Machine Learning Approaches for Geospatial Modeling of Urban Land Surface Temperature: Assessing Geographical Compactness, Interpretability, and Causal Inference

Abstract

This study presents a data-driven framework for modeling urban heat in a highland region of Quang Ngai Province, Vietnam—an area with limited prior research on heat stress. Using advanced machine learning methods, including Category Boosting (CatBoost) and deep convolutional neural network (CNN), the spatial distribution of urban land surface temperature (LST) is predicted based on topographical, land use/land cover, urban morphological, proximity, and compactness features. Our findings show that incorporating urban compactness metrics significantly enhances prediction accuracy, with CatBoost explaining 89% of LST variance. Based on Shapley Additive Explanations, built-up density, bare land density, distance to river, green space density, and built-up cluster compactness are identified as the most influential factors. Machine learning-based causal analysis further clarifies the direct effects of key urban features on LST. The proposed framework helps reveal distinct characteristics of the study area with respect to urban heat properties. The research findings can support sustainable urban planning and heat stress alleviation in the study area.