A Dataset and Model for Imputing Water Surface Elevation on a Large and Extremely Sparse Spatiotemporal Graph
Ruben Cartuyvels, Karim Douch, Gabriele Bertoli, Mounia El Baz, Artemis Vrettou, Sébastien Lefèvre, Diego Fernandez Prieto
cs.LG
Quick verdict
Introduces 19K-node 99%-sparse Amazon WSE benchmark + bidirectional Mamba beating STGNNs and Reach-Reg (18–39% RMSE cut); strong, but single-seed protocol.
Summary
The paper introduces AmazonWSE, a benchmark for spatiotemporal imputation of daily water surface elevation (WSE) across 19,172 SWORD reaches of the Amazon basin from 2016 to 2026 (3,774 daily bins). It merges quality-filtered satellite altimetry from SWOT RiverSP, HydroWeb virtual stations, and ICESat-2 (about 1.9M observations; under 1% of reach-days observed), holding out 375 ANA in situ gauges purely for evaluation. The dataset is far larger and sparser than existing imputation benchmarks (Table 1) and has a directed acyclic river topology. The authors show that transductive and inductive STGNN baselines (GRIN, SPIN-H, IGNNK, KITS, ImputeFormer, plus kNN and LSTM controls) struggle at this scale and sparsity, and propose a bidirectional selective state space model (Mamba) that samples connected river subgraphs, flattens observed measurements into a single token sequence, and injects temporal, source, coordinate, and tree-positional metadata encodings. Ablations attribute the gains to subgraph context, metadata encodings, and multi-source tokens. Against Reach-Reg, a state-of-the-art SWOT densification method, the model reports 18–39% lower RMSE against in situ gauges (Table 5: SWOT-era RMSE 0.55 vs 0.90, KGE 0.94 vs 0.85, on 184/284 overlapping gauges) alongside much higher coverage. Limitations include a single-seed protocol, a post hoc decoding/de-normalization evaluation configuration, reduced baseline capacities, and code deferred to acceptance.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- AmazonWSE is a new imputation benchmark at 19K-node scale, 99% sparsity, and directed-acyclic river topology, with in situ gauges held out for evaluation.
- Existing transductive and inductive spatiotemporal imputation methods underperform at this scale and sparsity, while a bidirectional selective SSM that flattens observations into one sequence with topology-aware encodings works better.
- River topology improves predictions when exploited through subgraph sampling and metadata encodings rather than explicit graph convolutions.
- The model reduces RMSE against in situ gauges by 18–39% versus Reach-Reg while predicting every river reach, exceeding Reach-Reg's ~60–65% gauge coverage.
Strengths
Genuine benchmark gap
Table 1: 19.2K nodes with 99% missing versus prior imputation benchmarks capped at ~1,200 nodes and mostly ≤20% missing; a real directed-acyclic river topology (SWORD v17b, 19,172 reaches) unlike road/air-quality grids.
Disciplined evaluation design
375 ANA in situ gauges (955K observations) are never used as model inputs or training targets; multi-source inputs (SWOT RiverSP 353K, HydroWeb 391K, ICESat-2 208K observations; 0.97% observed, Table 2) with expert-standard filters detailed in A.2 (width ≥50 m, 15 km cross-track, wse_r_u >0.5 m, reach_q flags; ~50% of SWOT measurements removed and justified with raw-vs-filtered plots, Figures 9/11/13).
Simple, well-matched method with supporting ablations
Candid reporting
Weaknesses
No variance or significance reporting visible
Post hoc evaluation configuration
Baseline fairness risks
Reach-Reg comparison partly out of scope
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This preprint (arXiv:2609.11580v2; ESA Φ-lab and partners) contributes both a dataset and a model. **AmazonWSE** assembles daily water surface elevation (WSE) over the Amazon basin on the SWORD v17b river graph: 19,172 reaches × 3,774 daily bins (2016-01-01 to 2026-05-01). Inputs are quality-filtered satellite altimetry — SWOT RiverSP (353K obs, 2023–26), HydroWeb virtual stations (391K obs), ICESat-2 ATL22 (208K obs) — totalling ~1.9M observations with 0.97% of reach-days observed (Table 2). 375 ANA in situ gauges (955K obs) are held out exclusively for evaluation, in two tracks: SWOT-era reconstruction (2023-07 to 2026-05) and a pre-SWOT hindcast. Claimed contributions (Introduction): (1) a benchmark for extreme-sparsity imputation — Table 1 shows prior benchmarks top out near 1,200 nodes and far lower missingness, while AmazonWSE has 19.2K nodes at 99% missing on a directed acyclic topology; (2) a Mamba-based baseline; (3) evidence it beats STGNN imputation methods and the hydrology state-of-the-art Reach-Reg by 18–39% RMSE with full coverage.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the AmazonWSE dataset?
AmazonWSE is a benchmark dataset for spatiotemporal imputation of daily water surface elevation in the Amazon basin. It covers 19,172 river reaches (SWORD v17b) on a daily grid from January 2016 to May 2026 (3,774 days), combining quality-filtered satellite altimetry from SWOT RiverSP, HydroWeb virtual stations, and ICESat-2 — about 1.9 million observations, with fewer than 1% of reach-days observed (99% sparsity). 375 in situ gauges from Brazil's ANA are held out exclusively for evaluation and never used for training.
What model does the paper propose for water level imputation?
A bidirectional selective state space model (bidirectional Mamba-1, 3 layers, dimension 192). Instead of graph convolutions on a huge sparse graph, it samples connected river subgraphs (up to ~300 km, 91-day windows), creates tokens only for actually observed measurements, flattens space and time into a single token sequence ordered by day and river-flow rank, and adds month, day, satellite-source, coordinate, and tree-structured positional encodings that are re-injected at every layer.
How much better is the proposed model than previous methods?
The authors report 18–39% lower RMSE against in situ gauges than Reach-Reg, the state-of-the-art method for densifying SWOT-based water levels, while predicting every river reach (Reach-Reg covers only about 60–65% of gauges). On the SWOT-era overlap subset, RMSE is 0.55 vs 0.90 and KGE 0.94 vs 0.85. The model also outperforms prior graph-imputation methods (GRIN, SPIN-H, IGNNK, KITS, ImputeFormer) in the paper's experiments.
Is the AmazonWSE data or code available?
The paper states that all data export, processing, loading, and model code will be made public upon acceptance, via a GitHub repository (github.com/rubencart/AmazonWSE) that will include a tiny demo dataset for training and evaluation. The underlying raw measurements come from public providers (SWOT, HydroWeb, ICESat-2, ANA) and are released in h5netcdf format with quality flags under each provider's terms of use.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The main baseline-results table (Table 3) and Figure 5 are referenced in the visible text but their contents fall within the excerpt's [...sampling gap...] markers; margins over GRIN/SPIN-H/IGNNK/KITS/ImputeFormer could not be verified from the provided text.
- Table 4 (ablations) and parts of Related Work and dataset construction were skipped by sampling; only their in-text descriptions were available, so ablation judgments rely on the authors' narration rather than the table itself.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.