Physics-Aware Multimodal Urban Heat Mapping with Open Web Imagery and Mobility Data

Extreme urban heat is intensifying worldwide and often falls hardest on vulnerable communities, posing growing challenges for climate adaptation and Sustainable Development Goal 11. Fine-grained land surface temperature (LST) estimates are essential for identifying local heat risks, yet most operational approaches still rely on satellite products alone, which are constrained by cloud cover, revisit cycles, and limited sensitivity to human-scale morphology and activity. Meanwhile, web-based resources such as online imagery and mobility data offer rich but underused signals for scalable heat-risk monitoring. We present AESPA, a physics-aware multimodal framework for tract-level urban LST estimation that combines satellite imagery, street-view panoramas, mobility-derived activity profiles, and interpretable physical proxies. AESPA trains a teacher model that jointly leverages imagery and mobility, then distills its predictions and representations into an imagery-only student, enabling deployment in data-poor cities. Physics-and proxyguided losses encourage consistency with basic urban-climate relationships and yield attributions linked to vegetation, impervious surfaces, shading, and surface reflectance. We evaluate AESPA across eight major U.S. metropolitan areas under within-city and cross-city transfer protocols: AESPA reduces mean absolute error by about 32% and increases Pearson correlation between predicted and observed tract-level LST by 0.15 compared with the strongest satellite-based baseline, and improves transfer correlations by roughly 0.05-0.10. Its proxy attributions recover physically coherent gradients associated with neighborhood-level heat-exposure inequality, illustrating how web-based imagery and mobility can support transparent, deployable urban heat-risk monitoring in practice 1 .