Microclimatic Heterogeneity and Nonlinear Environmental Associations Across Contrasting Spaces in a Historical District: An Exploratory THI–XGBoost–SHAP Framework

Historical districts contain spatial units with markedly different local microclimates, yet the temporal dynamics of their thermal-humidity environments and their associations with surrounding environmental conditions remain insufficiently resolved. This study examined three analytical settings in the Ciqikou Historical District, Chongqing: an Alley, the Main Street, and a Square. Using daytime multi-variable data collected on 4 August 2026, we integrated the temperature-humidity index (THI), an exact bivariate Shapley decomposition, and an XGBoost-SHAP framework to characterize thermal-humidity states and nonlinear environmental associations. Main Street and Square data were obtained from valid field measurements. Because the original Alley observations were lost owing to equipment failure, a stochastic reconstruction was generated within contemporaneous observational ranges and alley-specific environmental constraints; Alley results are therefore used only for exploratory comparison. Mean air temperature and THI were highest on Main Street (32.81 °C and 28.46 °C), followed by the Square (32.01 °C and 28.04 °C); the reconstructed Alley series yielded 31.16 °C and 27.73 °C, respectively. Exact Shapley decomposition showed that temperature accounted for 71.70–74.20% of the absolute contribution across the three analytical series, while the directions of the temperature and humidity contributions changed over time. These contribution shares depend on a common reference derived from both measured and reconstructed series and are not estimates based exclusively on field observations. Under grouped blocked cross-validation, out-of-fold R2 values were 0.80 for Main Street and 0.73 for Square, compared with 0.57 for the reconstructed Alley series. SHAP analysis indicated setting-specific predictive association structures and nonlinear responses. The high importance assigned to atmospheric pressure more plausibly reflects background meteorological and temporal covariation within the monitoring window than a direct causal effect. Overall, the results reveal pronounced context dependence in historical-district microclimates and underscore the importance of data provenance, temporal dependence, and explicit limits on model interpretation in short-duration field studies.