Answers are grounded in this paper and cite their sources.
Geospatial foundation models (GFMs) reach competitive downstream performance on multispectral remote sensing tasks, yet whether their representations are spectrally consistent with the discriminative structure of the data and with each other remains an open question. This article presents a band-level Shapley audit on four representative GFMs (Panopticon, DOFA, SMARTIES, and CROMA) over the native 12-band Sentinel-2 configuration of EuroSAT. The audit evaluates the per-band Shapley attributions of each GFM through a Spearman alignment metric with bootstrap-propagated uncertainty, applied both against a data-derived spectral baseline and pairwise between the audited models. Four convergent observations are obtained. First, the per-class alignment of each GFM with the baseline falls into three class-level regimes (coherent, partial, and inverted). Second, all four GFMs exhibit a sign flip between industrial and residential, consistent with contrastive discrimination across spectrally adjacent classes. Third, the overall alignment ranking against the baseline does not coincide with the downstream-accuracy ranking, with the highest-accuracy GFM showing the lowest alignment. Fourth, the audited GFMs are pairwise more aligned with each other than with the baseline, pointing to a shared band-attribution structure that exceeds what the data-derived signal alone explains. Together, these observations support band-level spectral auditing as a diagnostic complementary to downstream-accuracy evaluation.
No takes yet. Share an insight, caveat, or question.
Fernández et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: