Spatiotemporal Flood Hazard Classification in Bangkok Using Graph Convolutional Network and Temporal Fusion Transformer

Urban flooding is a significant and increasing phenomenon in Bangkok that threatens human life. High population density, low elevation, and seasonal monsoons contribute to increased vulnerability to flooding. Traditional flood prediction models often fail to capture spatial correlations across districts and the temporal patterns within different types of features. To address this problem, this study proposes a hybrid deep learning framework combining Graph Convolution Network (GCN) and the Temporal Fusion Transformer (TFT) for predicting flood hazard levels in 50 Bangkok districts. The GCN component learns spatial dependencies from a graph constructed based on district relations, while the TFT learns temporal sequences based on spatiotemporal inputs. The model is trained on a historical 22-year dataset incorporating meteorological, hydrological, and socioeconomic variables. Experimental evaluation demonstrates that the proposed GCN-TFT model achieves superior classification performance compared to the existing models. Interpretability is achieved using attention-based variable importance and regime shift detection. The predicted flood hazard levels are mapped into a high-resolution map produced through QGIS that indicates a strong correlation with flood-prone areas. This study offers a robust, interpretable, and spatially consistent framework for urban flood hazard assessment with practical applications in disaster mitigation and urban resilience planning.