Decoding seasonal urban heat dynamics at neighborhood-scale using explainable deep learning for climate-resilient, digital twin-ready green planning

Understanding the spatiotemporal dynamics of urban surface temperatures is essential for climate-resilient green planning under intensifying heat stress and rapid land-use change. This study investigates seasonal thermal behavior at a microscale analytical resolution (200 × 200 m) across Bologna, a densely urbanized city with diverse morphology. Satellite-derived land surface temperature (LST) was statistically downscaled to 10 m resolution and subsequently aggregated to the 200 m grid, and integrated with urban morphological and socioeconomic variables, including built surface, tree cover, population density, land-use classes, and settlement zones. A twofold approach was developed: (1) unsupervised k-means clustering to identify thermally distinct clusters per season, revealing consistent spatial patterns where tree-dominated areas remain cooler and dense urban cores warmer; and (2) an explainable deep learning model based on a convolutional neural network (CNN) to quantify variable contributions using permutation-based feature importance, benchmarked against Support Vector Machine, Random Forest, and XGBoost models. Model performance differences were evaluated using paired statistical tests with significance assessed at accepted thresholds. The CNN demonstrated competitive performance, effectively capturing nonlinear relationships in the data and, when combined with subsequent analysis, providing insights into feature relevance. Built surface was the most influential driver (>37% across all seasons), while tree cover (7%–11%) and population density (5%–35%) showed pronounced seasonal variability, reflecting shifts in vegetation activity and energy demand. This integrative, interpretable modeling framework establishes a foundational step toward a neighborhood-scale urban climate digital twin, capable of linking multi-source data to simulate and explain intra-annual heat dynamics. The findings provide critical insights for adaptive, fine-scale heat mitigation and climate-resilient urban green planning.