AmazonWSE Dataset and Model for River Water Elevation Imputation

What changed

Researchers have introduced the AmazonWSE dataset, designed for training and evaluating large-scale spatiotemporal graph imputation methods for water surface elevation (WSE). This dataset integrates processed satellite altimetry measurements, including data from the SWOT sensor, and covers over 19,000 river sections in the Amazon river basin spanning 10 years (2016-2026). A key challenge addressed is the extreme sparsity of observations, with fewer than 1% of river sections observed daily, making it significantly sparser than existing imputation benchmarks. The dataset also features a directed acyclic river topology that is larger and structurally different from those in current datasets.

To tackle these challenges, a simple bidirectional selective state space model is proposed. This model outperforms prior spatiotemporal graph imputation methods by sampling connected subgraphs and flattening space and time into a single token sequence, enhanced with topology-aware positional encodings. The model demonstrates significant improvements over the state-of-the-art method for SWOT-based WSE densification, reducing RMSE against in-situ gauges by 18-39% while providing predictions for all river sections, not just those with sufficient nearby satellite coverage.

Why it matters for builders

The AmazonWSE dataset and the accompanying model offer a new benchmark and a potential solution for a critical problem in hydrology and environmental monitoring: the lack of continuous WSE data. Builders working on flood forecasting, water resource management, or global water cycle research can use this dataset to train and test imputation models. The proposed sequence model, which handles extreme sparsity and complex river topologies effectively, provides a strong baseline for developing more robust WSE estimation systems.

Practical impact

Developers can explore the AmazonWSE dataset to build and refine WSE imputation models, particularly for large and complex river systems where data is scarce. The proposed model's ability to outperform existing methods by a significant margin (18-39% RMSE reduction) suggests its potential for practical application in real-time monitoring and forecasting systems. The approach of flattening space and time into a single token sequence with topology-aware encodings offers a novel direction for handling sparse spatiotemporal graph data.

Caveats and source limits

The source material describes a research paper and its contributions. Specific details regarding the availability of the AmazonWSE dataset and the code for the proposed model are not explicitly provided in the excerpt. The performance improvements are reported against a specific state-of-the-art method (Reach-Reg) and benchmarked against in-situ gauges, but broader independent validation across diverse geographical regions or different types of river systems is not detailed. The exact computational requirements or scalability for extremely large-scale real-time applications are also not elaborated upon.