Geospatial urban heat mapping with interpretable machine learning and deep learning: a case study in Hue City, Vietnam

Land Surface Temperature (LST) is considered a critical variable for assessing heat stress in urban environments. Understanding LST and its spatial variation is essential to comprehending the interactions between human activity and urban areas. This study investigates the impact of geospatially derived factors-namely built-up density, Normalized Difference Built-up Index (NDBI), road density, Normalized Difference Vegetation Index (NDVI), Bare Soil Index, distance to water bodies, elevation, slope, and aspect- on LST in Hue City, Vietnam, a region with limited prior documentation on this subject. Landsat 8 imagery data, collected in early 2024 during an exceptional heatwave, is utilized for this purpose. Advanced machine learning techniques, including deep neural networks, random forests, and XGBoost, are employed to model the relationship between LST and these explanatory variables. To deepen the understanding of the factors contributing to LST, the study uses the state-of-the-art Shapley Additive Explanations (SHAP) method. Experimental results show that the machine learning approach can accurately estimate the spatial variation of LST. The coefficient of determinations (R2) achieved by deep neural networks, random forests, and XGBoost are 0.83, 0.83, and 0.85, respectively. Sensitivity analysis based on SHAP reveals that built-up density, road density, and the Bare Soil Index are the most crucial variables that positively affect the LST. The factors of distance to water and slope negatively influence the LST. The established data-driven approach, coupled with SHAP, provides a valuable tool for understanding the spatial distribution of LST as well as mapping hot spots that experienced the highest level of urban heat stress. This tool also supports the analysis of mitigation measures for regulating temperature and reducing the impacts of the urban heat island effect.