An interpretable machine learning framework for mapping hotspots and identifying their driving factors in urban...

Mapping and identifying urban hotspots are essential for effective heatwave mitigation. Moreover, understanding the critical factors driving hotspot formation is crucial for developing targeted heat stress mitigation strategies. To address this need, a two-step approach is adopted: first, Extreme Gradient Boosting (XGBoost) machine is used to develop a classification model that accurately identifies and delineates hotspots. Second, leveraging the established model, Shapley Additive exPlanations (SHAP) is utilized to investigate how various physical indices influence hotspot formation. The urban center of Da Nang, in the central region of Vietnam, experienced severe heatwaves in the early periods of 2024 and was selected as the study area. Land surface temperature (LST) data was retrieved from Landsat 8 during the intense heatwave period, spanning from March 1, 2024, to May 30, 2024. Experimental results show that XGBoost is a highly capable model that achieves outstanding performance with a classification accuracy rate of 0.925, a precision of 0.926, a recall of 0.923, a F1 score of 0.925, and a Cohen's Kappa coefficient of 0.85. The research findings also reveal that the Normalized Difference Built-up Index, built-up density, green space density, elevation, and distance to water bodies are the five most influential factors in determining urban heat hotspots. Additionally, the Bare Soil Index, distance to coastlines, Normalized Difference Vegetation Index, and distance to roads moderately affect the XGBoost's predictions. The developed machine learning model, combined with SHAP, offers a robust tool for understanding the spatial variation of hotspots. This tool is crucial for urban planners and policymakers in establishing mitigation strategies to reduce the adverse impacts of the urban heat island effect.