Geospatial Representation Learning

Geospatial representation learning studies how to encode locations and geographic data so that useful spatial structure can be reused across tasks such as image geolocation, land-cover mapping, urban prediction, and mobility modeling. Common approaches include contrastive and masked self-supervision, coordinate and graph encoders, and models that align satellite imagery with maps, text, sensor data, or human activity. These representations can reduce the need for task-specific labels, but their usefulness depends on the data modality, spatial scale, and downstream task. A central challenge is reliable transfer: performance can change substantially across regions and under distribution shifts, so spatially blocked evaluation, calibration, and robustness testing matter alongside accuracy.