Mapping urban thermal vulnerability across urban Heat Island spatial patterns in Tehran: A machine learning-based comparative analysis

In recent decades, intensifying urban heat has emerged as a major challenge for Tehran, yet neighborhood-scale assessment of thermal vulnerability and the relative contributions of key urban components remain limited. This study develops an integrated machine learning (ML) framework to assess extreme daytime heat vulnerability across 367 neighborhoods in Tehran. A total of 25 spatial and environmental variables were derived from multi-source datasets and grouped into five components. Random Forest (RF), Generalized Linear Model (GLM), and Logistic Regression (LR) models were used to classify neighborhood vulnerability and evaluate predictor contributions. SHAP and standardized regression coefficients (SRCs) were employed to provide complementary interpretation at the variable and component levels, while SHAP dependence and interaction analyses were used to examine nonlinear and context-dependent contribution patterns. Results indicated that the RF model achieved the highest predictive performance (AUC = 0.87), with 223 neighborhoods (60.76%) classified as highly or very highly vulnerable and 87 (23.71%) as low or very low vulnerability. NDVI was the most influential individual predictor in the RF model, with SHAP values ranging from approximately −0.15 to +0.05 and increasingly negative contributions at higher NDVI levels. At the component level, urban morphology showed the strongest overall contribution based on SRCs. SHAP interaction analysis further indicated that the contribution of NDVI varied across urban-form conditions, particularly SVF, BD, and BH. These findings highlight the need for context-sensitive eco-morphological planning that integrates vegetation with urban form and spatial configuration. The proposed framework provides an interpretable, spatially validated approach for neighborhood-scale climate adaptation in hot and dry cities.