Bug in code used to produce s2cloudless results

[DomagojKorais]

There is a bug in the code used to produce s2cloudless results. The problem is that the order of the bands given as input to the model is not defined.

The bug originates here. The os documentation states that the output of listdir is "in arbitrary order". Using bands in the wrong order leads to poor model performance.

I did a test on the left lower corner of tile T21LXH, image from March 28, 2017 and obtained the following results using the same parameters that you have published:	correct bands order	published cloud mask (from Figure 4.c)	reference area (from Figure 4.a)

• "correct bands order" shows the cloud mask, computed using BANDS properly ordered, over imposed to the area of interest (the lower left corner of the full tile).
• "published cloud mask" is a crop of the left lower corner of the published mask in the paper, Figure 4.c
• "reference area" is a crop of the left lower corner of the reference image published in the paper, figure Figure 4.a

As you can see the results are totally different, so published results should be corrected once that the above issue is resolved. Additionally default threshold value is 0.4, and not 0.7 as stated in the paper:

[albhasan]

Dear @DomagojKorais,

thanks for helping us improving our code. We are reviewing this issue and as soon as we have a result, we will inform you.

Best regards,

[DomagojKorais]

Dear @albhasan, is there any news regarding the results obtained using the fixed code? I am really curious to see the difference. Thanks.

[albhasan]

Good morning,

yes, there is. We recomputed the s2cloudless mask and we are preparing a manuscript updating the results. The preliminary results are displayed in the table below.

Bests,

[DomagojKorais]

Nice to see the improvements, thanks!