Mixture of Earths: Leveraging the Complementary Strengths of Geospatial Foundation Models

Geospatial Foundation Models (GFMs) are designed to address a wide array of critical Earth observation applications, from flood mapping to deforestation tracking; yet no single GFM performs uniformly well across these tasks, with different GFMs excelling at different domains. Rather than building yet another GFM, we propose Mixture of Earths (MoEarths), a framework that identifies and dynamically fuses complementary cross-layer features from any set of frozen GFMs. We examine two feature fusion strategies built on top of i) an automatic top-$k$ router and ii) an interpretable, similarity-based selection algorithm and assess them across five highly diverse Earth Observation tasks. We find that the automatic router focuses on the single strongest model, stacking near-duplicate layers, whereas similarity-based selection retains genuinely different layers and matches or beats both router-fusion and the individual models on all five tasks. This suggests that beneficial fusion stems from combining representations that are inherently complementary. Our results motivate a shift from building monolithic GFMs towards the principled combination of those that already exist, and offer a low-cost path for incorporating new GFMs as they emerge.