Predicting Fluvial Flood Propagation using Graph Neural Networks

Climate change is expected to increase the likelihood of heavy precipitation events, leading to increased risks of riverine or urban flooding. This work investigates the application of two graph neural networks on predicting the propagation of a flood in its spatial and temporal dimension. The two models were trained to approximate the shallow water equations and evaluated on a synthetic and real world dataset. This study aims to determine the transferability of learned knowledge to new unseen topologies and investigates the impact of regular versus irregular grids on the model performance. During training the models learned basic skills of flood propagation which they were able to transfer to unseen topographies. Overall both models failed to accurately predict the inundation process on a real world dataset and were negatively or not at all impacted by regular over irregular grids. Error accumulation poses a challenge over long prediction intervals threatening the models of becoming unstable, which might be mitigated by larger training datasets. Despite these results, this research contributes to our knowledge in predicting and understanding floods, aiding in the planning of mitigation strategies.