Accelerating flood warnings by 10 hours: the power of river network topology in AI-enhanced flood forecasting

Abstract The increasing frequency and intensity of floods, exacerbated by climate change, necessitates the development of accurate and timely flood forecasting models. Although AI-based approaches have demonstrated promise, the effectiveness of graph neural networks (GNNs) in modeling the intricate dynamics of river networks remains contested. Despite the natural alignment between river topology and graph-based structures, recent studies reveal that GNNs often underperform in fully utilizing this structural information. This research aims to identify the underlying factors contributing to this limitation. Our analysis reveals that the tree-like configuration of river networks leads to over-squashing in GNNs, a problem caused by high resistance distances between nodes. To address this, we propose a novel method that transforms the topological graph into a dense, reachability-based graph, which reduces resistance distances. Empirical results demonstrate that GNNs applied to the transformed graph outperform EA-LSTM, particularly in predicting rare and extreme flood events. Furthermore, the incorporation of graph information significantly enhances long-term forecasting capabilities, as evidenced by the fact that, on average, GNN predictions of water levels at 24 h after using the dense graph match the accuracy of EA-LSTM’s 14-h forecasts. These findings highlight the potential of graph-based methodologies to improve flood prediction systems and contribute to more effective early warning mechanisms.