Earth Embeddings: Learning Mental Maps in Neural Nets

What makes AI models robust when reasoning about geography, climate, or human activity? While explicit geospatial data—like temperature, precipitation, or land cover—can inform models, it’s often incomplete, sparse, or too coarse to generalize well. Instead, we explore how to create implicit, robust environmental embeddings that capture the complex and diverse character of places directly from imagery. This talk introduces a novel approach to learning these embeddings by training AI to play Satellite GeoGuessr—contrastively learning what makes each location unique through satellite images. We present two complementary methods: First, we introduce location encoders—neural networks that learn to map coordinates to rich geospatial representations, storing knowledge in their weights. Second, we describe SatCLIP Earth Embeddings, which trains image and location encoders jointly to produce embeddings that reflect both the visual and contextual identity of a place.