SWOT-based spatiotemporal deep learning for reach-scale forecasting of river and reservoir stages in the Mekong

Machine learning-based hydrological forecasting models are conventionally developed as lumped systems that predict watershed outflow using basin-averaged weather forcings. These approaches neglect the spatial information of meteorological inputs and their interactions with river network topology, limiting their ability to produce detailed reach-scale forecasts, i.e., for individual river reaches within the watershed. Here we introduce a new framework that provides reach-scale forecasts trained directly on SWOT water surface elevation (WSE) observations rather than outlet gauges. Incorporating spatially continuous SWOT observations allows the ML model to learn across the river network instead of only at gauged locations. The framework delivers 0- to 9-day forecasts of river and reservoir stages across an entire watershed. Our model integrates a graph neural network (GNN) with a long short-term memory (LSTM) network. The GNN encodes gridded meteorological forcings into the river network in a manner consistent with the runoff generation process. The LSTM component captures temporal dependencies and produces stage forecasts at key reservoirs and river reaches. Additional model inputs include 0- to 9-day Multi-Source Weather (MSWX) forecasts and river attributes from the Global River Topology (GRIT) dataset. The framework is implemented over the Mekong River Basin to generate forecasts for cascading reservoirs. Results demonstrate improved predictive performance relative to baseline Random Forest and LSTM models, highlighting the value of incorporating hydrological connectivity and satellite-based observations to improve forecasting in data-scarce regions.