Functional Zone LST Patterns, Green-Space–LST Associations, and 2D/3D Urban-Form Relationships Across Taiyuan, Chengdu, and Yinchuan

Abstract

Using selected summer Landsat 8 scenes for Taiyuan, Chengdu, and Yinchuan, this study compared functional zone land surface temperature (LST) patterns, green-space–LST associations, and the relationships of 2D landscape configuration, blue-green composition, 3D building form, and terrain with LST. Separate city-specific Light Gradient Boosting Machine (LightGBM) models with SHapley Additive exPlanations (SHAP) interpretation characterized within-model predictor contributions, SHAP dependence patterns, and selected representative pairwise interactions. Polygon-level LST distributions differed among functional zones within every city (Kruskal–Wallis, all p < 0.001), and Commercial zones had higher LST than Park/green-space zones in all three cases with large pairwise effect sizes. Other pairwise contrasts were city-specific. Most city-zone regressions showed negative green-space–LST associations, whereas Chengdu Industrial showed no clear association. Building volume ranked among the Top-3 predictors in every city model, but the co-dominant structures differed: building volume–green-space proportion–elevation in Taiyuan, mean building height–building volume–green-space proportion in Chengdu, and building volume–aggregation index–elevation in Yinchuan. Five-fold fixed-specification spatial block cross-validation provided a more conservative spatially held-out sensitivity assessment than ordinary random splitting, while reference and repeated spatial-block out-of-fold SHAP analyses showed stable within-model predictor rankings and contribution shares. An 85-scene multi-date Landsat comparison with selected-scene leave-one-out sensitivity supported broad summer spatial consistency in Taiyuan and Yinchuan, with partial, context-dependent support in Chengdu. After excluding each city’s selected 2019 scene from its own reference median, the relative–LST Spearman correlations remained 0.895, 0.861, and 0.909 for Taiyuan, Chengdu, and Yinchuan, respectively. These findings distinguish recurring comparative anchors from context-dependent predictor structures, supporting context-sensitive planning hypotheses rather than universal predictor rankings.