SCL cloud mask bug over water bodies

[Fahdben]

@kvantricht noticed some strange patterns on the cloud mask provided over the tile 36MUD.

Note that the cloud mask shown above is the result of the binarization of the SCL from sen2cor. When comparing with the binary mask based on the SCL available on aws (ESA processing), we can clearly see that the patterns are no longer present (image on the right).

Cloud mask using ESA SCL

Cloud mask using Sen2cor SCL

The image shown in this example is S2B_L2A_20190331T075609_N0211R035T36MUD_36MUD_MASK .

Here are some experiments (to run on 36MUD 20190331) that can help us find the origin of these patterns:

• Run the full atmospheric correction of the L1C product
• Run the cloud mask generation with sen2cor using the scl_only option Compare the results of the two runs above with the SCL available through aws (before and after binarization).

[mbattude-cs]

Results before the binarization :

• Sen2Cor SCL obtained with and without the sc_only option are identical.
• Sen2Cor SCL obtained is very different from ESA SCL. In Sen2Cor SCL we got many class 2 pixels (dark area pixels) and class 3 pixels (cloud shadows) that are not visible in ESA SCL (pixels mainly correspond to water = class 6). In particular, the dark gray elongated patches correspond to the clouds shadows.

ESA SCL

Sen2cor SCL with and without sc_only option

Difference between ESA SCL and Sen2cor SCL

• Areas above land do not seem to be impacted.

ESA SCL (entire tile)

Sen2cor SCL (entire tile)

[mbattude-cs]

@Fahdben suggested to use ESA CCI and it solved the issue. https://www.esa-landcover-cci.org/?q=node/164

SCL values

L1C RGB image

ESA SCL

Results for Sen2Cor v2.9 (DEM = srtm)

Sen2cor v2.9 SCL without CCI

Sen2cor v2.9 SCL with CCI

Results for Sen2Cor v2.10 (DEM = copdem)

Sen2cor v2.10 SCL without CCI

Sen2cor v2.10 SCL with CCI

[savmickael]

It is confirmed by this presentation which expose the differences between the local run and ESA production:

• Since Oct. 2018 they use CCI data package in ESA production.
• Between May 2019 and February 2020, the ESA production center use the sen2cor 2.8 version.

Therefore if we want mix L2A data produced by ESA and L2A produced by us with Sen2Cor 2.9 in a time series, we need to do that carefully to have homogenized time series:

• don't mix ESA L2A product generated with other version than 2.9.
• support the CCI auxiliary data in our local run in order to have the same results on the SCL than ESA production center.

[Fahdben]

• [ ] Try to build an ewoc_s2c docker image with CCI to test the final image size
• [ ] Try running sen2cor 2.9 with a subset of the ESA CCI tiff files (clipped to S2 tiles)