From Thermal Diagnosis to Spatial Allocation: A Remote-Sensing and Explainable Machine Learning Framework for Heat-Resilient Planning in Semi-Arid Grassland Towns

Urban heat and extreme-temperature risks increasingly constrain sustainable planning in semi-arid grassland towns. Taking Xi Ujimqin Banner as a case study, this research develops a remote-sensing-informed diagnosis-to-allocation framework for heat-resilient spatial planning. The framework first uses remote-sensing-derived land surface temperature (LST) and multi-source spatial indicators to identify model-explained associations between LST and ecological, morphological, land-use, accessibility, and socioeconomic variables. A CatBoost model with SHAP, PDP, and ICE interpretation was applied to explain dominant LST-related patterns. Spatial units were then classified by integrating LST intensity, sky view factor (SVF)-related spatial openness, land-use structure, ecological sensitivity, and economic output characteristics. Finally, a Quasi-oppositional Learning and Levy-flight enhanced Cheetah Optimization Algorithm (QLA-COA)-based constrained optimization model was used to generate alternative allocation schemes. In the framework, LST reduction and land-development economic benefit were treated as the two primary planning objectives, while SVF-related openness was incorporated as an auxiliary spatial form and radiative geometry indicator rather than as a direct temperature-control objective. Results show that CatBoost achieved the best LST prediction performance (R2 = 0.834, RMSE = 1.762 °C, MAE = 1.371 °C). Cooling-priority, openness-priority, economic-priority, balanced-development, and ecological-restricted units accounted for 18.7%, 15.8%, 12.9%, 26.9%, and 25.7% of all units, respectively. The balanced-development scheme reduced mean LST by 1.30 °C while maintaining SVF-related spatial openness and improving economic benefits. This study provides a practical decision-support framework for climate-adaptive spatial governance in semi-arid grassland regions by linking LST diagnosis, spatial response translation, ecological constraints, auxiliary openness evaluation, and allocation-scheme comparison. The source code and processed spatial unit dataset used for model training, SHAP interpretation, and PDP/ICE visualization are publicly available.