SHAP-Driven Machine Learning Approaches for LST Modeling Using Surface Indices in Chuadanga

The rapid urbanization and climate change has increased the surface heating of urban peri urban areas in Bangladesh. This paper analyzes the land surface temperature within Chuadanga District using Landsat 8 imageries of April 2024 and discusses the impact of vegetation, built-up surfaces and water bodies based on NDVI, NDBI and MNDWI measures. The spatial indices were as NDBI -0.26 to 0.02, NDVI 0.14 to 0.42, MNDWI -0.24 to -0.05, and LST 31.27 to 39.25 °C. There existed strong correlations among NDBI NDVI ( -0.89), NDBI NDVLST (0.68), and NDVI LST ( -0.54), which shows that urban expansion increases the surface temperature, and vegetation alleviates the temperature. To predict LST, four machine learning models were used, which are the Random Forest (RF), the Support Vector Machine (SVM), the Extreme Gradient Boosting (XGBoost), and the Artificial Neural Network (ANN). The best model in terms of r = 0.63, R ² = 0.393, RMSE 1.71 C and MAE 1.02 C and the largest AUC of 0.8115 was Support Vector Machine. Analysis using SHapley Additive exPlanations (SHAP) found NDBI to be the strongest effectively causing surface temperature, NDVI, and MNDWI came in second and third. The Index of Urban Thermal Field Variance (-0.11 to 0.49) indicated the presence of high thermal stress that was localized in the urban cores and industrial areas that revealed the role of urbanization in amplifying the Urban Heat Island effect and the buffering effect of vegetation and water bodies. The paper has shown that remote sensing, spectral indices, and machine learning can be used to offer a solid framework of LST evaluation as well as provide arguments of evidence-based urban climate adaptation, such as LST-based zoning, nature-based mitigation plans, and policy-making in rapidly developed cities.