No One Knows the State of the Art in Geospatial Foundation Models

A review of 152 papers found major inconsistencies in how geospatial foundation models are trained, evaluated, and reported.

Because of these gaps, the field cannot yet reliably rank models or compare results across studies.

The authors propose six community standards, including shared benchmarks, clearer reporting, and more consistent release of model weights.

The goal is to make comparisons more reproducible and useful for high-stakes uses like Earth observation, disaster response, and food-security monitoring.