New paper ! An active learning cloud detection tool to generate reference cloud masks for Sentinel-2. Application to the validation of MAJA, Sen2cor and FMask cloud masks

Example of reference cloud mask generated by ALCD, and comparison with the cloud masks generated by three operational processors (Sen2cor, FMask and MAJA). True positive invalid pixels appear in blue, true negative in green, false negative in red and false positive in purple..

It is not that frequent when the work of a trainee ends up as a peer reviewed publication, but Louis Baetens was a brilliant trainee. In a six months training period at CESBIO, funded by CNES, here is what Louis Baetens did:

  • developed an active learning method to generate reference cloud masks for Sentinel-2, using multi-temporal data as input
  • validated the quality of the produced masks (around 99% overall accuracy)
  • generated cloud and shadow masks covering 32 entire Sentinel-2 images
  • produced these same scenes with Sen2cor 2.5.5, FMask 4.0 and MAJA 3.3
  • evaluated the results using ALCD masks
  • wrote a report and a user manual for ALCD
  • released the masks and tools on open access platforms
  • And wrote (with Camille and myself) a scientific publication

 

The publication was just released by remote sensing :

Baetens, L.; Desjardins, C.; Hagolle, O. Validation of Copernicus Sentinel-2 Cloud Masks Obtained from MAJA, Sen2Cor, and FMask Processors Using Reference Cloud Masks Generated with a Supervised Active Learning Procedure. Remote Sens. 2019, 11, 433.

 

The remaining of the post provides a plain language summary (but it's better to read the paper !)

Continue reading

A pause in MUSCATE production, end of February

The MUSCATE production centre will be offline for a week from February 25th to March 4th (or maybe the week after, please see Theia's RSS feed to see the exact date). This pause will be necessary to upgrade the processing centre to V2.5. As a result, we will not be able, for a week, to process Sentinel-2 data in real time, and hope it will not be too inconvenient to you. In case you need some data urgently, you can ask PEPS to produce MAJA L2A data for you during that period.

 

The main change regards the internal format used by MUSCATE for Sentinel-2 Level 2A products. This internal format is different from the external format that we distribute, and this results in unnecessary product format conversions, as well as the necessity to develop drivers for the various formats (internal and external) for the processors that use Sentinel-2 L2A data within MUSCATE. To enable this modification, it will be necessary to convert all the L2A data from the internal format to the external format, which will take a whole week.

 

MUSCATE V2.5 will bring other improvements, such as the integration of MAJA 3.1, with possibility to use Copernicus Atmosphere data, or a new version of LIS, the processor that delivers the snow maps.

MUSCATE  V2.6 is also ready and queuing to be installed, with MAJA 3.2, WASP and the possibility to process Venµs L2A data within MUSCATE and not externally on Venµs ground segent.

 

 

 

Diffusion d'un premier lot des données Sentinel-2A de niveau 2A sur le Sahel

Il y a quelques semaines, nous annoncions la sélection d'une nouvelle zone de production de données Sentinel-2 au niveau 2A par Theia, au Sahel. La production a démarré, et Theia a déjà produit les tuiles de la zone UTM28 (à l'ouest). Les tuiles en vert foncé existaient déjà, mais nous avons rajouté celles en vert clair, qui permettent de couvrir l'ensemble du Sénégal, la Gambie,  une partie de la Guinée Bissau, de la Guinée, et le nord de la Sierra Leone.

 

Les données disponibles ont été traitées du premier janvier 2017 à hier, soit plus de deux ans de données. Les nouvelles données seront maintenant traitées en temps réel au fur et à mesure de leur arrivée.

Nous procéderons de même avec les différentes zones  de l'ouest vers l'est : UTM29, UTM30...

 

Our blog's audience in 2018

A seventh year begins for the "Séries temporelles" blog, and as usual, it is an opportunity to review its audience, and to get a little self-satisfaction. We usually publish this post early January, but it seems there was no January this year (or was I too busy ?). The blog is still receiving more visits every year, with a sharp growth this year : +35% of visits ...

Blog traffic from December 2012 to January 2019

Blog traffic from December 2012 to January 2019 (the trends were computed using the Theil–Sen estimator), computed by Simon Gascoin.

So, if we look at the trends on the plots above, the audience growth is remarkably linear, but if we sum-up everything per year, we see a sharp increase. The cause is that outlier in the top-right corner or each graph above, related to a big buzz in Japan for SImon's article about Xe-Namnoy lake dam failure in July, that flooded several villages, and killed too many people.

 

2013 2014 2015 2016 2017 2018
Number of visits 13985 22928 34723 47773 57692 79243
Number of viewed pages 30922 46940 66947 89555 105846 131846

 

French visitors only counted for 25% of visits, much less than the other years. Japan ranked second with 16%,  United States ranked third, followed by European countries (UK, Germany, Italy, Spain) and by India, Canada and Morocco.

Continue reading

Improvement of water vapour retrieval in MAJA

Similarly to the aerosol retrieval, the retrieval of water vapour in MAJA atmospheric correction has also been improved, thanks to the work of Elsa Bourgeois (Cap Gemini) and Camille Desjardins (CNES). An accurate estimation of water vapour is not necessary to perform an accurate atmospheric correction, because water vapour absorption in most of Sentinel-2 bands is much lower than 5%. But the Sentinel-2 water vapour product could also prove useful, and when we plot validation results, showing a large bias for high water vapour contents is not nice.

 

 

 

Here is the kind of results we have been having with MAJA from the beginning, with a large bias when water vapour content is high :

Our very simple method uses the ratio between Sentinel-2 B9 and B8a bands to estimate the water vapour. B9 is located within a water vapour absorption band at 940 nm, while B8a serves as reference and is only moderately affected by water vapour. The ratio is converted thanks to the use of a Look-up table, which is obtained using radiative transfer calculations. Our method assumes that the water vapour is above the scattering layer, which is obviously not true. The errors due to this assumption increase with the amount of water vapour.

 

Elsa and Camille just empirically computed a new water vapour LUT to cancel this bias, and it works! As you can see, the RMS errors have been divided by a factor 2, from 0.2 g/cm2 to 0.1 g/cm2.

We will put this new parameter set in production in January within Theia, and make it available to the users of MAJA processor.

 

 

 

 

MAJA 3.1.2 with CAMS option finally validated

We had announced quite a long time ago the coming availability of MAJA 3.1 to correct for the atmospheric effects on Sentinel-2, Landsat 8 or Venµs satellites. This version brings a significant improvement in the estimation of Aerosol Optical Thickness, thanks to the use of Copernicus Atmosphere Monitoring Service (CAMS) data to constrain the aerosol type. The details of the methodscan be found here. Bastien Rouquié obtained them on our python prototype of MAJA.

 

We then implemented them in the operational and fast version of MAJA. If the validation tests of MAJA 3.1 were correct on the two test products we had defined, a large scale validation using 10 sites over two year time series showed that instead of improving, using the CAMS option was degrading the results. We had to search for the cause (a bad interpolation of CAMS data in space and time), and correct the errors and perform again a large validation.

 

This time, the validation results are improving a lot, as it may be seen on the figures below.

Without CAMS option With CAMS option

On the left column, we provide the results without activating CAMS option, while on the right, it is activated. The top row corresponds to the comparison between Aeronet AOT used as reference, and MAJA AOT, for eight sites in diverse landscapes. The bottom row provide an example on the well known validation site in Mongu, Zambia.The blue dots correspond to good quality aerosol measurements (no clouds, level 2.0 aeronet values), while red dots correspond to degraded conditions (with either clouds or not quality assured aeronet data (level 1.5 data)

 

Using CAMS to constrain the aerosol type improves the results by 25%, compared to the use of a continental aerosol model everywhere. Errors for the quality assured validation pixels decrease from 0.085 to 0.065 on the 8 sites, and from 0.143 to 0.094 on Mongu site in Zambia. This site has various types of aerosols depending on the season, including dust, biomass burning and continental aerosols. The results are still far from perfect, and we have work for the next 5 years, but it is still good to have them improved !

 

MAJA 3.1.2 is available starting from this link on github, as an executable program for linux. To be allowed to use it, you will have to sign the licence first, from this site.  If you want to use it for commercial applications, you should ask for a special licence (still for free), sending me an email. In January, I will provide the parameters to allow activate the CAMS options.

 

Regarding the production of Theia, our ground segment has been adapted to use MAJA version 3.1.2, and will soon be able to fetch the CAMS products from Copernicus Atmosphere. Then we will have an operational qualification phase, to check that we can download CAMS products in time for real time production. We should be able to start using in in February or March.  And after a few months, if the results are good, yoohoo, we will reprocess everything !

 

Many thanks to Bastien Rouquié, CESBIO, who did the scientific work, to Camille Desjardins w ho helpled with the validation, to Aurelien Bricier and Benjamin Esquis, at CS-SI for coding the operational version, and Peter Kettig (CNES) and Bruno Angeniol (Cap Gemini), and Bastien, for checking the consistency between prototype and operational versions.

 

 

 

Les séries temporelles de niveau 3A de Sentinel-2, de Juillet à Novembre

=>

Comme chaque mois, Peter Kettig du CNES a produit les synthèses de niveau 3A à partir de Sentinel-2 du mois précédent (Merci Peter !). Les données 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 oeil aux mosaïques de ces produits disponibles sur la France depuis Juillet. Chaque mosaique est accessible à partir des liens ci-dessous :

 
Une chouette interface de visualisation (merci à Michel Lepage !), est aussi disponible ci-dessous, pour comparer la synthèse d'octobre avec celle des mois précédents.
 
En novembre , en France, nous avons eu une météo française de novembre, et plusieurs régions sont restées couvertes à chaque passage de Sentinel-2, sur la période de synthèse de 45 jours, centrée sur le 15 Novembre. Dans ce cas, nous fournissons les réflectances de la date où la réflectance dans le bleu est minimale, et nous indiquons dans les produits, pas sur la mosaïque, que le pixel est nuageux. Donc, la synthèse du mois de novembre n'est pas aussi belle que les précédentes. Comme en octobre (voir ci-dessous), des bords d'orbite deviennent visibles.
 
Ceci dit, les résultats restent corrects sur de nombreuses régions, et on peut observer les sols plus humides et plus sombres, le démarrage des cultures d'hiver, la chute des feuilles dans les forêts de feuillus, et les sommets enneigés.
 
En octobre, pour la première fois, un artefact de bord d'orbite apparaît clairement du côté de Cambrai. Même si c'est un endroit où l'on peut faire des bêtises, cet artefact est dû au changement de date. la partie Ouest est brune, et la partie Est est verte. A cause de la couverture nuageuse importante, la date moyenne de la partie Est se situe bien après celle de la partie Ouest. Le seul moyen d'éviter ce genre d'artefacts sera d'a jouter un ou deux satellites Sentinel-2 de plus, pour faire des synthèses sur des données moins longues (ici nous utilisons 45 jours).

 

See it full screen
Continue reading

Sentinel-2 Level3A time series (July to November 2018)

=>

As every month, Peter Kettig from CNES processed the Sentinel-2 L3A composites of France from the Month before. 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 have a look to the mosaic of Sentinel-2 monthly syntheses for each month since July over France. Each monthly synthesis is accessible using the following links :

Or you may also use the nice viewer below (merci Michel Lepage !) to compare with the previous months.

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.

See it full screen
Continue reading

[MUSCATE News] Unavailability of CNES HPC center on 4-5 December

Le centre de calcul (de haute performance) du CNES sera en maintenance les 4 et 5 décembre, dans le but d'accroitre sa robustesse et ses performances. En conséquence, MUSCATE ne sera pas en mesure de produire les données de Sentinel-2, Landsat et Venµs en temps réel. Pour une fois, nous ne serons donc pas en mesure de tenir nos engagements de production en moins de 2.5 jours, et nous espérons que cela ne vous dérangera pas trop dans vos travaux. Les données déjà produites resteront disponibles.

Pour la même raison, les traitements avec MAJA sur PEPS seront suspendus.

 

CNES High Performance Computing center will be on maintenance on the 4th and 5th of December, to improve its robustness and capacity. As a result, MUSCATE will not be able to produce Sentinel-2, Landsat and Venµs data during these days. Exceptionally, we will not meet our target to process these data in less than 2.5 days, and we hope it will not cause too much inconvenience. The already produced data will still be available.

For the same reason, MAJA processing on PEPS will be suspended.