One city, two heats: An LLM-enabled comparative analysis of heat perception, thermal environment, and health pathways in Beijing, China

Urbanization intensifies urban heat islands and health risks, yet most studies prioritize physical heat exposure. Emerging work notes that perceived heat can diverge from thermal measures, but rigorous comparison is limited. We address this gap by using a Large Language Mode (LLM) to extract Social Media Heat Perception (SMHP) from social-media posts about Beijing's urban parks and integrating it with Physical Thermal Environment (PTE) derived from remote sensing and a population-weighted model. Coupled with subdistrict indicators of physical health and well-being, we apply Partial Least Squares–Structural Equation Modeling to identify direct and mediating pathways and Multiscale Geographically Weighted Regression to map spatial heterogeneity. We show that a LLM can robustly detect semantic SMHP, achieving a 94.5% true positive rate for extreme heat-discomfort cases. The results indicate that PTE is a stronger predictor of health outcomes and is shaped by macro-scale landscapes and regional context, whereas SMHP is associated with micro-scale, in-park features. The biophysical cooling effect has a greater impact on PTE than SMHP, suggesting a “perception lag” in which expressed sentiment is less responsive than objective measurements. Built and socioeconomic factors exhibit a “heat paradox,” where greater resources coincide with higher exposure. Spatial mismatches between SMHP and PTE reveal that temperature-only metrics can miss hidden vulnerabilities. We advocate a coordinated, dual-dimension, multi-scale strategy to support evidence-based climate adaptation and health equity.