This still needs to be fleshed out. But the crux of it is to build a habitat class classifier on top of geospatial foundation model embeddings like AlphaEarth/Tessera, that overcomes weaknesses in the existing habitat maps (Jung et al’s decision tree framework, and Lumbiarres et al’s landcover-habitat crosswalk table approach).

The goal will be to build a better global habitat map that performs better against the AOH validation scheme produced in Research Idea: Validating AOHs: Can we do better than Dahal et al?, thus improving the accuracy of LIFE and STAR.