Change BigEarthNet class mapping order to match original code

[lucastao]

It seems like the mapping from 43 to 19 classes in the BigEarthNet dataset code is incorrect.

Example
The below image corresponds to the RGB visualization of S2B_MSIL2A_20180502T093039_42_25. The provided textual labels are 'Broad-leaved forest', 'Natural grassland', and 'Transitional woodland/shrub'. Per the mapping provided here , this corresponds to indices 5, 24, and 39. After converting these from 43 class labels to 19 class labels using the label converter, we end up with 5 being discarded, 24 (Natural grassland) mapping to 10 (Mixed forest) and 39 (Transitional woodland/shrub) mapping to 17 (Inland waters). All three of these are incorrect. 5 (Broad-leaved forest) should be mapped to 8 (Broad-leaved forest), 24 (Natural grassland) should actually map to 11 (Natural grassland and sparsely vegetated areas) and 39 (Transitional woodland/shrub) should map to 13 (Transitional woodland, shrub).

Root Cause and Fix
The root cause of this issue appears to be an alphabetically sorted list here: https://github.com/microsoft/torchgeo/blob/main/torchgeo/datasets/bigearthnet.py#L145-L190 when it should instead match the order provided by the BigEarthNet code: https://git.tu-berlin.de/rsim/BigEarthNet-S2_19-classes_models/-/blob/master/label_indices.json#L2-46.

1. Look at the BigEarthNet code for 43 to 19 class mappings provided at https://bigearth.net/. Specifically, this file: https://git.tu-berlin.de/rsim/BigEarthNet-S2_19-classes_models/-/blob/master/label_indices.json#L2-46. Here, the mapping are not sorted in alphabetical order.

2. Look at the BigEarthNet code for 43 to 19 class mappings in torchgeo: https://github.com/microsoft/torchgeo/blob/main/torchgeo/datasets/bigearthnet.py#L145-L190. These are sorted in alphabetical order.

3. However, the mappings from 43 to 19 classes remain index based and are the same as the indexes used in: https://github.com/microsoft/torchgeo/blob/main/torchgeo/datasets/bigearthnet.py#L193-L226 and https://git.tu-berlin.de/rsim/BigEarthNet-S2_19-classes_models/-/blob/master/label_indices.json#L47-67. To fix this, either the ordering of the 43 classes or the indexing dictionary must be changed to fix this mismatch.

This commit adjusts the order of the 43 classes to match the original. After testing, I find that the above example now has the correct mapping:

Textual label: ['Broad-leaved forest', 'Natural grassland', 'Transitional woodland/shrub']
43 class: [22, 25, 28]
19 class: [8, 11, 13]

[lucastao]

@microsoft-github-policy-service agree

[adamjstewart]

Good catch, thanks for fixing this!

[calebrob6]

Thanks @lucastao!