Explainable Spatio-Temporal Graph Neural Network for Urban Land Surface Temperature Prediction and Thermal Hotspot Assessment

Abstract

The relationships between land cover, vegetation, urban morphology, and the atmosphere play a significant role in shaping urban land surface temperatures. Precise prediction of such spatiotemporal variations in thermal patterns is essential for assessing health risks and planning for urban climate resilience. The study introduces a novel explainable spatio-temporal graph neural network (X-STGNN) for forecasting Land Surface Temperature (LST) and conducting thermal hotspot analysis in three cities in India: Chennai, Hyderabad, and Delhi. The framework combines remotely sensed environmental indicators, Local Climate Zone (LCZ) data, urban morphological features, and meteorological data to capture the spatial and temporal heterogeneity of urban thermal environments. Spatial relationships between urban units are represented using graph convolutional learning and temporal relationships using Long Short-Term Memory (LSTM) learning, followed by feature fusion and the application of a temporal attention mechanism for prediction. The experimental data span the period from 2018 to 2023, with 270 temporal observations at 8-day intervals. The proposed model achieved an RMSE of 0.82 °C, an MAE of 0.61 °C, and an R2 of 0.96 for the predefined evaluation set. The proposed model further outperformed the best-performing comparative model, DCRNN, with an RMSE of 0.82 °C compared with 0.97 °C for DCRNN. Ablation experiments also highlighted the significance of spatial graph learning, temporal learning, LCZ information, remote-sensing indices, and meteorological variables. The LOCO evaluation yielded an average RMSE of 0.87 ± 0.03 °C, which showed beneficial cross-city transferability but also demonstrated the influence of geographical domain differences. Model predictions were interpreted using SHAP, and NDVI, NDBI, and LCZ were some of the most influential predictors. The results demonstrate the potential of the proposed structure as an explainable approach to providing predictive decision support in urban thermal assessment.