What or Where? Class–Site Geometry and Geographic Transfer in Geospatial Embeddings
Abstract
Geospatial foundation models are commonly compared by downstream accuracy, but similar accuracy does not imply similar embedding geometry. We study 387 hand-labelled forest and water points across 16 geographic sites using TESSERA v1.0, v1.1, v2-beta, and AlphaEarth. We introduce a pairwise analysis to determine whether embedding distances are organized more strongly by semantic class or geographic site. Although all four embeddings achieve perfect leave-one-site-out classification and same-class nearest-neighbour retrieval, their geometries differ substantially. TESSERA v2-beta is strongly organized by semantic class (P = 0.851, 98.75% CI [0.681, 0.982]), whereas AlphaEarth lies near the pooled boundary (P = 0.521) and exhibits strong class-specific asymmetry: water is class-organized (P = 0.947), while forest is site-organized (P = 0.263). Site-sensitive geometry does not necessarily imply poorer geographic transfer, as farther sites are not consistently harder to predict. Overall, the tested models show strong geographic coordinate dependence in their embedding spaces, while semantically separable classes remain separable across large geographic distances.