Exploring Spatial Contributions to Flood Generation Using Graph Neural Networks: A Case Study on the Upper Medway Catchment

Deep learning approaches are increasingly being used in hydrological modelling due to their ability to represent the nonlinear relationships that characterise rainfall–runoff processes. Despite this growing interest, their use for improving hydrological understanding remains limited. In particular, issues related to interpretability, spatial attribution, and model robustness persist, especially in catchments with sparse or uneven data coverage. Moreover, many deep learning applications represent catchments in a lumped manner, making it difficult to identify how different subcatchments contribute to flood generation. This study investigates spatial runoff contributions during flood events by representing hydrological connectivity with graph neural networks (GNNs). The graph-based rainfall–runoff modelling framework is applied to the Upper Medway catchment (~220 km²), located south of London. The catchment is conceptualised as a directed graph, where the nodes are represented by 34 subcatchments, generated from a digital elevation model, alongside their static features (area, slope, land use), and the edges encode the downstream hydrological connections of the river network. Rainfall inputs are aggregated at the subcatchment scale from 10 rain gauges using sub-hourly (15min) data, while sub-hourly discharge observations from two gauging stations provide the basis for model training and evaluation. Additionally, the model's robustness and information redundancy were explored through a sensitivity analysis involving the omission of certain rainfall gauges. Finally, the model behaviour is assessed through event-based simulations and compared to established hydrological modelling approaches in the catchment. Instead of focusing on predictive accuracy, the aim of this study is to investigate the learned graph representations, especially how information from upstream subcatchments propagates through the network and influences simulated responses at the catchment outlet. The limitations related to data resolution, event definition, uncertainty representation, and transferability are discussed, and future work will focus on refining model architecture and addressing evaluation strategies.