Graph based urban flood prediction with GNNs

Urban flooding is becoming an increasing problem in quickly growing cities, demanding precise and prompt forecasts to ensure effective disaster management. Traditional hydrological models require a large number of physical parameters and are frequently too slow for real-time applications. We suggest a data-driven method based on graphs, utilizing Graph Neural Networks (GNNs), to predict urban flooding. The urban area is depicted as a graph, with nodes representing monitoring stations or spatial grid cells that include information on rainfall, elevation, and land use, and edges indicating geographic closeness and drainage connections. Historical rainfall and water-level data are converted into time-based sequences and integrated with static geographic characteristics as inputs for the nodes. We use a spatio-temporal GNN that combines temporal encoding with graph convolution to forecast future flood conditions (such as water depth or whether there is a flood) at each node. The model is compared with persistence and non-graph baselines using RMSE, MAE, and F1-score, and the predictions are displayed on urban maps to analyse spatial patterns. The results indicate that the GNN effectively captures the temporal dynamics and spatial relationships involved in flood propagation, showcasing a scalable, data-driven approach for near real-time prediction of urban flood risks.