Understanding Urban Heat Islands in Dhaka City Through Explainable GeoAI

Urban Heat Islands (UHIs) are intensifying across rapidly urbanizing megacities in the Global South, posing serious threats to public health, infrastructure, and environmental equity. Understanding the drivers of urban heat is particularly challenging due to the nonlinear and spatially heterogeneous nature of land surface temperature (LST) dynamics. This study presents the first application of a spatially explainable GeoAI framework to analyze LST variation across Dhaka, one of the most densely populated and climate-vulnerable megacities in the Global South. Leveraging multi-source geospatial and remote sensing data, a 500 m grid-based machine learning model was developed using AutoML (FLAML) with LightGBM as the selected estimator. To enhance interpretability, the study employed both SHAP and GeoShapley methods to quantify global feature importance and spatially varying effects, enabling a critical comparison of their explanatory capacities in capturing geographic heterogeneity. Results reveal that built-up intensity, bare surfaces, and vegetation structure significantly influence urban thermal patterns, with strong spatial heterogeneity in their effects. GeoShapley decomposition highlights localized cooling thresholds for green and blue infrastructure and identifies spatial clusters of intrinsic thermal drivers not explained by observed variables. This integrated approach not only improves predictive performance but also supports spatially targeted heat mitigation strategies. The proposed framework is scalable and transferable, offering a practical template for climate-resilient urban planning in other data-constrained cities.