A Hybrid Graph Neural Networks for Coastal Flood Detection and Short-Horizon Forecasting Using Hydrological Data

Flooding remains one of the most destructive natural disasters worldwide, causing extensive socio-economic damage, environmental/ degradation, and displacement of vulnerable populations, especially in coastal and low-income regions. Traditional hydrological models and existing machine learning techniques often face challenges such as high computational complexity, limited spatial representation, and poor modelling of the non-Euclidean structure of hydrological networks. To address these limitations, this study proposes a Hybrid Graph Neural Network–Graph Attention Network (GNN–GAT) framework for coastal flood detection and short-horizon forecasting using hydrological and environmental data. The framework combines the spatial message-passing capability of Graph Neural Networks with the adaptive neighbour-weighting mechanism of Graph Attention Networks to improve prediction accuracy and interpretability. The study utilized a publicly available Kaggle hydrological dataset containing 1,117,957 records with 20 environmental and infrastructural predictor variables and one target variable representing Flood Probability. Data pre-processing involved validation, feature normalization using z-score scaling, and graph construction to represent hydrological connectivity among nodes. The hybrid framework employed graph-based learning and multi-head attention mechanisms to model upstream-downstream flood propagation and identify influential environmental factors contributing to flood occurrence. Model performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), coefficient of determination (R²), confusion matrix analysis, and Area Under the ROC Curve (AUC). Experimental results demonstrated strong predictive capability across all evaluated models. The Graph Neural Network achieved an AUC of 0.941, the Graph Attention Network achieved 0.948, while the Hybrid GNN–GAT model achieved 0.947, confirming the effectiveness of graph-based deep learning for flood risk prediction. The system successfully classified both high-risk and low-risk flood scenarios and generated interpretable probabilistic outputs through an interactive Stream lit-based dashboard. The findings indicate that integrating graph neural learning with attention mechanisms significantly improves the modelling of hydrological dependencies and flood propagation patterns compared to traditional approaches. The study concludes that the proposed Hybrid GNN–GAT framework provides a reliable, scalable, and interpretable solution for coastal flood detection and short-horizon forecasting. The framework supports disaster risk reduction, early-warning systems, and climate resilience initiatives aligned with Sustainable Development Goal 11 (Sustainable Cities and Communities) and Sustainable Development Goal 13 (Climate Action).