What on Earth is AlphaEarth? Hierarchical structure and functional interpretability for global land cover

A team of researchers has published a paper providing a major interpretability breakthrough for Google's AlphaEarth Foundations (GAEF), a powerful but opaque geospatial foundation model. By reverse-engineering how the model's 64-dimensional embeddings represent global land cover, they discovered the space is not random but functionally organized. Dimensions act along a spectrum from 'specialists' for specific land cover classes (e.g., forest, urban) to 'generalists' capturing shared characteristics or broad environmental gradients.

The most striking finding is the model's extreme redundancy. The team's framework, which combines large-scale experimentation with feature importance analysis and progressive ablation, showed that accurate land cover classification can be maintained using a tiny fraction of the available dimensions. For many classes, 98% of the baseline predictive performance was achievable with just 2 to 12 dimensions. This demonstrates that much of the embedding space is not critical for core tasks.

This work moves GAEF from a 'black box' to a more interpretable tool for scientists. The hierarchical functional map provides practical guidance for dimension selection, allowing users to prune unnecessary computational overhead. For operational tasks like real-time deforestation tracking or climate impact assessment, this could enable significant cost reductions and faster processing without sacrificing accuracy, making high-level geospatial AI more accessible and efficient.