What's in an Earth Embedding? Probing Geospatial Foundation Models Beyond Benchmarks

Geospatial foundation models (GFMs) have rapidly expanded Earth observation modeling, but selecting among them remains difficult because published benchmark evidence is fragmented across tasks, papers, and evaluation protocols. We ask whether information recoverable from frozen GFM representations can provide a complementary signal of downstream performance. We extract representations from 19 released GFMs and probe their spectral, spatial, temporal, geolocation, and elevation content by measuring how well each can be reconstructed or recovered under a common evaluation framework. We then compare probe performance with literature-derived model strength estimated using a Plackett--Luce model over more than 2,000 published results spanning 189 benchmarks. Reconstruction performance is consistently associated with published strength, reaching $r=0.61$ for single-acquisition probes, while geolocation shows essentially no relationship. Annual time-series probes exhibit even stronger correlations, although across fewer models. This provides preliminary evidence that lightweight probing of frozen representations can provide a signal of benchmark-based model ranking while also revealing what information different GFMs encode.