A novel GCN-based framework for uncovering diurnal patterns and drivers of the surface urban heat island (SUHI) in hyper-dense cities

Amid accelerating global climate change and rapid urbanization, the surface urban heat island (SUHI) effect has emerged as a contributor to urban thermal stress in hyper-dense cities. However, accurately modeling SUHI remains challenging due to pronounced spatial dependence and nonlinear interactions among urban features. To address these challenges, this study develops a novel framework based on a graph convolutional network (GCN) that integrates a SUHI dataset with built and natural urban features to predict daytime and nighttime SUHI in Hong Kong. The results show that the GCN outperforms traditional baselines, achieving a test R 2 of 0.9173 by day and 0.8789 by night. Interpretability analysis using integrated gradients (IGs) reveals that the normalized difference vegetation index (NDVI) dominates daytime cooling (mean absolute IG of 0.0987), whereas building density (0.0331) and coastal proximity (0.0217) are more influential at night. Notably, these effects vary considerably across districts. The cooling potential of vegetation within dense urban cores can be compromised by its ornamental form and surrounding impervious surfaces, whereas the impact of built features is highly dependent on local morphological characteristics. Moreover, Friedman's H 2 interaction analysis reveals that vegetation cools most effectively when co-located with dense building clusters during the daytime ( H 2 = 0.10), whereas coastal proximity enhances nighttime vegetation cooling ( H 2 = 0.09). Overall, the study underscores the spatiotemporal heterogeneity of SUHI, highlighting the need for context-sensitive mitigation strategies and thereby advancing spatially interpretable artificial-intelligence applications that support adaptive planning for climate resilience in hyper-dense urban environments. • GCN outperforms baselines by modeling spatial dependency and nonlinearity. • SUHI drivers vary across time and space, with spatiotemporal heterogeneity. • Key factors are context specific: vegetation, building density, shoreline. • Feature interactions shape and sometimes moderate cooling effects. • Explainable spatial AI informs context-sensitive heat mitigation.