Built-Environment Quality Buffer Urban–Rural Connectivity Risk? A SHAP-Based Multi-Method Assessment in Guangzhou, China

Composite environmental risks accumulate unevenly along urban–rural gradients, yet the conditional and nonlinear interaction between built environment quality (BEQ) and urban–rural functional connectivity (URFC) remains poorly quantified at fine resolution. This study aims to determine whether, and under what conditions, BEQ moderates the relationship between URFC and a population-weighted composite risk index (CRI), and to translate the result into spatia targeted green-infrastructure priorities. We use 744,714 grid cells at 100 m resolution over Guangzhou, China. The framework couples entropy-weighted BEQ from satellite and street-view imagery, gravity-model URFC computed on the real road network, and a two-stage population-weighted CRI of heat and air hazards. We apply nested ordinary least squares with incremental F-tests, spatial-lag and spatial-error models, generalised additive models with B-spline bases, gradient-boosted trees with SHAP interaction values, and Baron–Kenny mediation analysis. The main BEQ × URFC estimates are negative across the parametric and machine-learning specifications. The interaction is, however, small: a spatial-lag model on a 10,000-cell subsample returns β = −5.4 × 10−4, but a scalable generalised-method-of-moments spatial regression on the full grid—where the spatial autoregressive coefficient reaches ρ ≈ 0.99—shows the coefficient to be negative yet not statistically significant, and a five-seed re-estimation confirms that the subsample-based significance is draw-dependent. We therefore interpret the buffering as directionally supported but small and not robustly significant once spatial autocorrelation is fully modelled. The buffering response is nonlinear in the GAM main effects, and BEQ buffers across the entire observed connectivity range rather than switching sign at an interior threshold; URFC functions predominantly as a moderator rather than a mediator. Population-stratified estimation shows that the buffering is exposure-conditional: it is strongest where population exposure is high and weakens or reverses in sparsely populated cells, consistent with the risk = hazard × exposure structure of CRI. Sensitivity tests across values of the distance-decay parameter, 100 entropy perturbations and spatial scales corroborate the buffering direction. The framework provides an evidentiary basis for prioritising green infrastructure in functionally connected, populated but environmentally degraded transition zones.