Spatially Validated Explainable GeoAI for Urban Heat Attribution in Benghazi, Libya Quantifying Validation Leakage, Built-Up Effects, and Coastal Moderation

Abstract

Urban-heat studies in data-scarce cities frequently use random train-test splits that can overstate machine-learning performance because neighboring cells share spatial structure. We develop a reproducible, explainable GeoAI framework for Benghazi, Libya, using Landsat Collection 2 Level-2 land surface temperature with NDVI, NDBI, MNDWI, local spectral heterogeneity, and distance from the Mediterranean coast. The primary 240 m analysis contains 12,381 valid non-water cells and is repeated at 150 and 510 m. Ridge, Random Forest, and histogram gradient boosting are evaluated under both random five-fold validation and geographically blocked five-fold validation. At 240 m, the best R² falls from about 0.958 under random validation to 0.829 under spatial validation, while RMSE increases from about 1.56 °C to 3.14 °C. The same optimism gap persists at all three scales. SHAP attribution identifies NDBI as the strongest model driver, followed by MNDWI and spatial-context variables; high-NDBI cells are 5.70 °C warmer in median LST than low-NDBI cells. The results show that spatial validation is a substantive requirement for credible urban-heat inference and that Benghazi’s thermal pattern reflects coupled built-up, moisture, and coastal controls rather than a single greenness indicator.