TTE geolocation encoder learns where to focus, beats SOTA benchmarks

Existing geolocation encoders project coordinates onto fixed bases (e.g., spherical harmonics), uniformly allocating representational capacity across the entire globe—wasting resources on empty ocean while under-serving complex urban areas. To fix this, researchers from multiple institutions introduce Tessellating the Earth (TTE). TTE uses learnable Spherical Voronoi partitions: each Voronoi site carries its own embedding and migrates during training toward discriminative regions. This fully differentiable, end-to-end approach concentrates capacity exactly where it's most valuable.

Beyond local spatial structure, TTE introduces global semantic tokens—a shared vocabulary of learnable concept tokens that distill semantic knowledge from satellite imagery. These tokens allow geographically distant sites that share similar environments (e.g., two different deserts) to reference the same learned semantics, bridging local detail with global understanding. On a suite of geospatial classification and regression tasks, TTE achieves state-of-the-art results. It performs especially well as a geographic prior for fine-grained species classification on iNaturalist-2018. The paper is accepted at ECCV 2026, and the code and weights are publicly available.