Poor masking on very cloudy imagery

[kegro]

We have been running Fmask in northern europe over the past few years and have noticed that if often fails on very cloudy imagery. It mostly happens in the Autumn and Winter, but occasionally in Summer. The latest example here, but there are many:

S2B_MSIL1C_20210918T101639_N0301_R065_T33UWB_20210918T122319.SAFE

Its a fully clouded scene:

When Fmask 4.3 is run we get this result (black has value == 0 and should be clear land pixels)

When the Sentinel-2 image is masked using Fmask we get this result, with lots of cloud not masked.

[qsly09]

Hi kegro,

I downloaded this image with L1C TOA product from scihub.copernicus.eu, and the mask generated by Fmask version 4.3 seems to be good (see below image, red color indicates the cloud)

I am not sure what happened on your side. If you still have the problem, please feel free to provide the details of how you processed the data, and then we can help to figure the issue out. Thanks

Best,

Shi

[kegro]

Thanks for checking this out. A bit weird you could not reproduce it. We will do some tests on Fmask 4.4 and see how it goes.

[kegro]

@qsly09 we tested Fmask 4.4 on the same image as above on the Linux installation (the installation we routinely use) and the Windows UI installation and got much different results. This test explains the differences we also see above.

[kegro]

@qsly09 could you test the same scence as above on your Linux installation to see if you can reproduce the faulty mask we get?

[qsly09]

@kegro I updated the Fmask tool yesterday, and I think the Linux installation sometimes may not locate the global auxiliary dataset, which may be the reason. In the updated Fmask 4.4 now, you will get the message regarding the auxiliary dataset, and if you find it still fails to locate the dataset, you can set up the path of the global auxiliary dataset by using the command <Fmask_4_4 "path of the auxiliary data">. Could you please try it again using the Linux standalone? Thank you very much.