A Comprehensive Analysis of Graph Neural Networks for River Discharge Prediction: High-Resolution Applications in the Danube Basin

Recent advances in Earth observation and data collection technologies have made high-resolution hydrological datasets increasingly accessible, enhancing our capabilities for monitoring and predicting hydrological processes. While a variety of artificial intelligence (AI) models can be employed to leverage these datasets, the challenges and opportunities of different AI approaches in the context of high-resolution data availability remain only partially explored. For instance, although Long Short-Term Memory (LSTM) networks are widely used for discharge prediction, the potential of Graph Neural Networks (GNNs)—which naturally represent river networks as graphs and capture spatial dependencies—has yet to be fully investigated.

In this study, we conduct a comprehensive analysis of GNN-based models for river discharge prediction in the Danube River Basin. Leveraging the LamaH-CE (Large-Sample Data for Hydrology and Environmental Sciences for Central Europe) dataset, we incorporate both dynamic features (e.g., daily precipitation, temperature, soil moisture) and static variables (e.g., digital elevation model, river density, basin area). Three architectures—GNN, a hybrid LSTM-GNN, and a standalone LSTM—are trained, validated, and tested at daily time steps from 2000 to 2017.

We further investigate the impact of network density and high-resolution (1km) soil moisture and precipitation data on discharge prediction accuracy. Our analysis reveals the potential advantages and limitations of these architectural approaches in river discharge prediction under high-resolution data availability and underscores the growing importance of harnessing graph-based deep learning methods for hydrological applications.