma Reservoir
(credits: A. Selles) Details of the internal structure of circled ditches of the Konda Pochamma reservoir (credits: A. Selles) The bottom of the reservoir has been removed with giant trucks from a mining company to build the circled wall. (Credits: A. Selles)
Those pictures illustrate a simple construction strategy entitled (by myself) "the sand box strategy": excavate the soil and saprolite in the middle to raise surrounding walls, with a layer of clayey soils (Black soils found in this region) in the middle of the wall for waterproofing. The walls are consolidated with granite blocks on their side.

Nothing is made at the bottom to prevent infiltration of the Godaveri water into the underlying aquifer. An expected result would be the raise of groundwater level resulting from the recharge below those big reservoirs.

In this area where depleting groundwater levels have been studied by the Indo-French Center for Groundwater Research since 1999, this infrastructure project raises exciting scientific questions!
However, where will Indian fast growing economy look for water in the next decades ?
Well, let us be more innovative in the options we could explore, this time...

On Mars?
[https:]]
09 September 2020
10:30
Orpaillage sur le fleuve frontalier entre la Guyane et le Suriname : télédétection des dommages causés par les barges aux rives du Maroni
sur Séries temporelles (CESBIO)

Le Suriname, comme l’ensemble du plateau des Guyanes, est touché depuis plus de trente ans par le fléau de l’orpaillage qui entraîne la destruction du couvert forestier, la pollution des cours d’eau, en impactant fortement l’environnement et les population locales. L’orpaillage y est dit primaire quand l’or est directement extrait de la roche-mère, alors qu’il est alluvionnaire quand il est exploité dans le lit des cours d’eau.

Le second type est le plus répandu au Suriname, où il prend une forme particulière dans les fleuves et rivières de taille importante. Des barges, véritables mini-usines flottantes, draguent les sédiments du fond des cours d’eau qui passent ensuite sur une série de rampes de lavage pour en extraire un concentré aurifère. L’or est finalement extrait par amalgame avec le mercure, hautement toxique (environ 1,5kg par kg d’or). Outre cette pollution, le procédé est particulièrement dévastateur pour les cours d’eau dans lesquels il rejette de grandes quantités d’eau boueuse qui nuisent à la flore et la faune aquatique. Les populations locales sont également directement touchées par cette pollution, le fleuve constituant leur bassin de vie, où elles s’y nourrissent et s’y lavent.

Barge sur le Maroni. Source : WWF Guyane

Ces barges, interdites en Guyane mais pas au Suriname, font régulièrement polémique sur le Maroni, fleuve transfrontalier entre les deux territoires. Depuis début 2019, on assiste à un nouveau développement de cette pratique qui s’attaque désormais aux rives surinamaises. Celles-ci normalement recouvertes de forêt, sont défrichées et la couche supérieure du sol est arasée à la pelle mécanique. Les barges entrent ensuite en action depuis le fleuve en creusant la terre des berges à la recherche d’or. En plus des boues toujours rejetées dans le fleuve en grandes quantités, cette pratique entraine le déboisement et l’affaissement des rives et les rend vulnérables à l’érosion dans une région où la pluviométrie est très importante. C’est ainsi le lit même du cours d’eau qui s’en trouve fortement modifié par endroits.

Images Sentinel-2 de berges intactes le 23/08/2018 (à gauche) et défrichées par les barges le 01/12/2019 (à droite)

Cette pratique est visible en quasi temps-réel grâce à la méthode de détection de la déforestation de Bouvet et al. 2018 et décrite ici. Cette méthode, basée sur les images Sentinel-1, est particulièrement pertinente dans cette région où la couverture nuageuse importante limite les systèmes de surveillance optique (voir l’évaluation ici).

Détection des berges défrichées à l'aide de Sentinel-1. La couleur correspond à la date de déforestation. Cliquer sur l'image pour agrandir et voir l'animation.

Comme on peut le voir sur cette animation, plusieurs chantiers de ce type se sont développés en 2019 entre Maripasoula et Papaïchton (environ 25 km de fleuve). Les barges attaquent les rives et creusent de véritables chenaux sur plusieurs dizaines de mètres. Sur cette seule zone, on estime une perte de plus de 36.7 ha de forêt seulement en 2019 !

Ces pratiques, régulièrement dénoncées par les autorités françaises et les organisations locales, dont le WWF Guyane, devraient finalement disparaître à terme. Le nouveau gouvernement surinamais s'est en effet récemment engagé dans un accord de coopération avec la France, à ne pas délivrer de nouvelles licences d'exploitations pour ces usines flottantes et ne plus renouveler celles arrivant à expiration. Un pas en avant pour le fleuve Maroni et ses habitants si cet accord est effectivement respecté. Le suivi précis que l'on peut faire grâce à Sentinel-1 permettra de voir si ces annonces sont suivies d'effets.
17 August 2020
11:05
[MUSCATE News] Theia production is stalled
sur Séries temporelles (CESBIO)

Update (18/08/2020): all the issues have been corrected, MUSCATE production and distribution have resumed.

Update (17/08/2020): production has resumed, but it will take us a couple of days to catch the backlog up. The download site is still offline, but this should be solved quickly.

Following the failure of an equipement which hosts the database of MUSCATE processing center, the production of Venµs and Sentinel-2 level 2 and 3 data is stalled since the 11th of August. The technical teams are working to solve the issue, and we hope to resume production this afternoon.
20 July 2020
21:43
Pic Saillant : la balade hydrologique du dimanche
sur Séries temporelles (CESBIO)

En partant de Boutx (780 m) on peut monter au Pic Saillant (1756 m) (voir le topo) et admirer un magnifique panorama hydrologique sur le haut bassin de la Garonne et de son affluent la Pique. Par temps clair, on peut deviner les sommets du Luchonnais encore enneigés au mois de juillet qui marquent la limite du bassin versant comme le Pic de Maupas (3109 m). L'Aneto et le glacier de la Maladeta sont visibles aussi un peu plus à l'est.

Panorama depuis le Pic Saillant (en regardant vers le sud)

Au niveau du Maupas, la limite du bassin de la Garonne correspond à la frontière entre la France et l'Espagne. Mais ce n'est pas le cas de tout le bassin, car la Garonne naît dans le Val d'Aran en Espagne (on sait que 23% des frontières du globe sont des rivières, mais quel est le pourcentage des frontières qui sont aussi des lignes de partage des eaux ?).

Vue satellite du bassin versant de surface de la Garonne à Chaum

A Chaum, la Garonne est jaugée depuis 1992 par la DREAL Occitanie, ce qui permet d'analyser son régime hydrologique. A cet endroit, la superficie du bassin de la Garonne est 1027 kilomètres carrés. Le débit annuel total mesuré ramené en lame d'eau a varié entre 671 mm (2005 - 06) et 1297 mm (2002 - 03) ce qui témoigne des quantités importantes de précipitations reçues par ce bassin de montagne. Le maximum mensuel est atteint à la fin du printemps et la crue annuelle a eu lieu pratiquement toujours entre avril et juin, ce qui est la signature d'un régime hydrologique nival. En d'autres termes, c'est la fonte de la neige qui dicte le tempo des débits, comme pour la plupart des rivières dans les Pyrénées.

Caractérisation hydrologique de la Garonne à Chaum (1027 km2). Débits mensuels (boites à moustaches) et VCX 10 : débits maximaux sur 10 jours consécutifs (1993 - 2019). La série de Chaum a été complétée entre 2009 et 2012 avec celle de la station de Fronsac située quelques centaines de mètres en aval (superficie du bassin versant : 1037 km2).

Bertrand Cluzet me rappelle l'existence du Forau d'Aigualluts, un gouffre karstique situé en dehors du bassin hydrographique de la Garonne, dans lequel s'écouleraient les eaux du glacier de l'Aneto pour rejoindre la Garonne au niveau du Guelh de Joèu... Cette circulation souterraine aurait été démontrée par le spéléologue Norbert Casteret en 1931 par un essai de traçage. Voilà qui ajoute encore un peu de sel à cette balade hydro(géo)logique !

L'Aneto est-il la source de la Garonne ?
17 July 2020
15:23
Sortie du bulletin Theia n°13
sur Séries temporelles (CESBIO)

Le 13e bulletin Theia est sorti, un grand merci à Isabelle Biagiotti !
08 July 2020

18:39

[Directional effects] how far can we extend Sentinel-2 field of view ? Conclusions

sur Séries temporelles (CESBIO)

This is the final post of a series of studies we did to find out how far we can extend the field of view of next generation of Sentinel-2 satellites. I hope I did not lose most readers in the length of the previous posts. For the readers who land on this page, which will spoil all suspense of the series, the blog thread starts here.

We tried to answer the following questions in our work for ESA and Airbus:

Until which zenith viewing angle is it possible to correct directional effects with a directional model independent from the land cover?
Until which zenith viewing angle is it possible to neglect the effect of orientation angles?
Until which angle can we avoid hot-spot ?
If the field of view is extended, should we keep the constant view angles that we currently have in Sentinel-2

0. Performance objective

It is very difficult to define a performance requirement for a mission such as Sentinel-2, with infinite applications, as each of them would yield different needs. For this study, we adopted a pragmatic rationale which is that the uncertainty brought by directional effects after a simple correction should be negligible when compared to other sources of errors, such as atmospheric correction, cloud detection, absolute calibration, noise from the instrument...

For instance, absolute calibration requirement is 5%, 3 sigma, while the additive noise of atmospheric correction is ~ 0.03 in reflectance unit, 3 sigma.

As a result, we defined a goal for the unknown part of directional effects $\delta$\delta to be lower than half the above mentionned errors :

\(\delta \delta

where $\rho$\rho is the surface reflectance

1. Until which zenith viewing angle is it possible to correct directional effects with a directional model independent from the land cover

The study reported in this post shows that a directional correction with a simple model agnostic of the land cover would already leave errors of +/- 5% at the edges of the current field of view of Sentinel-2 (12° expressed in view zenith angle). This is already not negligible, it is already greater than the requirement we defined above. However, as readers of numerous papers in the literature based on Sentinel-2 data, we have witnessed that users tend to use both swaths in the overlap regions of Sentinel-2 orbits, despite the differences in viewing angles. It means that a lot of users give priority to the revisit rather than to the accuracy of their retrieval.

The study also showed that the deviation of directional effects between various types vegetation covers is almost linear in the range of the studied view zenith angles [+/- 24°]. This means that extending the field of view to 24 degrees would give variations of +/- 10% around the mean directional model. It would prevent accurate uses of time series without an accurate directional correction. And the amount of data to perform an accurate correction is probably too high to allow to process it at country scale (an accurate correction requires to know the type of plant, the surface reflectance of the ground, the development stage of the plant, the architecture of the canopy...). As a result, our recommendation is to only extend a little the field of view, until 15 or 16°, which will increase the uncertainty by 25 % at the very end of the field of view.

2. Until which zenith viewing angle is it possible to neglect the effect of orientation angles?

As shown in this post, the answer is not straightforward. We have shown that with the current field of view of Sentinel-2, that reaches 12° in VZA, the row orientation effects can already be much higher than the noise due to the instrument or to the early stages of processing. However, we noticed that despite this effect, users find that Sentinel-2 fits well to their needs. We decided to proceed further and to analyze the dependency of directional effects with zenith viewing angle, after discarding the row orientation roughly aligned to the sun azimuth., which accoun to 15% of the observations, assuming a random orientation of crops globally. Of course, this decision is questionable.

Early maize	Medium maize
Early wheat	Medium wheat
Vineyard	Summary of the directional effect variations (3 sigma) due to orientation, as a function of zenith view angle (colour of the plot), and solar zenith angle. The horizontal blue line corresponds to the average value of the performance goal $ \delta $ \delta . Values are expressed in reflectances $\times 100$\times 100.

In summary, our performance objective delta is not met :

when the rows are aligned with the sun direction (with a difference up to 20°).
for crops with a large spacing between rows such as vineyards
in the SWIR band for Vineyard and crops at early stage
above 16° of view zenith angle in the other cases, which represents most of the cases, after all

So this study has two conclusions

directional effects due to orientation can be quite large, and should be taken into account already with the current field of view of Sentinel-2.
increasing the filed of view does not degrade the situation until 16° of maximum View zenith angle.

3.Until which angle can we avoid hot-spot ?

On this aerial image taken from an helicopter, the shadow of the helicopter is visible just under the T of "wants". It can be noticed that the image is brighter around the helicopter, and that the brightness decreases as the angle to the helicopter shadow increases. This phenomenon is called hot-spot. It is due to the fact that in that direction, the observer only sees sunlit faces of objects. Over developped vegetation, the phenomenon is even more accute, but this image is more fun

With the current configuration of Sentinel-2,a study from D.Roy's team showed that the exact backscattering direction (hot-spot) does not enter in the field of view of Sentinel-2 [1]. At the time of overpass of Sentinel-2, the minimum sun zenith angle is 18°, and this does only happen on a few dates. With a 20° field of view, the hot spot would start to enter more regularly in the field of view if it reached 20° of zenith viewing angle. With a margin of 4 and 5 degrees, this directs us also to a maximum viewing angle of 15° or 16°.

4. Should we keep the constant view angles that we currently have in Sentinel-2 ?

All the above studies point to a maximum view zenith angle of ~16°, that would still allow accurate retrievals without requiring a large complexity of the processing. However, this limit is only a magnitude, it is not based on a strict comparison to a performance requirement, but more on knowledge on how Sentinel-2 data are used in its large community. Such a result is therefore highly questionable, and we would be happy to recive feedback in the comments section of this post.

Increasing the field of view has some advantages

better revisit frequency with the same number of satellites,
possibility to estimate the BRDF from the data themselves thanks to the enhanced revisit

Should Copernicus and ESA decide to increase the field of view, the next question is whether we should keep or not the current constraint of constant viewing angles, except in the overlap regions. The main advantage is that successive acquisitions can be compared easily on the short term. However, it does not mean that the acquisitions are not affected by directional effects, but just that directional effects result in a bias that depends on the view zenith angle (and even some variations due to the slow evolution of sun ables along the season.

But there are drawbacks, suppose that your application requires a viewing angle below 12°, well, some place sinthe world will be always observed with 15°, and it will be difficult for you to get good results there

Trishna's orbit has a a repeat cycle of 8 days, with a subcycle of 3 days. The graphic shows the amount or revisits per 8 days. 4 revisits in 8 days mean that the site is seen every second day on average. So trishna has a revisit of 3 days on the equator, but 2 days or better in Europe.

. Relaxing the constant angle constraint will enable the user to select only the suitable angles, at the expense of some revisit.

Moreover, with the current constant viewing angles, here in Toulouse, we have only one revisit every 5th days, but in Carcassonne, 100 km South East, there are 2 revisits every 5 days, that's unfair. With a longer orbital cycle, the revisits would be better distributed, as shown on the Trishna example on the right.

Although I have always been in favor of constant view angles in the past, my current feeling is that with a greater field of view, it is interesting to shift to longer orbital cycles with subcycles allowing revisit under different viewing angles. But here again, feel free to comment if you disagree, or even if you agree ;).

References

[1] Li, Zhongbin, Hankui K. Zhang, and David P. Roy. "Investigation of Sentinel-2 bidirectional reflectance hot-spot sensing conditions." IEEE Transactions on Geoscience and Remote Sensing 57.6 (2018): 3591-3598.

07 July 2020
17:04
[Directional effects] What field of view for the next generation of Sentinel-2 ?
sur Séries temporelles (CESBIO)
The European Union is planning a new generation of the Sentinel-2 mission in 2030 or soon after. The main request from users is to improve the revisit frequency from 5 days to 2 to 3 days. There are two ways to improve the revisit:

Artist view of Sentinel-2 (courtesy ESA)
- increase the number of satellites
- increase the field of view
According to ESA and Airbus, the second solution is cheaper, and as a result, it is worth trying to use a field of view as large as possible. But the field of view is limited by directional effects. The observation of a given pixel under different viewing angles yields different reflectance, and therefore adds unwanted noise (or at least variations) to time series. And of course, the greater the angle variations, the greater the reflectance differences.

It is possible to correct for directional effects, but the residuals after correction will probably increase with the viewing angles.

In a series of posts, we will provide the results of a study we did at CESBIO to address this issue, in the framework of an ESA contract led by Airbus. For that, we used the DART software, which is an accurate and fast 3D radiative transfer simulator able to simulate complex scenes with inhomogeneous vegetation, that correspond to realistic landscapes, such as maize or wheat fields, vineyards or forests.

In a series of posts, we will show the results of our study in several parts :

DART very high resolution simulations for 5 stages of wheat development
Although I was responsible for the study design and conclusions, most of the work was done by Nicolas Lauret (CESBIO), who is DART lead developer, and with the constant help, advice and verification by Jean-Philippe Gastellu-Etchegorry, the father and godfather of DART. Gérard Dedieu also gave us valuable advice. The whole Sentinel-2 NG study was supervised by Laure Brooker at Airbus, and the project officer at ESA was Armin Loescher.
15:55
[Directional effects] simple BRDF normalization
sur Séries temporelles (CESBIO)

Introduction
To perform a correction of directional effect, directional models are required. A large number of directional models have been developed, such as the ones of Roujean [1] or Ross-Li [2], that try to reproduce the directional variations as a function of viewing angles and solar angles, with a rather good accuracy for most types of surfaces. Here is how they look like :

$ \rho= \rho_0 (1 + K_1. F_1(angles), + K_2. F_2 (angles))$ \rho= \rho_0 (1 + K_1. F_1(angles), + K_2. F_2 (angles))

where $ \rho$ \rho is the reflectance for the actual viewing and solar angles, $ \rho_0 $ \rho_0 is the reflectance for a given angular condition chosen to standardise the data (for instance viewing at nadir and solar angle at 45 degrees), F1 and F2 are the directional functions that depend on the angles, and $ K_1 $ K_1 and $ K_2 $ K_2 are the coefficients of the directional model, that depend on the type of surface of the observed pixel.

For wide field of view sensors, with a daily revisit, it is possible to estimate the model parameters from the satellite acquisition themselves using a period of 2 to 4 weeks, and assuming that the BRDF does not change fast [3], [4]. However, given its lower revisit and its narrower filed of view, such a method is not well adapted to Sentinel-2. The directional effects are smaller, while the time to collect 3 to 4 cloud free samples is generally greater than a month. The risk is high to confuse time variation with directional variations.
Common solution for Sentinel-2
Fortunately, in the case of S2, the angle differences are low, no more than +/- 12 degrees. It is therefore possible to use constant values of the model parameters to perform the corrections. This approach has been used in [5] and [6]. In [5], we used a few acquisitions from SPOT4 (Take5) experiment obtained from two different angles to obtain constant coefficients. Roy et al did a more comprehensive study and compared Sentinel-2 observations on overlapping regions above whole countries to estimate the model parameters.

Comparison of reflectances on each side of Sentinel-2 swath, before correction (top), for red (left) and NIR (right) band, and after correction (bottom), using the Roy et al model [6]. (figure from the original article by Roy et al)We are using the Roy et al model to normalize directional effects in our surface reflectance syntheses obtain with WASP software. The swath edges are almost invisible in the syntheses obtained in summer, which shows a good success of the Roy model.

However, the Roy model does not perform as well in winter, when the sun elevation is much lower, and the edges of the swaths become visible, as shown in next figure, in which it is possible to see the limits of the swaths.

France monthly synthesis for March 2019. The Swath edges are starting to be noticeable.

It is not fully a surprise that a model with only three parameters is not fully able to account for the complexity of directional effects. However, reading again the paper from Roy et al, I noticed that it had been built on data from South Africa, for which, even in winter, the sun is always quite high on the horizon. It means a better fit of the data based on a better sample of reflectances is possible. However, it will still be a very (too) simple model.

Verification with DART model
Using the simulations performed with the DART 3D radiative transfer model, detailed on this post, we compared the bidirectional variations, normalized at Nadir, of different types of vegetation covers for 4 spectral bands : green, red, NIR and SWIR (1.6 µm). The figures below show that while the Roy et al model agrees with DART simulations for solar zenith angles between 15 and 45 degrees, they start to disagree at 55 degrees, and the Roy et al model is completely wrong, as suggested by what we observed on the L3A syntheses in winter with the Roy correction.

Green band Red band

NIR band
SWIR Band

Comparison for diverse geometrical configurations of BRDF variation with regard to nadir, for DART simulations of 4 different vegetation covers, and also for a turbid vegetation, i.e a layer of uniform smashed vegetation on top of a uniform ground. The absciss is the view zenith angle, with positive angles towards backscatering (or towards the East for a morning observation satellite, and negative values towards forward scattering. The maximum viewing angle, 24° corresponds to twice the field of view of Sentinel-2.

So as a conclusion, for thef ield of view of Sentinel-2 (12°), and for sun zenith angles greater than 50°, the Roy et al model can correct fairly well directional effects for the field of view of Sentinel-2. We can also notice that the different types of vegetation already have differences of +/- 5% at the edge of the swath, which might be insufficient for some applications. The simulations also show that the scattering increases almost linearly with view angles. Finally, it is also worth noticing that the turbid simulations with a LAI of 4 is not far from an average of all the vegetations covers. As a result, it should be possible to replace the Roy model by a LUT obtained from a turbid simulation.

As said above, a new version of Roy model could be obtained by fitting it over data from higher latitudes, in order to better cover the range of sun zenith angles. But it is a huge amount of data to process ...
References [1] J.-L. Roujean, M. Leroy, and P.-Y. Deschamps, “A bidirectional reflectance model of the Earth’s surface for the correction of remote sensing data,” J. Geophys. Res., vol. 97, no. D18, pp. 20455–20468, Dec. 1992, doi: 10.1029/92JD01411. [2] W. Wanner, X. Li, and A. H. Strahler, “On the derivation of kernels for kernel-driven models of bidirectional reflectance,” J. Geophys. Res., vol.100, pp. 21 077–21 089, 1995. [3] C. B. Schaaf et al., “First operational BRDF, albedo nadir reflectance products from MODIS,” Remote Sensing of Environment, vol. 83, no. 1–2, pp. 135–148, Nov. 2002, doi: 10.1016/S0034-4257(02)00091-3. [4] O. Hagolle, A. Lobo, P. Maisongrande, F. Cabot, B. Duchemin, and A. De Pereyra, “Quality assessment and improvement of temporally composited products of remotely sensed imagery by combination of VEGETATION 1 and 2 images,” Remote Sensing of Environment, vol. 94, no. 2, pp. 172–186, Jan. 2005, doi: 10.1016/j.rse.2004.09.008. [5] Claverie, M., Vermote, E., Franch, B., He, T., Hagolle, O., Kadiri, M., Masek, J., 2015. Evaluation of Medium Spatial Resolution BRDF-Adjustment Techniques Using Multi-Angular SPOT4 (Take5) Acquisitions. Remote Sensing 7, 12057–12075. [https:]] [6] Roy, D.P., Li, J., Zhang, H.K., Yan, L., Huang, H., Li, Z., 2017. Examination of Sentinel-2A multi-spectral instrument (MSI) reflectance anisotropy and the suitability of a general method to normalize MSI reflectance to nadir BRDF adjusted reflectance. Remote Sensing of Environment 199, 25–38. [https:]] [7] Gastellu-Etchegorry J.P., Lauret N., Yin T., Landier L., Kallel A., Malenovský Z., Al Bitar A., Aval J., Benhmida S., Qi J., Medjdoub G., Guilleux J., Chavanon E., Cook B., Morton D., Nektarios N., Mitraka Z. DART: recent advances in remote sensing data modeling with atmosphere, polarization, and chlorophyll fluorescence. IEEE JSTARS, 10(6), 2640-49. 2017.

15:55

[Directional effects] Effects of crop orientation on surface reflectance

sur Séries temporelles (CESBIO)

Introduction

In the framework of our research for ESA concerning the maximum field of view of next generation of Sentinel-2 satellite, we searched for a way to estimate directional effects, without using a particular directional model to correct for these effects. The current directional models are imperfect, and we can assume that progress will be made in the coming years. We did not want to get a result that would depend on the directional model correction that we use in the study.

We tested two ideas: until which maximum viewing, can we correct BDRF effects

with the Roy et al model, which assumes all types of bare soil or vegetation covers have similar directional effects
without knowing the type of vegetation cover which is observed, but depending on no directional model
knowing the type of crop, and even its development stage, but with no knowledge on the orientation of the crop

The items 1 and 2 have been studied in this article, while item 3 is the subject of this post. Well, as you will see in the results, this criterion did not deliver the results we expected, but this is what is interesting.

Simulations

For this, using the DART 3D radiative transfer model[1], Nicolas Lauret simulated 4 types of vegetation covers, a forest, corn, wheat and a vineyard. Corn and wheat were simulated at three development stages, and different ground reflectances were used. The number of simulated reflectances almost reached 1 million !

	Orange Grove	Wheat	Maize	Grapevine
Growth stage	1	3	3	1
Grounds	1	2	2	3
Sun zenith	7	7	7	7
Scene Orientations	18	18	18	18
Nb of simulations	126	756	756	378
View zenith	25	25	25	25
View azimuth	18	18	18	18
Nb of simulated reflectances	54558	327348	327348	163674

Here are a few view of the 3D mock ups we used as input of DART images. These images are then replicated infinitely on each side.

Orange grove	Vineyard
Wheat (early stage)	Maize (early stage)

The orange grove has no rows, it will serve as a reference to compare with the other cover crops.

Results

Here are some of the results we obtained with DART for all these cases. We only show here the results for the green band of Sentinel-2, but the results in the red band are similar, and those of the NIR and SWIR bands show even greater variations.

Orange grove

Vineyard

Wheat (early stage)

Wheat (medium stage)

Maize (early stage)

Maize (medium stage)

All these figures show the variation of surface reflectance in the green band, as a function of the orientation of the rows (the color of the plots). The X axis is the zenith viewing angle, with positive values in the backscattering direction and negative values in the forward direction. Reflectances are expressed in %. The dashed red line corresponds to 3 times the standard deviation of the reflectance variation as a function of orientation for a given viewing angle. The orientation of 0° corresponds to the case when the sun direction and the rows are aligned.

What stroke us at first is the fact that the orientation of rows has already a large effect even at nadir (View zenith angle=0). The orientations close to the sun direction (from -20° to +20°) often provide results which differ from the other orientations. This is explained by the next figure, which shows the position of shadows for a Maize field as a function of orientation (the Sun direction is to the North in these graphs.

Projection of Maize shadows on the ground as a function of orientation. The ground, which is brighter than the leaves in the green, red and SWIR bands, is visble from Nadir for relative angles to the sun azimuth between -20° and 20°

This explanation sounds logical when we are aware of this effect, but I have not seen it much mentioned in the literature (although a few papers have already addressed the effects of rows orientation ([2] and [3] ). These effects are quite large, sometimes greater than 10% on reflectance in relative, and have also effects on the NDVI.

Effects of orientation on NDVI for a vineyard

Effects of row orientation on Maize at medium stage (not the worst case)

For the vineyard at least, but also for crops at early stages, users should be aware that they need to know the orientation of rows before trying to invert biophysical variables such as leaf area index. This result is consistent with what was observed on Venµs data over vineyards.

References

[1] Gastellu-Etchegorry J.P., Lauret N., Yin T., Landier L., Kallel A., Malenovský Z., Al Bitar A., Aval J., Benhmida S., Qi J., Medjdoub G., Guilleux J., Chavanon E., Cook B., Morton D., Nektarios N., Mitraka Z. DART: recent advances in remote sensing data modeling with atmosphere, polarization, and chlorophyll fluorescence. IEEE JSTARS, 10(6), 2640-49. 2017.

[2] Meggio, Franco & Zarco-Tejada, Pablo & Miller, John & Martín, P. & González, M.R. & Berjón, Alberto. (2008). Row Orientation and Viewing Geometry Effects on Row-structured Crops for Chlorophyll Content Estimation. Canadian Journal of Remote Sensing. 34. 220-234. 10.5589/m08-023.

[3] Kuester, T.; Spengler, D. Structural and Spectral Analysis of Cereal Canopy Reflectance and Reflectance Anisotropy. Remote Sens. 2018, 10, 1767.

26 June 2020
16:26
Impact du COVID-19 sur la pelouse du Centre Spatial de Toulouse
sur Séries temporelles (CESBIO)

Le CNES a publié sur LinkedIn ces jolies photos du Centre Spatial de Toulouse après le confinement.

Grâce aux satellites d'observation de la Terre, les ingénieurs du CNES sont capables d'étudier la végétation aux quatre coins du globe. Mais, les cordonniers sont souvent les plus mal chaussés... Pourraient-ils étudier leur propre pelouse par satellite ?

Pour le savoir j'ai extrait le NDVI des image multispectrales du satellite Sentinel-2. Le NDVI est un indice qui reflète la vigueur et à la quantité de végétation. Plus le NDVI est élevé, plus la végétation est active et abondante.

Pour cela j'ai sélectionné la grande pelouse à proximité du poste de garde de l'entrée nord.
Zone d'étude : la pelouse devant le poste d'entrée nord du CNES
Voici la série temporelle du NDVI sur cette pelouse depuis le lancement de Sentinel-2. Les faibles valeurs indiquent une contamination par les nuages. En suivant la courbe qui relie les points du haut on peut noter une variation saisonnière avec un maximum qui a lieu à la fin du printemps de chaque année, ce qui reflète bien le cycle de la végétation en région toulousaine. Enfin on remarque un pic à la fin de la période, avec plusieurs dates ayant un NDVI supérieur à 0,6 entre le début avril et fin mai 2020, ce qui n'avait jamais été observé auparavant dans cette chronique.
Série temporelle du NDVI Sentinel-2 L1C (moyenne dans la zone d'intérêt) du 06 juillet 2015 au 24 juin 2020
Il semble donc que l'on puisse voir l'effet du confinement sur l'état de la végétation dans le centre spatial de Toulouse ! Pour gagner du temps j'ai tracé ce graphique à partir des données Sentinel-2 L1C non-corrigées des effets de l'atmosphère. Pour bien faire il faudrait utiliser les données L2A produites par Theia ce qui permettrait de masquer les nuages et d'avoir un NDVI moins bruité. Enfin, il faudrait s'assurer aussi que les conditions météo particulièrement favorables du printemps 2020 n'ont pas contribué à ce regain de végétation, par exemple en regardant la tendance du NDVI sur des prairies naturelles en région toulousaine !

Au fait, ce n'est pas la première fois qu'on voit les effets du COVID-19 depuis l'espace dans ce blog..
23 June 2020
3:48
Debris flow in the Ishkoman Valley
sur Séries temporelles (CESBIO)

In July 2018, the Pamir Times has reported that a lake formed due to a debris flow in the Ishkoman Valley, Pakistan. This event caused the destruction of dozens of housed in Badswat and Bilhanz villages.

The flood debris running down the slope hit the mountain on the opposite side, and formed a barrier, blocking the gorge through which the Kurumbar River has been flowing for centuries.

Aerial photos show the fresh debris flow deposit and the flooded area (more pictures are available here).
Source: Pamir Times
The development of the lake can be clearly seen from Sentinel-2 imagery.

The debris flow originated from the Badswat glacier valley after the snow melt season.

Interestingly a close-up near Blackswat glacier tongue shows that the glacier was undergoing a rapid movement, maybe a surge. It also shows that the debris flow came from the detachment of a proglacial terrain in an adjacent valley (and not from a glacier lake outburst flood near Blackswat glacier as written in local newspapers).
The debris flow pulled off a small fragment of the Bakswat glacier terminus Glaciers in this region are prone to surging, which can cause the formation of ice-dammed lakes and eventually destructive outburst flood. To mitigate this risk, the people of this region had deployed a network of beacon fires to warn the settlements in case of glacier lake outburst hazard even in the most remote valleys. This fascinating early warning telecommunication system was described by Iturrizaga (2019). The beacons fires of Gondor in the movie "The Lord of the Rings: The Return of the King"
21 May 2020
2:08
Chaos theory, Rössler cake and satellite remote sensing
sur Séries temporelles (CESBIO)

Chaos theory, Rössler cake and satellite remote sensing
A chaotic attractor was obtained from satellite remote sensing data for the first time in 2011 giving a strong argument for chaos in real environment. This result has a special sound today May 20th 2020, since it is the 80th birthday of Otto E. Rössler -- one of the fathers of the chaos theory -- to whom a Focus Issue entitled "Chaos: from theory to applications" is opened for submission in Chaos: an Interdisciplinary Journal of Nonlinear Science to celebrate his birthday.

The first chaotic attractor from satellite remote sensing
The first chaotic (and hyperchaotic) attractors (Lorenz-1963, Rössler-1975, -1976 and -1979, Ueda [1-5]) and maps (Gumowski-Mira-1968, -1973, Hénon-1976, Lozi-1978, Rössler-1979 [6-9;4]) were based on theoretical developments (see reference [10] for a historical perspective). Their dynamic is said chaotic since it is --- at the same time --- deterministic and unpredictable at long term. This can be considered as a major epistemological rupture: after the definition of determinism given Laplace in the early 19th century, such an association (determinism and unpredictability) was completely unexpected.

After such an unexpected theoretical result, important experimentations were carried out in the 1980s to investigate if such a paradoxical situation could be met under experimental conditions. The work by John Hudson and his group are among the most notable and important ones, which gave a definitive proof of experimental chaos (see [11] for historical details).

To detect chaos in real environmental world was another difficult challenge. Various algorithms have been developped to detect some of the properties of chaos but most of the developments were not able to ensure the presence of all the properties of chaos at the same time. In particular, very few methods were proven able to detect determinism. The global modelling technique enables to obtain sets of equations directly from observational time series. It is one of the rare approaches able to capture the whole -- dynamical, geometrical and topological -- properties of chaos in a single synthetic algebraic and deterministic formulation.

The global modelling technique was applied to the cycles of cereal crops in North Morocco, considering vegetation index aggregated on a large region. A chaotic model could be obtained. It is plotted in Figure 1. It is one rare example of a chaotic attractor directly obtained from real environmental conditions; and it is very likely the first chaotic attractor directly obtained from satellite remote sensing data. Therefore, it makes sense to remind this result on the multitemp website today.

Figure 1: Cereal crops attractor obtained with the GPoM package from the Normalized Difference Vegetation Index (AVHRR data) of the cycles of cereal crops aggregated in North Morocco [12-14].

The Rössler cake
To celebrate Otto's 80th birthday, it was decided to prepare the Rössler cake ! (more precisely, the Rössler-1976 cake!), that is, a cake diffeomorphic to the Rössler-1976 attractor. To do so, we did refer to the receipe of the Bûche de Noël. You are free to interpret the choice of this receipe as a kind of messianic choice (hou!). But one should not forget two other useful arguments. One argument relies on physics: the paste of the bûche de Noël is well adapted to apply smooth foldings and the chosen attractor has at least a single folding in its simplest regime. Another argument is a personal argument: it is a cake the author is able to cook because he has to prepare it each year (familly obligation, no matter what!), at the purposive date (by the end of December). The result we got is shown in Figure 2.

Figure 2: Prototype of the Rössler cake -- topological reconstruction of the 1976 version -- with its Poincaré section (top panel). The cake is bounded by a genus-1 torus (there is a single hole in the center). The paste flow is characterized by a positive first Lyapunov exponent (just consider the diverging trajectories in the right bottom part of the cake, we must admit we were not able to derive the Jacobian matrix from the paste flow) and by a single folding (in the left bottom part) which is a necessary condition for chaos. The Poincaré section to the cake (golden brown colour on the upper part of the cake) was cooked separately: this trick was used to get the proper structure for the Poincaré section to the cake! In this sense, this cake is not exactly a reconstruction of the Rössler-1976 attractor, but, more to the point, a topological suspension of the Poincaré section to the Rössler-1976 attractor, hooo! The cooking receipe was previously explored theoretically [15-16] considering the bidimensional chaotic map introduced by Hénon in 1976 [8] and the three-dimensional walking stick hyperchaotic map introduced by Otto in 1979 [4]. The original Rössler attractor is also provided (bottom panel) for visual validation.
Before serving, Poincaré sections should be carefuly sliced piece by piece. These sections are theoreticaly infinitely thin. In practice, to keep the result more presentable, the Poincaré sections should not be sliced too thin (recommendation of the chef).

FIGURE OF THE POINCARE SECION STILL TO BE ADDED BECAUSE I BIT IN IT BEFORE TAKING THE PHOTO... I'LL HAVE TO COOK IT AGAIN!
Figure 3: Poincaré section to the Rössler cake. A careful look at it should enable to detect several iterations of the horseshoe map [17].
Because of the current travelling restrinctions due to the epidemic of Covid-19 (see the GPoM-epidemiologic Bulletins), it was unfortunately not possible to share the taste this cake practically with Otto. But we could share its frendship taste with him and other friends virtualy during a somehow surealist video session.

Figure 4: Video session on May 20th 2020 at 5 pm (German time) with Otto Rössler and several friends of his (Gerold Baier, Vasileios Basios, Niels Birbaumer, Sven Sahle, Andrey Shilnikov, Ursula Kummer, Doyne Farmer, Kunihiko Kanako, Lars Folke Olsen, Plamen Ivanov, Johachim Peinke, Achim Kitell, Christophe Letellier and myself) to celebrate his 80th birthday. Reimara Rössler did also briefly appear during the session. Christian Mira could not participate to the session but told us he will send a message to Otto by email. At the end of the session, one of us recalled us that Konrad Lorenz, with whom Otto was close to, did consider Otto as an astonishingly promising young researcher; He was right! This session was kindly organized by Gerold Baier on a suggestion by Christophe Letellier, thank you so much. The birthday photo was recomposed from different screen copies by Vasileios and myself (thank you Vasileios) in order to have everybody on the same picture, and with a joyful face (actually we had special expressions when the connection was getting bad. But we could manage!). This celebration was an opportunity for me to wear my Vinod Gaur tie (friendly awarded by Vinod himself in 2017, I could even choose the color).
This cake reconstruction of the Rössler-attractor enabled us to unveil a new property of this chaotic attractor: this attractor is not just an attractor of simplest possible structure (characterized by a single folding); it is neither just a very beautiful attractor; It is also a (potentially delicious) frienship attractor.

Happy birthday Otto!

Invitation to submit to special focus
We are pleased to announce the forthcoming publication of a new Focus Issue entitled “Chaos : From Theory to Applications” in Chaos : An Interdisciplinary Journal of Nonlinear Science to celebrate the 80th birthday of Otto E. Rössler.

Otto submitted his first paper on chaos in 1976. The Rössler system quickly became one of the prototypes on which the paradigm of chaos theory is constructed. He also introduced hyperchaos and suggested a classification of chaotic attractors. Otto E. Rössler investigated applications of chaos in various fields : chemistry, heart rhythm, biology, electronics, astrophysics, fluid mechanics.

In order to celebrate his contribution to chaos theory, and in honor of his 80th birthday, this forthcoming focus issue in Chaos is devoted to showing how chaotic behaviors help to understand the reality of the world in its complexity. Contributions should try to show how the concepts of state space, attractor, first-return maps Poincaré sections, bifurcations, etc. allow for better views of the possible dynamics produced by simple as well as complex systems. Contributions can be either theoretical, numerical or experimental.

For those who met Otto E. Rössler and who would have particular recollections, we offer the opportunity to contribute to a co-authored paper in which each contributor will write about one half or one page (possibly with one picture or one figure) for providing to the community how Otto was perceived along his career. For this type of contribution, please contact Christophe Letellier.

This Focus Issue is scheduled for publication in December 2020. All submitted articles should conform to the Chaos author guidelines. Please be sure to write a lead paragraph – see guidelines for more detail. Papers will be published immediately after acceptance and will available online in the journal and in a virtual collection dedicated to this Focus Issue.

Your submission should be sent by AUGUST 21, 2020 (extended date). It will be reviewed within 4-6 weeks of submission.

Guest editors :
Christophe Letellier (Rouen Normandie University, France)
Sylvain Mangiarotti (Centre d’Etudes Spatiales de la Biosphère, Toulouse, France)
Lars Folke Olsen (University of Southern Denmark, Odense )

References
[1] E. N. Lorenz, Deterministic nonperiodic flow. Journal of the Atmospheric Sciences, 20, 130-141, 1963.

[2] O. E. Rössler, An equation for continuous chaos. Physics Letters A, 57 (5), 397-398, 1976.

[3] O. E. Rössler, Chaotic behavior in simple reaction system. Zeitscrift für Naturforschung A, 31 (3-4), 259-264, 1976.

[4] O. E. Rössler, An equation for hyperchaos. Physics Letters A, 71, 155-157, 1979.

[5] Y. Ueda, The road to chaos. Aerial Press, Santa Cruz, 1993.

[6] C. Mira, Étude de la frontière de stabilité d’un point double stable d’une récurrence non linéaire autonome du deuxième ordre. Proceedings of the IFAC Symposium on Pulse-rate and Pulse-number Signals in Automatic Control, D 43-7II, 1968.

[7] J. Bernussen, Liu Hsu & C. Mira, Quelques exemples du second ordre. Collected preprints of Colloque International du CNRS, 229, Transformations ponctuelles et applications (Toulouse, September 1973), Proceedings Editions du CNRS (Paris), pp. 195-226, 1976.

[8] M. Hénon, A two-dimensional mapping with a strange attractor. Communications in Mathematical Physics, 50, 69-77, 1976.

[9] R. Lozi, Un attracteur étrange du type attracteur de Hénon. Journal de Physique, 39 (8), C5-9 - C5-10, 1978.

[10] C. Letellier, Chaos in Nature. World Scientific Series on Nonlinear Science Series A: Volume 94, 2019.

[11] C. Letellier, S. Mangiarotti, I. Sendiña-Nadal & O. E. Rössler, Topological characterization versus synchronization for assessing (or not) dynamical equivalence. Chaos, 28, 045107, 2018.

[12] S. Mangiarotti, L. Drapeau, R. Coudret & Jarlan L., Modélisation par approche globale de la dynamique du blé pluvial observée par télédétection spatiale en zone semi-aride, 14^e Rencontre du Non Linéaire, 14, 103-108, Université Pierre et Marie Curie, Paris, France, ISBN 978-2-9538596-0-7, 2011.

[13] S. Mangiarotti, R. Coudret, L. Drapeau & L. Jarlan, Polynomial search and Global modelling: two algorithms for modeling chaos. Physical Review E, 86(4), 046205, 2012.

[14] S. Mangiarotti, L. Drapeau & C. Letellier, Two chaotic global models for cereal crops cycles observed from satellite in Northern Morocco. Chaos, 24, 023130, 2014.

[15] S. Mangiarotti, Modélisation globale et Caractérisation Topologique de dynamiques environnementales – de l'analyse des enveloppes fluides et du couvert de surface de la Terre à la caractérisation topolodynamique du chaos. Habilitation à Diriger des Recherches, Université Toulouse 3, France, 180 pp. , 2014.

[16] S. Mangiarotti & C. Letellier, Topological analysis for designing a suspension of the Hénon map. Physics Letters A, 379, 3069-3074, 2015.

[17] S. Smale, Differentiable dynamical systems. Bulletin of the American Mathematical Society, 73 (6), 747-817, 1967.
13 May 2020
1:18
Détection de la neige sur les routes avec Sentinel-2
sur Séries temporelles (CESBIO)

France Bleu nous signale une impressionnante avalanche en Savoie près du Col du Glandon qui bloque la route Départementale 927.
Cliché pris le 11 mai 2020 par Hervé Girard - Data-avalanche.org
Les cartes d'enneigement distribuées par Theia permettent d'évaluer la présence de neige sur les routes de montagne grâce à leur haute résolution (pixel de 20 m). Une carte qui date du 09 mai (soit deux jours avant la photo ci-dessus) couvre le secteur de la D927.

Depuis QGIS on peut récupérer le tracé de la route à partir de la base OpenStreetMap et le convertir en une série de points espacés de 20 m. Avec ces points on peut échantillonner la carte Theia du 9 mai et ainsi visualiser les tronçons couverts de neige sur une carte.

On peut même exporter ces points et tracer le profil topographique de la route. Pratique pour ceux qui préfèrent le vélo !

Merci à César pour l'info !

PS. si à l'inverse vous cherchez absolument à toucher la neige tout en restant à moins de 100 km de chez vous ceci pourrait vous aider :

Trouvez la neige à moins de 100 km de chez vous !
11 May 2020
12:00
Benefits of SAR for forest loss alert detection in the tropics
sur Séries temporelles (CESBIO)
The tropical forest is an important carbon sink which also hosts high biodiversity. Monitoring forest loss is therefore a major issue which requires efficient and accurate tools. In fact, the major challenge is to detect forest loss as precisely and quickly as possible, despite cloud cover which limits the availability of usable optical satellite imagery. SAR (synthetic aperture radar) sensors are an efficient tool in this context at it can penetrate clouds.

We used the Sentinel-1 based method developed at CESBIO (Bouvet et al., 2018), and explained in this post, to produce forest alerts (FA) maps every week at 10-m resolution over the whole French Guiana (83?534 km²), for the years 2016 to 2019. The maps are available there: [cesbiomass.net]

The term "Forest Alerts" is used because the purpose of producing the maps is the rapid detection of forest cover loss. However, we could have referred to it as "Forest loss", because the areas are accurately detected.

We assessed the performance of these maps using a reference dataset of 1 867 plots covering 2 124.5 ha, from in situ field data, helicopter surveys and photo interpretation, and describing various sources of forest loss. We would like to thank the ONF, EMOPI and PAG for providing these data. The Sentinel-1 FA maps showed 84.4% of producer’s accuracy and 96.3% of user’s accuracy, using the Olofsson et al. (2013) method for validation.
Forest loss at 10-meter resolution between 2016 and 2019, over the whole French Guiana and over 8 × 12 km areas, highlighting the various sizes and spatial distributions of disturbed areas
The Sentinel-1-based FA maps were also compared with an optical-based system that raises FA at the pan-tropical scale: the Global Land Analysis and Discovery (GLAD) FA by the University of Maryland. The optical-based FA showed 62.3% of producer’s accuracy and a systematic temporal delay relative to radar, up to one year. The temporal comparison regarding forest loss during the dry season showed little difference. In contrast, the optical-based FA showed a median temporal delay of 143 days (more than 4.5 months) regarding forest loss during the whole year, due to strong delays during the wet cloudy season.
Number of detected reference samples per month for Sentinel-1 based (green) and for Landsat based (yellow) forest alerts (left), and histogram of the difference in detection date of the reference samples (right), where negative values indicate an earlier detection in Sentinel-1 based forest alerts.
A field campaign in collaboration with the Statistics Service and the Prospective Department of the Ministry of Agriculture, which we thank, made possible the evaluation of the average absolute detection delay using the SAR-based method on 26 plots : 8 days.

These results highlight the benefits of SAR for forest loss alert detection in the tropics.

Subsequently, we plan to produce the map of the Guiana Shield (Suriname, Guyana, Amapa) and adapt and apply the method to other tropical countries.

References:
- Ballère, B., Bouvet, A., Mermoz, S., Koleck, T., Bedeau, C., Forestier, E., André, M., Le Toan, T., Frison, P.L. and Lardeux, C. SAR data for tropical forest disturbance alerts: Benefit over optical imagery - Study case in French Guiana. Submitted to Remote Sens. Environ. on Jan. 2020
- Bouvet, A., Mermoz, S., Ballère, M., Koleck, T., Le Toan, T., 2018. Use of the SAR Shadowing Effect for Deforestation Detection with Sentinel-1 Time Series. Remote Sens. 10, 1250.
- Hansen, M.C., Krylov, A., Tyukavina, A., Potapov, P.V., Turubanova, S., Zutta, B., Ifo, S., Margono, B., Stolle, F., Moore, R., 2016. Humid tropical forest disturbance alerts using Landsat data. Environ. Res. Lett. 11, 034008.
- Olofsson, P., Foody, G.M., Stehman, S.V., Woodcock, C.E., 2013. Making better use of accuracy data in land change studies: Estimating accuracy and area and quantifying uncertainty using stratified estimation. Remote Sens. Environ. 129, 122–131.
08 May 2020
15:23
Trouvez la neige à moins de 100 km de chez vous !
sur Séries temporelles (CESBIO)

Voici un outil qui permet de trouver de la neige à moins de 100 km de chez vous ou dans votre département !

[https:]]

Cette app utilise les cartes de neige journalières produites par la Nasa à partir du capteur Modis sur le satellite Terra (MOD10A1.006). L'observation non-nuageuse la plus récente dans les dix dernières cartes disponibles est affichée pour chaque pixel de 500m. En général les données affichées datent de moins d'une semaine.

Le code source de mes apps peut être récupéré ici :

git clone [https:]
22 April 2020
15:17
Lucky Sentinel-2 still can visit Italia, Spain and Portugal
sur Séries temporelles (CESBIO)

That's unfair ! Sentinel-2 is a lucky satellite who still can visit Spain, Italy or Portugal every 5th day, while we are locked at home. Thanks to the hard work and good organisation at ESA (for the Level-1 products) and CNES (for the advanced L2A and L3A products), I have been able to generate mosaics of Sentinel-2 products over these 3 countries, and I spent quite a lot of time dreaming of the places I should go to.

So to help you go through these difficult times, here are the winter animations of L2A Mosaics over the Iberian peninsula and Italia.

Copernicus Sentinel-2 level 3A products from Theia over Italy, from September 2019 to March 2020 (click to enlarge)

Copernicus Sentinel-2 level 3A products from Theia over Italy, from September 2019 to March 2020 (click to enlarge)
These products were generated with MAJA atmospheric correction and cloud screening processor and WASP, the weighted average synthesis processor. The level of artefacts is quite low, except for dates and regions when the cloud cover was high, such as Galicia in November. However, as it may be seen in Northern Italy, the directional correction model that we use in WASP to normalise for the varying observation angles should be improved when the sun is low.

Usually, in my publications on monthly syntheses, I also provide a tool to zoom to full resolution, select a couple of syntheses and compare them, as I do for instance for France. I will only be able to do it after the lockdown is loosened, as I do not have access to the visualization server from home. Anyway, you still can have access to the data on Theia distribution site, and download them tile by tile at full resolution.
18 April 2020
11:28
A new method from CNES to improve VENµS spectral band registration
sur Séries temporelles (CESBIO)

=>

A new attitude estimation method has been developed for the VENµS satellite, based solely on image content. The registration quality can thus be improved by a factor of 10, paving the way for new applications.
Multi-spectral registration performance with the old (top) and new (bottom) methods.
Since its launch in August 2018, the Franco-Israeli Venµs pushbroom satellite has been affected by attitude restitution noise that hinders the registration of the 12 spectral bands and of time series of images. The efforts made during flight acceptance tests have significantly improved the geometric quality of the images to make them usable. However, this correction is insufficient for 20% of the images produced.

In order to solve this problem, a new attitude estimation method has been developed. VENµS successively acquires images in 12 spectral bands between 390 and 950 nm. By exploiting sub-pixel deformation measurements of 3 images corresponding to 3 spectral bands whose acquisition is shifted by 2.8s, we obtain attitude information with an angular accuracy and time frequency which is much higher than those of VENµS' AOCS (Attitude and Orbit Control Systems) equipment. The accuracy achieved on the satellite's attitude is 20 times greater than that of the AOCS, and 10 times greater than the correction implemented in the initial processing chains of the VENµS mission.

The figure illustrates the improvement in superposition over 4 pairs of spectral bands. The residual overlay error is 0.05 pixel maximum. This unparalleled accuracy opens the way to new applications that require a registration quality that exceeds the specifications of 0.1. These new applications include the creation of Digital Terrain Models, which exploit the native stereoscopy capability of the focal plane, and coastal zone bathymetry (a study at LEGOS laboratory is in progress) which exploits the time delays of the spectral bands on the swell.
16 April 2020
16:48
La nouvelle méthode du CNES pour superposer les bandes de VENµS
sur Séries temporelles (CESBIO)

Une nouvelle méthode d'estimation d'attitude a été développée pour le satellite VENµS, basée uniquement sur le contenu des images. La qualité de superposition peut ainsi être améliorée d’un facteur 10, ouvrant la voie à de nouvelles applications.
Performance de recalage multi-spectral avec l'ancienne (haut) et la nouvelle (bas) méthode.
?Depuis son lancement en aout 2018, le satellite pushbroom franco-israélien Venµs est affecté par un bruit de restitution d'attitude qui nuit à la superposition des 12 bandes spectrales et des séries temporelles. Les efforts menés lors de la recette en vol ont amélioré significativement la qualité géométrique des images pour les rendre exploitables. Cependant cette correction est insuffisante pour 20% des images produites.

Afin de résoudre ce problème, une nouvelle méthode d'estimation d'attitude a été mise au point. VENµS acquiert successivement des images dans 12 bandes spectrales entre 390 et 950 nm. En exploitant les mesures de déformation sub-pixeliques de 3 images correspondant à 3 bandes spectrales décalées de 2,8s, la précision angulaire et la fréquence temporelle dérivées de ces mesures sont bien supérieures à celles des équipements SCAO (Systèmes de Contrôle d'Attitude et d'Orbite) de VENµS. La précision atteinte sur l'attitude du satellite est 20 fois supérieure à celle du SCAO, et 10 fois supérieure à la correction implémentée dans les chaînes de traitement initiales de la mission VENµS.

La figure illustre l'amélioration de superposition sur 4 couples de bandes spectrales. L'erreur résiduelle de superposition est de 0,05 pixel maximum.

Cette précision inégalée ouvre la voie à de nouvelles applications qui exigent une qualité de superposition supérieure aux spécifications. Parmi ces nouvelles applications, citons la création de Modèles Numériques de Terrain, qui exploite la capacité de stéréoscopie native du plan focal, et la bathymétrie des zones côtières (étude au LEGOS en cours) qui exploite les délais temporels des bandes spectrales sur la houle.
02 April 2020
21:43
60 years of snow from space
sur Séries temporelles (CESBIO)

Suomi/VIIRS had a nice view of snow in the Alps today, 60 years after the first TIROS image.

Thanks to all space agencies for providing us great data to study snow. And thanks to Richard Essery for the tip!

Snow cover in the High Mountain Asia on 13 May 1960
30 March 2020
18:15
Le désherbage chimique est-il visible depuis l’espace ?
sur Séries temporelles (CESBIO)

Quand on se balade à la campagne (c’était surtout avant le Covid-19...), ça pique toujours un peu les yeux de voir dans les champs, ces « mauvaises herbes » (ou adventices pour les intimes) devenues toutes rousses, presque du jour au lendemain. C’est pas que je n’aime pas les rousses, mais je préfère quand même les couleurs naturelles !

Déformation professionnelle, comme probablement beaucoup des lecteurs de ce blog, la question arrive d’emblée : « Est-ce qu’on le voit sur les images Sentinel-2 ? ». En rentrant, on s’empresse de demander à Sentinel-Hub, notre meilleur ami.

Et bien OUI ! Il est assez aisé de détecter visuellement dans les parcelles agricoles les endroits où des produits phytosanitaires (herbicide en l’occurrence) ont été pulvérisés, en pré-émergence de la culture. Dans les images en fausses couleurs infra-rouges, cela apparaît très nettement : la végétation qui était rouge initialement passe à l’orange, puis marron. Et c’est confirmé par une chute dans le NDVI (assez logique…). On peut donc faire la différence avec une parcelle simplement labourée. En cas de désherbage chimique, le changement de couleur est progressif (et non abrupt) et il est observable dans les séries temporelles. De là à dire qu’on peut l’identifier automatiquement, c’est une autre histoire. Ça demanderait un vrai travail. Sur les 8 semaines d’images analysées, il y en avait assez peu sans nuage, ce qui complique la tâche. Par ailleurs, si les images sans nuage sont trop éloignées dans le temps, l’effet du désherbage chimique devient moins perceptible (empiriquement, d’après les images analysées et la connaissance de terrain, un décalage de 3 à 4 semaines est encore suffisant). Bien sûr, cela dépend aussi de la superficie de la zone traitée. Souvent, le traitement ne concerne pas l’entièreté de la parcelle. Il peut se limiter aux bordures ou à des endroits envahis d’adventices, bien circonscrits dans la parcelle.

Les images ci-dessous illustrent ces observations. Elles sont localisées dans le Lauragais, une région de coteaux composée principalement de grandes cultures (blé, maïs, colza, tournesol). Dans la première figure, on peut voir une grande parcelle pentue envahie de végétation au 20 février 2020 (couleur rouge dans l’image IRC et valeurs élevées de NDVI en vert foncé à la même date). Après pulvérisation, les adventices en ont pris un coup (!), ce qu’on observe sur le terrain (photo du 3 mars 2020 prise depuis le point blanc localisé dans les images). Le changement est bien visible dans l’image Sentinel-2 du 13 mars 2020 (enfin une sans nuage !) ainsi que dans le NDVI associé. La couleur orange-marron ne saute pas au yeux dans cette capture mais elle est plus évidente depuis le site Sentinel-Hub, selon la date dispo sans nuage... Sur le terrain, les traces du traitement n’ont pas disparu au 28 mars 2020 (oui : j’avais bien mon attestation – le sport individuel est permis pendant le confinement ).

Ci-dessous, des animations (cliquez dessus) mettent aussi en évidence le désherbage chimique sur deux zones différentes, en comparant les images avant/après desherbage :

Si vous aimez jouer à "Où est Charlie ?", vous trouverez un équivalent ci-dessous (extrait d'image S2 du 13 mars 2020) : "Où sont les parcelles traitées ?" Enjoy !
29 March 2020
0:25
China's NO2 emissions now rising
sur Séries temporelles (CESBIO)

NO2 emissions are rising again over China, showing the slow recovery of Chinese industry after a two-month shutdown. The graph below was obtained from Sentinel-5p data using the NO2 time series Google Earth Engine app.
Total vertical column of NO2 (ratio of the slant column density of NO2 and the total air mass factor) from 01 January to 27 March. Daily average over Mainland China. Data source: Copernicus Sentinel-5P NRTI NO2: Near Real-Time Nitrogen Dioxide.
Global NO2 monitor
23 March 2020
7:03
Madagascar, the new zone processed by MAJA and WASP for Theia
sur Séries temporelles (CESBIO)

Nouvelle zone traitée par Theia (les zones en rouge étaient déjà traitées).
=>

Here it comes ! Following numerous demands, we found ressources to add the great island of Madagascar to the list of zones processed by Theia. The processing with MAJA started a few days ago, with the data acquired in December 2016 by Sentinel-2. The start was only slightly delayed by the current lockdown in France, which forces thz Theia production team to telework. We will also trigger the production of monthly cloud free syntheses with WASP as soon as possible.

It will of course take some time to produce the four years of data until we reach near real time processing, but the first 6 months are already available, as you may see on the animation below, over tile T38JPT, South South East of Madagascar.
22 March 2020
11:29
Madagascar, la nouvelle zone traitée par MAJA et WASP
sur Séries temporelles (CESBIO)

Nouvelle zone traitée par Theia (les zones en rouge étaient déjà traitées).
=>

Voilà voilà, suite à de nombreuses demandes, nous avons trouvé le moyen de rajouter la grande île de Madagascar dans les zones traitées par Theia, notamment pour les besoins de Geodev. Les traitements avec la chaîne MAJA ont démarré il y a quelques jours, avec les données acquises fin décembre 2016 par Sentinel-2. Ils ont été à peine retardés par le confinement actuel en France, qui oblige la plupart d'entre nous, et notamment l'équipe d'exploitation de Theia à télétravailler. Nous enclencherons aussi le plus tôt possible la génération des synthèses mensuelles avec WASP.

Il va nous falloir bien sûr un peu de temps pour traiter les quatre ans de données jusqu'au temps réel, mais les 6 premiers mois sont déjà disponibles comme le montre l'animation ci dessous, sur la tuile T38JPT, au sud sud est de l’île.
17 March 2020
22:49
MUSCATE 20 - COVID 19
sur Séries temporelles (CESBIO)

Yes, so far MUSCATE is winning against COVID !

MUSCATE is THEIA land data production center in CNES. Although everyone in the production and development team is working from home, in good health so far, the production is still on, as you may see on the counters. We even passed the 10 000 products milestone on Landsat 8 (we produce France and overseas territories only), and the next milestone should be for Sentinel-2 : 200 000.

We should thank a lot the production team, because it must be quite hard to manage, with the small screens of laptops and the reduced network bandwidth at home. Thanks ! The production team even added a new Sentinel-2 site, the whole Madagascar island, but it is worth another post.
16 March 2020
16:10
Contextual classification for Theia's land cover (OSO) map
sur Séries temporelles (CESBIO)
Although the operational production of Theia's OSO map was transferred to CNES' MUSCATE team, at CESBIO we keep on working on the improvement of the algorithms ans the tools (the iota2 processing chain).

We have recently implemented into iota2 a new contextual classification method developed by Dawa Derksen during his PhD (the slides of his oral defense are available here). The details of the method have been published in this paper, where he shows that it performs as well as Deep Learning approaches, but with much lower computational (and energy) costs.

As most of the researches who publish in technical journals, we could have stopped there, but Dawa wrote code using the Orfeo Toolbox and made it free software (public money, public code), then Arthur Vincent (iota2's guru') made everything needed so that iota2 can call this code and use it to process entire countries!

See below this small animation comparing the results of the "classic OSO" map and the contextual version.

In order for you to make your own opinion, we have made available for download the map over the southern 1/3 of Metropolitan France (around 30 Sentinel-2 tiles). The GeoTiff file is available here:
For reminders, the classic OSO version is available here:
- OSO classic
You will be able to compare both versions and give us your point of view. The 2019 map is on its way, but it will not use the contextual version of the processing chain. We need users' feedback to decide whether this version is preferred to the previous one. If this was the case, upcoming products could use the new algorithm.

Don't hesitate to send us your feedback.

Thanks!
13 March 2020
11:15
The TRISHNA mission is decided
sur Séries temporelles (CESBIO)

12 mars 2020 CP042-2020

Here is a translation of the CNES press release following the meeting of CNES board (I translated only the parts concerning Trishna). This is great news !

=>

On Thursday 12 March 2020, the 362nd session of the CNES Board of Directors was held at the CNES Headquarters in Paris Les Halles. France's commitment to the development of the Franco-Indian Trishna programme was approved as well as the activities related to the new flexible satellite industry Space Inspire and the development of platform and payload equipment for shared use.

Trishna is a high-resolution space-time imaging mission in the thermal infrared for observation of the Earth's surface. Trishna's observations will contribute to the understanding of the water cycle and to the improvement of the management of the planet's water resources, in the context of climate change whose impacts are increasingly visible at the local scale. This programme was identified as a priority during the CNES Scientific Foresight Seminar held in Le Havre in 2019 and will strengthen CNES cooperation with ISRO (Indian Space Research Organisation), as agreed in 2018 during the State visit of the President of the Republic to India. The development by Airbus Defence and Space of the satellite's thermal infrared instrument was approved by the Board of Directors.

(...)
12 March 2020
17:30
La mission Trishna est décidée !
sur Séries temporelles (CESBIO)

=>

De : Le Service Presse du CNES
Envoyé : jeudi 12 mars 2020 13:37
Objet : CP : 362ème séance du Conseil d’Administration du CNES, décisions sur TRISHNA et Space Inspire

12 mars 2020 CP042-2020

362^ème séance du Conseil d’Administration DU CNES

Décisions sur TRISHNA ET Space inspire

Jeudi 12 mars 2020, s’est tenue au Siège du CNES à Paris Les Halles, la 362^ème séance du Conseil d’Administration du CNES. L’engagement de la France dans le développement du programme franco-indien Trishna a été approuvé ainsi que les activités liées à la nouvelle filière de satellites flexibles Space Inspire et au développement des équipements plateforme et charge utile à usages mutualisés.

Trishna est une mission d’imagerie à haute résolution spatio-temporelle dans l’infrarouge thermique pour l’observation de la surface terrestre. Les observations de Trishna contribueront à la compréhension du cycle de l’eau et à l’amélioration de la gestion des ressources en eau de la planète, dans le contexte du changement climatique dont les impacts sont de plus en plus visibles à l’échelle locale. Ce programme a été jugé prioritaire au cours du Séminaire de Prospective Scientifique du CNES, tenu au Havre en 2019 et il renforcera la coopération du CNES avec l’ISRO (Indian space Research Organisation), comme cela avait été acté en 2018, lors de la visite d’Etat du Président de la République en Inde. Le développement par Airbus Defence and Space de l’instrument infrarouge thermique du satellite a été approuvé par le Conseil d’Administration.

Space Inspire est la nouvelle filière de satellites flexibles, développée par Thales Alenia Space, pour répondre aux futurs enjeux des télécommunications spatiales. Hautement flexible et compétitive, cette nouvelle gamme de satellites s’accompagne d’importantes ruptures technologiques autour du numérique, des antennes actives et des plateformes, avec des répercussions transverses. Le CNES a décidé de participer au développement de cette nouvelle génération de satellites et de ces technologies, tout en veillant aux synergies entre ses investissements et ceux de la Défense, qui pourront ainsi bénéficier de l’avance technologique acquise. La signature d’un accord-cadre avec l’industrie, concernant ces activités, a été approuvée par le Conseil d’Administration.

Le Conseil d’Administration a aussi renouvelé pour une durée de trois ans, le mandat de Paul-Henri Ravier en qualité de Président de la Commission Interne des Marchés du CNES.

À l’issue de cette 362^ème séance, Jean-Yves Le Gall, Président du CNES a déclaré : « Cette séance du Conseil d’Administration a été déterminante pour l’avenir de nos activités, avec l’engagement de deux nouveaux programmes à très fort potentiel. Au-delà de leur caractère particulièrement innovant, Trishna, développé avec Airbus Defence and Space et Space Inspire, développé avec Thales Alenia Space, vont permettre au CNES de valoriser des synergies sur de nombreuses thématiques, pour le plus grand bénéfice de notre programme spatial et de notre industrie. »
06 March 2020
18:37
MAJA V4 will soon become an open source software
sur Séries temporelles (CESBIO)
CESBIO is the home of a highly motivated bunch of open source software advocates ! Let's have a look at some the main softwares to which CESBIO members contributed, and which are available on various git platforms :
- the father of Orfeo Toolbox is at CESBIO
- WASP (monthly syntheses of Sentinel-2 images),
- LIS (Snow cover from Sentinel-2
- Iota2 (land cover classification)
- S1-Tiling (Ortho-rectification of Sentinel-1 tiles onto Sentinel-2 grid
WASP IOTA2 LIS S1-Tiling

But one of our softwares was only distributed freely as a binary package, and not as an open source package : MAJA. As far as Sentinel-2 is concerned, MAJA is the most reliable processor for cloud detection, and one of the good ones for atmospheric correction. This is due to the use of multi-temporal methods, that use more information than single image methods. MAJA was developped by CNES, even if most of the algorithms were defined at CESBIO, and a couple of them at DLR.
Sentinel-2 image time series over Mopti (Mali), left, L1C (TOA reflectance) from ESA, right L2A from MAJA (Surface reflectance)
Why was MAJA not open source ? Well, CNES used to be quite reluctant to distribute its software as open source. My own interpretation is that it comes several reasons :
- a long military tradition, and therefore the habit of keeping some things secret
- a strong pressure from our government to value our developments
- CNES mandate to contribute to building of a strong French space industry. there used to be a belief that protecting our developments could give an advantage to French companies in the European competition.
Maybe I'm wrong on my interpretation, but anyways, four years ago, our proposal to release MAJA as open-source was rejected, even if CNES accepted to distribute MAJA as a free binary package. But we filed the same demand in 2020, and this time, last friday, the release of MAJA as an open source software was finally accepted !
Number of users downloading Theia products every month. With a total exceeding 2000 so far. Cumulative number of Theia products downloaded. On average, each product delivered by Theia has been downloaded 1.5 times.
MAJA's products within Theia have been downloaded by 2023 persons (as of February 2020). MAJA's binary code has been downloaded more than 500 times per year, 1530 times so far. MAJA is also used by DLR, by Venµs project, within the Sen2Agri and Sen4Cap systems, or in the EEA snow and ice project. Although the core of MAJA code is in advanced C++ (with templates, functors...), we know some of our users are able to bring contributions. Moreover, in MAJA 4.0, the interfaces management is written in python. Users will have the opportunity to modify the read/write drivers for a better integration in their workflow. Moreover, as MAJA becomes open source, users have the guarantee to be able to go on using MAJA, even if one day, CNES ends its development. But don't worry, MAJA will be used for Trishna mission, and should therefore live and evolve for 10 more years.

This is why my CNES colleagues asked CS-SI, the company in charge of MAJA's development, to do a large mondernization of MAJA's code, introducing an orchestrator and input/output drivers writtent in python instead of C++. The core modules are still written in C++, but their interfaces are now those of Orfeo Tool Box applications. This large refactoring was the main purpose of MAJA V4. it is currently under validation, and the last bugs are being corrected. We should be releasing MAJA's source code and executable version in a few weeks.
05 March 2020
16:15
Fort déficit de neige dans le Haut Atlas
sur Séries temporelles (CESBIO)
Il n'y a presque pas de neige dans le Haut-Atlas en ce moment.
Image du 04 mars 2020 (Terra/MODIS Band 7/Band 2/Band 1). La neige apparaît en bleu clair.
Même les cimes du massif du Toubkal (4167 m) sont pratiquement déneigées.
26 février 2020, Sentinel-2A L1C
Toutefois, le manteau neigeux dans le Haut Atlas est connu pour fortement fluctuer d'une année à l'autre (Marchane et al. 2015; Baba et al. 2018). Comment se compare ce manteau neigeux à celui des années précédentes ?

L'archive satellite MODIS nous fournit des images journalières depuis l'année 2000. J'ai sélectionné une image avec peu de nuages au début du mois de mars de chaque année.
Sélection d'images Terra/MODIS acquises entre le 03 and 09 mars de chaque année.
Ensuite j'ai extrait l'indice "normalized snow difference index" (NDSI) du produit MOD10A1 correspondant pour calculer la surface enneigée dans cette région par seuillage des pixels où le NDSI est supérieur à 0.2.
Images du NDSI Terra/MODIS (surface enneigée entre parenthèses)
Cette analyse montre que la surface enneigée en 2020 (113 km²) est bien plus faible que la surface enneigée la plus faible ayant été enregistrée par MODIS (325 km² in 2000).

Cet enneigement déficitaire est une mauvaise nouvelle pour les habitants de la région du Haouz car la fonte des neiges alimente les rivières et les réservoirs au printemps. Cette région est actuellement en proie à une sécheresse consécutive à deux années de faible pluviométrie.

Références
- Marchane A., Jarlan L., Hanich L., Boudhar A., Gascoin S., Tavernier A., Filali N., Le Page M., Hagolle O., Berjamy B., (2015) Assessment of daily MODIS snow cover products to monitor snow cover dynamics over the Moroccan Atlas Mountain range, Remote Sensing of the Environment, 160, 72-86, doi:10.1016/j.rse.2015.01.002
- Baba, M. W., S. Gascoin, L. Jarlan, V. Simonneaux, L. Hanich (2018), Variations of the snow water equivalent in the Ourika catchment (Morocco) over 2000-2018 using downscaled MERRA-2 data, Water, 10(9), 1120, doi:10.3390/w10091120.
15:46
Snow drought in the High Atlas of Morocco
sur Séries temporelles (CESBIO)
There is almost no snow in the High Atlas mountains now.
Image captured on 2020-Mar-04 by Terra/MODIS (corrected reflectance in false color composite: Red = Band 7, Green = Band 2, Blue = Band 1). The snow appears in light blue.
Even in the Toubkal massif which culminates at 4167 m above sea level is almost clear of snow.
2020-02-26, Sentinel-2A L1C, True color
However, the snow cover area is known to be highly variable in the High Atlas (Marchane et al. 2015; Baba et al. 2018). How does it compare to previous years?

The MODIS archive provides daily images since 2000. I selected one image with low cloud cover in early March for each year.
Time series of Terra/MODIS images. All images were acquired between 03 and 09 March.
Then I extracted the normalized snow difference index (NDSI) from the corresponding MOD10A1 product to compute the snow cover extent over this region (pixels where NDSI>0.2).
Time series of Terra/MODIS NDSI for the same dates (snow cover area in parenthesis)
This analysis shows that the snow cover area in early March 2020 (113 km²) is significantly lower than the previous worst year on record (325 km² in 2000).

This low snowpack is very bad news for the people in the Haouz region where lakes and rivers are already dry due to the low precipitation in the past two years. In normal years, snow melt runoff sustains the river discharge and replenish the reservoirs in spring.

References
- Marchane A., Jarlan L., Hanich L., Boudhar A., Gascoin S., Tavernier A., Filali N., Le Page M., Hagolle O., Berjamy B., (2015) Assessment of daily MODIS snow cover products to monitor snow cover dynamics over the Moroccan Atlas Mountain range, Remote Sensing of the Environment, 160, 72-86, doi:10.1016/j.rse.2015.01.002
- Baba, M. W., S. Gascoin, L. Jarlan, V. Simonneaux, L. Hanich (2018), Variations of the snow water equivalent in the Ourika catchment (Morocco) over 2000-2018 using downscaled MERRA-2 data, Water, 10(9), 1120, doi:10.3390/w10091120.
01 March 2020
22:36
Impact du coronavirus sur la qualité de l'air en Chine
sur Séries temporelles (CESBIO)

La Nasa a publié des cartes satellite de la concentration en dioxyde d'azote (NO2) au dessus de la Chine qui indiquent une forte réduction de la pollution atmosphérique en février 2020 liée au ralentissement économique chinois consécutif à l'épidémie de coronavirus. Ces images sont issues des observations du satellite européen Sentinel-5p.
NASA Earth Observatory images by Joshua Stevens, using modified Copernicus Sentinel 5P data processed by the European Space Agency (source: [https:]
L'article mentionne d'autres raisons qui ont pu contribuer à cette chute de NO2 : les vacances du Nouvel An Chinois (24-30 janvier) et la mise en place d'une réglementation environnementale plus stricte en 2020. Pour essayer d'y voir plus clair j'ai tracé les séries temporelles des concentrations moyennes sur la province du Hubei (dont la capitale Wuhan est l'épicentre de l'épidémie) et sur une région au nord comprenant la province du Hebei, la municipalité de Pékin et son port Tianjin.
Evolution des concentrations en NO2 données par le produit Copernicus Sentinel-5P "Near Real-Time Nitrogen Dioxide" à 0.01 arc degrees de résolution et une revisite de deux jours. J'ai fait cette figure avec 3 lignes de code dans Google Earth Engine.
On observe effectivement un contraste entre les années 2019 et 2020, et ce contraste est bien plus marqué dans la province du Hubei, ce qui suggère que la baisse de NO2 est bien liée aux mesures de confinement. Le 22 janvier, le gouvernement chinois a placé sous quarantaine trois villes de la province de Hubei soit vingt millions d'habitants, puis l'ensemble de la province a été confinée soit 56 millions d'habitants.

Et si cette amélioration de la qualité de l'air avait permis de sauver des vies ? Par exemple, en 2008 les mesures prises pour améliorer la qualité de l'air à l'occasion des Jeux Olympiques de Pékin ont entrainé une baisse de 8% de la mortalité mensuelle toutes causes confondues dans les villes concernées.

La pollution de l'air causerait 1,2 millions de morts prématurés par an en Chine selon un rapport du Health Effects Institute en 2017. Au 1er mars 2020, l'épidémie de COVID-19 aurait coûté la vie à 3030 personnes, essentiellement en Chine.

Suivez en temps réel l'évolution du dioxyde d'azote atmosphérique avec le Global N02 Monitor :

Global NO2 monitor
29 February 2020
22:00
6 thèses soutenues au CESBIO entre décembre et février
sur Séries temporelles (CESBIO)

En France, les nouveaux docteurs ne portent pas de chapeaux ridicules, mais au CESBIO, la célébration de l'avènement d'un nouveau docteur fait toujours l'objet d'un sérieux buffet, organisé par Laurence Keppel et financé par le laboratoire. Si vous avez constaté que certains membres du CESBIO ont pris du poids récemment, ne cherchez pas plus d'explication : nous avons eu six soutenances de thèses en deux mois, et donc 6 pots de thèse, seulement entrecoupés par les fêtes de fin d'année.

Trois de ces thèses portent sur l'eau (irrigation (Luis Oliviera), humidité des sols (Safa Bousbih), bilan hydro-météorologique (Jordi Etchanchu)), deux sur l'occupation des sols (utilisation de l'information historique (Benjamin Tardy), utilisation du contexte spatial(Dawa Derksen)), et enfin, une sur l'estimation des variables forestières (David Morin). Toutes utilisent des données des satellites Sentinel du programme Copenicus, ou des données similaires plus anciennes.

Vous pouvez cliquer sur les images pour accéder aux diaporamas des thèses. J'en profite pour féliciter les brillants docteurs !

Luis Enrique Olivera

Jordi Etchanchu

Dawa Derksen

Safa Bousbih

Benjamin Tardy

David Morin
27 February 2020
11:55
Bulletin GPoM-epidemiologic
sur Séries temporelles (CESBIO)

The epidemic of Covid-19 and its bulletins
In this post, a brief analysis of the epidemic of Coronavirus Covid-19 that broke out in Wuhan (China) on last December 2019 is sketched. More specifically, this analysis originally focused on the period January 31st to February 27th and was then extended by Bulletins.
Bulletins
Last Bulletin:

Bulletin no11 (May 14)

Paper:

Mangiarotti, S., Peyre, M., Zhang, Y., Huc, M., Roger, F., & Kerr, Y. (2020). Chaos theory applied to the outbreak of Covid-19: An ancillary approach to decision-making in pandemic context. Epidemiology & Infection, 148, E95, 1-9. doi:10.1017/S0950268820000990

Supplementary Materials 1-9 MangiarottiEtAl_EpidInf2020

Archives:

Bulletin no1 (February 9), Bulletin no2 (February 26), Bulletin no3 (6 March) erratum, Bulletin_no4_(March 9), Bulletin_no5 (March 15), Bulletin no6 (March 26), Bulletin no7 (March 29), Bulletin no8 (April 2), Paper preprint (April 15), EpidInfSuppMat (April 15), Bulletin no9 (April 29), Bulletin no10 (May 6)

Context
The early development of the outbreak of Corovavirus Covid-19 (initialy called nCov) was first monitored by the Wuhan Municipal Health Commission. By the end of January, while the epidemic had started to spill over into other provinces in China, the data were then monitored at national scale and made available by the Chinese National Health Commission. Most of the data presented in the present study were gathered from this site for the present analysis.

Three time series have been considered from this data set to model the propagation of the epidemic: (1) the number of cumulated confirmed cases C?(t), (2) the current number of severe cases s(t), and (3) the cumulated deaths D?(t). Their time series are shown hereafter. Note that a new methodology including the clinic cases was used after February 11. For this reason, a correction was applied to the time series on the period ranging from Day of Year 21 to 42 (around +30% for C?(t), +10% for s(t) and D?(t)).

The global modelling technique was used to obtain a model from these three variables. This technique aims to obtain sets of equations of the dynamic directly from observational time series (https://doi.org/10.1063/1.5081448, see also Actualité de l'INSU mars 2019). This technique has been applied recently to model the epidemic-epizootic coupling between human and rats for plague in Bombay (1896-1911) and for the epidemic of Ebola virus desease in West Africa (2013-2016). However, the results could be shared only after these outbreaks had stopped. Interestingly, its application to the present epidemic of Covid-19 enabled to obtain a set equations during the outbreak.
Results
The first results were obtained on February 6 and shared with other colleagues from CESBIO and ASTRE labs while the outbreak had not reached its highest peak of propagation. At this date, this preliminary analysis suggested that the epidemic was close to reach its stationnary regime and that the earlier behavior did correspond to a transient. The highest peak of propagation was attained the day after, as confirmed later. The first bulletin was diffused on February 9 (Bulletin no1) while the decrease had just started. An update of this bulletin was published online two days ago (Bulletin no2).

The simulations using the model obtained on February 6 are shown (in colour) on the following Figure where observations are also reported in black. The first time series shows that the peak of confirmed cases per day (after corrections applied) was reached on February 6 (Day of Year 37) after which a decrease started. Obviously, the maximum number of cases per day was overestimated by the model; nonetheless, the model clearly detected that the end of the transient was very close. The second time series corresponds to the number of additional severe cases per day. This reveals large oscillations that are reproduced by the model (although smaller in amplitude and speed). Finally, the third time series corresponds to the number of deaths per day. Although their maximum level is slightly underestimated in magnitude by the model, simulations appear consistent with the observations.

For the three variables, it is observed that a small difference in the initial conditions can lead to very different trajectories, even using fully deterministic equations. This is characteristic of chaos: a behavior at the same time deterministic and unpredictable at long term.

Another representation can be used to monitor the epidemic. The phase space (or state space) enables to represent all the states of a dynamical system. When this space is reconstructed from time series, it provides a trajectory presenting the successive states visited by the studied system. Three different projections of this phase space reconstructed from the model trajectories (colour lines) are presented on the next Figure, each starting from slightly different initial conditions. The simulations are the same as the ones presented in the last Figure, only the representation differs here. Although the time evolution was obviouly different in a temporal projection, it is clearly observed in this representation that, after a short transient, these three trajectories (in red, orange and purple) do clearly converge to the same geometrical structure. It is one typical interest of this representation: it enables a geometrical representation independent to the initial conditions. The initial conditions are different, but the dynamic is strictly the same (same deterministic equations); for this reason, the three trajectories do converge to the same geometrical structure, here, a chaotic attractor. Although these three trajectories will visit the same chaotic attractor, they will visit it in a different way, leading to an unpredictable behavior in the time space.

The observed trajectory is also superimposed in black which shows that the transient was modelled correctly. Although the model drawbacks previously evoked can be retrieved here (maximum number of infections per day overestimated at the maximum peak, maximum number of daily deaths slightly underestimated), it can also be noticed that the number of deaths per day did experience a succession of oscillations as suggested by the chaotic attractor (see the third projection). Differences between observed and modelled dynamics remain large. In particular,the observed signal appear much noisier than the modelled one which underlines the difficulty to handle this type of dynamical behavior. It may be noted that the model was obtained rather early, it could not take into account the most recent human action to slow down the disease, which may contribute to explain the higher maximum of infections simulated by the model.
Geographical differences
To compare the behaviors geographically, another data set made available by the Center for Systems Science and Engineering (CSSE, Johns Hopkins University) was used which provides almost online information about the spill over of the Covid-19, at global scale, and province by province for China.

Very important differences are observed from one place to another. Actually, most of the cases of Covid-19 are found in the Hubei province (65596 cumulated confirmed cases presently). The provinces the most affected after Hubei are the Gouangdong (1347), Henan (1272), Zhejiang (1205), Hunan (1017), Anhui (989) and Jiangxi (934) provinces where the number of confirmed cases is close to fifty times lower, that is, relatively much more moderate. Note that for five other provinces (the Shandong, Chongqing, Jiangsu, Sichuan and Heilongjiang), the cumulated number of confirmed cases is ranging from 400 to 800 whereas it is lower in all the other chinese provinces.

Forecastings
A focus of infection by the Covid-19, characterized by a quick increase of confirmed cases and deaths, has recently broken out in Italy (February 21-22). The phase space has been used to compare the early evolution in Italy to the evolution observed in the six chinese provinces previously mentioned. Two projections of this reconstructed space are provided here after for a period up to February 26 (Italy in red). The observed dynamics are very similar in amplitude and patterns. Based on the present data available, a similar behavior might have been expected in Italy in the days to come.

Based on this preliminary observation, the time evolution of the number of infections could be forecasted using the dynamical models obtained for the epidemic of Covid-19 in the Zhejiang, Henan, Guangdong, Anhui and Jianxi provinces. Started on February 24 these simulations suggest that the end of the epidemic may not be reached before three to four weeks, with a cumulated number of infections that would reach 800 to 1400 cases. The very last updates of confirmed cases (for February 27) however suggest that these models may probably underestimate the outbreak in Italy.

Authors
Sylvain Mangiarotti (CESBIO), Marisa Peyre (ASTRE), François Roger (ASTRE), Yàn Zhang, Mireille Huc (CESBIO), Yann Kerr (CESBIO)
Acknowledgements
This work was supported by the French programmes Les Enveloppes Fluides et l’Environnement (CNRS-INSU), Défi Infinity (CNRS) and Programme National de Télédétection Spatiale (CNRS-INSU). The extension of this work is also supported by Montpellier université d’excellence (MUSE).
References
S. Mangiarotti, M. Huc (2019) Can the original equations of a dynamical system be retrieved from observational time series? Chaos 29, 023133 (2019); [https:]]

S. Mangiarotti (2015) Low dimensional chaotic models for the plague epidemic in Bombay (1896-1911). Chaos Solitons & Fractals, 81(A), 184-196 (2015); [https:]]

S. Mangiarotti, M. Peyre & M. Huc (2016) A chaotic model for the epidemic of Ebola virus disease in West Africa (2013–2016). Chaos 26, 113112 (2016); [https:]]
24 February 2020
0:47
Enneigement de la station de ski Superbagnères en février 2020
sur Séries temporelles (CESBIO)

Vous avez sans doute entendu parler de l'Affaire de l'Hélicoptère :

"Le bilan carbone de la France ne se joue pas ici" : la station de ski de Superbagnères dépassée par la polémique de la neige livrée par hélicoptère [https:]] pic.twitter.com/It9dedwwgZ

— franceinfo (@franceinfo) February 20, 2020

Les cartes d'enneigement du pôle Theia réalisées à partir des images du satellite européen Sentinel-2 (en accès libre ici) permettent de regarder l'état du domaine skiable de Superbagnères avec objectivité : il y a une prise de vue sans nuage le 15 février, c'est-à-dire le lendemain du fameux héliportage. Ci-dessous j'ai superposé la surface enneigée (en bleu cyan) sur la carte de la station. D'après la vidéo la neige aurait été déposée en haut du téléski du Baby. En effet l'enneigement était très limité dans ce secteur.

Cela saute aux yeux si l'on compare avec la carte de l'enneigement de l'année dernière à la même époque (20 février 2019)

Et si l'on compare l'enneigement au lendemain de l'héliportage avec l'enneigement 10 jours auparavant dans ce secteur en particulier, on voit que cela n'a pas enrayé la fonte qui avait largement érodé le manteau neigeux dès le début du mois de février.

Cet évènement médiatique témoigne du fait que l'enneigement est largement déficitaire cette année dans le Luchonnais et en Ariège. Dans les Pyrénées orientales en revanche la tempête Gloria a apporté d'importantes chutes de neige. A la fin de la saison (31 avril) j'essayerais de faire un bilan sur l'ensemble du massif.

En attendant... à vos smartphones !
05 February 2020
11:34
Aidez-nous mesurer l'enneigement par satellite avec votre smartphone Android !
sur Séries temporelles (CESBIO)

Nous sommes à la recherche de volontaires pour nous aider à valider un algorithme de détection de la neige avec le satellite Sentinel-2 ! Toute personne équipée d'un smartphone Android peut participer. Pour cela il suffit de remplir le formulaire ci-dessous pour vous inscrire et d'installer ODK Collect sur votre smartphone (tutoriel ci-dessous). Ensuite c'est très simple de faire des relevés !
[contact-form-7]
Votre mission, si vous l'acceptez, consiste à estimer visuellement le pourcentage de la surface couverte par la neige dans un cercle de 10 m de rayon autour de votre position. (en forêt, c'est la surface du sous bois, couverte de neige qui compte)

Exemples d'estimation du pourcentage enneigé. Il est possible de marquer 0% ou 100% mais de préférence les relevés doivent être faits dans une zone partiellement enneigée. Nous avons prévu des cadeaux pour récompenser ceux qui auront fait le plus de relevés ! - Un lot de deux entrées pour une journée à l'observatoire du Pic du Midi - Des paires de raquette - Des goodies ! Les gagnants seront informés avant l'été 2020. Voici la procédure détaillée pour installer le formulaire Neige : (en cas de problème d'affichage télécharger le tutoriel au format PDF).
20 January 2020
19:08
My remote sensing analyses on Twitter. Part 2: landslides
sur Séries temporelles (CESBIO)
I started to look at landslides in satellite images when I saw this tweet by @RemotePixel:

July 11th 2016 Sentinel-2 Image over Glacier Bay, AK massive landslide! pic.twitter.com/pTnHQ7el8G

— Remote Pixel (@RemotePixel) July 14, 2016

Like floods, landslides can be quickly identified using a before/after image comparison. In the case of the Glacier Bay landslide I used this method with Sentinel-1 and Sentinel-2 to compute the area of the debris tongue (about 20 km², 10 km long).

Mapping the Glacier Bay landslide using Sentinels [https:]] @davepetley @RemotePixel @theAGU pic.twitter.com/2bpVa0cYZr

— Simon Gascoin (@sgascoin) July 19, 2016

But Sentinel-1 was not particularly useful since the Glacier Bay landslide was a dark slide on a snow covered glacier, so it was easy to map it using natural color Sentinel-2 images. These kind of landslides are called "supraglacial landslides" and can be found in Landsat or Planet images too. Hashtag #lessbears

I think I spotted a similar supraglacial landslide in Baffin Island#armchairfieldwork [https:]] pic.twitter.com/eLr81Lc6xw

— Simon Gascoin (@sgascoin) December 3, 2017

Although the Glacier Bay landslide was one of the most impressive landslide of the decade, it did not cause any damage. The most devastating landslides in the recent years occurred in Palu, Indonesia after an earthquake.

Two articles show that rice irrigation caused the Palu landslide in Sep 2018 [https:]]
We showed images of the fault line and landslides here [https:]] pic.twitter.com/WujBIL4OiG

— Simon Gascoin (@sgascoin) October 11, 2019

Palu's shallow earthquake left a scar in the city which was visible from space

The fault line running through #Palu captured by #Sentinel2 @ESA_EO @CopernicusEU pic.twitter.com/PPYDbbJRZU

— Simon Gascoin (@sgascoin) October 2, 2018

This animation astonished me but other ~~grumpy~~ sharped-eye scientists noticed the poor muti-temporal registration of the images..

I saw a better example from planet images. That's sad #Sentinel2 does not do as well. I hope accurate registration comes soon @CopernicusEU

— Olivier Hagolle (@OHagolle) October 2, 2018

... and even fixed it!

A @CopernicusEU #Sentinel2 satellite view of #Palu before and after the #earthquake.

Because geolocation inaccuracy makes it difficult to compare the two raw images, this animation stabilises the west side of the image to highlight the 5m+ relative shift along the fault: pic.twitter.com/Sa3aDKJNv4

— Dr Robbi Bishop-Taylor (@EarthObserved) October 2, 2018

Only a couple of days after the disaster, a surface displacement map made from Planet images was tweeted by @SotisValkan:

This is the most sharp and linear earthquake surface displacement I've seen since 20130Balochistan eq! More than 20km of major strike-slip displacement for Palu fault segment, Sep 28 M7.5 #earthquake. Optical correlation w/MicMac & @planetlabs imagery pic.twitter.com/23ItZnQUNQ

— Sotiris Valkaniotis (@SotisValkan) October 1, 2018

After the Kaikoura earthquake in New Zealand, it was also possible to see both the ground displacement and the landslides caused by the earthquake in the same Sentinel-2 image:

@davepetley Some landslides are also visible [https:]]

— Simon Gascoin (@sgascoin) November 16, 2016

Earthquakes often trigger landslides. The 2018 Sep 15 earthquake in Hokkaido, Japan literally triggered hundreds of landslides (see also this animation made by @zzsylvester with Planet images) !

On 18 Sep #Sentinel2 captured the many #landslides in Hokkaido, Japan
Before/after image comparison here [https:]] pic.twitter.com/WVaXsucH1K

— Simon Gascoin (@sgascoin) September 17, 2018

In the images above I used a false color band combination with the near-infrared band of Sentinel-2 to highlight the contrast between the detachment zone and the vegetation. Otherwise landslides in temperate regions can be difficult to see in natural colors:

Willow creek landslide by Sentinel-2. Imagery suggests that it occurred between May 11 and 31 @davepetley see here [https:]] pic.twitter.com/Hyt3SUEhsi

— Simon Gascoin (@sgascoin) July 15, 2017

Whereas, with the near infrared band even small landslides can be identified:

#Landslide #Deslizamiento Kantutani near La Paz Bolivia captured by #Sentinel2 @CopernicusEU [https:]] pic.twitter.com/LfqPRSJpXX

— Simon Gascoin (@sgascoin) May 15, 2019

Most of my landslides tweets were inspired by the landslide blog, where the author Dave Petley often comments Planet images. However, Sentinel-2 is also very useful to make quick landslide analyses because of these characteristics:
- 10 m spatial resolution enable to detect most landslides
- near-infrared band enable to distinguish bare earth and mud from vegetation
- systematic acquisitions with a short revisit time enable to create before/after image comparison even in cloudy regions
Planet provide images with similar characteristics and even a better spatial resolution (albeit with a lower image quality). But a very important aspect of the Sentinel-2 mission is the open data policy of the Copernicus program and the availability of efficient tools like the EO-Browser to explore the data with an Internet Browser.

Next part will be about glacier hazards! In the meantime here are more landslides..

"The landslide crossed the river Hítará, damming the river and causing a lake to form above the debris tongue" pic.twitter.com/RMV2uub6c5

— Simon Gascoin (@sgascoin) January 30, 2020

#Sentinel2 shows the extent of the Pizzo Cengalo #landslide
check out @sentinel_hub [https:]] @davepetley pic.twitter.com/s5VV1zpwUP

— Simon Gascoin (@sgascoin) August 26, 2017

The recent 3.5 km long Kurbu-Tash landslide in Kyrgyzstan captured by Sentlnel-2 [https:]] pic.twitter.com/hpAsIttpTQ

— Simon Gascoin (@sgascoin) May 18, 2017

Erratum to "Landslides in Kyrgyzstan captured by Sentinel-2" @davepetley [https:]] pic.twitter.com/5TDBOxjnW5

— Simon Gascoin (@sgascoin) May 23, 2017
17 January 2020
13:44
[VenµS News-4] Multi-angular views, every second day
sur Séries temporelles (CESBIO)
Venµs is a research satellite, and a rather agile one (at least for a microsatellite). It enables us to make experiments useful for users like us who like to study physics of optical measurements (which is an allowed perversion, as fas as I know). We have started last year to acquire multi-angular observation, under 3 different view angles from the same orbit, over one site in California, close to Sacramento. AND, these multi-angular views are available every second day !

This data set is available from Theia website, you will find the data under the site names Gallo, GalloM30, GalloP30:
- Close to nadir view: [https:]]
- Northward view, Venµs looks backward: [https:]]
- Southward view, Venµs looks forward: https://theia.cnes.fr/atdistrib/rocket/#/search?page=1&collection=VENUS&location=GALLOP30
Here is a nice example captured by Gerard Dedieu, on a part of the image. Do you see the change within the colour images ?. The "bacward image is much brighter".

Let's explain a little bit what happens, with a zoom on another prt of the image :

The evaporative towers of this power plant allow to understand what happens. Of course, due to the height of the towers, you probably noticed how their apparent orientation changes with the observation angle. But you probably also noted that the towers colour changes from black to white. In the Northward observation angle, which is close to the backscattering direction, VENµS sees the the sunlit faces of the objects, it is the case for the towers as well as for the plants or the clods over bare soil. This effects explains why the backward image is brighter that the two other ones. This is what we call "directional effects" and this data set is a perfect one to study them !

To summarize, the Gallo site is a unique experiment to study directional effects thanks to VENµS :
- with three viewing directions separated by 30 degrees along the track
- with images acquired every second day
- with a good atmospheric correction performed by MAJA
- and 5m resolution
There are in situ measurements on that site and we will try to access them.
15 January 2020
18:44
[VENµS News 3] Venµs L2A products are now produced by Theia, at 5m resolution
sur Séries temporelles (CESBIO)
1/ New 5m Level 2 products

The Venus L2 production has been moved to a different processing center. From January 2020, the VENµS L2 production is only performed by the MUSCATE processing center, instead of the VIP (VENµS Image Processing) center, with MAJA processor also used for Sentinel-2.

The aim is to improve the integration of Venus products in the MUSCATE center which also produces Sentinel 2, Landsat and other satellite products.

For the users, the main features are:
- Level 2 (atmospheric corrections) are available at 5m ground resolution. Formerly, only Level 1 products (TOA reflectances) were available at 5 m resolution and L2 were provided at 10m resolution. . From February 7^th, 2020, Level 2 (and Level 1) products will be distributed with 5m resolution only
- The new 5m L2 products are identified with an added "_XS_" in the filename. The new products feature different metadata keywords from the previous ones and one independent file per band. The format change iis due to the homogeneisation of formats within MUSCATE. The new format is described here :
[https:]]
- Access to the data does not change : [https:]]
2/ Reprocessing

The full archive is being reprocessed for providing L2 products at 5 m resolution. Another improvement deals with cloud detection. As a consequence you might have more products available over your site for both recent and ancient acquisitions.
18:10
[VENµS news-2] Mission extension and next phases
sur Séries temporelles (CESBIO)
About one year ago, we had submitted a call for ideas for the use of VENµS imager after the end of its nominal phase. CNES also organised a brain storming meeting, which came up with a few ideas, and then our proposals were discussed with our israeli partners. Here is what was agreed between CNES and the Israeli Space Agency :
- VM1 extension: The current scientific imaging mission (VM1) will be extended from the initial date of August 31st, 2020 to the end of October 2020 in order to better cover the end of the main vegetation season in the Northern hemisphere.
=> Therefore the acquisitions will continue until October 2020 over the sites that are currently acquired.
- VM2 From November 2020, VENµS will go down from the altitude of 720 km to the altitude of 410 km. This phase (VM2) of lowering the orbit will take 6 months and will use the Israeli Hall Effect Thruster (IHET).
- VM3 The IHET will be used for 3 months to maintain the 410 km orbit, the revisit period is 2 days. During this phase, only a few images will be acquired for demonstration and technological purposes.
- VM4 Then VENµS will raise up to 560 km. The duration of this phase will be three months.
- VM5 VENµS will stay at 560 km (VM5) for at least 1 year from about November 2021 to November 2022, and hopefully more. On this orbit, the revisit period is one day, the overpass time will be mid-morning. However, not all the current scientific sites can be observed from this orbit. Therefore a new call of proposal for new sites will be issued in the coming weeks.
Clearly this plan is ambitious, but given the good health of the satellite, its "fuel" reserves and the skills of the teams we are confident we will achieve it.

In summary:
- extension of current acquisition until late October 2020, with 2 days revisit
- no image acquisition from November 2020 to November 2021 (except to check the health of the instrument)
- Images every day for scientists, starting from November 2021, for at least one year. But sites are to be renewed partially.
17:49
[Venµs News 1] One month without images
sur Séries temporelles (CESBIO)

There will be no VENµS image acquisition from January 7th to February 8th 2020. This period (called IHET month) is devoted to the test of the Israeli electric thruster (IHET)developed by Rafael Ltd, which is one of the missions of the satellite.

(in order to be kept informed of such events and other news, please subscribe to the RSS feed: [https:]] )
14 January 2020
17:40
Floods in southeast Iran
sur Séries temporelles (CESBIO)

The region of Sistan and Baluchestan was hit by very heavy rain causing widespread flooding near Chabahar city in southeastern Iran.

Sentinel-2 captured the plume of sediments from rivers Bahu Kalat (Iran) and Dasht (Pakistan) in the Gwadar Bay, located in the Gulf of Oman on the maritime border of Pakistan and Iran.

In false colors using the SWIR band of Sentinel-2, flooded areas around the Bahu Kalat river are well visible in turquoise blue.
13 January 2020
9:11
My remote sensing analyses on Twitter. Part 1: floods
sur Séries temporelles (CESBIO)

S. P. Hicks (@seismo_steve) wrote for Nature Geoscience a nice piece "Geoscience analysis on Twitter" where he shares his experience on using Twitter for the real time analysis of geohazards.

Actively using Twitter for science may be perceived by some as a significant time-sink — the demands on individual researchers and their institutions at times of newsworthy events can be very high — but it is ultimately rewarding.

I fully agree with his viewpoint and I would like to share my experience here too.

As a trained hydrologist I'm inclined to tweet about water-related hazards like floods and droughts. However, I have mostly tweeted on floods because flooded surfaces can be easily characterized by remote sensing using a simple before/after image comparison. In particular, water surfaces are strong reflectors of the electromagnetic waves emitted by SAR satellites such as Sentinel-1. In addition SAR sensors can see through the clouds which is a key asset for near-real time analysis during flooding events.

(1/n) Annotated map of the Xe-Namnoi dam breach. The flood came from a saddle dam breach and followed the Vang Ngao river down to the villages. @CopernicusEU #Laosdam pic.twitter.com/Vtay6iXrCY

— Simon Gascoin (@sgascoin) July 25, 2018

I remember that I made this analysis of the Xe-Namnoi dam failure overnight, just hours after the first Sentinel-1 image of the flood was available on the Sentinel Hub. My tweets triggered many interesting discussions with other flood specialists, civil engineers and NGO staff, but also with close colleagues at LEGOS who provided water elevation data from Sentinel-3 to expand the analysis. My animation and other images were featured in ESA's website (here and here, without acknowledging my work though). This post was the most visited in 2018 (11,371 visits, 18% of the 2018 visits).

SAR images are great, but optical sensors like Sentinel-2 can also provide useful observations of flooded areas as shown below in the case of the recent flood in southwest France:

Les inondations dans le Sud-Ouest de la France vues par le satellite #Sentinel2 [https:]] @sudouest @ladepechedumidi @CNES @ESA_fr pic.twitter.com/vAA3SxJ4WP

— Simon Gascoin (@sgascoin) December 16, 2019

To go a bit further with optical data, the NDWI is a simple a simple multispectral index, which makes it is easy to map the extent of the water surface:

Using the NDWI to map #flooding in the Central US
Blog post: [https:]] @CopernicusEU @esascience pic.twitter.com/dQmPT8u57P

— Simon Gascoin (@sgascoin) May 4, 2017

I used the same method to show the Oroville dam lake extent when it was under the threat of spillway failure, causing the evacuation of 188'000 northern California residents.

Lake Oroville in January 2016 vs. 2017 from #Sentinel2 imagery #OrovilleDam @CopernicusEU [https:]] pic.twitter.com/R5nusxOpJH

— Simon Gascoin (@sgascoin) February 13, 2017

In the case of the dramatic 2019 Omaha flood, I tried to go beyond the visualization of the flooded areas ..

Just a comparison with last year helps figure out the huge extent of the flooded areas around Omaha, Nebraska #NebraskaFlood #OmahaFlood Imagery: @CopernicusEU #Sentinel2 pic.twitter.com/8XFjSnWpoL

— Simon Gascoin (@sgascoin) March 17, 2019

.. using additional images and weather data it was obvious that the flood was caused by a rain-on-snow event:

1/2 #Sentinel2 shows extensive snow melt around Omaha, Nebraska.. in only 5 days! #nebraskaflooding pic.twitter.com/izaPfoRoQX

— Simon Gascoin (@sgascoin) March 17, 2019

To conclude this post, an interesting fact is that my least successful tweet about flood is the one below (only 1 "like" at the time of writing), probably because I was a newbie in Twitter at that time. However, the linked article is one of the most visited page (if not the most visited) in this blog. Twitter isn't the only way to reach out to people!

Mapping flooded areas using Sentinel-1 in Google Earth Engine [https:]]

— Simon Gascoin (@sgascoin) June 10, 2016

The next article will be about landslides [here]. In the meantime, here are some more flood tweets:

In France...

Image prise par le satellite #Sentinel2 des panaches de sédiments dans la Méditerranée entre Béziers et Agde après les fortes pluies qui se sont abattues sur le sud de la France les 22 et 23 octobre. Source: @CopernicusEU @lindependant @ladepechedumidi pic.twitter.com/hB1lvLBDkJ

— Simon Gascoin (@sgascoin) October 27, 2019

Italy...

Flood in #Livorno #Sentinel2 on Sep 04 and Sep 11 pic.twitter.com/9pqbONLIkX

— Simon Gascoin (@sgascoin) September 12, 2017

Mozambique...

Flooded areas near Beira on 19 Mar 2019 computed from #Sentinel1 #MozambiqueFloods @CopernicusEMS @CopernicusEU pic.twitter.com/2OdE6Ixzig

— Simon Gascoin (@sgascoin) March 20, 2019

Australia...

The after/before of the 2017 Queensland floods in #Sentinel2 imagery [https:]] #QLDfloods @BOM_Qld pic.twitter.com/0SozZ3VzzU

— Simon Gascoin (@sgascoin) May 2, 2017

USA...

Flooded areas near Highway 90 and Dayton TX (July 24 vs. Aug 29) #Sentinel1 #HarveyFlood pic.twitter.com/z3S1SvYcE5

— Simon Gascoin (@sgascoin) August 29, 2017

India...

The NDWI layer in the @sentinel_hub reveals the extent of the flooded areas in Kerala @CopernicusEU @ESA_EO #KeralaFloods pic.twitter.com/qNKbgM2fRF

— Simon Gascoin (@sgascoin) August 28, 2018

Spain... Sometimes remote sensing is not useful:

There will be .. flood!
The latest report just released by the Ebro river basin agency shows that the 1st April snow water equivalent is very large in the #Pyrenees catchments. It is actually the highest value since the beginning of the record in 2002. pic.twitter.com/FGPUwgcIQy

— Simon Gascoin (@sgascoin) April 3, 2018

Prediction was correct...

Fotos: El río Ebro da una tregua tras una noche en vela [https:]] Las imágenes de la crecida del río Foto de @Charlie_photo pic.twitter.com/3Cb1lpCEdB

— EL PAÍS Fotografía (@elpais_foto) April 14, 2018
09 January 2020

22:05

The audience of the Multitemp blog pathetically drops: -9%

sur Séries temporelles (CESBIO)

In this period of festivities, feasts, gifts and good resolutions, the contrast is strong with the gloomy balance sheet of the year 2019 for our blog, which for the first time shows a pathetic drop in traffic: -9%.

Faced with such a worsening situation, the editorial board must react and take remedial action, and is expected to point fingers and knock heads. Our investigation quickly revealed the name of the culprit; a few years ago, in order to boost our publication and increase its support, we hired a social networking star, a first of its kind in communications, Simon Gascoin. But is it a drop in motivation ? a small stroke of fatigue? Simon did not perform as well as expected in 2019. Disasters, fires, floods, avalanches have occurred that have not been mentioned in this blog, and we have not renewed the previous year's BUZZ and its 11,371 visits (18% of the 2018 visits). However, Simon noticed the problem at the end of the year and managed to bring up our statistics at the end of 2019 with another flood, in the South West of France.

Comparison of the number of visits to our blog in 2019 and 2018

But in these times of Darwinism in the research community, where researchers are selected for their "excellence" with a short-term vision and where "losers" are left without a budget, some people might have thought that we would give in to the trend. That's knowing us badly! This blog is interested in the whole time series, and not a couple of isolated dates, so we decided to continue our efforts without knocking heads, but hoping that many other contributors will want to join us, and, who knows, make the buzz. This blog is open, writing an article doesn't take a lot of time (especially the second, and the third...), so we're waiting for you, whether you're CESBIO members or not.

We also have a little excuse, the blog had to move (from the university servers to the Observatoire Midi Pyrénées servers), and if it has been renovated and has an up-to-date interface, the move may have caused some loss of links from other sites... In short, we should start again from better in 2020.

So here is the list of the most read pages this year, after having removed the lists of articles, like of course the home page, the Sentinel-2 or Landsat pages, the authors' names (I'm still ahead of Simon, but it won't last... )

Mapping flooded areas using Sentinel-1 in Google Earth Engine	5141
Radiometric quantities : irradiance, radiance, reflectance	4457
A python module for batch download of Sentinel data from ESA	3723
MACCS/MAJA, how it works	3130
Les séries temporelles de niveau 3A de Sentinel-2	2043
Sentinel-2 + OpenStreetMap = ?	1954
Les inondations dans le Sud-Ouest vues par satellite	1950
The product level names, how they work ?	1848
The Sentinel-2 tiles, how they work ?	1734
S1-Tiling, on demand ortho-rectification of Sentinel-1 images on Sentinel-2 grid	1695
THEIA’s L2A product format	1678
How to automatically download Sentinel data from PEPS collaborative ground segment	1589
La première carte d'occupation des sols 2016 de la France avec Sentinel-2	1557
S2 Tile Viewer/ Voir les tuiles Sentinel-2	1299
Khumbu icefall in 4D	1073

So what may we conclude ?

the distribution of small free software is at the top of the list (and we get several questions a week...)
Simon's geophysics articles (with the associated ad on social networks) attract crowds
the description of the MAJA channel is nice, but half of my articles point to this page
the example on how to use Google Earth Online attracts much more than the articles that denounce its dangers (it's sad)
the "How It Works" series continues to be a success...
Level 3A products were very popular (whether in France, Spain or Italy, Maghreb or Sahel).
the news of Theia's production and product formats are well tracked
two articles on Sentinel-1 ranked in the top 15 (and a third on deforestation ranked in the top 30)
for the first time, a series of VENµS images enters the top 15, with the magnificent video on the Everest glacier created by Simon

05 January 2020

22:55

La fréquentation du blog Séries temporelles dévisse : -9% !

sur Séries temporelles (CESBIO)

En cette période de réjouissances, d'agapes, de cadeaux et de bonnes résolutions, le contraste est fort avec le sombre bilan de l'année 2019 pour notre blog dont la fréquentation affiche, pour la première fois, une pathétique baisse de sa fréquentation : -9 %.

Devant une telle dégradation de la situation, le comité de rédaction se doit de réagir et de prendre des mesures de redressement, et l'on s'attend à le voir désigner des coupables et faire tomber des têtes. Notre enquête a rapidement révélé le nom du fautif ; nous avions embauché il y a quelques années, pour dynamiser notre publication et accroître son aidience, une star des réseaux sociaux, un premier de cordée de la communication, Simon Gascoin. Mais est-ce une baisse de la motivation, un petit coup de fatigue ? Simon n'a pas eu en 2019 le rendement escompté. Des catastrophes, des incendies, des inondations, des avalanches ont eu lieu qui n'ont pas été citées dans ce blog, et nous n'avons pas renouvelé le BUZZ de l'année précédente, et ses 11 371 visites (18 % des visites de 2018). Simon s'est cependant aperçu du problème en fin d'année et a réussi à faire remonter nos statistiques en fin d'année 2019 avec une autre inondation, dans le Sud Ouest de la France.

Comparaison du nombre de sessions de visite du blog, entre 2018 et 2019.

Mais en ces temps de darwinisme dans le milieu de la recherche, où l'on sélectionne, avec une vision à court terme, les chercheurs par leur "excellence" et où l'on abandonne sans budget les "losers", certains ont pu penser que nous allions céder à la tendance. C'est mal nous connaître ! Ce blog qui s'intéresse à la série temporelle en son entier, et pas à un couple de dates isolées, et nous avons donc décidé de poursuivre nos efforts sans faire tomber de têtes, mais en espérant que de nombreux autres contributeurs voudront nous rejoindre, et, qui sait, faire le buzz. Ce blog est ouvert, l'écriture d'un article ne prend pas beaucoup de temps (surtout le deuxième, et le troisième...), et nous vous attendons donc, que vous soyez membres du CESBIO ou pas.

Nous avons aussi une petite excuse, le blog a du déménager (des serveurs de l'université aux serveurs de l'Observatoire Midi Pyrénées), et s'il a été rénové et présente une interface au goût du jour, le déménagement a pu occasionner quelques pertes de liens en provenance d'autres sites... Bref nous devrions repartir de plus belle en 2020.

Voici donc le palmarès des pages les plus lues cette année, après avoir ôté les listes d'articles, comme bien sûr la page d'accueil, les pages Sentinel-2 ou Landsat, les noms d'auteurs (je devance encore Simon, mais ça ne va pas durer)...

Mapping flooded areas using Sentinel-1 in Google Earth Engine	5141
Radiometric quantities : irradiance, radiance, reflectance	4457
A python module for batch download of Sentinel data from ESA	3723
MACCS/MAJA, how it works	3130
Les séries temporelles de niveau 3A de Sentinel-2	2043
Sentinel-2 + OpenStreetMap = ?	1954
Les inondations dans le Sud-Ouest vues par satellite	1950
The product level names, how they work ?	1848
The Sentinel-2 tiles, how they work ?	1734
S1-Tiling, on demand ortho-rectification of Sentinel-1 images on Sentinel-2 grid	1695
THEIA’s L2A product format	1678
How to automatically download Sentinel data from PEPS collaborative ground segment	1589
La première carte d'occupation des sols 2016 de la France avec Sentinel-2	1557
S2 Tile Viewer/ Voir les tuiles Sentinel-2	1299
Khumbu icefall in 4D	1073

Qu'en déduire ?

la distribution de petits logiciels libres occupe le haut de la liste (et plusieurs questions par semaine...)
les articles de géophysique de Simon (avec la pub associé sur les réseaux sociaux) attirent les foules
la description de la chaîne MAJA plait, mais la moitié de mes articles pointent sur cette page
l'exemple d'utilisation de Google Earth Online attire bien plus que les articles qui dénoncent ses dangers (c'est triste)
la série "Comment ça marche" continue à avoir du succès
les produits de Niveau 3A ont beaucoup plu (que ce soit en France, en Espagne ou en Italie).
les nouvelles de la production par Theia et les formats des produits sont bien suivis
deux articles sur Sentinel-1 se classent dans les 15 premiers (et un troisième, sur la déforestation figure dans les 30 premiers)
pour la première fois, une série d'images VENµS entre dans les 15 premiers, avec la magnifique vidéo sur le glacier de l'Everest créée par Simon

04 January 2020
19:23
Best wishes for 2020
sur Séries temporelles (CESBIO)
=>

This is the eighth time this blog wishes you a Happy New Year! We send you and your loved ones our best wishes for personal, family and professional happiness for the year 2020.

On our program for this new year:
- the beginning, at last, of the ortho-rectification with ground control points of the Sentinel-2 data. Originally planned 1 year after the launch of Sentinel-2A, it has only been announced, for the last 6 months, to start within 3 months (yes, this sentence may need to be read twice).
- The Sentinel-2C satellite is expected to be almost ready later this year, and this will be an opportunity to call for an early launch to improve the system's revisit frequency. It would be very useful, especially in periods of relatively cloudy weather, to have three passes every 10 days.
- Version 4 of MAJA will enter into service, and with it, finally, the processing of data with the types of aerosols from CAMS.
  On the research side, we will work on modelling and improving the correction of adjacency effects, and we will try to better understand what causes the appearance of tile limits observed on the mosaics of our Level 3A products.
- We will install, on the Lamasquère site, a ROSAS surface reflectance measuring station, similar to those at La Crau and Gobabeb, to measure the accuracy of Level 2A products in real and difficult conditions on sites that are not entirely uniform and have variable reflectances over time.
- The B,C,D phase of the Trishna satellite (Thermal Infrared at 60m resolution) starts. The French-Indian mission (CNES, ISRO), should be launched in 2024 or 2025.
- We will start the production of the COSIMS project of the European Environment Agency, to follow, with Sentinel-2, the snow cover of Europe. The project is led by Magellium and CESBIO's methods (MAJA and LIS) will be used.
- 2020 will also be a year of transition of the CESBIO management, during which Mehrez Zribi will gradually replace Laurent Polidori as director, and where I will try to replace Mehrez as "Observation System" team leader. Gilles Boulet will still lead the modelling team. Three transversal thematic axes will be created, devoted to water (Anim. Olivier Merlin), eco-agrosystems (Anim. Eric Ceschia) and regional vegetation dynamics (Anim. Lionel Jarlan). We will also have two assistant directors, Valérie Démarez and Lionel Jarlan.
  At CNES, the development of the hydrological treatment centre project (Hysope-2) should begin. In the framework of Theia and in anticipation of the launch of the NASA/CNES SWOT satellite, HYSOPE 2 should notably generalize a large part of the MUSCATE treatments to Europe and Africa at least, hopefully to the world.
And of course :
- we will continue the animation of this blog and try to recruit new editors to assist the two chief editors, O.H. and S.G. before they start rambling too much (if it's not already too late ...) ! If you are a user of Earth observation time series, you will certainly have some interesting things to tell in this blog. Please contact us!
19:08
Meilleurs voeux pour 2020 !
sur Séries temporelles (CESBIO)
=>

C'est la huitième fois que ce blog vous souhaite la bonne année ! Nous vous adressons, à vous et vos proches, nos meilleurs souhaits de bonheur personnel, familial et professionnel pour l'année 2020.

Au programme de cette nouvelle année :
- les débuts, enfin, de l'ortho-rectification avec points d'appuis des données Sentinel-2. Prévue à l'origine 1 an après le lancement de Sentinel-2A, elle n'est plus annoncée, depuis 6 mois, que dans un délais de 3 mois (oui, cette phrase peut nécessiter une relecture)
- le satellite Sentinel-2C devrait être presque prêt en cours d'année, et ce sera l'occasion de réclamer un lancement anticipé pour améliorer la répétitivité du système. Il serait très utile, notamment lors des périodes assez nuageuses, d'avoir trois passages tous les 10 jours.
- la version 4 de MAJA entrera en service, et avec elle, enfin, le traitement des données avec les types d'aérosols issus de CAMS.
- côté recherche, nous travaillerons sur la modélisation et l'amélioration des effets d'environnement, et nous essaierons de mieux comprendre à quoi sont dues les limites de tuiles que l'on observe sur les mosaïques de nos produits de Niveau 3A.
- nous allons installer, sur le site de Lamasquère, une station de mesure de réflectances de surface ROSAS, similaire à celles de La Crau et Gobabeb, pour vérifier la qualité des produits de niveau 2A en conditions réelles et difficiles sur des sites pas entièrement uniformes et de réflectances variables avec le temps.
- la phase B,C,D du satellite Trishna (Infra-rouge thermique à 60m de résolution) démarre. La mission Franco-Indienne (CNES, ISRO), devrait être lancée en 2024 ou 2025.
- nous commencerons la production du projet COSIMS de l'Agence Européenne de l'Environnement, pour suivre, avec Sentinel-2, la couverture neigeuse de l'Europe. Le projet est mené par Magellium et ce sont les méthodes du CESBIO (MAJA et LIS) qui y seront utilisées
- 2020 sera également une année de transition de la direction du CESBIO, au cours de laquelle Mehrez Zribi remplacera progressivement Laurent Polidori, et où j'essaierai de remplacer Mehrez en tant qu'animateur de l'équipe "Système d'observation". Gilles Boulet animera toujours l'équipe modélisation. Trois axes thématiques transverses verront le jour, consacrés à l'eau (Anim. Olivier Merlin), les éco-agro-systèmes (Anim. Eric Ceschia) et les dynamiques régionales de la végétation (Anim. Lionel Jarlan). Nous aurons aussi deux directeurs adjoints, Valérie Démarez et Lionel Jarlan.
- Au CNES, le développement du projet de centre de traitement hydrologique (Hysope-2) devrait débuter. Dans le cadre de Theia et en prévision du lancement du satellite NASA/CNES SWOT , HYSOPE 2 devrait notamment généraliser une bonne partie des traitements de MUSCATE à l'Europe et l'Afrique au moins, au monde si tout va bien.
Et bien sûr :
- Continuer l'animation de ce blog et recruter de nouveaux rédacteurs pour seconder les deux rédac'chefs, O.H. et S.G. avant qu'ils ne radotent trop (si ce n'est pas déjà trop tard ...) ! Si vous êtes utilisateurs de séries temporelles d'observations de la terre, vous aurez certainement des choses intéressantes à raconter dans ce blog. N'hésitez pas à nous contacter !
20 December 2019

10:20

Sentinel-2 also sees grassland taxonomic diversity !

sur Séries temporelles (CESBIO)

Sentinel-2 image time series are widely used to map land cover, see for instance here or here. However, they can be used to extract other relevant information about the observed landscapes.

In a recently published article ( [https:]] ), we have used fifteen months of Sentinel optical and radar data to predict taxonomic and functional diversity at the pixel scale.

We conducted field survey on various type of grasslands (from semi-natural to temporary ones) to collect plant diversity. The figure below shows the histogram of the Shannon index. A high Shannon index value indicates high biodiversity.

The field plots were located among 4 different Sentinel-2 tiles. We used iota2 to manage and pre-process Sentinel 1 and Sentinel 2 time series. Then we investigated several machine learning regression algorithms to relate the pixel’s spectro-temporal profile to these diversity indices. Among the several diversity indices tested, Simpson and Shannon indices were best predicted with a coefficient of determination around 0.4 using a Random Forest predictor and Sentinel-2 data. The use of Sentinel-1 data was not found to improve significantly the prediction accuracy. The table below shows the best results obtained (R-IR means that the features used were the red and near-infra red bands, S2 means all the spectral bands at 10m, S2-NDVI means that the NDVI profiles were added and S1-S2 means VV, VH and HH S1 polarization and S2 spectral bands were used together).

Data	Model	R2	Index
R-IR	Random Forest	0.45 (+- 0.13)	Simpson
S2	Random Forest	0.44 (+- 0.08)	Simpson
R-IR	Random Forest	0.43 (+- 0.13)	Shannon
S2-NDVI	Random Forest	0.43 (+- 0.17)	Simpson
S2-NDVI	Random Forest	0.41 (+- 0.17)	Shannon
S1-S2	Random Forest	0.41 (+- 0.13)	Simpson
S2	Random Forest	0.40 (+- 0.14)	Shannon

Thanks to iota2, we then predicted the Simpson index for all the grasslands of the region, at the pixel scale. It was the first time such predictions were performed at a resolution of 10m per pixel. Figures below show some predicted parcels (high resolution ortho image & Simpson index predicted).

The contribution of this work does not lie in selecting the best set of dates nor the best machine learning regression methods but in assessing the potential of Sentinel-1 and -2 dense times series as explanatory variables for the prediction of grasslands biodiversity indices. We believe that our results, using large scale data with various agricultural practices for different meteorological and topographic conditions, demonstrate the capacity of such data to monitor grasslands from an ecological viewpoint. In particular, intra-parcel variability was highlighted in this work and can be monitored over large areas. Yet, accuracy scores are still limited and methodological works need to be done to improve them.

9:00
Quelle est la probabilité d'avoir de la neige à Noël dans votre commune ?
sur Séries temporelles (CESBIO)

La carte ci-dessous donne la probabilité d'avoir de la neige d'après les observations des trois dernières années.

Voir en plein écran

Pour produire cette carte j'ai calculé le nombre de fois où il y a eu de la neige le 25 décembre d'après les cartes d'enneigement Theia. Ces cartes sont établies à partir des images Sentinel-2 et Landsat-8 à une résolution de 20 m. Comme il n'y pas forcément une image disponible le 25 décembre, j'ai utilisé les cartes interpolées au pas de temps journalier pour faire ce calcul. Cela donne un score compris entre 0 et 3 depuis le 25 décembre 2016. Ensuite j'ai agrégé par commune en affectant à chaque commune la valeur majoritaire des pixels dans l'emprise du polygone de la commune. Pour cet exercice j'ai limité la couverture spatiale aux massifs des Alpes, du Jura et des Pyrénées.

Pour ce post j'ai été inspiré par cette carte de la NOAA :

For those who are dreaming of a white Christmas. from MapPorn

Joyeux Noël !
18 December 2019
17:31
Theia's christmas present: Sentinel-2 L2A and L3A on all Sahel zone are now available in NRT
sur Séries temporelles (CESBIO)

=>

I take the advantage of this post to wish you all nice holidays, a merry christmas and a happy new year !

It will have taken us almost a year to process the 300 Sentinel-2 tiles in Theia's area in Sahel, from the beginning of January 2017. This processing ends next week, and all the tiles will be now processed in near real time. The region of Conakry in Guinea, and the east of the zone in Chad are the two last ones, and they will join the stream around Christmas.

Level 2A data can be selected using a request such as the one below (in the case of Niger, for instance):

[https:]]

Santa Claus (personified by MUSCATE exploitation team) offered us a nice Christmas present: the production of Level 3A for the whole Sahel zone has started, except for Conakry region and the Easternmost region in Chad. You may have a look to the Mosaic by clicking on the image below:
Click on image to zoom to 50m resolution..

Because of the size of the region, we processed a mosaic at 50m, but the Level3A monthly syntheses are available at 10m. Unlike the mosaics of the Level 3A syntheses obtained in theau Maghreb, France, Italy or the Iberian Peninsula, the seams between tiles or between orbits are quite visible in the southern area of the image. We do not yet know how to explain this behaviour well. The syntheses are carried out with the WASP software, which uses as input the level 2A data (atmospheric corrections and cloud masks) provided by MAJA. The period is at the end of the rainy season and becase of that we have had much less cloud free input data, so some of the defects are due to differences in dates on either side of the limits. The defects may also be related to atmospheric corrections, which assume that we have the same type of aerosols all over the world (we should change this in a few months), and we will also have to check the directional effects correction model, which is also constant worldwide.

To access the Level 3A products, available at 10m de résolution, please use a request as follows (here for Burkina Faso) :

[https:]]
16:27
Special Noel THEIA: passage au fil de l'eau pour les traitements Sentinel-2 L2A et L3A sur la zone Sahel
sur Séries temporelles (CESBIO)

=>

Je profite de cet article pour vous souhaiter à tous de bonnes vacances, un joyeux Noel, et une excellente nouvelle année 2020 !

Il nous aura fallu près d'un an pour traiter les 300 tuiles de la zone Sahel de Theia, à partir de début Janvier 2017. Ce traitement s'achève ces jours-ci, il doit nous rester juste deux ou trois mois de données à traiter, avant de rejoindre le fil de l'eau sur les deux dernières zones, la région de Conakry en Guinée, et l'Est de la zone au Tchad. Ces deux zones rejoindront le fil de l'eau aux alentours de Noël.

Les données de Niveau 2A peuvent être téléchargées en utilisant une requête telle que celle ci-dessous (pour le Niger, par exemple):

[https:]]

Le père Noël (ou plutôt le père Bernard Specht et l'équipe d'exploitation MUSCATE) nous a par ailleurs offert un magnifique cadeau de Noël : le démarrage de la production des produits de Niveau 3A, les synthèses mensuelles sans nuages, sur la totalité de la zone, excecppté, encore, Conakry et le zone UTM 33 (Tchad, Cameroun). Vous pouvez jeter un coup d'oeil aux données en cliquant sur l'image ci-dessous.
Cliquer sur l'image pour zoomer jusqu’à 50m de résolution.

En raison de sa taille, la mosaïque a été produite à 50m de résolution, mais les données de niveau 3A sont disponibles à 10m. Contrairement aux mosaïques des synthèses de niveau 3A obtenues au Maghreb, en France, en Italie ou dans la péninsule ibérique, les coutures entre tuiles ou entre orbites sont assez visibles sur la zone sud de l'image. Nous ne savons pas encore bien expliquer ce comportement. Les synthèses sont réalisées avec le logiciel WASP, qui utilise en entrée les données de niveau 2A (corrections atmosphériques et masques de nuages) fournies par MAJA. La période se situe à la fin de la saison des pluies et nous avons eu beaucoup moins de données en entrée, une partie des défauts sont donc dus à des différences de dates de part et d'autre des limites. Les défauts peuvent être aussi liés aux corrections atmosphériques, qui supposent que nous avons le même type d'aérosols partout dans le monde (nous devrions changer cela dans quelques mois), et il faudra que nous vérifiions aussi le modèle de correction des effets directionnels, qui est lui aussi constant dans le monde entier.

Pour télécharger les produits de Niveau 3A, disponibles à 10m de résolution, utiliser une requête comme ci-dessous (ici pour le Burkina Faso) :

[https:]]
16 December 2019
0:50
Les inondations dans le Sud-Ouest vues par satellite
sur Séries temporelles (CESBIO)

La Garonne a largement débordé entre Agen et Marmande comme le montre la comparaison ci-dessous entre une image prise par le satellite Sentinel-2 en décembre 2018 et une autre prise ce dimanche 15 décembre 2019 à la faveur d'une éclaircie (plein écran, animation):

Les inondations sont aussi visibles le long de l'Adour entre Dax et Bayonne (plein écran, animation):

Ces images ci-dessus sont des compositions colorées en "fausses couleurs" qui utilisent les propriétés multispectrales des images Sentinel-2 pour faire ressortir le contraste entre la végétation et l'eau de surface souvent boueuse. Voici ce que donne l'image du 15 décembre en couleur naturelle centrée sur Dax comparée à une composition colorée des bandes 11/8/3 (SWIR=1610 nm/NIR=842 nm/Green=560 nm) :

Enfin on peut comparer avec l'image acquise le même jour par le satellite radar de la flotte Copernicus, Sentinel-1. Les zones sombres correspondent aux surfaces en eau libre sur lesquelles rebondissent les ondes radar (cliquer ici pour voir en plein écran) :

Photo en tête d'article : La Garonne en crue à Tonneins (crédits: Amandine Gasparotto). Pour en savoir plus: [https:]]
13 December 2019

22:21

Sentinel-2 L2A (MAJA) and L3A (WASP) products, produced by DLR, available over Germany as WPS

sur Séries temporelles (CESBIO)

The DLR starts distributing Sentinel-2 L2A products produced with MAJA over Germany

In September 2019, the German Aerospace agency, DLR, started delivering Level-2A products (surface reflectances with a good cloud mask), using the MAJA L2A processor, to which DLR contributed with CNES and CESBIO. The system received a major upgrade recently: it now includes L3A data (seamless monthly syntheses generated with our WASP processor), and both L2A and L3A are accessible with a browser and a WCS interface (but don't ask me how WCS works, or even better, please explain !). If you are old fashioned as I am, you may also download the data.

If you want to discover the L2A products, use the following link : [https:]] , while for the L3A products, use the following link : https://geoservice.dlr.de/web/maps/sentinel2:l3a:wasp

And finally, to download L2A, use [https:] , data are sorted by tile number, and then year, month and date.

Here is a nice example of a level 3A mosaic is provided below, with a zoom to DLR Oberpfaffenhofen site, near Munich, where the Earth Observation Center (EOC) is, and where Pablo d'Angelo integrated MAJA and WASP to the EOC production center, and Torsten Heinen integrated it to the EOC Geoservice portal. Congrats to this great team !

The DLR team produced all the monthly composites since June 2015. At that time, and until July 2017, we had only one Sentinel-2 satellite, and the data gaps due to clouds are much more frequent. In 2019, Germany was almost cloud free throughout the whole year, except the winter months. Here is an animation over South Bavaria from February to October 2019, Munich is at the centre top of the image.

What I also like a lot with the representation chosen by DLR is the possibility to select two annex data sets provided with the syntheses, the flags and the averaged dates (click on images to zoom). This representation should be very useful, and I should do the same in my own representation. You may notice that the extension of flagged pixels is greater than the extension of apparently cloudy pixels. When a pixel has been cloudy for the whole period, WASP provides the minimal reflectance in the blue, which tends to select the thinnest cloud cover.. or the thickest cloud shadow.Moreover, as we perform a cirrus cloud correction, some of the clouds may disappear, but we are not confident enough on the correction to unflag these pixels.

Some coloured artefacts still appear in the level 3A where tiles overlap. This is a bug which is not present in the syntheses, but only in the viewer. DLR colleagues are currently hunting for that bug.


Weighted average surface reflectance, Nov 2019, with some remaining clouds	You may check that these clouds are flagged as such	This filed gives you the weighted average of dates, which may vary significantly within the image

0:05
Persistent snow cover area in western Europe
sur Séries temporelles (CESBIO)

These maps show the area that remained covered by snow from May 01 to Sep 01 over the past three years. It was computed from Theia Sentinel-2 snow products at 20 m resolution to capture the smallest snow patches.

Persistent snow cover area from May 01 to Sep 01. Each dot corresponds to the total area within a region of 20 km by 20 km, i.e. the aggregation of 1 million pixels.
In average there are about 1300 km2 (Lake Genava: 680 km2) which remain snow-covered until September (less than 1% of the Alps, depending on the definition of the Alps). The persistent snow cover area was highest in 2018, reflecting the exceptional snow accumulation in west Europe mountains this year.

The spatial distribution of the persistent snow is stable from one year to another. In fact, it is mostly found on glacier, as shown here over the Bernina range near St Moritz for 2018.
Persistent snow map and Randolph Glacier inventory 6.0
This is normal: if the snow persisted every year at the same place, it would become a glacier eventually.

For the next Snow & Ice Copernicus service, the European Environmental Agency requested the distribution of a map indicating the permanent snow line (area) for each year. The algorithm to compute this layer called "PSL" is straightforward (see the script in the LIS repo). The PSL corresponds to the pixels which have a snow probability of one over a hydrological year, i.e. pixels which were always detected as snow or cloud (no data). Because the snow detection can fail during the winter season due to shadows in steep slopes, and also to save CPU time, I reduced the period of computation to May 01 - Sep 01.
06 December 2019
18:43
How to find the closest snow to you?
sur Séries temporelles (CESBIO)

If you live in Pamiers near Toulouse, France, where is the snow-covered place that is the fastest accessible by car or by bike?

There is a beautiful snow map of the Pyrenees mountains near Pamiers from a Sentinel-2 image that was acquired yesterday (at the time of writing) [link to the data].
Theia snow cover map on 19 Nov 2019
In QGIS with the OpenRouteService API it is possible to calculate isochrones for a car trip from Pamiers:
This map shows the areas accessible by car from Pamiers in a given period of time (yellow 30-40 min, orange 40-50 min, red 50-60 min)
Finally if we make a spatial join between this map and the vertices of the snow polygons, we obtain the following map:

You can see in yellow the snow-covered places that are the fastest accessible according to OpenRouteService. Here is a zoom near Foix, there is a small road to reach the snow half an hour from Pamiers!

There are other APIs like OpenRouteService, for example that of Google enables to integrate traffic density in the calculation, but this service is not free. Beware that the time to mount the chains on the tires of the car is not taken into account ...

You can even change the mode of transportation to calculate the travel time! Here the travel time is calculated for a cyclist from the city center of Foix.
05 December 2019
23:49
20 m resolution snow maps, why bother?
sur Séries temporelles (CESBIO)
"Snow varies." (Snow scientist proverb)

Sentinel-2 enables to generate snow maps at 20 m spatial resolution. But what's the use of such high resolution snow maps? Because one picture is worth a thousand words, I resampled some Sentinel-2 images of snow-covered regions to coarser resolutions using the average reflectance of the contributing pixels. I did not use snow maps but true color Sentinel-2 composites to better show the context of the selected regions, but it should be easy to mentally compute the snow maps from these examples. The 500 m resolution is close to the resolution of MODIS, Sentinel-3, VIIRS.

First example, a ski resort in the Dolomites, Corvara (Italy).

Second example, urban and agricultural landscape in Saint-Jean-sur-Richelieu, Quebec (Canada)

Third example, in the Sierra Nevada National Park

There are better arguments in the scientific literature. In hydrology, the spatial variability of patchy snow cover area has an impact on the melting rates [1]. In ecology, the spatial patterns of the snow cover duration at high resolution is an important predictor of the plant biodiversity [2]. The drawback of high resolution snow maps is the high computing cost to process them. However, previous studies have shown that snow reanalyses assimilating 90 m resolution maps of the snow cover fraction can be done at the scale of entire mountain ranges [3].

References
1. Liston, G.E. (1995) "Local advection of momentum, heat, and moisture during the melt of patchy snow covers." Journal of Applied Meteorology 34.7, 1705-1715.
2. Carlson, B.Z.; Choler, P.; Renaud, J.; Dedieu, J.P.; Thuiller, W. (2015) Modelling snow cover duration improves predictions of functional and taxonomic diversity for alpine plant communities. Annals of Botany, 116, 1023–1034.
3. Margulis, S.A.; Girotto, M.; Cortés, G.; Durand, M. (2015) A particle batch smoother approach to snow water equivalent estimation. Journal of Hydrometeorology, 16, 1752–1772.
28 November 2019
7:58
Juste une meilleure image de couverture pour l'article de Simon ci-dessous
sur Séries temporelles (CESBIO)

Juste une meilleure image de couverture pour l'article de Simon ci-dessous
21 November 2019
3:22
Theia vous aide à trouver la neige la plus proche de chez vous
sur Séries temporelles (CESBIO)

Si vous habitez à Pamiers, où se trouve la neige la plus rapidement accessible en voiture ?

Theia nous fournit une belle carte de l'enneigement du 19 novembre 2019 dans les Pyrénées ariégeoises (lien vers les données).
Surface enneigée le 19 Nov 2019 d'après le produit neige Theia Sentinel-2
Dans QGIS avec l'API OpenRouteService on peut calculer les isochrones pour un trajet en voiture depuis Pamiers :
Cette carte indique les zones accessibles en voiture depuis Pamiers en un laps de temps donné (jaune 30-40 min, orange 40-50 min, rouge 50-60 min)
Enfin si on effectue une jointure spatiale entre cette carte et les sommets des polygones "neige" du produit Theia, on obtient la carte suivante.

On peut voir en jaune les endroits enneigés les plus rapidement accessibles selon OpenRouteService ! Voici un zoom vers Foix, il y a une petite route pour aller toucher la neige à une demi-heure de Pamiers !

Il existe d'autres APIs comme OpenRouteService, par exemple celle de Google permet d'intégrer les embouteillages dans le calcul, mais ce service est payant. Il y a aussi celle de l'IGN en France qui est gratuite pour les particuliers. Attention le temps de monter les chaînes sur les pneus de la voiture n'est pas pris en compte...

PS. Comme suggéré par Samuel Morin, on peut changer le mode de transport pour calculer le temps de trajet ! Ici le temps est calculé pour un cycliste au départ du centre ville de Foix.
19 November 2019
14:40
Theia starts producing Sentinel2 L2A products over Maghreb
sur Séries temporelles (CESBIO)

=>

Since October 2019, the Theia land data center started producing monthly cloud free and almost seamless syntheses, using Sentinel-2 data acquired on the Maghreb coastal regions, from Agadir to Djerba, passing through Rabah, Alger and Tunis. These syntheses, also called Level3A products, will be delivered every month, thans to the WASP processor, which was integrated within the %MUSCATE production center at CNES. The products can be freely downloaded from Theia distribution website, using requests such as this one :

[https:]]

To download the data, you just need to register to Theia website. Then you can either use the website interface, or a download tool. As for the Sentinel2 L1C and L2A products, the syntheses are cut into tiles of 110x110 km2 and are provided with a 10m resolution for bands B2,B3,B4,B8, and 20m for the other bands. The syntheses use all the cloud free data acquired by Sentinel-2 over a period of 46 days, centred on the day 15th of each month. In the eventually of a constant cloud cover during all the Sentinel-2 overpasses in the 46 days duration, a mask is provided to show where clouds or shadows are remaining. This mask is however more useful in North of France than in Maghreb.

Click on the image above to view the mosaic of Theia L3A products over Maghreb for October 2019. Given the large size of the region, we produced the mosaic at a resolution of 40m, but Theia L3A products have a full resolution of 10m.

Flight over Maghreb to join the sites monitored by CESBIO (with its Moroccan and Tunisian partners), the Haouz plain in Morocco, and the Merguelil catchment in Tunisia.

To have a detailed description of the methods behind WASP, please read the link provided below. In a few words, WASP (Weighted average Synthesis Processor) computes a weighted average of Sentinel-2 cloud free reflectances gathered over a duration of 46 days, centred on the 15 th day of each month. This processing relies on the quality of atmospheric correction and above all cloud detection obtained from MAJA L2A processor.

Theia's Sentinel-2 L3A monthly cloud free syntheses
12:17
Démarrage de production des N3A Sentinel2 de Theia au Maghreb
sur Séries temporelles (CESBIO)

=>

Depuis octobre 2019, Theia a commencé à produire les synthèses mensuelles Sentinel-2 sans nuages et quasiment sans coutures apparentes, sur la zone côtière du Maghreb, d'Agadir à Djerba en passant par Rabat, Alger et Tunis. Ces synthèses, dites de niveau 3A, seront désormais produites tous les mois, grâce à la chaîne WASP qui a été intégrée au centre de production MUSCATE. Les produits sont disponibles, gratuitement, sur l'atelier de distribution de Theia, au CNES, en utilisant une requête comme celle-ci :

[https:]]

Pour télécharger les données, il vous suffit de vous inscrire. Comme pour les produits de niveau 1C ou 2A de Sentinel-2, les données sont découpées par tuiles de 110 par 110 kilomètres, et sont fournies à une résolution de 10m pour les bandes B2,B3,B4,B8, et 20m pour les autres bandes. Les synthèses utilisent, pour chaque pixel, toutes les observations sans nuages disponibles sur une période de 46 jours centrée sur le 15 du mois. Au cas où des nuages auraient été présents sur la totalité des données acquises pendant la période de 46 jours, les données sont marquées comme nuageuses dans le masque fourni avec les données. Ce masque sera cependant plus utile dans le Nord de la France qu'au Maghreb.

Pour visualiser la mosaïque des produits L3A Theia sur le Maghreb, cliquez sur l'image. Etant donnée la taille de la zone, la mosaïque a été réalisée à une résolution de 40m, mais les produits d'origine ont une résolution de 10m.

Survol du Maghreb joignant les deux sites étudiés par le CESBIO, la plaine du Haouz au Maroc et le bassin du Merguellil en Tunisie.

Si vous souhaitez mieux connaître la méthode WASP, consultez sa page de description en utilisant le lien à la fin de la page. En quelques mots, WASP (Weighted average Synthesis Processor) calcule une moyenne pondérée des réflectances de surface sans nuages mesurées par Sentinel2 sur une durée de 46 jours centrée sur le 15 de chaque mois. Ce traitement s'appuie fortement sur la qualité de la correction atmosphérique et de la détection nuageuse du processeur de niveau 2A MAJA.

Theia's Sentinel-2 L3A monthly cloud free syntheses
05 November 2019
14:45
[Theia/MUSCATE] CNES cluster offline for two days
sur Séries temporelles (CESBIO)

There will be no production within MUSCATE, or using the on-demand MAJA tool on PEPS, until wednesday 6th of November, as CNES HPC cluster is off for two days. An hardware and software upgrade is taking place.

Il n'y aura pas de production dans MUSCATE, ou pour MAJA à la demande sur PEPS les 5 et 6 Novembre. Le cluster du CNES est en maintenance pour des améliorations matérielles et logicielles.
10:07
The flow of Pine Island Glacier in Antarctica as seen by Sentinel-1
sur Séries temporelles (CESBIO)

Enclosure: [download]

The clip below shows the evolution of Pine Island glacier ice tongue in West Antarctica, from June 2017 to October 2019. I made it with 111 quicklooks of Sentinel-1 radar images processed by the Alaska Satellite Facility and available via the vertex data portal. All images were acquired in interferometric wide swath mode (polarization HH, only ascending passes, path: 65) and projected to ground range.

During this short period, two "Manhattan-sized" icebergs calved off the glacier tongue, and the next one is coming soon!
document.createElement('video'); [https:]]
This area was imaged every 6 days by Sentinel-1A and Sentinel-1B with a constant look angle.

PS. this post is just a throwback of this one about the Thwaites ice shelf:

Watching ice shelves break up with Sentinel-1
31 October 2019
10:18
For your eyes only, New Caledonia monthly cloud free synthesys.
sur Séries temporelles (CESBIO)

=>

I will probably never go to New Caledonia, as the plane trip would spoil my carbon budget for the next 10 years. But I just had the chance to admire it from 700 km above sea level, thanks to the monthly Level 3A cloud free syntheses produced by Theia with Sentinel-2 data. As I am not selfish, here is a link to visualize the mosaic of these images obtained for the month of August 2019. For once, I made the mosaic at 10 meters resolution, and it's worth it.

If you look closely, you will find some artefacts on this cloudy region, but most of them are above water, and sadly over the magnificent lagoon. Our method has been designed to work on lands, but is somewhat degraded above water becuase water colours change very fast with time. to know more about our methodology please see this post.

A little zoom on the Heart of Voh

To start visiting New Caledonia, click on the image below. I am taking the risks to make you want to go to New Caledonia, but, I will not accept to take charge of your carbon budget.
Cliquez sur l'image pour visiter la Nouvelle-Calédonie

Crossing New Caledonia from West coast to East coast
30 October 2019
18:26
Pour le plaisir des yeux, la synthèse mensuelle Theia sur la Nouvelle Calédonie
sur Séries temporelles (CESBIO)

=>

Je n'irai probablement jamais en Nouvelle Calédonie, le voyage en avion plomberait mon bilan carbone pour les 10 prochaines années. Mais je viens d'avoir la chance de l'admirer depuis 700 km d'altitude, grâce aux synthèses mensuelles de Niveau 3A produites par Theia avec les données Sentinel-2. Comme je ne suis pas égoïste, voici un lien pour visualiser la mosaïque de ces images obtenue pour le mois d’août 2019. Pour une fois, j'ai fabriqué la mosaïque à 10 mètres de résolution, et ça en vaut la peine.

En cherchant bien, vous allez trouver quelques artefacts sur cette région très nuageuse, mais la plupart sont sur la mer, et malheureusement sur le lagon. Notre méthode est faite pour fonctionner sur les terres : si vous voulez en savoir plus, consultez cet article.

Un petit zoom sur le fameux cœur de Voh, immortalisé par Yann Arthus-Bertrand

Pour commencer la visite, cliquez sur l'image ci-dessous. Je prends le risque de vous inciter à aller voir sur place, mais si c'est le cas, je refuse d'endosser votre bilan carbone. Nous dirons que c'est la faute de Sentinel-2.
Cliquez sur l'image pour visiter la Nouvelle-Calédonie

Une traversée Ouest-Est de la grande Terre
28 October 2019
12:28
It has never been so easy to obtain Sentinel2 L2A data processed with MAJA.
sur Séries temporelles (CESBIO)
Since October 2018, it has been possible to run MAJA on PEPS, the French Copernicus collaborative ground segment, to produce L2A products. but until now, the way MAJA had been implemented within PEPS was rather inefficient, as the implementation only produced one single L2A at a time with MAJA, even if 8 products were generated to obtain that product. Producing at least 8 products is necessary because MAJA, which is a multi-temporal processor, requires to process time series to improve the quality of the output L2A products. But in its usual implementation, the 8 or more products are provided as output, while in PEPS, only the last one was provided. So if you wanted to obtain a whole time series this way, it turned out each product had to be produced 8 times. It had been implemented this way because of constraints of the PEPS on line processing interface.

Thankfully, my colleague from CNES, Christophe Taillan, who works on PEPS, recently improved the implementation of MAJA with PEPS.
Time series of L2A images generated with MAJA on PEPS. Clouds are circled in green, cloud shadows are outlined in yellow. and snow in pink. MAJA implements a cirrus cloud correction, which explains why you might not see any clouds in regions outlined in green.
With the new FULL-MAJA implementation, we have abandoned the clickable interface, and only offer a python interface, but the processing is much more efficient. With only one command line, you can start the processing of a time series of MAJA images. You can get the python interface program from my github repository. Here is how we obtained the nice images presented to the right of this text, that correspond to Sebastopol in Crimea (I picked a random tile)

You can get the python interface program from my github repository. You will also need to create an account on PEPS. Here is an example of how it works, for more information, please go the the github page. The examples below show how I obtained the nice images presented above, that correspond to Sebastopol region (I picked a random tile).
1. Request processing
To start processing a 6 month time series starting in July 2017, on Sebastopol tile (31TCJ), using only orbit 51, just use the following command line :

python ./full_maja_process.py -a peps.txt -t 36TWQ -o 64 -d 2019-03-01 -f 2019-09-01 -g 36TWQ_2019-03-01.log

In this command line, -a, gives the password file, and -g is a log file, that will be used subsequently to download the data. Do not use the same log file for two different processings, as you will not be able to download the data from the first. The time period to process must be longer or equal to two months, and shorter that 12 months.
2. Check processing status, and if finished, download data
To check if the processing is finished, and retrieve the data, just use the following command line, and the generated products will be stored in the directory provided in with -w option.

python full_maja_download.py -a peps.txt -g 36TWQ_2019-03-01.log -w /path/to/36TWQ

As MAJA only provides data for dates with less than 90% of cloud cover, some of the dates are not downloaded. With Sentinel-2, in some cases, products are split in two, with eigher the upper or the lower part with no_data values. The implementation of MAJA within PEPS only processes one of those two, so far. This corresponds to products which are stated at missing in the log-file
```
Warning: 4 products have not been processed For more information, 
please check Full_MAJA_Sebastopol_2019.json
```
How long does it take ?
If PEPS is not too busy, it may take up to 24 hours to process a one year time series. But, as the computing resource is limited, the number of simultaneous processes with MAJA on PEPS is limited to 10. Above that number, processes will be queued. As a result, your request could be "pending" for a while. It will mean our on-demand processing facility is a success To see the progress, use the full_maja_download_tool.py tool, or have a look at the processing page (the gear icon) on [https:] (after you have logged in).

A big thank you to the PEPS team, and especially to Christophe Taillan !
27 October 2019
11:37
Panaches de sédiments dans la Méditerranée suite aux fortes pluies dans le sud de la France
sur Séries temporelles (CESBIO)

Cette animation faite à partir de deux images Sentinel-2 montre l'apparition des panaches de sédiments dans la Méditerranée entre Béziers et Agde après les fortes pluies qui se sont abattues sur le sud de la France les 22 et 23 octobre.
22 October 2019
17:58
[New] Time series of Sentinel-2 monthly syntheses in Italy
sur Séries temporelles (CESBIO)

=>

As announced a few months ago, our monthly synthesis processor for Sentinel-2 data, WASP, has been integrated to THEIA's land data center processing facility at CNES, in spring 2019. We are therefore now extending the zones where we process the monthly synthesis (the so-called Level-3A products).

Here is how to look for Level 3A in Italia in THEIA's catalog. Substitute "Italy" by "Italie" if your browser's language is French.

The first countries that we added are Spain and Portugal, and we are happy to release the first high quality monthly synthesis of surface reflectance of the whole Italian peninsula, valid for September 2019. It is not yet a time series, but we will produce this product every month from now on. As for the standard Level-1C and Level-2A products of Sentinel2, the level-3A product is split in tiles of 110*110 km, and is available for all Sentinel-2 spectral bands, except the atmospheric bands (B1, B9, B10). This data set can be downloaded, for free, from Theia web site, using the following link :

https://theia.cnes.fr/atdistrib/rocket/#/search?q=Italy&collection=SENTINEL2&processingLevel=LEVEL3A

If your browser language is French, please use the following link:

https://theia.cnes.fr/atdistrib/rocket/#/search?q=Italie&collection=SENTINEL2&processingLevel=LEVEL3A

We also usually create a mosaic of these tiles, for visualisation purposes, using only the Red, Green and Blue spectral bands, at 20m resolution (the Theia Level-3A products are available at 10m resolution). This mosaic shows the quality of the synthesis, as usually, a monthly synthesis is plagued with artefacts, but the ones we have here are very faint, but you might find some tile limits.

Click on the image to zoom to 20m resolution

To have a detailed description of the methods behind WASP, please read the link provided below. In a few words, WASP (Weighted average Synthesis Processor) computes a weighted average of Sentinel-2 cloud free reflectances gathered over a duration of 46 days, centred on the day 15th of each month. This processing relies on the quality of atmospheric correction and above all cloud detection obtained from MAJA L2A processor.

Theia's Sentinel-2 L3A monthly cloud free syntheses
21 October 2019
18:49
[Nuovo !] Serie storiche delle sintesi mensili di Sentinel-2 in Italia
sur Séries temporelles (CESBIO)

=>

Come annunciato qualche mese fa, il nostro processore di sintesi mensile per i dati Sentinel-2, WASP, è stato integrato nel centro di elaborazione dei dati terrestri THEIA presso il CNES, nella primavera del 2019. Stiamo quindi ampliando le zone di elaborazione delle sintesi mensili (i cosiddetti prodotti Level-3A).

Come consultare il catalogo Theia per trovare i prodotti Level 3A in Italia. Sostituire "Italy" con "Italie" se la lingua predefinita del browser è il francese.

Doppo la Francia, I primi paesi che abbiamo aggiunto sono l'Italia, la Spagna, ed il Portogallo, e siamo felici di pubblicare la prima sintesi mensile della riflettanza di superficie, ad alta qualità per tutta la penisola italiana, valida per il mese di settembre 2019. Come per i prodotti standard Level-1C e Level-2A di Sentinel2, il prodotto Level-3A è diviso in settori di 110*110 km, ed è disponibile per tutte le bande spettrali di Sentinel-2, ad eccezione delle bande atmosferiche (B1, B9, B10). Questo set di dati può essere scaricato gratuitamente dal sito web di Theia, utilizzando il seguente link:

https://theia.cnes.fr/atdistrib/rocket/#/search?q=Italy&collection=SENTINEL2&processingLevel=LEVEL3A

Per i browser la cui lingua è il francese, si prega di utilizzare il seguente link:

https://theia.cnes.fr/atdistrib/rocket/#/search?q=Italie&collection=SENTINEL2&processingLevel=LEVEL3A

A scopo illustrativo, di solito, realizzamo anche un mosaico di questi settori, a scopo di visualizzazione, utilizzando solo le bande spettrali Rosso, Verde e Blu, con una risoluzione di 20 metri (i prodotti Theia Level-3A sono disponibili con una risoluzione di 10 metri). Questo mosaico ha l'obiettivo di mostrare la qualità della sintesi : ovviamente, una sintesi mensile è afflitta da artefatti, ma quelli delle nostri prodotti sono molto deboli.

Clicca sull'immagine per ingrandirla a 20 metri di risoluzione

Per avere una descrizione dettagliata dei metodi alla base di WASP, si prega di leggere il link riportato di seguito. In poche parole, WASP (Weighted average Synthesis Processor) calcola una media ponderata della riflettanza (senza nuvole) di Sentinel-2 raccolta su un periodo di 46 giorni, centrata sul 15 di ogni mese. Questa elaborazione si basa sulla qualità della correzione atmosferica e soprattutto sul rilevamento delle nuvole ottenuto dal processore MAJA L2A.

Grazie mille a DeepL translator, e supratotto a Manuel Salvoldi (BGU) per la aiuta a traducire (cuesta ultima sentenza fu scritta senza aiuta)
13 October 2019

3:24

Khumbu icefall in 4D

sur Séries temporelles (CESBIO)

This video shows all available Venµs images since launch over the Khumbu icefall near Mt Everest (109 images). The images are RGB composites of Level 1C bands 7,4,3 (5 m resolution) that were draped onto the HMA 8m DEM.

document.createElement('video'); [https:]]

How to do that? First download the Venµs images using theia_download.py, extract the region of interest and make color composites using gdal_translate as explained here. With a recent GDAL installation you can even read ZIP archives on-the-fly without decompressing them. Then, prepare an elevation raster from the HMA 8m DEM so that it has the same resolution and extent as the extracted Venµs images. Finally make the 3D view in R using rayshader.

library(rayshader)
library(abind)
library(raster)
p="./"
fdem=paste0(p,"HMA_DEM8m_MOS_20170716_tile-677_UTM45_filled_crop_Venus.tif")
pimg=paste0(p,"/VIS1C/")

dem=raster(fdem)
elmat=matrix(extract(dem, extent(dem), buffer = 1000),nrow = ncol(dem), ncol = nrow(dem))

files=list.files(path=pimg, pattern="jpg$", full.names=FALSE, recursive=FALSE)

# rotating view  
th=200
k=0
for (x in files) {
    fimg=paste0(pimg,x)
    fout=paste0(p,"/3Drot/",x)
    print(x)
    txt=substr(x, 36, 43)
    k=k+1
    if (k<length(files)/2){
    th=th+0.5
    }else{
    th=th-0.5
    }
    imgr = raster(fimg,band=1)
    imgg = raster(fimg,band=2)
    imgb = raster(fimg,band=3)
    r = t(matrix(extract(imgr, extent(imgr), buffer = 1000),nrow = ncol(imgr), ncol = nrow(imgr)))
    g = t(matrix(extract(imgg, extent(imgr), buffer = 1000),nrow = ncol(imgr), ncol = nrow(imgr)))
    b = t(matrix(extract(imgb, extent(imgr), buffer = 1000),nrow = ncol(imgr), ncol = nrow(imgr)))
    col=abind(r,g,b,along=3)/255
    plot_3d(col, elmat, zscale = 5, fov = 0, theta = th, zoom = 0.7, phi = 45, windowsize = c(1000, 800))
    render_snapshot(fout,clear = TRUE,title_text = txt, title_color = "black",
               title_font = "Arial", gravity = "NorthEast", title_offset = c(0,0))
    }

Merci à Gérard Dedieu pour l'idée !

Here is the same animation without rotation:

[https:]]

26 September 2019
18:43
Cartographier l'occupation des sols d'une partie de l'Australie à 10 m de résolution
sur Séries temporelles (CESBIO)
(English version here)

Peu de pays sont plus éloignés de la France que l'Australie. La carte d'occupation des sols dont nous parlons aujourd'hui concerne l'État du Victoria au Sud-Est de l'Australie. Cet État est quasiment diamétralement opposé à la Turquie avec des latitudes autour de 37°S et des longitudes autour de 145°E. Nous avons reproduit la méthode qui est utilisée pour créer la carte de la France par l'équipe OSO du CESBIO. Dans ce post, je parlerai des differentes étapes que nous avons suivies pour créer notre carte (qui, dans les grandes lignes, reprend les idées de la chaîne IOTA-2), et j'en profiterai pour vous montrer de belles images Sentinel-2 et leurs cartes associées.
La carte

Commençons par la fin, et donc par la carte complète ci-dessous.
Carte d'occupation des sols de l'État du Victoria en Australie

Victoria fait à peu près la moitié de la taille de la France métropolitaine. Au Sud, vous pouvez voir la côte de la mer de Tasmanie avec la grande ville de Melbourne (en rose) à peu près au milieu, le long de la côte. Pour l'anecdote, la ville de Melbourne est très étendue et à peu près la même surface que Shanghai mais avec 5 fois moins d'habitants. Au Nord-Est de Melbourne, vous pouvez voir une large surface boisée (en vert plutôt foncé) : ce sont les Alpes victoriennes (eh oui, nous avons aussi nos Alpes !). Au Nord-ouest, on peut voir une large région agricole (en orange et en jaune) puis, à l'extrème Nord-Ouest, le désert.

Notre carte est la première carte d'occupation des sols créée à 10 m de résolution - la précédente avait été créée à 250 m (la résolution de MODIS). Je ne résiste pas à vous présenter un effet avant/après autour du Lac Eildon :
Comparaison de la précédente carte à 250 m avec notre carte à 10 m Comment cette carte a-t-elle été créée ?
Dans les grandes lignes, nous avons suivi les idées de la chaîne IOTA-2 du CESBIO, mais nous avons développé nos propres scripts d'appel à la librairie OTB, en utilisant notre algorithme de classification développé par Dr Charlotte Pelletier (une ancienne du CESBIO) pendant son séjour à Monash University : TempCNN. Voici les principales étapes du processus :
1. Téléchargement de toutes les images Sentinel-2 de juillet 2017 à août 2018 (d'hiver à hiver) au niveau-2A. Pour cette étape, notre boulot a été grandement facilité par le service PEPS qui nous a permis de télécharger les produits Sentinel-2 au niveau 2A : nous avons donc collecté 4,000+ images, recouvrant 37 tuiles, et ce directement avec les corrections atmosphériques et les masques de nuages (grâce au traitement à la demande par le logiciel de corrections atmosp?ériques MAJA !).
2. Prétraitement. Nous avons d'abord ré-interpolé à 10 m les bandes à 20 m (et n'avons pas utilisé les bandes à 60 m), puis construit deux cubes de données (un pour toutes les images de la série et un pour tous les masques), et finalement créé un cube de données ré-interpolées temporellement (pour avoir les mêmes dates pour toutes les tuiles et enlever les nuages). Tout ça a été fait grâce à l'OTB. Petite note en passant : notre NFS est tombé en panne et nous avons dû nous résoudre à utiliser 7 disques de 4To...
  Notre NFS est tombé en panne...
3. Création du jeu d'entraînement. Nous avons pu bénéficier des campagnes de terrain de l'équipe d'Agriculture Victoria ; données qui avaient été collectées pour la création de la carte à résolution MODIS ci-dessus. Nous y avons ajouté quelques polygones d'urbain, puis mis en relation ces polygones aux données images de nos cubes ré-interpolés (un cube par tuile) et enregistré ces 'pixels' étiquetés (une série temporelle multi-variée par pixel) dans une base de données Sqlite.
4. Apprentissage du modèle. Nous avons utilisé notre algorithme TempCNN basé sur un réseau convolutionnel, ayant pour objectif d'extraire des motifs de dynamiques temporelles dans l'évolution des bandes spectrales. Nous avons amélioré la précision de 1,5% de précision par rapport au Random Forest (et +3% pour le F-Score). L'apprentissage va assez vite pour nos 3,8M d'échantillons, il a pris moins de deux heures.
5. Création de la carte. Une fois le modèle entraîné, nous l'avons utilisé pour classifier les tuiles une par une, puis fusionné ces résultats (les tuiles ayant un recouvrement), et finalement recoupé la carte pour correspondre à l''mprise de l'État de Victoria.
Et voilà ! Nous avons mis un serveur à disposition ; vous pouvez donc allez voir le résultat sur [MonashVegMap.org]
Retour d'expérience
Du début à la fin, la création de la carte nous aura pris.e.s environ 6 mois : 3 mois pour obtenir un premier résultat puis 3 mois pour faire une deuxième passe et corriger quelques problèmes (principalement liés à des données manquantes - il n'est pas toujours facile de savoir que des données manquent quand on gère tant de données). Je pense que notre surprise principale a été que nous avions probablement sous-estimé le temps passé à pré-traiter les données : téléchargement et prétraitement nous ont pris 7 semaines alors même que nous récupérions directement les données au niveau 2A. Notre procédure aurait également pu être optimisée pour éviter les entrées/sorties sur les disques, ce que IOTA fait (pas de stockage de résultat intermédiaire). Globalement, nous étions surpris·e·s de nous rendre compte que le processus développé au CESBIO marcherait si facilement : nous nous attendions à devoir réinventer des traitements pour telle ou telle partie du processus, mais non (sauf peut-être la fusion des résultats des différentes tuiles). Donc, un grand merci s'impose pour l'équipe de Jordi Inglada qui a mis au point le processus.

La carte a été reçue très positivement avec de nombreux contacts et téléchargements du fichier GeoTiff final, notamment pour la modélisation des feux (un thème important en Australie, et la gestion des bassins versant et de l'agriculture.

Je ne pouvais pas finir ce post sans vous montrer une belle image de la région viticole australienne ; ci-dessous la ville de Shepparton dans laquelle nous cultivons du Shiraz (pensez Côte du Rhône poussant dans une terre ocre). On voit d'ailleurs bien les exploitations viticoles sur la carte en violet (et je recommande une visite si vous êtes dans la région).
Shepparton, as viewed by Sentinel-2 Shepparton, notre carte
24 September 2019
17:36
Mapping (a part of) Australia at 10 m resolution
sur Séries temporelles (CESBIO)
(French version here)

It would be difficult to be farther away from France than Australia. In fact, the map we're talking about today is of the State of Victoria in Australia, which is roughly diametrically opposed to Turkey with latitudes around 37°S and longitudes around 145°E. What we have done is reproduce the methodology that has been used to create the land-cover map of France to map the State of Victoria in Australia. In this post I will talk about the different steps we have followed to produce the map (which basically follow the IOTA-2 processing chain) as well as show some pretty Sentinel-2 pictures and associated land-cover maps over Australia.
Let's look at the map

Let's start with the resulting map - below.
Land-cover map of the State of Victoria in Australia

Victoria is roughly half the size of Metropolitan France. You can see the outline of the Tasman sea in the south, with the city of Melbourne and its bay along the coast in the middle. You can also see a large forested area to the east of Melbourne and up to the border of the State (in the darker shade of green), which corresponds to the Victorian Alps (yes we have Alps as well!). In the north west, a very large agricultural area (orange and yellow) and in the tip of the north-west corner: the desert!

This is the first map of Victoria created at 10 m resolution - the previous one had been created from MODIS data, ie at 250 m resolution.This is what the before/after looks like.
Comparison of the previous map at 250 m vs the map we created at 10 m How did we do it?
Basically, we followed the work that had been done at CESBIO with IOTA-2, but did redevelop our own scripts around OTB, and used our house-developed TempCNN algorithm, developed by Dr Charlotte Pelletier (formely at CESBIO) while at Monash University. Here are the main steps we followed:
1. Download all Sentinel-2 images from July 2017 to August 2018 (winter to winter). Our job here was made much easier than it should have been, thanks to the PEPS team. We wanted good-quality LEVEL-2A data, and the PEPS servers made it possible to download 4,000+ images, across 37 tiles, directly with atmospheric corrections and cloud masks, using the on-demand MAJA processing.
2. Preprocessing. All 20 m bands were re-interpolated at 10 m (we left the 60 m ones out), stacked into 2 cubes (one for the images, one for the masks), then we interpolated the cloud-covered values using linear temporal interpolation. All of this was done using the wonderful OTB. As an aside here, our NFS had issues during the production, and so we had to resort to using 7x4TB hard drives...
  Our NFS died...
3. Creating the training data. We were lucky to have data from field campaigns from Agriculture Victoria, which they used during the creation of the MODIS map showed above. We added a few polygons of 'urban' and simply mapped those polygons to our gapfilled stacks of images to create one .sqlite file containing all the examples we could use for training our deep learning model.
4. Train the model. We used our TempCNN algorithm, which is a temporal convolutional neural network, which basically extracts features on the temporal dynamics of the bands. We showed is to be 1.5% more accurate (F-measure +3%) than Random Forest. Training is fast for our 3.8M examples and takes less than a couple of hours.
5. Create the map. Once the model is trained, we classify each pixel of each tile independently in the different types of land-cover in our nomenclature, then merge the overlap between the different tiles (using some fancy recommendations by Olivier!) and cut the map to the shape of the State.
Tada! You can have a play with the map yourself at [MonashVegMap.org]
Reflecting on the exercise
From start to finish, creating the map took us about 6 months, with a first iteration of the map created within 3 months and then slowly iterating through a second version. Overall, we felt that we didn't have a good idea of the time preprocessing would take: download and preprocessing took us about 7 weeks, and this was while we were already getting L2A images. Definitely some time was lost in inputs/outputs to disks, which I believe IOTA doesn't do as much as what we've done. Overall, it was quite amazing to see that this process 'just worked' - very well done to the team led by Jordi a few years ago and now Sen2agri. The uptake has been very positive here with numerous government bodies already using the map in their GIS, be it for fire modelling, water catchment or agricultural planning.

I cannot finish this post without a beautiful picture of the city of Shepparton, north of Melbourne where we grow some beautiful Shiraz (in purple) - a big recommendation if you travel to Australia and want to visit some wineries.

Shepparton, as viewed by Sentinel-2 Shepparton, our map
23 September 2019
16:31
The DLR starts distributing Sentinel-2 L2A products produced with MAJA over Germany
sur Séries temporelles (CESBIO)

A good news for German users of Sentinel-2 data !

You might have already read on this blog that the German Aerospace Center, DLR became a member of MAJA development team in 2016. For instance, they provided the nice cirrus correction method. One of the team objectives was also to implement MAJA within DLR's CATENA production facility, and we are very happy to announce that it worked ! The DLR team has been processing all the images acquired by Sentinel-2 over Germany with MAJA for some time now and they just started to make them available, freely and very easily. The animation below shows, for instance, a few of the images available over the east of Munich, in Bavaria for summer 2019.
Time series of Sentinel-2 images on tile T32UQU, east of Munich, downloaded from DLR server. Detected clouds are circled in green, and detected shadows in yellow.

The dataset is described here. The service is easily accessed without any registration: the data sets are stored geographically per tile, and then temporally per year and month. The root of the product tree is : [https:]] . For instance, to see the available data from tile T32TNT in September 2019, just search for the following page: [https:]] and you will get a list of the products available for that tile and that month.

If you navigate through the folders, you will see that the number of products for a given folder is variable: to save processing time and storage, and download, MAJA discards the images for which the cloud coverage is above 90%, which seems to be quite frequent in Germany.

The image format is the same as for the products generated by Theia, as a result, the same software will be able to work across the French German border.

Many thanks to our colleagues from DLR, Pablo d'Angelo, and Stefan Auer who took this implementation in charge. They are also planning to deliver monthly cloud three syntheses (Level 3A products) using WASP, and Pablo already started working on it.
18 September 2019
17:03
Series temporales de síntesis mensuales Sentinel-2 en La Península Ibérica
sur Séries temporelles (CESBIO)

=>

Como se anunció hace unos meses, nuestro procesador de síntesis mensual para datos Sentinel-2, WASP, se ha integrado a las instalaciones de procesamiento del centro de datos terrestres THEIA en el CNES, en la primavera de 2019. Por lo tanto, ahora estamos ampliando las zonas donde procesamos la síntesis mensual.
How to query Theia catalog to find Level 3A products over Spain
Después de Francia, los primeros países que añadimos son España, Portugal y Andorra, y ahora Italia, nos complace presentar nuestras primera síntesis mensual de nivel 3Ade alta calidad de reflectividad superficial de toda la Península Ibérica, válidas para agosto y septiembre de 2019. Vamos a producirlas cada mes. En cuanto a los productos estándar de Nivel 1C y Nivel 2A de Sentinel2, el producto de nivel-3A se divide en granules de 110*110 km, y está disponible para todas las bandas espectrales de Sentinel-2, excepto las bandas atmosféricas (B1, B9, B10). Este conjunto de datos puede descargarse gratuitamente del sitio web de Theia, utilizando la siguiente solicitud:

[https:]]

Si el idioma de su navegador es el francés, utilice el siguiente enlace:

https://theia.cnes.fr/atdistrib/rocket/#/search?q=Espagne&collection=SENTINEL2&processingLevel=LEVEL3A

También, cada mes, solemos crear un mosaico de estos granules, con fines de visualización, utilizando únicamente las bandas espectrales Rojo, Verde y Azul, con una resolución de 20m (los productos Theia Level-3A están disponibles con una resolución de 10m). Este mosaico muestra la calidad de la síntesis: de costumbre, una síntesis mensual está plagada de artefactos, pero los que tenemos aquí son muy tenues, gracias al método de media ponderada contenido en WASP. Sin embargo, el mosaico no está completamente libre de nubes, ya que Sentinel2 no consiguió obtener una observación libre de nubes en Asturias, pero, bueno, la gente que ha estado allí sabe lo difícil que es, incluso de agosto.

Haga clic en la imagen para ampliarla a una resolución de 20 metros y comparar los mosaicos de los diferentes meses. ¿Encontrarás las 1000 diferencias?

Para tener una descripción detallada de los métodos detrás de WASP, por favor lea el enlace que se proporciona a continuación. En pocas palabras, el WASP (Weighted Average Synthesis Processor) calcula una media ponderada de la reflectividad superficial de las observaciones sin nubes de Sentinel-2 recogidas durante 46 días, centradas en el día 15 de cada mes. Este proceso se basa en la calidad de la corrección atmosférica y, sobre todo, en la detección de nubes obtenida del procesador MAJA L2A .

Theia's Sentinel-2 L3A monthly cloud free syntheses

(He traducido en Castellano con mis pocos recuerdos de mis cursos, con la ayuda de DeepL translator, y sobre todo de Milena Planells del CESBIO).
16:37
[NEW !] Time series of Sentinel-2 monthly syntheses in Spain
sur Séries temporelles (CESBIO)

=>

As announced a few months ago, our monthly synthesis processor for Sentinel-2 data, WASP, has been integrated to THEIA's land data center processing facility at CNES, in spring 2019. We are therefore now extending the zones where we process the monthly synthesis (the so-called Level-3A products).
How to query Theia catalog to find Level 3A products over Spain
After France, the first countries that we added are Spain, Portugal, and Andorra, and now Italia. So, we are happy to release out first high quality monthly Level-3A syntheses of surface reflectance of the whole Iberian peninsula, valid for August and September 2019. As for the standard Level-1C and Level-2A products of Sentinel2, the Level-3A product is split in tiles of 110*110 km, and is available for all Sentinel-2 spectral bands, except the atmospheric bands (B1, B9, B10). This data set can be downloaded, for free, from Theia web site, using the following request:

[https:]]

If your browser language is French, please use the following link:

https://theia.cnes.fr/atdistrib/rocket/#/search?q=Espagne&collection=SENTINEL2&processingLevel=LEVEL3A

We also usually create a mosaic of these tiles, for visualisation purposes, using only the Red, Green and Blue spectral bands, at 20m resolution (the Theia Level-3A products are available at 10m resolution). This mosaic shows the quality of the synthesis, as usually, a monthly synthesis is plagued with artefacts, but the ones we have here are very faint. However, the mosaic is not completely cloud free, as Sentinel2 did not manage to obtain a cloud free observation in Asturias, but, well, people who have been there know how difficult it is even in August.

Click on the image to zoom to 20m resolution, and to compare the syntheses of the different months. Will you spot the 1000 or more differences ?

To have a detailed description of the methods behind WASP, please read the link provided below. In a few words, WASP (Weighted average Synthesis Processor) computes a weighted average of Sentinel-2 cloud free reflectances gathered over a duration of 46 days, centered on the 15 th day of each month. This processing relies on the quality of atmospheric correction and above all cloud detection obtained from MAJA L2A processor.

Theia's Sentinel-2 L3A monthly cloud free syntheses
17 September 2019
14:19
India water crisis: Sentinel-1 detects surface water, ForEarth app restitutes and shares water stresses
sur Séries temporelles (CESBIO)

In the framework of an open call Science4society funded by ESA, researchers at CESBIO have implemented a surface water detection algorithm from radar data Sentinel-1 on a cloud computing system. An API developed by Geomatys and JeoBrowser, two Information Technologies companies in France, is used to send the resulting surface water masks to a smartphone App named ForEarth, which display the surface water fluctuation in time for Indian regions. The beta version is available on PlayStore and iOS displays statistics and water masks for the Hyderabad region, Telangana, India.
Region of interest already processed since 2016.

The radar data are able to monitor for any weather the numerous small surface water reservoirs that are used for agriculture as well as the biggest dammed reservoirs used for both irrigation and domestic water (Figures below). This technical aspect is essential in India as the replenishment of surface reservoirs occur during the monsoon season, a cloudy season lasting for 4 months.

The Singur reservoir used for domestic water of Hyderabad area. The blue mask corresponds to the surface water detected from Sentinel-1 image on the 3rd Sept. 2019 (up) and 2nd Sept. 2018 (down). A satellite optical image is used in background. The surface water area detected for the full ROI is provided. The global surface water area extents are similar in 2018 and 2019 even if the Singur lake is empty nowadays which will lead to domestic water shortages soon.
Contrary to existing data base for largest reservoirs, this app deliver surface water extent fluctuation in the water harvesting system, a typical network of small dams maintained by local farmers and village communities to catch monsoon runoff. It provides fresh water for rice irrigation, cattle and is supposed to enhance local aquifer recharge (figure below).
Maximal surface water area extent observed after the heavy rainfall in Sept. 2016. The water harvesting system is full.
From this automated Sentinel-1 detection, the smartphone application aims at easily providing a domestic or agricultural water shortage alert system based on surface water fluctuation for lay users (figure below).
Surface water area fluctuation, averaged for the ROI, since 2016. The fourth monsoon season is monitored in near real time, providing a quantitative estimate of monsoon impact on surface water replenishment.
This second version of the App will be enriched over the ROI with surface water volume estimates. It is planned to deploy the surface water detection algorithm over South India.

We’re thankful to Marc Gorman who highly contributed to the obtainment of the ForEarth project and will not have the opportunity to see the outcome.
12 September 2019
22:43
Upgrade of let-it-snow to compute the fractional snow cover
sur Séries temporelles (CESBIO)

In preparation for the next Pan European high resolution snow and ice monitoring service, I have upgraded the let-it-snow (LIS) processor to enable the computation of the fractional snow cover (FSC) within the pixel.
Binary vs. fractional snow cover in the Pyrenees (tile T31TCH on 12 June 2019). Here the FSC was computed using the MOD10A1.005 equation (SCF=?0.01+1.45×NDSI) by Hall et al. (2001). Source of the binary product: Theia.
So far LIS was only used to compute a "binary snow mask" (in fact ternary: snow, no-snow, cloud). However, the EEA requested that the service provides the snow fraction in every 20 m Sentinel-2 pixel. The EEA also specified that the FSC should be computed from the Normalized Difference Snow Index. Hence, the latest version of the LIS processor (currently in the develop branch) can now read a custom function in the configuration file to compute the FSC from the NDSI (see this template).

In forest areas, the function f(NDSI) provides a “top-of-canopy” FSC (FSCTOC), which can be adjusted based on the tree cover density (TCD) to estimate the “on-ground” FSC (FSCOG) (Raleigh et al., 2013).

$FSC_{OG} = FSC_{TOC} / (1-TCD)$FSC_{OG} = FSC_{TOC} / (1-TCD)
"On ground" snow fraction computed with LIS (Sentinel-2 on 2019-06-12 tile T31TCH)
The TCD is given at 20 m resolution by the [https:] density">Copernicus High Resolution Layer Tree Cover Density status map (current release is 2015). The example below illustrates how this correction increased the snow fraction near the edges of the snow cover area, where the forest partially masked the snow cover.
Snow fraction: top-of-canopy vs. on-ground

We will now focus on the calibration of the FSC=f(NDSI) function using ground truth data (read this if you're interested in contributing!).
10 September 2019
17:33
Le traitement des données Sentinel-2 L2A au Sahel progresse
sur Séries temporelles (CESBIO)

=>

Comme nous l'avions annoncé en Novembre, le centre de production MUSCATE de Theia rajoute progressivement des zones au Sahel, qui sont affichées sur l'image ci-dessous. Les données sont traitées à partir de décembre 2016, ce qui nous fait une grande quantité de données à traiter. Nous avons donc commencé avec les tuiles les plus à l'ouest, au Sénegal sur la zone UTM28, puis de proche en proche vers l''Est.

En rouge, les zones déjà disponibles en temps réel et depuis fin 2016, en bleu, les zones qui le seront bientôt.
Depuis quelques jours, Theia a terminé le traitement des tuiles de la zone UTM30 et l'ouest de la zone UTM31 qui couvrent principalement le Burkina Faso, le Sud du Mali, l'Ouest du Niger et le Nord du Benin. Le traitement de la zone UTM 32 qui couvre principalement le Nigeria est aussi avancé. N'hésitez pas à jeter un coup d’œil de temps en temps à la carte des zones couvertes de MUSCATE. Les tuiles en bleu deviennent rouges dès que l'on passe au traitement au fil de l'eau. Les données peuvent être téléchargées, gratuitement, depuis l'adresse ci-dessous :
[https:]]

Animation sur la région de la ville de Mopti, au Mali, avec environ une image par mois en 2017. La série temporelle s'étend entre deux saisons des pluies et couvre la saison sèche. De nombreuses cicatrices d'incendies sont visibles pendant la saison sèche. Quelques ombres apparaissent, qui correspondent en fait aux ombres des cirrus corrigées par MAJA. Les ombres sont marquées dans les produits.
13:45
THEIA's Sentinel-2 L2A processing on Sahel is progressing
sur Séries temporelles (CESBIO)

=> As we had announced in November, the MUSCATE production centre in Theia is gradually adding areas in the Sahel, which are shown in the image below. The data are processed from December 2016 onwards, which means that we have a large amount of data to process. We started with the most westerly tiles, in Senegal on the UTM28 zone, then progressed from one zone to another towards the East.

Progress of the processing of tiles in Theia's Sahel zone. The tiles in red are available since January 2017, and processed in real time. The tiles in blue are being processed or will be shortly.
In the recent days, Theia has completed the processing of tiles in the UTM30 zone and in most of UTM31 area. The added region includes Burkina Faso, South of Mali, West of Niger, North of Benin. The processing of the east of UTM31 and the whole UTM 32 tiles has started. Feel free to take a look from time to time at the map of areas covered by MUSCATE. The blue tiles turn red as soon as we switch to near real time processing. The data can be downloaded from here:
[https:]] Animation in the region of the city of Mopti, Mali, with about one image per month in 2017. The displayed time series extends between two rainy seasons and covers the dry season. Many fire scars are visible during the dry season. Some shadows appear, which actually correspond to the shadows of cirrus clouds corrected by MAJA. Shadows and cirrus are marked in the products.
09 September 2019
16:08
The multi-temp blog is moving soon/Le blog "Séries temporelles" déménage
sur Séries temporelles (CESBIO)

New address: [https:]]

Early 2019, the University Paul Sabatier warned us we would be expelled before the end of the year (after 6 years of an excellent service, thanks !). We've been very lucky to quickly find a new shelter at the Observatoire Midi Pyrénées. But that's not a simple shelter, that's a luxury home, and we hope you'll like the new clean and modern decoration, designed by Pierre Vert, our interior architect and talented craftsman, in collaboration with Simon Gascoin.

As always, there are still a few details to fix, so please tell us if you find bugs, if some pages look strange, if some links are broken, or if everything is not fully convenient, .

Début 2019, l'Université Paul Sabatier nous a prévenus que nous serions expulsés avant la fin de l'année (après 6 ans d'un service d'une fiabilité sans faute, merci !). Nous avons eu la chance de trouver rapidement un nouvel abri à l'Observatoire Midi Pyrénées. Mais ce n'est pas un simple abri, c'est une maison de luxe, et nous espérons que vous aimerez la nouvelle décoration propre et moderne, conçue par Pierre Vert, notre architecte d'intérieur et artisan de talent, en collaboration avec Simon Gascoin.

Comme toujours, il y a encore quelques détails à corriger, alors dites-nous ce que vous en pensez, n'hésitez pas à nous prévenir si vous trouvez des bugs, ou si certains des 700 articles ne s'affichent pas correctement, ou si vous trouvez des liens qui n'aboutissent plus.
04 September 2019
22:45
[Comment ça marche] Les réflectances peuvent-elles être supérieures à un ?
sur Séries temporelles (CESBIO)
Attention, ce document contient des équations

A force d'écrire des articles sur ce blog, sur twitter ou linkedIn, ou parce que je publie des codes sur github, je reçois près d'une dizaine de questions par semaine, auxquelles j'essaie de répondre.

De nombreuses questions portent sur les valeurs des réflectances, comme par exemple:
- "when dividing the images by 10,000 the maximum value I saw was around 1.6, I guess the values are supposed to be within 0 and 1?"
- "I am trying to figure out if the TOA Reflectance range should be limited to 0 and 1 or not. I have read in two different posts that the TOA Reflectance should not be bigger than 1.The part that confuses me is that for many images I am using, the maximum pixel value is above 1"
- ou, de la part d'un utilisateur de Venµs, moins diplomate : "It means the CNES radiometric calibration process for Venus imagery has many bugs, starting from L1 all the way to L2 products with max surface reflectance values > 2.0"
Le fait que les réflectances doivent être inférieures à un est une croyance très répandue, mais... elle est fausse !

Photometrie

Revenons à la physique :
Conventions et définitions angulaires, N indique le Nord. Dans le texte, nous utilisons l'indice v pour les angles de visée plutot que la lettre r dans le schema).
une surface froide et passive n'émet pas de lumière dans le spectre solaire (ou seulement de façon négligeable) et ne réfléchit qu'une partie de la lumière incidente. Le rapport entre le flux lumineux entrant et le flux sortant dans toutes les directions est connu sous le nom d'albédo, et il ne peut être supérieur à un (voir ce post pour une définition des flux, de l'irradiance, de l'éclat et de la réflexion). Mais la réflectance et l'albédo ne sont pas la même chose.

La réflectance Bidirectionelle est définie par :

$\rho(\theta_{s},\phi_{s},\theta_{v},\phi_{v}) =\frac{\pi.L(\theta_{s},\phi_{s},\theta_{v},\phi_{v})}{Es.cos(\theta_{s})}$\rho(\theta_{s},\phi_{s},\theta_{v},\phi_{v}) =\frac{\pi.L(\theta_{s},\phi_{s},\theta_{v},\phi_{v})}{Es.cos(\theta_{s})}

où L est la luminance, Es est l'éclairement solaire $\theta_s$\theta_s est l'angle zenithal solaire $\phi_s$\phi_s est l'azimut solaire, $\theta_v$\theta_v est l'angle zénithal de visée, and $\phi_v$\phi_v est l'azimut de visée.

L'albédo par ciel noir((ou réflectance hémisphérique directionnelle) est le rapport entre la lumière entrante et sortante d'une source ponctuelle, telle que le soleil, sans atmosphère. La BSA est l'intégrale de la réflectance dans toutes les directions (pondérée par le cosinus de l'angle de visée, car on ne tient compte que de la surface qui est perpendiculaire à la direction d'observation, et par le sinus de l'angle de vision, lorsque l'on intègre les couronnes de l'hémisphère).

$Albedo(\theta_{s},\phi_{s})=\int_{0}^{\pi/2} \int_{0}^{2. \pi} rho(\theta_{s},\phi_{s},\theta_{v},\phi_{v}) cos(\theta_v).sin(\theta_v)d\theta_v d\phi_v$Albedo(\theta_{s},\phi_{s})=\int_{0}^{\pi/2} \int_{0}^{2. \pi} rho(\theta_{s},\phi_{s},\theta_{v},\phi_{v}) cos(\theta_v).sin(\theta_v)d\theta_v d\phi_v

Bien entendu, l'éclairement total réfléchi par une surface dans toutes les directions est inférieur à l'éclairement solaire entrant, ce qui signifie que l'albédo est inférieur à un. Supposons que vous ayez une parfaite
surface lambertienne, qui disperse toute la lumière entrante de manière uniforme dans toutes les directions vers le haut (imaginez une feuille de papier blanc parfaite). Pour une telle surface, la réflectance est égale à un dans toutes les directions et l'albédo est égal à la réflectance.

Imaginez maintenant un miroir parfait, qui réfléchit également toute la lumière entrante, mais seulement dans la direction spéculaire. Dans ce cas, la réflectance n'est pas du tout constante. Elle est nulle dans toutes les directions, sauf dans un petit cône qui correspond à l'angle solide du soleil. L'albédo en est toujours un, mais la réflectance dans le cône spéculaire est bien supérieure à un :

$\rho(\theta_{s},\phi_{s},\theta_{s},\phi_{s}+\pi) = \frac{2.\pi}{\omega_s}$\rho(\theta_{s},\phi_{s},\theta_{s},\phi_{s}+\pi) = \frac{2.\pi}{\omega_s}

où $\omega_s$\omega_s est l'angle solide du soleil, et $2.\pi$2.\pi est l'angle solide d'un hémisphère.

Comme $\omega_s=6.8 \times 10^{-5}$\omega_s=6.8 \times 10^{-5}, nous obtenons une réflectance qui vaut à peu près $10^{5}$10^{5} pour un miroir parfait, ce qui est beaucoup beaucoup plus grand que un ! En pratique, à cause des effets atmosphériques, des miroirs imparfaits et des détecteurs qui saturent, nous n'observons pas de réflectances aussi élevées avec nos satellites, mais nous pouvons encore trouver des réflectances supérieures à 1.
Quand rencontre t'on des réflectances supérieures à un ?
Réflections spéculaires

Un miroir orienté dans la bonne direction peut avoir une réflectance beaucoup plus élevée qu'un. Cela se produit par exemple fréquemment sur des fenêtres ou lucarnes installées sur des toits de maisons. C'est le cas des images ci-dessous, pour lesquelles un utilisateur s'est plaint à tort que nos, produits fournissaient des valeurs de réflectance de surface maximales supérieures à 2, et que donc nos traitements étaient pourris (ce qui n'est pas le cas, ou pas tout à fait )
Exemple de réflectances de surface supérieures à 1, observées sur des images VENµS. il s'agit ici de fenêtres sur les toits qui jouent le rôle de miroirs. Exemple d'image Sentinel-2, où un toit vitré, une serre, ou des panneaux solaires, saturent complètement les détecteurs, avec des réflectances supérieures à un. On observe même que la saturation produit des artefacts perpendiculairement et parallèlement aux détecteurs, hors de la zone saturée sur plus d'une centaine de mètres. Effet des pentes sur les montagnes
Un autre phénomène produit des réflectances supérieures à un :
Définition de l'angle zénithal solaire et de
Un pixel enneigé dans une montagne face au soleil peut aussi atteindre des valeurs supérieures à un. C'est parce que la réflectance est normalisée par le cosinus de l'angle zénithal du soleil (angle entre la direction du soleil et la verticale). Mais dans le cas d'une pente, elle devrait en fait être normalisée par le cosinus de l'angle d'incidence, qui est l'angle entre la direction du soleil et la direction normale à la pente. Sen2cor et MACCS disposent d'une "correction des effets topographiques" qui doivent réduire les réflectances..

Les faces enneigées orientées vers le soleil peuvent présenter des réflectances apparentes supérieures à 1 (les points rouges, par exemple). La correction des effets topographiques disponible dans MAJA permet d'annuler (ou au moins réduire) ces effets. Effet des pentes des nuages
Les nuages épais peuvent aussi avoir des réflectances très élevées, et le même effet que ci-dessus peut être observé pour les pentes à la surface des nuages qui font face au soleil. Mais nous ne sommes pas en mesure de le corriger car cela signifierait avoir un modèle d'élévation du nuage.
References:
Nicodemus, F.E.; J.C. Richmond; J.J. Hsia (October 1977). "Geometrical Considerations and Nomenclature for Reflectance" (PDF). NATIONAL BUREAU OF STANDARDS. Retrieved 8 April 2013.

Schaepman-Strub, G., Schaepman, M. E., Painter, T. H., Dangel, S., & Martonchik, J. V. (2006). Reflectance quantities in optical remote sensing—Definitions and case studies. Remote sensing of environment, 103(1), 27-42.
03 September 2019
18:41
[How it works] Can reflectances be greater than 1 ?
sur Séries temporelles (CESBIO)
Beware, this document contains equations

Having written a lot of posts in this blog, twitter or linkedIn, or due to the distribution of products and softwares on Theia or Github, I get about 10 questions per week in one of these channels, which I try to answer, even if not always at once.

Many questions are about reflectance values, such as:
- "when dividing the images by 10,000 the maximum value I saw was around 1.6, I guess the values are supposed to be within 0 and 1?"
- "I am trying to figure out if the TOA Reflectance range should be limited to 0 and 1 or not. I have read in two different posts that the TOA Reflectance should not be bigger than 1.The part that confuses me is that for many images I am using, the maximum pixel value is above 1"
- Or, in a less diplomatic way from a VENµS user : "It means the CNES radiometric calibration process for Venus imagery has many bugs, starting from L1 all the way to L2 products with max surface reflectance values > 2.0"
But the fact that reflectances should be lower than one is a common belief which is... false !
Photometry

Let's go back to physics :
Angle conventions, N is the North direction, in the text, we used the v subscript for viewing angles instead of r subscript).
a cold and passive surface does not emit light in the solar spectrum (or only negligibly), and only reflects some part of the incoming light. The ratio between the incoming light flux and the outgoing flux in all directions is known as the albedo, and it cannot be greater than one (see this post for a definition of fluxes, irradiance, radiance, and reflectance). But reflectance and albedo are not the same thing.

The Bidirectional reflectance is defined as :

$\rho(\theta_{s},\phi_{s},\theta_{v},\phi_{v}) =\frac{\pi.L(\theta_{s},\phi_{s},\theta_{v},\phi_{v})}{Es.cos(\theta_{s})}$\rho(\theta_{s},\phi_{s},\theta_{v},\phi_{v}) =\frac{\pi.L(\theta_{s},\phi_{s},\theta_{v},\phi_{v})}{Es.cos(\theta_{s})}

where L is the radiance, Es is the sun irradiance and $\theta_s$\theta_s is the solar zenith angle, $\phi_s$\phi_s is the sun azimuth angle, $\theta_v$\theta_v is the viewing zenith angle, and $\phi_v$\phi_v is the viewing azimuth angle.

The black sky albedo BSA, (or directional Hemispherical reflectance (DHR) is the ratio of incoming and outgoing light for a point source, such as the sun, with no atmosphere. The BSA is the integral of the reflectance in all directions (weighted by the cosine of the viewing angle, because we only account for the surface which is perpendicular to the observation direction, and by the sinus of viewing angle, as we integrate along crowns on the hemisphere).

$BSA(\theta_{s},\phi_{s})=\int_{0}^{\pi/2} \int_{0}^{2. \pi} rho(\theta_{s},\phi_{s},\theta_{v},\phi_{v}) cos(\theta_v).sin(\theta_v)d\theta_v d\phi_v$BSA(\theta_{s},\phi_{s})=\int_{0}^{\pi/2} \int_{0}^{2. \pi} rho(\theta_{s},\phi_{s},\theta_{v},\phi_{v}) cos(\theta_v).sin(\theta_v)d\theta_v d\phi_v

Of course the total irradiance reflected by a surface in all directions is lower than the incoming sun irradiance, which means that the albedo is lower than one. Suppose you have a perfect lambertian surface, that scatters all the incoming light in a uniform way in all the upward directions (imagine a perfect white paper sheet). For such a surface, the reflectance is equal to one in all directions, and the albedo, is equal to the reflectance.

Now imagine a perfect mirror, which also reflects all the incoming light, but only in the specular direction. In that case, the reflectance is not at all constant. It is null in all directions, except in a little cone that corresponds to the solid angle of the sun. The albedo is still one, but the reflectance in the specular cone is much greater than one :

$\rho(\theta_{s},\phi_{s},\theta_{s},\phi_{s}+\pi) = \frac{2.\pi}{\omega_s}$\rho(\theta_{s},\phi_{s},\theta_{s},\phi_{s}+\pi) = \frac{2.\pi}{\omega_s}

Where $\omega_s$\omega_s is the sun solid angle (and $2.\pi$2.\pi) is the hemispherical solid angle.

As $\omega_s=6.8 \times 10^{-5}$\omega_s=6.8 \times 10^{-5}, we get a reflectance of about $10^{5}$10^{5} for a perfect mirror, which is much greater than one. In practice, we have atmospheric effects, imperfect mirrors, and detectors that saturate, so we do not observe such high reflectances wth our satellites, but we can still find reflectances greater than 1.
Usual cases of observations of reflectances greater than one : Specular reflections
For instance, a window roof that reflects the sun right right to the satellite camera, can have a reflectance much higher than one. It is the case for the images below, for which a user wrongly complained that our products were providing maximum surface reflectance values greater than 2, and therefore were completely bugged (which is not the case, or not completely the case )
Example of reflectances greater than one observed over roof windows in a VENµS image Slope effect on snow mountains
Another phenomenon contributes to observing reflectances larger than one :

A snow covered pixel in a mountain facing the sun can also reach values above one. This effect is due to the fact that reflectance is normalized by the cosine of sun zenith angle (angle between sun
Definition of sun zenith angle and sun incidence angle
direction and vertical). But in the case of a slope, it should be in fact normalized by the cosine of the incidence angle, which is the angle between sun direction and the normal direction to the slope. Sen2cor and MACCS have a “terrain correction” to correct for this effect.
Slope effect on clouds
Thick clouds can also have very high reflectances, and the same effect as above can be observed for slopes in the cloud surface that face the sun. But we are not able to correct it as it would mean having an elevation model of the cloud.
References:
Nicodemus, F.E.; J.C. Richmond; J.J. Hsia (October 1977). "Geometrical Considerations and Nomenclature for Reflectance" (PDF). NATIONAL BUREAU OF STANDARDS. Retrieved 8 April 2013.

Schaepman-Strub, G., Schaepman, M. E., Painter, T. H., Dangel, S., & Martonchik, J. V. (2006). Reflectance quantities in optical remote sensing—Definitions and case studies. Remote sensing of environment, 103(1), 27-42.
26 August 2019
14:52
Les Sentinels capturent les pierres ponces flottant dans l'océan Pacifique
sur Séries temporelles (CESBIO)

L'éruption d'un volcan sous-marin dans l'océan Pacifique a provoqué la formation de curieux îlots de pierres ponces au large des Tonga. Ces roches volcaniques très poreuses ont une densité inférieure à celle de l'eau de mer ce qui leur permet de flotter à la surface de l'océan.

Ces "radeaux de pierres ponces" (pumice rafts) sont nettement visibles sur les images Sentinel-1 (radar) et Sentinel-2 (optique).

Image Sentinel-2 des radeaux de pierres ponces le 11 août 2019

Sur l'image radar du même jour il est plus facile de discerner leurs contours :

Image Sentinel-1 des radeaux de pierres ponces le 11 août 2019

Enfin on peut estimer que ces radeaux volcaniques se sont déplacés de 45 milles marins (~85 km) en dix jours à partir de deux images prises par le satellite Sentinel-2B :

Images Sentinel-2 des radeaux de pierres ponces les 11 et 21 août 2019
14:52
Les Sentinels capturent les pierres ponces flottant dans l'océan Pacifique
sur Séries temporelles (CESBIO)

L'éruption d'un volcan sous-marin dans l'océan Pacifique a provoqué la formation de curieux îlots de pierres ponces au large des Tonga. Ces roches volcaniques très poreuses ont une densité inférieure à celle de l'eau de mer ce qui leur permet de flotter à la surface de l'océan.

Ces "radeaux de pierres ponces" (pumice rafts) sont nettement visibles sur les images Sentinel-1 (radar) et Sentinel-2 (optique).
Image Sentinel-2 des radeaux de pierres ponces le 11 août 2019
Sur l'image radar du même jour il est plus facile de discerner leurs contours :
Image Sentinel-1 des radeaux de pierres ponces le 11 août 2019
Enfin on peut estimer que ces radeaux volcaniques se sont déplacés de 45 milles marins (~85 km) en dix jours à partir de deux images prises par le satellite Sentinel-2B :
Images Sentinel-2 des radeaux de pierres ponces les 11 et 21 août 2019
20 August 2019
17:51
Sentinel-2 Level-3A time series
sur Séries temporelles (CESBIO)
=>

Update: we have added August 2019, but the synthesis corresponds, so far, to the 1st of August, while that of 2018 was for the 15th of August, please take that into account for your comparisons. We will update with the data from the 15th of August around the 20th of September.

The loop is closed, and we have generated monthly reflectance summaries for France for more than a year (with the exception of winter months that are too cloudy), so we can compare the images from one year to the next. This is particularly interesting in this summer heatwave year (but there was already a drought in northern and eastern France in 2018). With the exception of Brittany and southern Aquitaine, drought and heat waves have almost turned all France to brown, particularly in the northern part of the Massif Central. Feel free to use the comparator below to check.

Comparison of mosaics of Theia monthly syntheses from July 2018 and 2019, obtained thanks to Copernicus Sentinel-2 satellite (click to enlarge, or use the viewer below)

In this production, there is something new. Until now, the Level 3A syntheses were processed by Peter Kettig from CNES using the previous month's Sentinel-2 L2A data. Now the syntheses are (almost) automatically produced by MUSCATE after the WASP processor has been implemented in Theia Ground segment. Well, as always, everything was not straightforward in this production, but the exploitation managed to get the production done. Due to these initial difficulties, we are not able yet to extend the production to other zones, but it is still our aim.

On the images of April and May 2019, you will find holes in the data on the tile column near the Greewich meridian, which passes near Bordeaux, and Normandy. Many Sentinel-2 data are missing, causing holes and artifacts in the data. The problem is not with MUSCATE, MAJA or WASP, but it is due to the download of Sentinel data by PEPS. This problem also affects Level 2A data, so we will have to reprocess everything when the problem is fixed (in the near future, from the latest news). The synthesis of June is also a bit special, as instead of being based on 45 days centered on the 15th of the month, it was based on only 30 days, due to a little design error in MUSCATE. Because of that, you will find a hole around Toulouse. Toulouse had a rather cloud free weather in June and even a heatwave, but had clouds every 5th day when Sentinel-2 was observing. We will reprocess the composite to use the data acquired end may and beginning of June and fill the gaps.
The full resolution data, and the corresponding data quality masks, can be downloaded from Theia's distribution server at CNES.

If you are not afraid to spend too much time while you have urgent things to do, you may also use the nice viewer below (merci Michel Lepage !) to compare with the previous months :

See it full screen

Here how to cite this data set :
made from Sentinel-2 images from the Copernicus program of the European Union
Processed by CNES for Theia data Center, based on methods developped ar CESBIO

In June 2019, , because of a little design error in the integration of WASP within MUSCATE, the exploitation team had to reduce the composite period to 30 days instead of 46. This synthesis will be reproduced when the error is corrected. Due to that, the remaining cloud cover is greater, and around Toulouse, there are no cloud free data, despite the area was mostly cloud free except every 5th day (bad luck).

In May 2019, The cloud cover in may was quite high, and the synthesis keeps several cloudy pixels (which are flagges as such). The column of tiles along the Greenwich meridian still suffers from missing L1C data, and is therefore degraded, with a loss of continuity with the adjacent tiles, and sometimes data gaps.

In April 2019, the weather was very cloudy, with a few spells of sunshine, and good weather at the end of March and beginning of May, which made it possible to produce a cloudless synthesis over 46 days, centred on April 15. Except on a column of tiles, where PEPS seems to have forgotten to retrieve half of the data from ESA servers, I hardly saw any artifacts. We will of course run again the L2A and the syntheses. Snow cover has been reduced, but modestly, due to late snowfall this year. On the other hand, the greening of crops, grasslands and forests is very clear. Like in the song my children learned in kindergarten, "houli houlà, le printemps est là" (spring is here) !

In March 2019 the weather remained nice and warm, and we were beginning to talk about drought. Fortunately for farmers, and unfortunately for the monthly syntheses, April was much more cloudy and rainy. On the March images, I found almost no artifacts, except on water and around the edges of snowy areas. There may be a visible swath edge between Paris and Brittany, but this is not obvious when zooming in. Snow cover has greatly reduced, particularly on the Massif Central, and crops have begun to turn green, especially in the South, much less in Alsace. The deciduous forests wisely waited until April to put their leaves on.

In February 2019 the weather was nice, hot and dry. Sentinel-2 was able to acquire several cloud free images in nearly all regions In February, snow is abundant, winter crops are green, and deciduous forests are brown. That's a normal february image, except for the clouds. Speaking of clouds, a low cloud escaped the detection within MAJA just North of Bourges. A good reminder to go on improving our methods. MAJA 3.3 should improve the results, and it's nearly ready. If you zoom on the snow covered regions, you will see atefacts, which come from WASP. We know how to improve them, we just need to find some time. And finally, because of the low sun elevation, directional effects are increased and our correction model is not perfect. Some swath edges are visible in the West of France. But all in all, it is a good February synthesis.
In November, in France, we had a... French November weather, and several zones stayed overcast for all Sentinel-2 overpasses during the synthesis period of 45 days. In that case, we try to provide a value, which is the minimum reflectance in the blue band. Of course this value is flagged as invalid. So the November synthesis is not as nice as the previous ones, due to the presence of remaining clouds. As in October (see below), we now also see artefacts at the edges of the swath.

Anyway, in many regions, the results are rather correct and they allow us to see the changes. Forests are now brown, soils are wetter and darker, winter crops have started, and the highest mountains are turning white.
In October, we had the first the opportunity to observe a neat swath edge effect in four months, near Cambrai, North of France. The Western part of the artefact is browner than the Eastern part. Because of the cloud cover, the average date used in the eastern part is several days before the average date of the western part. , due to the observation at very different dates on each side of the swath. So the only way to improve that with the current method would be to add a third or even a fourth Sentinel satellite.
Methodology, caveats and artefacts
The monthly syntheses are produced using the WASP processor, which is described here.. In a few words, it computes a weighted average of the cloud free surface reflectances from L2A products processed with MAJA. It is therefore very sensitive to the quality of cloud detection (which is good, thankfully).

By comparing the various syntheses, you will see the evolution of the landscape, generally a little greener in October, but this representation will also help you spot the composite artefacts. These are not very numerous, but you will see them :
- on some web browsers (firefox V58), geometrical differences appear even at a low resolution. Other browsers and versions do not have this defect. It is really not due to Sentinel-2 or Theia products
- above water and snow (we must work on this defect)
- where clouds have covered a place during the whole synthesis period. These pixels are flagged as invalid in the products (but not on the mosaic), and for the mosaics, we provide the minimum of blue value.
- where clouds or shadows were not properly detected by MAJA
- at the edges of Sentinel-2 swath (but the effect is really faint)
- some tile edges in July, due to the fact that Level 3A products were not all generated for the 15th of July, but for dates between the 8th and the 26th. This has been corrected for August and September.
Direct acces to monthly mosaics
Each monthly synthesis is accessible using the following links :
15:46
[How it works] Can reflectances be greater than 1 ?
sur Séries temporelles (CESBIO)
Beware, this document contains equations

Having written a lot of posts in this blog, twitter or linkedIn, or due to the distribution of products and softwares on Theia or Github, I get about 10 questions per week in one of these channels, which I try to answer, even if not always at once.

Many questions are about reflectance values, such as:
- "Also when dividing the images by 10,000 the maximum value I saw was around 1.6, I guess the values are supposed to be within 0 and 1?"
- "I am trying to figure out if the TOA Reflectance range should be limited to 0 and 1 or not. I have read in two different posts that the TOA Reflectance should not be bigger than 1.The part that confuses me is that for many images I am using, the maximum pixel value is above 1"
- Or, in a less diplomatic way from a VENµS user : "It means the CNES radiometric calibration process for Venus imagery has many bugs, starting from L1 all the way to L2 products with max surface reflectance values > 2.0"
But the fact that reflectances should be lower than one is a common belief which is... false !
Photometry

Angle conventions, N is the North direction, in the text, we used the v subscript for viewing angles instead of r subscript).

Let's go back to physics : a cold and passive surface does not emit light in the solar spectrum (or only negligibly), and only reflects some part of the incoming light. The ratio between the incoming light flux and the outgoing flux in all directions is known as the albedo, and it cannot be greater than one (see this post for a definition of fluxes, irradiance, radiance, and reflectance). But reflectance and albedo are not the same thing.

The Bidirectional reflectance is defined as :

$\rho(\theta_{s},\phi_{s},\theta_{v},\phi_{v}) =\frac{\pi.L(\theta_{s},\phi_{s},\theta_{v},\phi_{v})}{Es.cos(\theta_{s})}$ \rho(\theta_{s},\phi_{s},\theta_{v},\phi_{v}) =\frac{\pi.L(\theta_{s},\phi_{s},\theta_{v},\phi_{v})}{Es.cos(\theta_{s})}

where L is the radiance, Es is the sun irradiance and $\theta_s$ \theta_s is the solar zenith angle, $\phi_s$ \phi_s is the sun azimuth angle, $\theta_v$ \theta_v is the viewing zenith angle, and $\phi_v$ \phi_v is the viewing azimuth angle.

The black sky albedo BSA, (or directional Hemispherical reflectance (DHR) is the ratio of incoming and outgoing light for a point source, such as the sun, with no atmosphere. The BSA is the integral of the reflectance in all directions (weighted by the cosine of the viewing angle, because we only account for the surface which is perpendicular to the observation direction, and by the sinus of viewing angle, as we integrate along crowns on the hemisphere).

$BSA(\theta_{s},\phi_{s})=\int_{0}^{\pi/2} \int_{0}^{2. \pi} rho(\theta_{s},\phi_{s},\theta_{v},\phi_{v}) cos(\theta_v).sin(\theta_v)d\theta_v d\phi_v$ BSA(\theta_{s},\phi_{s})=\int_{0}^{\pi/2} \int_{0}^{2. \pi} rho(\theta_{s},\phi_{s},\theta_{v},\phi_{v}) cos(\theta_v).sin(\theta_v)d\theta_v d\phi_v

Of course the total irradiance reflected by a surface in all directions is lower than the incoming sun irradiance, which means that the albedo is lower than one. Suppose you have a perfect lambertian surface, that scatters all the incoming light in a uniform way in all the upward directions (imagine a perfect white paper sheet). For such a surface, the reflectance is equal to one in all directions, and the albedo, is equal to the reflectance.

Now imagine a perfect mirror, which also reflects all the incoming light, but only in the specular direction. In that case, the reflectance is not at all constant. It is null in all directions, except in a little cone that corresponds to the solid angle of the sun. The albedo is still one, but the reflectance in the specular cone is much greater than one :

$\rho(\theta_{s},\phi_{s},\theta_{s},\phi_{s}+\pi) = \frac{2.\pi}{\omega_s}$ \rho(\theta_{s},\phi_{s},\theta_{s},\phi_{s}+\pi) = \frac{2.\pi}{\omega_s}

Where $\omega_s$ \omega_s is the sun solid angle (and $2.\pi$ 2.\pi ) is the hemispherical solid angle.

As $\omega_s=6.8 \times 10^{-5}$ \omega_s=6.8 \times 10^{-5} , we get a reflectance of about $10^{5}$ 10^{5} for a perfect mirror, which is much greater than one. In practice, we have atmospheric effects, imperfect mirrors, and detectors that saturate, so we do not observe such high reflectances wth our satellites, but we can still find reflectances greater than 1.

Usual cases of observations of reflectances greater than one : Specular reflections
For instance, a window roof that reflects the sun right right to the satellite camera, can have a reflectance much higher than one. It is the case for the images below, for which a user wrongly complained that our products were providing maximum surface reflectance values greater than 2, and therefore were completely bugged (which is not the case, or not completely the case )

Example of reflectances greater than one observed over roof windows in a VENµS image

Similar image from a Sentinel-2 image. With reflectances above one (and with detector saturation and smearing on the adjacent pixels)

Slope effect on snow mountains
Definition of sun zenith angle and sun incidence angle

Another phenomenon contributes to observing reflectances larger than one :

A snow covered pixel in a mountain facing the sun can also reach values above one. This effect is due to the fact that reflectance is normalized by the cosine of sun zenith angle (angle between sun direction and vertical). But in the case of a slope, it should be in fact normalized by the cosine of the incidence angle, which is the angle between sun direction and the normal direction to the slope. Sen2cor and MACCS have a “terrain correction” to correct for this effect.

L2A Image acquired by Sentinel-2 in the pyrenees i March, where surface reflectances above 1 can be observed in south oriented slopes, for instances in places identified by red dots. After terrain correction, these reflectances are closer to one (SRTM DEM at 90m is a little too coarse to be exact, and our correction makes a lot of approximations)..
Slope effect on clouds
Thick clouds can also have very high reflectances, and the same effect as above can be observed for slopes in the cloud surface that face the sun. But we are not able to correct it as it would mean having an elevation model of the cloud.

References:
Nicodemus, F.E.; J.C. Richmond; J.J. Hsia (October 1977). "Geometrical Considerations and Nomenclature for Reflectance" (PDF). NATIONAL BUREAU OF STANDARDS. Retrieved 8 April 2013.

Schaepman-Strub, G., Schaepman, M. E., Painter, T. H., Dangel, S., & Martonchik, J. V. (2006). Reflectance quantities in optical remote sensing—Definitions and case studies. Remote sensing of environment, 103(1), 27-42.

11:51

Sentinel-2 Level-3A time series

sur Séries temporelles (CESBIO)

A few days ago, the Theia land data centre released the July 2020 monthly cloud free synthese of surface reflectances from Sentinel-2, for France, but also Italy, Spain, Switzerland, Belgium, Netherlands, the Northern part of Maghreb and Madagascar, and a large part of Sahel The distribution site of Theia products at CNES now provides more than three years of L3A products. Over France, for the third year in a row, the July 2020 synthesis shows the effects of an intense drought which set up this summer. It probably worsened in August, we will see that in a few days.

Animation of monthly syntheses of Sentinel-2 data over France for the month of July, from 2017 to 2020. Among these for years, the normal year should be 2017, but it looks like normality is changing. The data are available at 10m resolution from Theia's distribution site: [https:]]

Summer 2017 was a rather wet and cloudy period, with some pixels still cloudy after 45 days. There is a deep contrast between the healthy vegetation observable in 2017, and the brown hue observable for 2018 , 2019 and 2020 very dry summers. This is also all the more observable in the zoom over the center of France, provided below.


July 2017	July 2018	July 2019

The full resolution data, and the corresponding data quality masks, can be downloaded from Theia's distribution server at CNES.

If you are not afraid to spend too much time while you have urgent things to do, you may also use the nice viewer below (merci Michel Lepage !) to compare with the previous months :

See it full screen

Here is how to cite this data set :
made from Sentinel-2 images from the Copernicus program of the European Union
Processed by CNES for Theia data Center, based on methods developped at CESBIO

In 2019, Pyrénées névés near vignemale mountain have almost disappeared.

Methodology, caveats and artefacts

The monthly syntheses are produced using the WASP processor, which is described here.. In a few words, it computes a weighted average of the cloud free surface reflectances from L2A products processed with MAJA. It is therefore very sensitive to the quality of cloud detection (which is good, thankfully).

With the average of surface reflectances, we also provide the average of the non cloudy dates. This information is important because these dates differ from place to place depending on the cloud cover of #Sentinel overpasses. As you may see on next figure, the weighted average date in July can change by 25 days depending on the regions and the Sentinel-2 orbital configuration. Neglecting this information can introduce an error of 10 to 15 days on the estimation of the pixel date.


July 2020 : weighted average of surface reflectances of cloud free observations	July 2020 : weighted average of the dates of cloud free observations. Blue : beginning of the month, white : End of the month.

By comparing the various syntheses, you will see the evolution of the landscape, generally a little greener in October, but this representation will also help you spot the composite artefacts. These are not very numerous, but you will see them :

on some web browsers (firefox V58), geometrical differences appear even at a low resolution. Other browsers and versions do not have this defect. It is really not due to Sentinel-2 or Theia products
above water and snow (we must work on this defect)
where clouds have covered a place during the whole synthesis period. These pixels are flagged as invalid in the products (but not on the mosaic), and for the mosaics, we provide the minimum of blue value.
where clouds or shadows were not properly detected by MAJA
at the edges of Sentinel-2 swath (but the effect is really faint)
some tile edges in July, due to the fact that Level 3A products were not all generated for the 15th of July, but for dates between the 8th and the 26th. This has been corrected for August and September.

Direct acces to monthly mosaics

Each monthly synthesis is accessible using the following links :

9:34

Les séries temporelles de niveau 3A de Sentinel-2

sur Séries temporelles (CESBIO)

Le centre de données Theia a produit en début de semaine la synthèse mensuelle des données Sentinel-2 pour le mois de juillet 2020 (avec nos chaines WASP et MAJA). Pour la troisième année consécutive, celle-ci est sans nuages et montre une forte sécheresse. Et celle-ci s'est probablement aggravée au cours du mois d'août.

Comparaison des synthèses mensuelles de données Sentinel-2 pour le mois de juillet, de 2017 à 2020. Sur ces 4 synthèses, l'année la plus "normale" est 2017, mais il semble que la normalité ait changé.

Les synthèses à pleine résolution, avec leurs masques de qualité, peuvent être téléchargées depuis le serveur de distribution Theia au CNES.

Si vous n'avez pas peur d'y passer trop de temps, alors que de nombreuses urgences vous attendent, vous pouvez jeter un œil aux mosaïques de ces produits disponibles sur la France depuis Juillet 2017. Une chouette interface de visualisation (merci à Michel Le Page !), est aussi disponible ci-dessous, pour comparer les différentes synthèses deux à deux.

Voir en plein écran

Pour bien citer nos données :
Images Sentinel-2 du programme Copernicus de l'Union Européenne traitées par le CNES pour le pôle Theia à partir de méthodes développées au laboratoire CESBIO

En septembre 2019, les névés des Pyrénées (ici, le Vignemale) ont presque tous disparu

Méthode, artefacts et limites

Les synthèses mensuelles sont produites avec le processeur WASP, qui est décrit ici en détails. En quelques mots, nos synthèses calculent une moyenne pondérée des réflectances de surface pour les observations non nuageuses, issues des produits de niveau 2A obtenus avec la chaîne MAJA. Cette méthode est très sensible à la qualité du masque des nuages, qui, dans le cas de MAJA, est heureusement plutôt bon.

Avec la moyenne pondérée des réflectances, nous fournissons aussi la moyenne pondérée des dates sans nuages utilisées dans la synthèse. Cette information est importante parce que ces dates peuvent varier en fonction de la nébulosité des régions à l'heure de passage du satellite. Comme on peut le voir dans la figure ci-dessous, la date moyenne en juillet peut varier de 25 jours d'un bout à l'autre de la France. Négliger cettee information conduit à introduire une erreur maximale de 10 à 15 jours sur la datation de chaque pixel.


Juillet 2020 : moyenne pondérée des réflectances de surface des pixels non nuageux	Juillet 2020 : dates moyenne d'acquisition des observations non nuageuses. En bleu, début de mois , en blanc, fin du mois.

En comparant les différentes synthèses vous verrez l'évolution du paysage avec le temps, mais cette représentation met aussi en évidence les artefacts dus au traitement et au mode d'acquisition de Sentinel-2. les artefacts ne sont pas très nombreux, mais vous en trouverez :

sur certains navigateurs (firefox V58), des différences géométriques qui apparaissent à basse résolutions. Chez d'autres navigateurs, l'effet n'est pas aussi fort. Ce défaut n'est vraiment pas dû à Sentinel-2 ni aux produits Theia
sur l'eau et la neige (là, c'est vraiment un problème de la méthode, nous testons une autre solution)
là où la couverture nuageuse est restée constante à chaque passage de Sentinel-2. Ces pixels sont indiqués dans les produits (mais pas sur la mosaïque).
là où des nuages ou des ombres n'ont pas été détectés par MAJA (en général, des nuages fins ou petits)
en bordure d'orbite, en raison du changement de date et à l'utilisation d'un modèle simpliste de correction des effets directionnels
sur quelques bordures de tuiles en Juillet 2018, car les produits de niveau 3 n'avaient pas été générés à la même date, mais du 15 juillet au 26 juillet). Cela a été corrigé pour les mois suivants ou précédents.

09 August 2019
21:26
VENµS monitors the resillience of Bouconne forest to caterpillar defoliation
sur Séries temporelles (CESBIO)

=>

I left you in mid-July with an unbearable suspense: would the Bouconne oak forest, the real lung of the Toulouse conurbation, resist the caterpillar attack of which it was the victim? The gypsy moth is a small butterfly whose voracious caterpillar appreciates oaks. This year, the moth caterpillars have flourished in some areas of the Bouconne forest, to the point where its damage has become visible from 700 km above sea level. Everything was explained in this post at the beginning of July.

But rest assured, our satellites have continued to watch over the forest, and it is with joy that we can see that everything is better, thanks to this beautiful series of images from the Venµs satellite of CNES and the Israeli Space Agency. In these images, we used false colors, and the vegetation appears in red. It is the observation in the near infra-red band that is most sensitive to vegetation strength, and it is this spectral band that appears here in red. On this series, we notice the defoliation effect is most significant in June, on two areas of the forest whose color turns brown, and that from mid-July, the situation improves. In the last image of August 4, there is only a slight difference between the areas attacked in June and the others.

Now that the caterpillars have metamorphosed into butterflies that enjoy their summer thinking only about gathering pollen and breeding, the oaks have been able to take advantage of this break to produce new leaves, and start storing reserves for the winter. Experts indicate that in years of moths attacks, which fortunately do not occur every year, the rings on the trunks are tighter, which proves that the attack has an impact on tree growth. If some of these specialists are passing through here, I would like to know why the attacks were concentrated on two bands of the forest, and not on the whole forest. Why did some parts resist better than others? Do we have different tree species, different ages, different soils?

By the way, you can note the large number of clear images available in this series, thanks to the temporal repetitiveness of Venµs, which makes observations every other day. These data have been corrected for atmospheric effects, this is what we call Level 2A products. You can see that the image colours change very slowly with time, without the variations due to the amount of aerosols in the atmosphere. This shows that our treatments, performed with MAJA, work quite well. If you wish to access the data, it is available, free of charge, on https://theia.cnes.fr.

You are now relieved, so I wish you an as good summer as the one enjoyed by the gypsy moths (in their butterfly form)!
15:21
VENµS monitors the resillience of Bouconne forest to caterpillar defoliation
sur Séries temporelles (CESBIO)

=>

I left you in mid-July with an unbearable suspense: would the Bouconne oak forest, the real lung of the Toulouse conurbation, resist the caterpillar attack of which it was the victim? The gypsy moth is a small butterfly whose voracious caterpillar appreciates oaks. This year, the moth caterpillars have flourished in some areas of the Bouconne forest, to the point where its damage has become visible from 700 km above sea level. Everything was explained in this post at the beginning of July.

But rest assured, our satellites have continued to watch over the forest, and it is with joy that we can see that everything is better, thanks to this beautiful series of images from the Venµs satellite of CNES and the Israeli Space Agency. In these images, we used false colors, and the vegetation appears in red. It is the observation in the near infra-red band that is most sensitive to vegetation strength, and it is this spectral band that appears here in red. On this series, we notice the defoliation effect is most significant in June, on two areas of the forest whose color turns brown, and that from mid-July, the situation improves. In the last image of August 4, there is only a slight difference between the areas attacked in June and the others.

Series of Venµs images in false colors, from May to early August 2019, above the Bouconne forest, west of Toulouse, France. The data are presented in false colors, the near infrared is in red, while the blue and green channels correspond to the blue and green bands. The more vigorous the vegetation, the more red the colour.

Now that the caterpillars have metamorphosed into butterflies that enjoy their summer thinking only about gathering pollen and breeding, the oaks have been able to take advantage of this break to produce new leaves, and start storing reserves for the winter. Experts indicate that in years of moths attacks, which fortunately do not occur every year, the rings on the trunks are tighter, which proves that the attack has an impact on tree growth. If some of these specialists are passing through here, I would like to know why the attacks were concentrated on two bands of the forest, and not on the whole forest. Why did some parts resist better than others? Do we have different tree species, different ages, different soils?

By the way, you can note the large number of clear images available in this series, thanks to the temporal repetitiveness of Venµs, which makes observations every other day. These data have been corrected for atmospheric effects, this is what we call Level 2A products. You can see that the image colours change very slowly with time, without the variations due to the amount of aerosols in the atmosphere. This shows that our treatments, performed with MAJA, work quite well. If you wish to access the data, it is available, free of charge, on https://theia.cnes.fr.

You are now reassured, so I wish you an as good summer as the one enjoyed by the gypsy moths (in their butterfly form)!
14:56
[VENµS] La forêt de Bouconne a résisté à l'attaque à la bombyx
sur Séries temporelles (CESBIO)

=>

Je vous avais laissés à la mi-juillet avec un suspense insoutenable: la forêt de chênes de Bouconne, véritable poumon de l'agglomération de Toulouse, allait-elle résister à l'attaque à la bombyx dont elle était la victime ? Le bombyx disparate est un petit papillon dont la chenille vorace apprécie les chênes, qui cette année a pullulé dans certaines zones de la forêt de Bouconne, au point où ses dégâts sont devenus visibles depuis 700 km d'altitude. Tout était expliqué dans ce billet de début Juillet.

Mais rassurez vous, nos satellites ont continué à veiller sur la forêt, et c'est avec joie que nous pouvons constater que tout va mieux, grâce à cette belle série d'images du satellite Venµs du CNES et de l'agence spatiale Israélienne. Sur ces images, nous avons utilisé des fausses couleurs, et la végétation apparaît en rouge. C'est l'observation dans le proche-infra rouge qui est la plus sensible à la vigueur de la végétation, et c'est cette bande spectrale qui apparaît ici en rouge. Sur cette série, on remarque que c'est en juin que l'effet de défoliation est le plus important, sur deux zones de la forêt dont la couleur tourne au brun, et que dès la mi juillet, la situation s'améliore. Sur la dernière image du 4 août, on ne distingue plus qu'une légère différence entre les zones attaquées en juin et les autres.

Série d'images Venµs en fausses couleurs, de mai à début août 2019, au dessus de la forêt de Bouconne à l'ouest de Toulouse. Les données sont présentées en fausses couleurs, le proche infra rouge est en rouge, alors que les canaux bleu et vert correspondent aux bandes bleues et vertes. Plus la végétation est vigoureuse, plus la couleur est rouge.

Maintenant que les chenilles se sont métamorphosées en papillons qui ne pensent plus qu'à butiner et se reproduire (vivent les vacances !), les chênes ont pu en profiter pour produire de nouvelles feuilles, et commencer à stocker des réserves pour l'hiver. Les spécialistes indiquent que les années d'attaques à la bombyx, qui heureusement n'ont pas lieu tous les ans, les cernes sur les troncs sont plus resserrées, ce qui prouve que l'attaque a un impact sur la croissance des arbres. Si certains de ces spécialistes passent par ici, j'aimerais bien savoir pourquoi les attaques se sont concentrées sur deux bandes de la forêt, et pas sur toute celle-ci. Pour quelles raison certaines parties ont-elle mieux résisté que d'autres ? Avons nous des espèces d'arbres différents, des ages différents, des sols différents ?

Au passage, vous pouvez noter le grand nombre d'images claires disponibles sur cette série, grâce à la répétitivité temporelle de Venµs, qui fait des observations tous les deux jours. Ces données ont été corrigés des effets atmosphériques, ce sont des produits de niveau 2A. Vous pouvez observer que les teintes changent très lentement, sans les variations dues à la quantité d'aérosols dans l'atmosphère, ce qui montre que nos traitements, réalisés avec MAJA, fonctionnent bien. Si vous souhaitez accéder aux données, celles-ci sont disponibles, gratuitement, sur [https:]] .

Vous voilà rassurés, vous pouvez donc passer d'aussi bonnes vacances que les bombyx disparates (sous leur forme de papillons) !
08 August 2019
22:39
[VENµS] La forêt de Bouconne a résisté à l'attaque à la bombyx
sur Séries temporelles (CESBIO)

=>

Je vous avais laissés à la mi-juillet avec un suspense insoutenable: la forêt de chênes de Bouconne, véritable poumon de l'agglomération de Toulouse, allait-elle résister à l'attaque à la bombyx dont elle était la victime ? Le bombyx disparate est un petit papillon dont la chenille vorace apprécie les chênes, qui cette année a pullulé dans certaines zones de la forêt de Bouconne, au point où ses dégâts sont devenus visibles depuis 700 km d'altitude. Tout était expliqué dans ce billet de début Juillet.

Mais rassurez vous, nos satellites ont continué à veiller sur la forêt, et c'est avec joie que nous pouvons constater que tout va mieux, grâce à cette belle série d'images du satellite Venµs du CNES et de l'agence spatiale Israélienne. Sur ces images, nous avons utilisé des fausses couleurs, et la végétation apparaît en rouge. C'est l'observation dans le proche-infra rouge qui est la plus sensible à la vigueur de la végétation, et c'est cette bande spectrale qui apparaît ici en rouge. Sur cette série, on remarque que c'est en juin que l'effet de défoliation est le plus important, sur deux zones de la forêt dont la couleur tourne au brun, et que dès la mi juillet, la situation s'améliore. Sur la dernière image du 4 août, on ne distingue plus qu'une légère différence entre les zones attaquées en juin et les autres.

Maintenant que les chenilles se sont métamorphosées en papillons qui ne pensent plus qu'à butiner et se reproduire (vivent les vacances !), les chênes ont pu en profiter pour produire de nouvelles feuilles, et commencer à stocker des réserves pour l'hiver. Les spécialistes indiquent que les années d'attaques à la bombyx, qui heureusement n'ont pas lieu tous les ans, les cernes sur les troncs sont plus resserrées, ce qui prouve que l'attaque a un impact sur la croissance des arbres. Si certains de ces spécialistes passent par ici, j'aimerais bien savoir pourquoi les attaques se sont concentrées sur deux bandes de la forêt, et pas sur toute celle-ci. Pour quelles raison certaines parties ont-elle mieux résisté que d'autres ? Avons nous des espèces d'arbres différents, des ages différents, des sols différents ?

Au passage, vous pouvez noter le grand nombre d'images claires disponibles sur cette série, grâce à la répétitivité temporelle de Venµs, qui fait des observations tous les deux jours. Ces données ont été corrigés des effets atmosphériques, ce sont des produits de niveau 2A. Vous pouvez observer que les teintes changent très lentement, sans les variations dues à la quantité d'aérosols dans l'atmosphère, ce qui montre que nos traitements, réalisés avec MAJA, fonctionnent bien. Si vous souhaitez accéder aux données, celles-ci sont disponibles, gratuitement, sur [https:]] .

Vous voilà rassurés, vous pouvez donc passer d'aussi bonnes vacances que les bombyx disparates (sous leur forme de papillons) !
25 July 2019
14:52
Quatre saisons de vélotaf sur le Canal du Midi
sur Séries temporelles (CESBIO)

Pour changer un peu des images satellites voici une série temporelle de 140 photos prises entre le 03 septembre 2018 et le 24 juillet 2019 depuis le pont situé en face des résidences de Supaéro. Chaque photo a été prise un jour différent (cliquer sur l'image pour l'agrandir).

J'ai essayé de prendre une photo chaque jour ouvré, mais parfois j'ai oublié, parfois j'étais en déplacement, en vacances...

L'heure de capture médiane de ces photos est 08h44 (j'ai l'habitude de partir au travail après avoir déposé les enfants à l'école à 08h30)

Je vais devoir interrompre la série car je pars en vacances une bonne partie du mois d’août ! A moins que d'autres vélotafeurs prennent le relais ?
18 July 2019
15:53
Synthèses mensuelles de la durée d'enneigement dans les Alpes françaises
sur Séries temporelles (CESBIO)

Pour le Conservatoire botanique national alpin j'ai généré les cartes de durée d'enneigement par mois (avril, mai, juin et juillet) pour 2016, 2017 et 2018 à partir de toutes les images Sentinel-2 et Landsat-8 disponibles chez Theia sur les Alpes (tuiles 31TGM 32TLS 31TGL 32TLR 31TGK 32TLQ). Normalement la méthode est conçue pour faire des synthèses annuelles donc j'étais curieux de voir le résultat au pas de temps mensuel. En guise de sanity-check, j'ai calculé la moyenne de la durée d'enneigement au dessus de 2000 m par tuile, par année et par mois. J'ai sélectionné les surfaces au-dessus de 2000 m car dans ce post j'avais estimé qu'il faut regarder au-dessus de 2000 m pour éviter les forêts dans les tuiles T31TGL, T31TGM, T32TLR et T32TLS. Ces graphes montrent bien l'effet des chutes de neige exceptionnelles de l'hiver 2017-2018 qui ont causé des durées d'enneigement moyennes élevées à haute altitude dans les Alpes du Sud. Autre sanity-check, les vignettes des durée d'enneigement plaquées sur un relief ombré pour visualiser un peu plus en détail les données. Cette fois aucun filtre n'est appliqué. La légende est en nombre de jours. On voit une erreur (nuage détecté comme neige) sur la tuile 31TGM en juillet 2017 qui a pu légèrement fausser le graphe ci-dessus, mais sinon les cartes semblent cohérentes.

Pour 2019 il faudra attendre mon retour de vacances ! PS. Un grand merci à Christophe Taillan au CNES pour sa disponibilité. En cette période de sécheresse, au moins le datalake du CNES ne désemplit pas
09 July 2019
10:34
Sentinel-2 sees caterpillars nibbling French forests
sur Séries temporelles (CESBIO)

=>

One of my colleagues from CESBIO and CNES, a specialist in automatic classification methods and creator of free software named after Monteverdi's opera, told me that during his last Sunday walk in the Bouconne forest, the trees were being devoured by caterpillars. By listening, you could even hear them nibbling on the leaves. Even if he doesn't have my exaggeration tendencies, I wanted to check with my favourite satellite, Sentinel-2.

The image below compares, in false colours (the more red the vegetation is, the more vigorous the vegetation), the images acquired on 5 July 2017 and 5 July 2019 (the same period was cloudy in 2018). The differences are impressive, the red colour fades and almost disappears in places.

False color comparison (vegetation appears red) of Sentinel-2 level 2A images (processed by Theia) acquired on July 5, 2017 and 2019 over Boucone Forest near Toulouse. Some forest areas appear darker in 2019, due to the appetite of gypsy moth caterpillar...

A little research has taught me that this phenomenon is due to the caterpillar of a butterfly, the disparate gypsy moth, of local origin. Some years, it swells. There is an Asian variant of it that was introduced in the United States in the 19th century. She caused serious damage there under the name of Gypsy moth. This butterfly is particularly fond of oak leaves. It is called "disparate" because the male and female have very different aspects:

It is its magnificent caterpillar which causes all the damage, when the conditions are favourable:

Par Didier Descouens — Travail personnel, CC BY-SA 4.0, Lien

The forest of Bouconne is not the only one in France to have been attacked this year, it is also the case of the forest of the Moors on the French Riviera, and there the effects are even more impressive, as shown in the article of var morning below, and especially the comparison of the images 2018 and 2019. Seen from above, you'd think the forest had burned down. It seems that the trees can recover from gypsy moth outbreaks, as the larvae turn into butterflies in July and let the leaves grow back.

False colour comparison (vegetation appears in red) of Sentinel-2 images acquired in late June 2018 and 2019. Some forest areas appear black due to the gypsy moth attacks, as if they were burned. The oblique line on the sea is due to observations from slightly different angles by Sentinel-2's detectors.
08 July 2019
23:21
Bombyx attaque
sur Séries temporelles (CESBIO)

=>

Un de mes collègues du CESBIO et du CNES, spécialiste des méthodes de classification automatique et créateur de logiciels libres, m'a raconté que lors de sa dernière ballade dominicale dans la forêt de Bouconne, les arbres étaient en train d'être dévorés par des chenilles. En tendant l'oreille, on pouvait même les entendre grignoter les feuilles. Même s'il n'a pas mes tendances à l’exagération, j'ai voulu vérifier avec mon satellite préféré, Sentinel-2.

L'image ci dessous compare, en fausses couleurs (plus c'est rouge plus la végétation est vigoureuse), les images acquises le 5 juillet 2017 et le 5 juillet 2019 (la même période était nuageuse en 2018). Les différences sont impressionnantes, la couleur rouge s'atténue et disparaît presque par endroits, signe que les chenilles ont consommé une bonne partie des feuilles.

Comparaison en fausse couleur (la végétation apparaît en rouge) d'images Sentinel-2 de niveau 2A (traitées par Theia) acquises les 5 juillet 2017 et 2019 sur la forêt de Bouconne, près de Toulouse. Certaines zones forestières apparaissent plus sombres en 2019, du fait de l’appétit des larves de bombyx..

Une petite recherche m'a appris que ce phénomène est dû à la chenille d'un papillon, le bombyx disparate, d'origine locale. Certaines années, il pullule. Il en existe une variante asiatique qui a été Introduite aux Etats-Unis au 19e siècle. Elle y cause de graves dégâts sous le nom de Gypsy moth. Ce papillon est particulièrement amateur de feuilles de chêne. Il est dit "disparate" car le mâle et la femelle ont des aspects très différents:

C'est sa magnifique larve qui lorsque les conditions sont favorables cause tous les dégâts :

Par Didier Descouens — Travail personnel, CC BY-SA 4.0, Lien

La forêt de Bouconne n'est pas la seule en France à avoir été attaquée cette année, c'est aussi le cas de la forêt des Maures sur la côte d'Azur, et là les effets sont impressionnants, comme le montre l'article de var matin ci dessous, et surtout la comparaison des images 2018 et 2019. Vu d'en haut, on pourrait croire que la forêt à brûlé. Il semble que les arbres puissent se remettre des attaques de bombyx, car la larve se transforme en papillon en juillet et laisse les feuilles repousser.

Voir l'article de var matin Voir l'article de la dépèche

Comparaison en fausse couleur (la végétation apparaît en rouge) d'images Sentinel-2 acquises fin juin 2018 et 2019. Certaines zones forestières apparaissent noires. Le trait oblique sur la mer est dû aux observations sous des angles légèrement différents par les détecteurs de Sentinel-2.

Skip to page: 1 2 3

← UrbaLine (le blog d'Aline sur l'urba,... Datafoncier, données pour les ... →

Communauté	Ressources	L'emploi	Le marché	L'association	Fils RSS
Forum Blogs Wiki Planet Carte des membres Sites amis	Agenda Biblio Netvibes Fiches AMO-CNIG Paris PCGIS Téléchargements	Offres d'emploi Banque de CV Carte des CV Formations Démarche Métiers	Geo-Entreprises Geo-Communiqué Appels d'offres	L'association Qui sommes-nous ? Nous aider Bienfaiteurs Partenaires Historique Statuts	Tout le site Job Géomatique Services GeoBlogs ...	Faire un don Mentions légales


Masque de nuage à 240m de résolution, exemple sur la Guyane Française	Masque de nuage à 120m de résolution. Les petits nuages et leurs ombres sont bien mieux détectés


Cloud mask generated at 240m resolution, example in French Guyana	Cloud mask generated at 120 m resolution. More small clouds and more shadows are detected.


WASP	IOTA2	LIS	S1-Tiling

Le portail francophone de la géomatique

Feeds

Séries temporelles (CESBIO)

Enclosure: [download]

Enclosure: [download]