Misaligned images and masks in SSL4EO-L benchmark dataset

[adamjstewart]

Description

There is a reprojection bug in the script we used to download SSL4EO-L. The result of this is that the bbox of each file is slightly off (up to 10 px). This doesn't really matter much for the SSL version of the dataset, but the benchmark version of the dataset (NLCD, CDL) has masks that are not properly aligned with the images.

TODO:

• Fix the reprojection bug in download_ssl4eo.py
• Re-download the images in our benchmark datasets
• Re-chip the mask products in our benchmark datasets
• Upload the fixed benchmark datasets to HF
• Update downloads and checksums in our TorchGeo datasets
• Re-download/upload/update the pure-SSL versions of the dataset?
• Re-run benchmark experiments and update SSL4EO-L paper?

Thanks @Vikki23 for reporting this issue and @AABNassim @wangyi111 for investigating it!

Steps to reproduce

Look at any two pairs of images in GIS software to see the misalignment.

Version

0.6.1