Harvesting AlphaEarth: Benchmarking the Geospatial Foundation Model for Agricultural Downstream Tasks
4 views
Skip to first unread message
Aaditeshwar Seth
Jan 23, 2026, 9:29:06 PMJan 23
to learning, ac...@googlegroups.com, CoRE stack: NRM, core-st...@googlegroups.com
This is a nice paper also on how to set up evaluation tasks to test for spatio-temporal robustness and transferability across different eco-regions.
We observed mixed patterns in the tillage mapping and cover crop mapping tasks... One possible reason is that agricultural practices are less influenced by differences among ecoregions than crop yields... AEF-based models trained in one region learn region-specific association between the predictors and the yield data, which can hinder transfer to the other region... During model training, AEF embeddings were decoded to not only reconstruct satellite images but also DEM and gravity fields, which are region-specific and relatively static, and may limit the embeddings to capturing localized information. Another possible reason is that geolocated articles from Wikipedia were used to provide text-based information as inputs to a text encoder, which were updated via contrastive learning with vision-based models. Such text-based information might be region-specific, resulting in distinct embeddings across space, particularly when comparing regions across countries or continents.
Co-founder, Gram Vaani; Co-founder, CoRE Stack
Reply all
Reply to author
Forward
0 new messages