Dependence of Geospatial Foundation Model Representations on Survey Region and Sensor: Analysis Code, Scale-Correction Table and Embeddings

Analysis code, scale-correction table, embeddings and figures supporting the study "Dependence of geospatial foundation model representations on survey region and sensor: a measurement protocol with zero calibration". This version supersedes versions 1-6, which supported an earlier study with uncorrected reflectance scales and a degenerate distance measure; see CORRECTION.md. The frozen encoder of Prithvi-EO-2.0-300M is applied to Sentinel-2 L2A tiles of the Aral region and of an arid control region in northern Turkmenistan, drawn from the Major TOM Core-S2L2A corpus. Shift is quantified in representation space by three measures, each read against its own calibrated zero: squared maximum mean discrepancy with a permutation null and subsampled confidence interval; a linear domain classifier; and a Mahalanobis distance to the control computed after reduction to principal components, the number of which is selected by the zero calibration itself rather than by a rule. Six controls separate the measured effect from properties of the measuring procedure: zero calibration on two halves of one domain, resolution control, time control, time control at equalised date gap, stratification by brightness, and an artifact check. A second axis is measured on 1153 pairs of near-simultaneous Sentinel-2 and Landsat acquisitions of the same location. The per-tile shift is related to surface salinity, brightness and vegetation. A scale discrepancy in the source corpus is identified and corrected: since processing baseline 04.00, Sentinel-2 L2A products store reflectance with an offset, and the corpus mixes both epochs. The processing epoch of each tile is determined by three independent routes with 97.7 % coverage and a measured accuracy of 0.42 %. Every published number is produced by a named script from the files deposited here, and verify_tables.py prints the whole of tables 2 to 4 in a single run. This answers correction 12: in earlier versions the values of one table came from separate runs at different settings, each correct alone and inconsistent together, with nothing in the archive to reveal it. Contents:- indices.py: the single definition of every spectral feature, imported by every other script, so that one name cannot denote two quantities- offset_fix.py, offset_table.csv: reflectance-scale correction, single point of application imported by every computation script- build_offset_table.py, propagate_within_product.py, band_bias.py, download_landsat_pairs.py: construction and validation of the correction table- mahalanobis_pca.py: the per-tile measure after dimensionality reduction, its zero calibration, and the sweep across numbers of components- verify_tables.py: every published value of tables 2 to 4, in one command- verify_classifier.py: the domain classifier and its own zero calibration- mmd_null.py: the maximum mean discrepancy, its permutation null saved in full, and the confidence interval by subsampling without replacement- compute_indices.py, partial_correlations.py, analyze_corrected.py: surface indices and the association analysis- cross_sensor.py: the sensor axis- figures_article.py, make_figs.py: figure generation- fixed_run_corrected.npz, fixed_run_corrected_tiles.csv: embeddings and measures- tile_indices_corrected.csv: per-tile surface indices- maha_k8.csv, maha_k16.csv, k_sweep.csv: the reduced measure per tile, and the measurement and its zero against the number of components- mmd_null.csv, mmd_result.json: all 1000 permutation values, and the result- cross_sensor_results.json, band_bias_landsat_pairs.csv: sensor-axis outputs- baseline_status.csv: processing epoch and product identifier per tile- fig1_calibration.png, fig2_distributions.png, fig3_mmd_null.png- rerun_fixed.py, mmd_local.py, mmd_geo.py: the scripts of versions 2 to 6, retained for provenance only; they reproduce superseded numbers- README.md, CORRECTION.md, CORRECTION_previous.md The regional imagery corpus is available separately at doi:10.5281/zenodo.21663900.