Dealing with Spatial Heterogeneity: A Graph Neural Network-based Domain Adaptation Framework in Urban Heat Prediction
Abstract
Land Surface Temperature (LST) is critical in understanding urban heat dynamics and their implications for sustainability, health, and planning. However, accurately predicting LST across diverse urban areas remains challenging due to spatial heterogeneity arising from variations in urban morphology, land cover, and local climate. This study introduces a Graph Neural Network-based Domain Adaptation (GNN-DA) framework that leverages graph-based spatial representations and adversarial learning to align feature distributions across cities. Trained on data from Singapore, the model demonstrates strong predictive performance in Malacca, achieving high R 2 scores and low Root Mean Square Error (RMSE), and shows reasonable generalisability to Melbourne and Bristol during summertime, underscoring the frameworks capacity to transfer knowledge across diverse urban contexts.
Pengyuan Liu
Research Fellow
Binyu Lei
PhD Researcher
