Spatial Generalization

Spatial generalization asks whether a model trained in one place can make reliable predictions elsewhere, rather than exploiting similarities between nearby training and test data. Approaches include geographically separated or buffered validation, geospatial foundation-model embeddings, graph and diffusion models, and adaptation methods that use geographic diversity or physical and biological priors. The field also includes cartographic generalization, where maps are simplified or aggregated while preserving important spatial relationships and constraints. A central tension is that fine local detail can improve predictions near training data but hurt transfer across larger geographic gaps, while spatial autocorrelation can make random-split evaluations overstate reliability.