From Heat Maps to Cooling Actions: AI-Driven Citywide Hourly Mapping of Thermal Stress, Drivers, and Targeted Cooling

Thermal stress is critical for sustainable, health-oriented urban planning, yet remains difficult to characterize at hourly resolution and citywide in dense cities. Using Hong Kong as a case study, this study develops a regression-based machine learning framework to predict hourly thermal stress, expressed as Net Effective Temperature (NET). Long-term meteorological data was combined with detailed geographic, urban-morphological, and landscape composition–configuration metrics in a multilayer perceptron artificial neural network (MLP-ANN). The model uses 37 input variables, while the target variable is the continuous hourly NET. The model attains high predictive skill (R² = 0.97, RMSE=1.1°C over the full dataset) and captures pronounced diurnal and seasonal patterns in thermal stress. Performance remains robust under hot (R² = 0.97, RMSE = 0.5°C) and cold (R² = 0.94, RMSE = 1.7°C) conditions. In summer, afternoon NET reaches about 30.3°C and remains elevated at night, whereas in winter nocturnal NET can fall below –4°C. Spatially, higher NET concentrates in dense urban and industrial districts, with summer hotspots around the airport, Northwest New Territories, and Kwai Tsing Container Terminals. Daytime NET is dominated by near-surface air temperature, followed by P_impervious and NDBI, along with configuration metrics such as SHAPE, CIRCLE, and AREA, while P_green and NDWI become especially influential in winter daytime. At night, configuration metrics including FRAC, CONTIG, CORE, and GYRATE gain importance. The framework delivers citywide, hourly NET maps and quantitative driver rankings to guide targeted, nature-based and morphological interventions in subtropical high-density cities.