Spatial Downscaling of Streamflow Data with Attention Based Spatio-Temporal Graph Convolutional Networks

Abstract

Abstract Accurate streamflow data is vital for various climate modeling applications, including flood forecasting. However, many streams lack sufficient monitoring due to the high operational costs involved. To address this issue and promote enhanced disaster preparedness, management, and response, our study introduces a neural network-based method for estimating historical hourly streamflow in two spatial downscaling scenarios. The method targets two types of ungauged locations: (1) those without sensors in sparsely gauged river networks, and (2) those that previously had a streamflow sensor, but the gauge is no longer available. For both cases, we propose the ScaleGNN, a graph neural network architecture. We evaluate the performance of ScaleGNN against a Long Short-Term Memory (LSTM) baseline and spatial persistence in estimating discharge values over a 36-hour period. Our findings indicate that ScaleGNN surpasses spatial persistence in the first scenario, while both neural network approaches demonstrate their effectiveness compared to spatial persistence in the second scenario.