AmazonSWE: A Dataset and Model for River Water Surface Elevation Imputation

What changed

Researchers have introduced AmazonSWE, a new dataset and a corresponding imputation model for estimating water surface elevation (WSE) in river networks. The dataset focuses on the Amazon river basin, spanning 10 years (2016-2026) and covering over 19,000 river sections. It integrates satellite altimetry data, including from the SWOT sensor, and holds out in situ gauges for evaluation. A key challenge highlighted is the extreme sparsity of the data, with less than 1% of river sections observed daily, making it significantly sparser than existing imputation benchmarks. The river topology is also noted as being structurally different and larger than those in current datasets.

To address these challenges, a simple bidirectional selective state space model was proposed. This model outperforms prior spatiotemporal graph imputation methods by sampling connected subgraphs and flattening space and time into a single token sequence with topology-aware positional encodings. The proposed model reduces Root Mean Square Error (RMSE) against in situ gauges by 18-39% compared to the state-of-the-art method for SWOT-based WSE densification.

Why it matters for builders

This development provides AI builders with a challenging, real-world dataset and a specialized model for handling extremely sparse spatiotemporal graph data. This is crucial for applications requiring continuous monitoring in environments with limited sensor coverage, such as hydrology, climate science, and disaster management. The proposed model's architecture offers insights into effectively processing complex graph structures and temporal dependencies in data-scarce scenarios.

Practical impact

The AmazonSWE dataset and model have the potential to significantly improve the accuracy of flood forecasting and water resource management. By enabling more reliable WSE estimations across vast river networks, these tools can support better decision-making in critical environmental contexts. The model's ability to produce predictions for every river section, rather than being limited by satellite coverage, enhances the utility of satellite altimetry data.

Caveats and source limits

The provided source is a research paper preprint, and the findings are presented as preliminary. The dataset and model are specific to river WSE imputation in the Amazon basin, and their generalizability to other geographical regions or environmental variables would require further validation. The exact implementation details and performance benchmarks beyond the reported RMSE reduction are not fully elaborated in the excerpt.