Modeling high-resolution air temperature variability with multi-source data: predictive insights for urban heat mitigation strategies

Understanding the spatial and temporal variability of urban air temperature is essential for addressing urban heat island (UHI) effects and supporting climate-responsive planning. However, modeling fine-scale air temperature variability remains a challenge due to the complex interplay of urban morphology, land cover, and anthropogenic factors. This study presents a scalable and spatially explicit machine learning (ML) framework to model near-surface air temperature at high spatial resolution, using multi-source geospatial and meteorological data. By integrating satellite-derived land surface temperature (LST), Internet-of-Things (IoT) sensor observations, urban morphology, and anthropogenic activity indicators, the model captures diurnal dynamics of urban heat. To account for spatial autocorrelation, spatial lag features were incorporated as model inputs, improving predictive performance and spatial coherence. The best-performing random forest model with spatial lag features achieved an R² of 0.9939 and an RMSE of 0.4363 °C. Compared to conventional interpolation or physical modeling approaches, the proposed framework offers reduced dependence on dense sensor networks, and enhancement in spatial detail. This approach enables detailed mapping of thermal exposure and supports the identification of priority areas for UHI mitigation strategies.