Measuring the Domain Shift of a Global Geospatial Foundation Model Over the Desiccated Aral Basin

Analysis code, embeddings, and figures supporting the study "Measuring the Domain Shift of a Global Geospatial Foundation Model Over the Desiccated Aral Basin: A Label-Free, Multi-Metric Assessment". The frozen encoder of Prithvi-EO-2.0 (300M parameters) is applied to 1628 Sentinel-2 tiles of the Aral basin (Karakalpakstan) and 400 tiles of an arid control region. Domain shift is quantified in representation space by three measures: maximum mean discrepancy with permutation and bootstrap testing, a linear domain classifier, and a Mahalanobis distance to the control. The per-tile shift is further related to the surface salinity signature. Contents:- mmd_local.py, mmd_geo.py: embedding extraction (with per-tile geolocation)- analyze_extended.py, analyze_salinity_link.py: geographic and salinity analysis- figures_mmd.py, figures_geo.py: figure generation- mmd_embeddings.npz, mmd_geo.npz: extracted embeddings- figures: mmd_figure_full.png, fig_map_shift.png, fig_shift_salinity.png The regional imagery corpus is available separately at doi:10.5281/zenodo.21663900.