Unraveling urban surface heat dynamics through deep ensemble machine learning

Abstract

Climate change intensifies urban heat challenges, necessitating a comprehensive investigation into urban thermal dynamics. This study presents an integrated machine learning framework to predict both daytime and nighttime heat distribution in Victoria, Australia, by incorporating a wide range of natural and anthropogenic factors. The models demonstrate high predictive accuracy, with R² values of 0.92 for daytime and 0.91 for nighttime, and RMSE values of 1.29 °C and 0.89 °C, respectively. The results indicate that daytime heat is predominantly influenced by the Normalized Difference Vegetation Index (NDVI), while nighttime heat is more strongly associated with population density. In low-altitude areas, moderate increases in NDVI contribute to more balanced heat distribution, although excessive vegetation coverage yields limited benefits and may even increase local temperatures. Nighttime temperatures in mid- to high-altitude regions are significantly affected by population density. The urban daytime thermal environment responds well to vegetation, suggesting that greening is an effective strategy for heat mitigation. However, in densely populated suburban areas, nighttime heat distribution exhibits substantial spatial variability. Furthermore, population density shows a clear seasonal effect on nighttime temperatures, with the influence being most evident in summer and tending to stabilize once the population exceeds 5 × 10⁵ people. This study reveals the underlying mechanisms influencing urban heat distribution and provides scientific guidance for urban planning strategies aimed at enhancing thermal comfort and promoting environmental sustainability.