Change finder

[favyen2]

Draft PR -- various ways to try and find changes in the world, building on the OPTIMUS work.

There are a lot of things here but currently the most relevant is summarized in README_landcover.md.

• Train land cover prediction model on WorldCover labels (data/land_cover_change/worldcover_change/config.yaml)
• Create rslearn dataset with many locations (some random over whole world, some random near urban areas), ten years of Sentinel-2 images (six monthly mosaics per year).
• Apply land cover model on each location / year.
• For each location, find change by looking for pixels / years where the model is confident in the three preceding years that the land cover category is X, but confident in the next three years that the category is Y. This gives us change polygons. We also get no-change polygons by finding pixels where the model is confident across all seven years that the category is X.
• Then do some annotation of these to prune out the incorrect ones, as well as assign more precise timestamps to when the change started / ended. This is done in the land_cover_change_viewer web app.
• Finally land_cover_time_series_change_model has code to take those annotations and convert it to another rslearn dataset that has twenty quarterly images (spanning five years). The model will input twelve quarterly images (three years) at a time and predict whether the pixel has change, along with the source/destination land cover category.