Street-Scale Nonlinear Associations Between 2D and 3D Plant Morphology and Land Surface Temperature

Urban streets are important heat-exposure environments, yet the relationships of two-dimensional (2D) planar plant morphology and three-dimensional (3D) vegetation structure with land surface temperature (LST) remain insufficiently integrated at a continuous street scale. We analyzed 42,603 street-scale study units in the central urban area of Wuhan, defined at 50 m sampling intervals with a 150 m radius. The 2D variables comprised green-space area (A), mean patch perimeter (P_mean), and perimeter-area ratio (P_A), while the 3D variables comprised green view index (GVI), mean 3D green volume (NV_mean), mean canopy height (CH_mean), and canopy-height variability (CH_sd). Anselin Local Moran's I identified High-High (HH) and Low-Low (LL) zones; linear regression (LR), random forest (RF), and SHAP characterized linear, nonlinear, and model-based contributions; and buffered spatial cross-validation and spatial resampling evaluated robustness. Under the original random 80-20% train-test split, LR/RF R² values were 0.2153/0.4002 for the overall study area, 0.1940/0.3539 for the HH zone, and 0.0488/0.4853 for the LL zone. Under five-fold buffered spatial cross-validation, the corresponding pooled out-of-fold R² values were 0.1940/0.2237, 0.1427/0.0965, and -0.0942/-0.0231, showing that the RF advantage weakened after spatial separation and did not persist in the HH and LL zones. In the original fitted RF models, A, P_mean, and NV_mean had the largest mean absolute SHAP contributions overall; A, P_A, and P_mean ranked highest in the HH zone; and GVI, CH_sd, and CH_mean ranked highest in the LL zone. Repeated buffered spatial validation showed no stable overall 2D predominance because the median 2D share of 52.1% had a 46.5-58.8% percentile range, but it supported stable 2D relative predominance in the HH zone (61.0% [55.5-66.7%]) and stable grouped 3D relative predominance in the LL zone (61.9% [52.3-70.3%]), despite unstable LL variable-level rankings. SHAP relationships were nonlinear and zone-dependent: A and P_mean showed clearer directional transitions overall and in the HH zone, whereas the LL zone and most 3D variables exhibited multiple directional changes. Spatial-block bootstrap analysis examined 40 full-sample zero-crossing candidates, of which 39 met the predefined stability criteria; these ranges represent model-derived directional transitions rather than ecological thresholds or causal planning standards. The findings demonstrate thermal-context-dependent, model-based associations between 2D and 3D plant morphology and street-scale LST, while emphasizing that model performance and some importance rankings are spatially sensitive and require local validation.