Impact of High-Density Urban Environment on Residents’ Heat Risk: Evidence from Shenzhen, China

This study explores the nonlinear effects of urban environmental factors on heat risk in Shenzhen, a typical high-density megacity. We established a multidimensional indicator system covering the built environment (building density, road density), ecological conditions (Enhanced Vegetation Index, solar radiation), and green space morphology (seven MSPA-derived patch types). The heat risk index is based on the hazard–exposure–vulnerability framework. It integrates mean air temperature, population density, and elderly population density data. It is modeled using the CatBoost (1.2.10), XGBoost (3.2.0), and LightGBM (4.7.0) models combined with SHAP analysis (0.46.0). The findings show a spatial distribution of heat risk characterized by high values in the southern and core areas and low values in the northern and peripheral areas, with fine-scale variation captured primarily through building density and green space morphology, while coarser variables represent background thermal conditions. The CatBoost model performed best (R2 = 0.782), significantly outperforming multiple linear regression (R2 = 0.567). Building density (BD, 33.62%) and Enhanced Vegetation Index (EVI) (29.32%) are the strongest predictors of heat risk in the model. Both have clear nonlinear thresholds. Predicted heat risk increases sharply when BD exceeds 0.120. Beyond a certain EVI threshold, the cooling benefit weakens. The highest heat risk occurs where high BD meets low vegetation. In contrast, intact core patches with high EVI produce a synergistic cooling effect. Although the impact of green space morphology is relatively limited, it still holds important research value. These predictive associations point to a three-dimensional strategy (development control, vegetation enhancement, and green space optimization) for heat-resilient urban planning. Methodologically, this study aligns with recent advances in heat risk assessment, employing high-resolution spatial analysis, multi-source data fusion, and interpretable machine learning; indicator weights were objectively determined via the entropy weight method, and sensitivity analyses confirmed the robustness of the index and its predictive relationships.