Wrong band order in EuroSAT dataset

[lhackel-tub]

Description

Hello,

During the loading of the EuroSAT dataset, the band order with [..., B08, B8A, B09, ...] as described here is assumed. However, in the original EuroSAT repository, this issue was raised regarding the band order and accepted by the original authors, which describes the Bandorder as [..., B08, B09, ..., B12, B8A]. I was therefore wondering if this might be a mistake in torchgeo that assumes a different band order or if there is any additional information available that shows that the original repository's assumption is incorrect.

Thanks!

Steps to reproduce

No code needed

Version

0.6.1

[adamjstewart]

Thanks for pointing this out! I ended up taking a deep dive to try to find the answer.

Things that suggest that B8A is in the 9th spot:

• The EuroSAT paper lists the bands in the same order we use

Things that suggest that B8A is in the 13th spot:

• The issue you linked to (although it wasn't answered by the EuroSAT developers, it was closed without disagreement by a EuroSAT developer)
• When I open a file in QGIS and compare the individual bands, the 9th band seems like it has 60m resolution, which is true of B9, not B8A
• The range of values for the 10th band is completely different for all other bands, which is true for B10 (cirrus), not B9 (water vapor)

I think you're probably right. Want to open a PR to fix this? There are at least a few places that need to be fixed:

• torchgeo/datasets/eurosat.py: simply change the order of all_bands
• torchgeo/datamodules/eurosat.py: the normalization values are in the correct order, but the band name keys are not
• docs/tutorials/transforms.ipynb: simply change the order of bands

There's a good chance this inconsistency exists in other datasets as well. If someone has time, it would be interesting to check all Sentinel-2 datasets.

[adamjstewart]

@lhackel-tub do you want to submit a PR to fix this or would you like one of us to fix this?

[lhackel-tub]

@adamjstewart I created a pull request to fix the above mentioned issues. I did not look at the other Sentinel-2 datasets.