Snow conditions in southern Africa ski resorts

When I present the potential of Sentinel-2 for snow science, I often tell that the spatial resolution of Sentinel-2 is sufficient to detect snow at the scale of the ski runs. Because a picture is worth a thousand words, here is the Sentinel-2 view of the only two ski resorts in southern Africa on July 11.

Sentinel-2 true color composites on 11 July 2018

The snow on these ski slopes is artificial but this region can get quite a lot of snow!

NDVI time series in 2018 World Cup stadiums

The figure below shows the evolution of the Normalized Difference Vegetation Index (NDVI) in the pitches of all the 2018 World Cup stadiums.

NDVI in the piches of the 2018 FIFA World Cup stadiums. Data from Copernicus Sentinel-2.

NDVI in the 2018 FIFA World Cup stadiums

I retrieved these data from Sentinel-2 observations using the new time series plotter in the EO Browser. I just drew a polygon in each of the 12 stadiums to get the average NDVI values on every available cloud-free date.

Sentinel-2 NDVI on June 27 in Kaliningrad Stadium (Arena Baltika)

Arena Baltika

Time series of the NDVI in Arena Baltika from Sentinel-2 observations in the EO Browser

Because the NDVI is a proxy of the vegetation health (here the grass on the pitch), these charts allow us to identify which stadiums were built for the 2018 World Cup (Volgograd Arena, Cosmos Arena). On the other hand, the Fisht stadium in Sotchi looks well maintained since 2016. It "served as the venue for the opening and closing ceremonies of the 2014 Winter Olympics (...) was originally built as an enclosed facility; it was re-opened in 2016" (wikipedia). Also it should be noted that the Krestovsky Stadium in Saint Petersburg is a retractable roof stadium. "As of May 2017, the stadium was 518% late and 548% over budget (...) At a cost of $1.1 billion at current exchange rates, it is considered one of the most expensive stadiums ever built." (wikipedia). Hopefully the grass will remain green in the next months, unlike some stadiums in Brazil after the Olympics.
 
In the meantime, as you can see by yourself, the grass is blue in the Kazan Arena!

Color composite of the Kazan Arena on June 27, three days before the first encounters of the Round of 16 (France vs. Argentina)

Three snow seasons in the Pyrenees through the eyes of Sentinel-2 and Landsat-8

On June 23 we will celebrate the third anniversary of Sentinel-2A in orbit. With three years of data we can start looking at the inter-annual variability of biophysical variables, like.. (random example), the snow cover.

 

This is what I attempted to do for the Theia workshop. I downloaded all available snow cover products from Theia over the Central Pyrenees (tile 31TCH) and I generated additional snow maps from the Theia Landsat-8 level-2A products using let-it-snow processor. Landsat-8 images enable to increase the frequency of observations when only Sentinel-2A was operational between 2015 to 2017.

 

I resampled the Landsat-8 snow maps to the same reference grid as Sentinel-2 at 20 m resolution using the nearest neighbor method. I cropped all snow maps to the intersection of the Sentinel-2 tile (green polygon) and Landsat-8 tile (red polygon).


When there was a snow map from Sentinel-2 (S2) and Landsat-8 (L8) on the same day, I merged them into a composite using a simple pixel-based rule:
Continue reading

La vectorisation du produit OSO, comment ça marche ?

Le produit vecteur d'OSO 2017 est enfin sorti ! Après plusieurs semaines de traitements, les vecteurs de chaque département sont disponibles ici. La production requiert la mobilisation d'une grande quantité de ressources de calcul et une stratégie de traitements un peu particulière. Nous voulions vous expliquer comment parvient-on à produire cette couche d'information.

Exemple du raster initial (10 m), régularisé (20m) et vectorisé

A priori, le plus simple serait de prendre la couche raster issue de la chaine de traitements iota² de l'intégrer dans notre logiciel SIG préféré et d'appuyer sur le bouton "Vectorisation" ! Mais les choses ne sont pas si simples, certaines contraintes et besoins nous obligent à quelques tours de passe-passe :

Continue reading

Are Sentinel-2 water vapour products useful ?

Atmospheric absorption: in blue, the surface reflectance of a vegetation pixel, as a function of wavelength. In red, the reflectance of the same pixel at the top of atmosphere. For a wavelength of 1.38 µm, water vapour totally absorbs the light that comes from the earth surface at sea level. At 0.94 µm (940nm), a weaker water vapour absorption band only partly absorbs the photons.

=>

Sentinel-2B has two channels centered on water vapour absorption bands: channel 9 (940 nm) and channel 10 (1380 nm). Band 10 corresponds to a very strong absorption, strong enough to prevent any photon to reach ground from the Sun without being absorbed in the atmosphere. This band is intensively used to detect and correct high clouds.

 

In this blog, we discussed much less band 9 (940 nm) yet. Here, water vapour absorption is not strong enough to catch all the photons which reach the surface. The proportion of absorbed photons depends on the water vapour atmospheric content, and also on the viewing and solar zenith angles. We use band 9 for atmospheric correction, but it could be useful to study convection phenomena within the atmosphere too.

Continue reading

Des applications pour la vapeur d'eau observée par Sentinel2 ?

Absorption atmosphérique. En bleu, la réflectance de surface pour un pixel couvert de végétation, en fonction de la longueur d'onde, en rouge la réflectance au sommet de l'atmosphère pour ce même pixel. A 1.38 µm, la vapeur d'eau absorbe totalement la lumière provenant de la surface au niveau de la mer. Autour de 0.94 µm, une autre bade d'absorption de la vapeur d'eau est visible, mais l'absorption y est moins intense

=>

Sentinel-2B dispose de deux canaux centrés sur des bandes d'absorption de la vapeur d'eau: la bande 9 (940nm) et la bande 10 (1380 nm). La bande 10 correspond à une très forte absorption, telle qu'en général, dans cette bande, les photons n'ont aucune chance d'atteindre le sol, et encore moins le satellite après réflexion sur le sol. Cette bande est utilisée pour détecter les nuages hauts.

 

Dans ce blog, nous avons jusqu'ici moins parlé de la bande d'absorption à 940 nm. Ici l'absorption des photons passant par la surface terrestre n'est pas totale, seule une forte proportion d'entre eux est absorbée, et cette proportion dépend bien sûr de la quantité de vapeur d'eau, mais aussi de l'inclinaison des directions solaires et d'observation. Cette bande nous sert à la correction atmosphérique, mais pourrait aussi, je pense, intéresser les spécialistes de l'atmosphère.

Continue reading

Machine learning for land cover map production - Follow-up on the TiSeLaC challenge

I discussed some important aspects to take into account when validating land cover maps in a previous post. In that same post I insisted on the fact that machine learning pipeline building using a blind optimisation of accuracy metrics can lead to unrealistic expectations about land cover maps produced using these approaches.

 

I cited as an example the TiSeLaC challenge, where 2 of the participating teams achieved FScores above 99%, which is an accuracy higher than the one we can expect from the reference data used for the validation.

 

I assumed that this unrealistic performances where due to over-fitting and the use of a validation set too similar to the training set. I have recently asked the challenge organisers about the procedure for splitting the reference data into train and test sets and they confirmed that the split was done at the pixel level and not at the polygon level. Therefore, nearly identical pixels coming from the same polygon could be used for training and validation.

 

Therefore, looking at the challenge results, one could expect that all the teams would have got similar high performances. Since this was not the case, I asked for references to the methods used. Two of the methods are published. I am assuming that these are the 2 winning methods.

 

One of the methods uses spatial nearest neighbour classification to decide the labels, that is, the class for a pixel is decided using the labels of the nearest pixels of the training set. Here, "nearest" means the closest in the image using an Euclidean distance on the spatial coordinates of the pixel. Indeed, the pixel coordinates where provided as a separate record, but I don't think they were intended to be used as features. And, yes, the best results are obtained if only pixel coordinates are used (no reflectances, no NDVI, nothing!). And 1 single neighbour works best than 2-NN or 10-NN.

 

This shows that indeed, neighbouring pixels were present in the training and test sets, and the fewer the information used (just the closest pixel) the better the result obtained.

 

To quickly check this, I ran a simple, out-of-the-box, Random Forest classifier using the coordinates as features and got 97.90% accuracy on the test set, while using the image features gives about 90%.

 

The second of the 2 winning methods (which is actually the first with an FScore of 99.29 while the method above obtains 99.03), uses 3 deep neural networks, 2 of which use temporal convolutions for each pixel. The third network is a multi-layer perceptron were the input features are statistics computed on all the pixels found in a spatial neighbourhood of the pixel to be classified. Different sizes of neighbourhoods between 1 and 17 are used. This is much more complex than using only the label of the closest pixel, but actually, contains the same information. Adding the information of the 2 first networks may allow to correctly classify the few pixels that the previous method got wrong. The performance difference between the 2 methods is less than 0.3%, which may probably fall within typical confidence intervals.

 

What can we learn from these results?

 

First of all, blind metric optimisation without domain knowledge can produce misleading results. Any remote sensing scientist knows that pixel coordinates only are not good predictors for producing a map. Otherwise, one could just spatially interpolate the reference data. Even when applying krigging, other variables are usually used!

 

Second, when organising this kind of contest, realistic data sets have to be used. The split between training and validation has to follow strict rules in order to avoid neighbouring pixels appearing in both data sets.

 

Third, map validation has to have a spatial component: are the shapes of the objects preserved, is there image noise in the generated map, etc. This is a tricky question which needs either to have dense reference data in some places or having specific metrics which are able to measure distortions without reference data. Obtaining dense reference data is very costly to and can even be impossible if some of the classes can't be identified by image interpretation (we are not tagging images of cats or road signs!). Developing specific metrics for spatial quality which don't need reference data is an open problem. Some solutions have been developed for the assessment of pan-sharpening algorithms, but the problem is rather different.

 

Finally, I hope that this critical analysis of the TiSeLaC contest will be useful for future contests, because I think that they may be very useful to get together the remote sensing and the machine learning communities.