Foundation-Model Embeddings for Land-Cover Mapping and Annual Change Detection in a Hyper-Arid Region: A Case Study of Saudi Arabia (2017–2024)

Spectral remote sensing struggles in hyper-arid environments because desert substrates share overlapping optical signatures. We evaluate AlphaEarth foundation-model embeddings—64-dimensional annual representations fused from Sentinel-1, Sentinel-2, Landsat, and LiDAR—for land-cover classification and change detection across Saudi Arabia (2017–2024). From an equal-allocation draw of ESA WorldCover 2021 strata, 25,241 labelled samples entered cross-validation; seven classes reached the requested 3000 and the two rarest returned their full national extent at the sampling scale. A Random Forest reproduces the WorldCover labels at a spatially blocked cross-validated overall accuracy of 0.815 ± 0.007, averaged over twenty independent assignments of the spatial blocks to folds, compared with 0.859 ± 0.005 under standard random cross-validation; this figure measures agreement with WorldCover rather than accuracy against independent ground truth, and is not directly comparable with WorldCover’s own globally validated accuracy; the 4.4 percentage-point gap quantifies spatial leakage and is reported transparently. Design weighting following Olofsson et al. corrects the distortion introduced by equal per-class allocation for 2021; because the reference labels are not independent of the training labels, the resulting fractions are reported as model-predicted national composition rather than accuracy-adjusted area estimates. UMAP visualisation of the embedding manifold reveals five sub-types within the single WorldCover bare/sparse vegetation class, consistent with geomorphologically distinct desert substrates. An indicative cross-feature benchmark produced an OA 14.1 percentage points higher for the foundation-model representation than for the strongest Sentinel-2 baseline under the same fold partition. Because the conditions were evaluated on non-identical samples, this difference cannot be attributed solely to feature representation. An independent probability sample of 374 points, interpreted on very-high-resolution imagery in two rounds by two analysts and reconciled to 97.2% agreement, gives a design-weighted overall accuracy of 0.868 for the 2021 map. The same interpretation confirms only 49.5% of the WorldCover class assignments at those points, with tree cover, grassland and herbaceous wetland largely reassigned to cropland, bare/sparse vegetation and shrubland; WorldCover and the Random Forest map agree with the independent reference at the same design-weighted rate of 0.868. The classifier therefore reproduces its label source closely, including where that source departs from independent interpretation, which shows directly that agreement with WorldCover and accuracy against land cover are distinct quantities in this landscape. Together, these results establish a methodologically transparent workflow for foundation model-based land-cover monitoring in data-scarce arid environments, with independent validation limited to the 2021 map and no independent annual reference data available across the complete 2017–2024 period. A separate 49-point cropland two-date sub-test found interpreted field-state change in nine of 25 flagged points (36%) versus two of 24 unflagged points (8%; Fisher’s exact p = 0.037); within this class the screening flag had precision = 0.360, recall = 0.818 and F1 = 0.500, but these class-specific metrics do not constitute national validation of the change layer.