bigearthnet compute s1 mean + std

[tchken]

for evaluation for bigearthnet -- compute the bands mean and std for s1 data

[tchken]

1. current results using compute-dataset-pixel-mean-std.py output
   tensor([1.0990e-05, 1.0990e-05, 1.0990e-05]) --- mean
   tensor([0.0005, 0.0005, 0.0005]) --- std

2. I calc directly in order to verify from file, upload as well to the branch, though it resulted the same mean and std with the channels as (1)

3. found out the issue may come from dataset.ImageFolder() that output 3 channels. As such, probably need to load custom as sen12ms . looking into

[tchken]

I've custom loaded pytorch the dataset for bigearthnet s1 in order to calc the mean and std s1 -- I have the updated results directly compute each of the bands for S1 on the separate load. Given the quick look at the raw data pixel levels, I think this should be correct info for S1 from the bigearthnet.

S1 bands: _VH (mean: 0.2542, std: 0.1720) _VV (mean: 0.4994, std: 0.1902)
will double check and reorganize the code to upload

wrong pixel values as the tensor transformation takes to divide 25 and clip with sen12ms, sort of messing up -- please see below updates

[tchken]

i forgot to transform to tensor and sort of mess up pixel values the computation, I've recalc and the S-1 bands mean and std for bigearthnet should be as following
band_names_s1 = ['VV', 'VH']
_VV (mean:-12.514466285705566, std:4.754617214202881)
_VH (mean:-18.64471435546875, std:4.298763751983643)

script update here at branch.
working also on the s2 verify

[tchken]

###Update

I've re-checked and remove the pixel value clip (by sen12ms team). Re-calced with the "raw" pixel values from the BigEarthNet. In addition, include S-2 bands. Also sent a pull request from branch for code upload.

S-1 (4dp)

band:VV` -- mean:-12.6200, std:5.1159
band:VH -- mean:-19.2904, std:5.4644

S-2 (4dp)

band:B01 -- mean:711.44852, std:1564.5292
band:B02 -- mean:783.1296, std:1562.5066
band:B03 -- mean:923.6776, std:1436.2452
band:B04 -- mean:912.0620, std:1527.7863
band:B05 -- mean:1263.0253, std:1531.7996
band:B06 -- mean:2044.0737, std:1583.0283
band:B07 -- mean:2292.6069, std:1630.4586
band:B08 -- mean:2445.4717, std:1732.2150
band:B09 -- mean:2445.7380, std:1587.6179
band:B10 -- mean:1492.5336, std:1080.1777
band:B11 -- mean:964.3022, std:807.7414
band:B12 -- mean:2460.2112, std:1644.2445

S-1 (full dp)

band:VV -- mean:-12.620004653930664, std:5.115894794464111
band:VH -- mean:-19.290447235107422, std:5.464434623718262

S-2 (full dp)

band:B01 -- mean:711.4485473632812, std:1564.5291748046875
bandB02 -- mean:783.1295776367188, std:1562.506591796875
bandB03 -- mean:923.6775512695312, std:1436.2452392578125
bandB04 -- mean:912.06201171875, std:1527.7862548828125
bandB05 -- mean:1263.025390625, std:1531.799560546875
bandB06 -- mean:2044.07373046875, std:1583.0283203125
bandB07 -- mean:2292.60693359375, std:1630.4586181640625
bandB08 -- mean:2445.4716796875, std:1732.2149658203125
bandB09 -- mean:2445.738037109375, std:1587.617919921875
bandB10 -- mean:1492.5335693359375, std:1080.177734375
bandB11 -- mean:964.30224609375, std:807.7413940429688
bandB12 -- mean:2460.211181640625, std:1644.2445068359375

[suryagutta]

While searching on a different topic, found S1 mean and std in tensor flow model module of bigearthnet code. Those values are almost similar to the one Ken has calculated. Added these values to the code and checked in as they might help performance. Link: https://git.tu-berlin.de/rsim/bigearthnet-models-tf/-/blob/master/BigEarthNet.py