Beyond Surface Temperature: Nonlinear, Interactive, and Spatially Heterogeneous Associations Between the Built Environment and Street-Scale Urban Heat Risk

Urban heat risk involves hazard, exposure, and vulnerability, but street-scale built-environment associations remain unclear. We assessed 9083 street-buffer units in Changsha, China, using a composite risk index, Random Forest, XGBoost, LightGBM, SHapley Additive exPlanations (SHAP), and multiscale geographically weighted regression (MGWR). All 13 predictors entered machine learning and SHAP, while seven SHAP-selected predictors (87.7% cumulative contribution) entered MGWR. Random Forest achieved the highest test-set R2 (0.583), while its mean spatial cross-validation R2 was lower (0.229). Functional activity contributed most to SHAP importance (57.1%); land use mix, point-of-interest density, distance to water bodies, and nighttime lights were leading predictors with nonlinear responses. Cross-dimensional pairs accounted for 73.6% of absolute SHAP interaction contributions but comprised 56 of 78 possible pairs (71.8%). Under the primary specification, MGWR indicated predominantly negative local canopy-height and positive point-of-interest-density associations. Functional activity remained the leading dimension across alternative buffer and risk-index specifications, although local patterns varied: canopy height showed a rank correlation coefficient of 0.195 with the baseline after excluding activity intensity and 73.9% sign agreement under entropy weighting; point-of-interest density retained high sign agreement across both index alternatives. These within-study associations support site-specific adaptation rather than universal planning thresholds.