Nonlinear relations of urban morphology to thermal anomalies: A cross-time comparative study based on Grad-CAM and SHAP

Abstract

Urban thermal anomalies profoundly impact human society, affecting daily life, public health, and residential comfort. Prior studies linked thermal anomalies to urban morphology evolution and land use change during urbanization based on multi-indicator quantification of urban morphology and linear regression modeling. However, it remained unclear which urban morphology elements predominantly dominate thermal anomalies and whether their impact is solely linear, and understanding on the diverse mechanisms through how urban morphology influences various thermal anomalies across seasons remains limited. Therefore, this study employed convolutional neural networks and interpretable machine learning (Grad-CAM and SHAP) to explore nonlinear relationships between urban morphology and thermal anomalies, focusing on comparisons between different types of anomaly events across time. The main findings indicated: (1) Grad-CAM's identification of pivotal hotspot pixels and SHAP's interpretability assessment highlighted that crucial urban morphology factors contributing to thermal anomalies include the area of green spaces, water spaces, the number of residential facilities, building floor area ratio, and the count of industrial production facilities. (2) Clear nonlinear relationships were observed between dominant urban morphology factors and the occurrence of thermal anomalies, which confirming the existence of multiple thresholds and activation levels, as demonstrated through SHAP's partial dependency analysis. The dynamic complexity of these associations significantly varied depending on the type of event and the timing of thermal anomalies. These findings offer actionable guidance for urban planners to refine climate-friendly strategies, revealing the heterogeneity of these relationships across time and seasons through multi-scenario analysis and providing tailored insights for climate-sensitive urban planning.

Keywords

• Urban morphology
• Thermal anomalies
• Interpretable machine learning
• Nonlinear relationship