A Dataset and Model for Imputing Water Surface Elevation on a Large and Extremely Sparse Spatiotemporal Graph

Monitoring water surface elevation (WSE) is vital for flood forecasting and water management, yet in situ gauges are sparse, and satellite altimetry suffers from low temporal frequency. The authors address this by creating AmazonWSE, a comprehensive dataset covering 19,000 river reaches in the Amazon basin over a decade (2016–2026). The dataset is characterized by extreme sparsity—fewer than 1% of river sections are observed on any given day—and a complex, directed acyclic river topology. To perform imputation, the authors propose a bidirectional selective state space model (Mamba) that treats the river network as a sequence. By sampling connected subgraphs and flattening space and time into a single token sequence, the model avoids the computational overhead of processing the entire graph while leveraging topological information through learned positional encodings.

The proposed model consistently outperforms existing transductive and inductive spatiotemporal imputation methods, including graph neural networks (GNNs) and attention-based models, across both the SWOT-era (2023–2026) and historical hindcast periods. While previous studies suggested that river topology offers limited benefits for discharge prediction, this paper demonstrates that topology significantly improves WSE reconstruction when used to guide subgraph sampling and metadata encoding. Compared to Reach-Reg, a state-of-the-art domain-specific method for SWOT-based densification, the authors' model reduces root mean square error (RMSE) by 18–39% and provides more comprehensive spatial coverage.

This work provides a robust benchmark for spatiotemporal imputation in extremely sparse, large-scale settings. By demonstrating that sequence-based models can outperform traditional GNNs on river networks, the study offers a new paradigm for integrating heterogeneous satellite data. The AmazonWSE dataset and the accompanying model provide a scalable solution for densifying water level observations, which is critical for improving flood risk assessment and understanding global hydrological cycles in data-poor regions.