Identifying morphological thresholds in spatio-thermal interactions

Abstract

Urban heat exhibits complex nonlinear relationships with morphological form, with thermal responses shifting across critical thresholds in building density, canyon geometry, and block configuration. While machine learning enables the detection of these interactions, existing studies predominantly report broad threshold ranges without identifying precise morphological breakpoints or validating findings across different models, thereby limiting translation into regulatory standards. This study benchmarks four tree-based AI algorithms – XGBoost, Gradient Boosting Machine (GBM), Random Forest (RF), and Light Gradient Boosting Machine (LGBM) – for predictive performance and interpretability in Istanbul's urban heat analysis. Shapley Additive Explanations (SHAP) quantify feature contributions across models, while hierarchical change-point detection identifies precise morphological thresholds where thermal effects shift between nonlinear regimes. Results reveal narrow differences in accuracy (R2 = 0.675–0.685), with computational efficiency and interpretability proving more decisive. LGBM trained 4.5 times faster than RF, and XGBoost exhibited the highest morphological sensitivity. Normalised Difference Vegetation Index and building count emerge as dominant thermal drivers across all models. Consensus-based threshold detection, quantified as inter-model standard deviation (SD) across the four algorithms, yielded 80 breakpoints. Three exhibit high cross-model agreement (SD < 0.02), indicating model-invariant regime shifts: verticality at height-to-footprint ratio (H/A) = 0.14 marks the onset of canyon shading effects that offset thermal mass penalties; urban block area thresholds at 2248 m2 and 6883 m2 indicate permeability constraints triggering heat retention. Four thresholds demonstrated moderate consensus (SD 0.02–0.04): building density at ∼16 buildings/ha corresponds to nonlinear heat intensification as impervious coverage reaches 49%; floor area ratio at 1.38 marks mid-rise regime shifts. By validating thresholds across models, this study moves beyond approximate value ranges toward robust thresholds applicable to climate-responsive zoning. This represents one of the first empirical contributions to systematically detect morphological breakpoints through cross-model validation.