Graph Neural Networks in Hydrology: Improving River Discharge Forecasts for Flood Early Warning Systems

Despite the effectiveness of flood early warning systems as mitigation tools, significant losses persist even in developed countries. Most operational systems provide discharge or water level forecasts at gauge locations, but their predictive accuracy remains limited. Recent advancements in earth observation and data collection technologies offer a wealth of high-resolution hydrological datasets, creating an unprecedented opportunity to apply Artificial Intelligence (AI) algorithms for improved monitoring and prediction of hydrological processes. While Long Short-Term Memory (LSTM) networks are widely used for discharge prediction due to their ability to model temporal dependencies in hydrological time series, they often overlook spatial dependencies within river networks.

This study explores whether Graph Neural Networks (GNNs), which represent river networks as graphs to capture spatial relationships, can improve discharge predictions when integrated with LSTM models. Leveraging the LamaH-CE dataset for the Danube Basin (1981–2017), we incorporate dynamic variables such as daily precipitation, temperature, and soil moisture, alongside 59 static catchment attributes, including digital elevation models, river density, basin area, and soil properties, and etc. We evaluate two approaches: (1) local discharge prediction at sub-basin scales using LSTM models, and (2) a hybrid LSTM-GNN framework where GNNs model water routing across the river network, accounting for upstream-downstream connectivity.

Our findings reveal that the hybrid LSTM-GNN approach significantly outperforms the standalone LSTM model, particularly in capturing spatial propagation of flows during high-discharge events. By explicitly modelling routing dynamics, GNNs enhance prediction accuracy in complex river systems, addressing a critical gap in current forecasting methods. These results underscore the value of integrating spatial context into hydrological modelling and highlight the transformative potential of graph-based deep learning for flood prediction. This framework offers a pathway to strengthen flood early warning systems, supporting more effective mitigation strategies.