Researchers Achieve 0.97 Accuracy in Flood Forecasting with Graph Neural Networks and 15x Speedup

Accurate flood forecasting presents a continuing challenge for effective water resource management, requiring models that capture both local runoff patterns and complex spatial interactions within river networks. Aishwarya Sarkar, Autrin Hakimi, and Xiaoqiong Chen, from Iowa State University, along with Hai Huang, Chaoqun Lu, and Ibrahim Demir from Tulane University, address this need by introducing HydroGAT, a novel approach that integrates graph neural networks and transformers to simultaneously learn from spatial and temporal data. The team’s model represents river basins as detailed graphs, allowing it to propagate information along hydrological flow paths and capture influences between different locations, even at high resolutions. Results demonstrate that HydroGAT significantly improves flood prediction accuracy, achieving high scores in key metrics and offering interpretable insights into how different parts of a basin influence each other, while a distributed computing pipeline enables efficient training on large datasets and powerful supercomputers.

Traditional data-driven approaches, including convolutional networks and sequence-based models, often overlook crucial topological information about the region. Graph Neural Networks (GNNs), however, propagate information directly along the river network, making them well-suited for learning how water moves through a landscape. Existing GNN-based flood prediction models often simplify the landscape by aggregating data to large areas, because training complex models requires substantial computational resources. Furthermore, many existing methods treat spatial and temporal dependencies separately, limiting their ability to fully capture the complexity of flood events.

Spatio-Temporal Hydrological Forecasting with Deep Learning

Researchers developed a new approach to hydrological forecasting that combines the strengths of Graph Neural Networks (GNNs) and Transformers to accurately predict streamflow and flood risk. The team addresses limitations in existing methods by explicitly modeling both the spatial relationships within a river basin and the temporal dependencies in rainfall and streamflow data. The model represents the hydrological system as a graph, where locations like river segments and land parcels function as nodes, and connections represent how water flows between them. GNNs then learn how water moves through this network based on the landscape’s topography and connectivity.

To capture long-range temporal dependencies, the team utilizes Transformers, originally developed for natural language processing. These models analyze time series data, such as rainfall and streamflow, to identify patterns and predict future conditions. By combining GNNs and Transformers, the researchers created a model that effectively integrates spatial and temporal dynamics, leading to more accurate predictions. A key innovation is the development of scalable training methods, allowing the model to process large datasets and high-resolution data. This is achieved through distributed training, where the computational workload is shared across multiple machines.

HydroGAT Improves Spatiotemporal Flood Forecasting Accuracy

Researchers developed HydroGAT, a novel spatiotemporal network designed to improve flood forecasting by accurately modeling complex hydrological processes within river basins. The team addressed limitations in existing methods that often treat spatial and temporal dependencies separately, or struggle with the computational demands of high-resolution data. HydroGAT utilizes a heterogeneous basin graph where every land and river pixel functions as a node, connected by physical flow directions and inter-catchment relationships, allowing for a more comprehensive representation of the hydrological landscape. Experiments conducted on the Cedar River Basin and Des Moines River Basin in the US Midwest demonstrate significant performance improvements over several baseline architectures.

HydroGAT achieves a peak Nash-Sutcliffe Efficiency (NSE) of 0. 97, alongside a Kling-Gupta Efficiency (KGE) of 0. 96, indicating a high degree of accuracy in predicting discharge levels. Notably, the model maintains low bias, with a Percentage Bias (PBIAS) consistently within 5%, demonstrating reliable predictions without systematic over or underestimation. To facilitate high-resolution basin-scale training, the researchers developed a distributed data-parallel pipeline capable of scaling efficiently up to 64 A100 GPUs on the NERSC Perlmutter supercomputer, achieving up to a 15x speedup compared to single-machine training. This advancement enables the processing of detailed hydrological data, leading to more accurate and nuanced flood forecasts. The results demonstrate that HydroGAT surpasses existing methods like DCRNN, GraphWaveNet, and RGCN, establishing a new benchmark for spatiotemporal flood prediction and offering a valuable tool for water resource management and disaster preparedness.

Spatiotemporal Graph Model Improves Flood Forecasting

HydroGAT demonstrably improves flood forecasting by treating river and land pixels as nodes within a heterogeneous graph, capturing both flow dynamics and inter-catchment dependencies. Evaluated in two Midwestern US basins, the model achieves higher accuracy, with peak Nash-Sutcliffe efficiency of 0. 97 and Kling-Gupta efficiency of 0. 96, and reduced bias in hourly discharge prediction compared to five baseline architectures. Detailed analysis confirms the importance of both spatial and temporal components in accurately modeling runoff-routing processes.

The research team also developed a distributed training pipeline that significantly accelerates model training, achieving up to a 15-fold speedup using 64 GPUs on a supercomputer. This advancement enables high-resolution, basin-wide flood forecasting, making real-world deployment more feasible. While the current implementation focuses on two specific basins, the authors acknowledge that integrating additional environmental features, such as impervious surface and land use, could adapt the model to more diverse landscapes. Beyond flood prediction, the principles behind HydroGAT are applicable to other spatiotemporal prediction tasks where network topology and long-range temporal dependencies are important, such as traffic flow or weather forecasting.