Fusion of AlphaEarth Embeddings and Sentinel-1 Time-Series for Conflict-Related Urban Damage Mapping

Abstract. Conflict-related urban damage mapping requires methods that remain reliable when field access is limited, cloud cover is frequent, and urban conditions vary strongly. Sentinel-1 SAR time series provide systematic all-weather coverage, but SAR-only indices are still sensitive to speckle, acquisition geometry, and heterogeneous urban background. AlphaEarth Foundations embeddings offer a complementary feature space that may stabilize change detection by encoding broader semantic context. This study evaluates scalar fusion of AlphaEarth embedding change and Sentinel-1 change indices for building-level conflict damage mapping.

We first compare multiple AlphaEarth-based change metrics and retain Earth Mover's Distance (EMD) as the most suitable embedding-space indicator. We then combine EMD with Sentinel-1 indices through simple scalar fusion rules and evaluate the results on 75,825 reference building footprints from seven Ukrainian cities with damage rates ranging from 0.2% to 33.7%. The strongest transferable method, EMD+PWTT_w40, achieves mean building-level AUC 0.676, outperforming both PWTT alone (0.657) and EMD alone (0.593). A lightweight learned fusion gives only a small within-city gain (mean AUC 0.686) and degrades under leave-one-city-out transfer (0.614). Sensitivity analysis further shows that the transferred fusion weight remains stable, with w_EMD = 0.4 staying within 0.0028 AUC of the city-specific optimum on average.

These results show that AlphaEarth and SAR capture complementary aspects of urban damage and that a simple scalar fusion rule can combine them effectively without sacrificing interpretability or transferability.