Real-time sensing of hourly human thermal comfort by integrating AI, satellite images, and ground observations

Accurate real-time monitoring of the thermal environment is essential for mitigating the impacts of thermal stress on human health and ecosystems, particularly in regions with limited high-resolution data, such as inland China. In this study, we utilized hourly meteorological observations to derive reference thermal indices for model training and validation. We then applied Himawari-8 satellite data, digital elevation models (DEMs), and time/geographic predictors to estimate 12 human thermal indices at a spatial resolution of 5 km across Jiangxi, China. The XGBoost model demonstrated the best overall performance, with mean testing metrics for the 12 human thermal indices of R 2 = 0.987, RMSE = 0.925 °C, and MAE = 0.656 °C. The retrieval results also show a high degree of consistency with in-situ observations across both spatial and temporal domains, effectively capturing the diurnal variations and spatial heterogeneity of the urban thermal environment. Our study offers a methodological foundation for the hourly estimation of human thermal indices and supports applications such as extreme-event monitoring, urban thermal environment analysis, and heat/cold risk assessments. • An hourly human thermal monitoring method driven by the fusion of satellite-ground observations has been constructed. • Among multiple AI models, XGBoost exhibits the best estimation performance and shows strong spatio-temporal transferring ability. • The proposed model achieves high-precision, seamless, near real-time refined monitoring and early warning capabilities.