3rd TAGRA

Land surface temperature (LST) is a key indicator of urban thermal environments. However, studies of its drivers often rely on linear models or black-box machine learning, and rarely address nonlinearity, interpretability, and spatial autocorrelation together. This study proposes to analyze the factors controlling LST in Ho Chi Minh City using Landsat 8/9 imagery, combined with spectral indices (NDVI, NDBI, MNDWI, albedo), topography, and urban characteristics (built-up fraction, road density, population density, distance to water). The approach has two complementary branches: (1) gradient-boosted trees (XGBoost) with SHapley Additive exPlanations (SHAP) to quantify variable importance, nonlinear responses, and effect thresholds; and (2) a generalized additive model (GAM) with a spatial smooth term to account for residual spatial structure and to estimate spatially varying effects. Models will be evaluated using spatial block cross-validation to avoid optimistic bias from spatial autocorrelation, and residuals will be checked with Moran’s I. The study will compare SHAP-based and GAM-based interpretations to assess their consistency, and will derive thresholds for vegetation cover, proximity to water, and built-up density to support urban heat mitigation planning.