Hapi: A Multivariable Land-Surface Transformer for Medium-Range Hydrological Forecasting at Continental Scale
Hong Zhang, John K. Hutchison, Rao Kotamarthi, Jeremy Feinstein, Haiwen Guan, Haiwen Guan, Romit Maulik, Vijay P. Ramalingam, et al.
cs.AI
Quick verdict
Claims SOTA continental multivariable flood forecasting; rigorous stats, but headline wins use perfect reanalysis forcing—operational transfer unproven.
Summary
Hapi is a U-Net Swin Transformer V2 (43.96M parameters) that forecasts four land-surface variables—river discharge, surface runoff, snow water equivalent, and soil wetness—over the contiguous United States at 0.05° resolution for 24–72 h leads. Trained on 2005–2022 GloFAS reanalysis and ERA5-Land data, it balances heterogeneous targets through learnable Laplacian task weights and fine-tunes on randomized-depth rollouts for autoregressive stability. On a 2024 holdout, with GloFAS reanalysis as verification target, Hapi achieved the highest flood-detection F1 in 20 of 21 lead–threshold comparisons against persistence, operational GloFAS, and RiverMamba, and its advantage over RiverMamba was significant in all 21 cases under a Holm-corrected paired moving-block bootstrap. Per-cell median NSE favored Hapi at every lead (0.776 vs 0.559 for GloFAS at 72 h), although GloFAS led pooled RMSE, NSE, and KGE at 48 and 72 h. Independent validation at 3,881 USGS gauges showed higher median NSE (0.160 vs 0.076 at 24 h) but modest absolute skill. Ablations attribute rare-flood gains largely to learned task weighting, and a Hurricane Helene case study found both Hapi and GloFAS reproducing less than half the reference peak magnitudes. Main limitations: the primary evaluation relies on perfect reanalysis forcing, the test covers a single year, and code is not yet released. A 72-h CONUS forecast takes 0.11 s on one A100.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 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
- A single hierarchical Swin Transformer V2 model forecasts discharge, runoff, snow water equivalent, and soil wetness over CONUS at 0.05° for 24–72 h using 43.96M parameters.
- On the 2024 reanalysis-conditioned holdout, Hapi attains the highest flood-detection F1 in 20 of 21 lead–threshold comparisons against persistence, GloFAS, and RiverMamba, with all 21 advantages over RiverMamba significant after Holm-corrected paired block bootstrap.
- Hapi achieves higher median per-cell NSE than GloFAS at every lead, and the advantage persists against 3,881 independent USGS gauges (median NSE 0.160 vs 0.076 at 24 h).
- Learnable Laplacian task scales, not merely multivariable training, improve discharge accuracy (MAE 19% lower than uniform weighting) and rare-threshold flood detection under matched training budgets.
- A four-variable, 72-h CONUS forecast runs in 0.11 s with 1.69 GB peak memory on a single A100 GPU.
Strengths
Exemplary uncertainty quantification
Paired moving-block bootstrap (30-day blocks, R=10,000) resampling identical days for all models, Holm correction across the 21 lead–threshold cells per pair and metric, block-length sensitivity over 7–60 days (Appendix N), and explicit refusal to equate non-significance with equivalence; all 21 F1 advantages over RiverMamba survive correction.
Honest multi-scale reporting
The paper reports GloFAS's pooled RMSE/NSE/KGE wins at 48/72 h (Table 3) alongside Hapi's per-cell median NSE wins (Table 4: 0.776 vs 0.559 at 72 h, ahead at 79.2% of cells) and explains the aggregation effect rather than cherry-picking one scale.
Independent observational validation
Controlled ablation of the core methodological claim
Operationally relevant sensitivity analysis
Weaknesses
Reanalysis-conditioned headline claim
Asymmetric baseline footing
Modest absolute skill at gauges
Conditional snow conclusion
Single-year, single-seed evidence
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Hapi (arXiv:2609.22702, Argonne) is a U-Net Swin Transformer V2 (43.96M parameters; Conv3D patches T=2/p=2, three-stage encoder (1,2,4), d=192, 8×8 windows, axial RoPE, gated attention) forecasting discharge, surface runoff, snow water equivalent (SWE), and soil wetness over CONUS at 0.05° for 24–72 h. Its claims: (1) highest flood-detection F1 in 20 of 21 lead–threshold comparisons vs persistence, operational GloFAS, and RiverMamba on a 2024 holdout; (2) all 21 F1 advantages over RiverMamba significant after Holm-corrected paired block bootstrap; (3) highest median per-cell NSE at every lead, confirmed at 3,881 USGS gauges; (4) learned Laplacian task scales drive rare-flood gains; (5) 0.11 s per 72-h four-variable CONUS forecast on one A100. Claims (1)–(3) are well supported within the paper's stated reanalysis-conditioned setting; the abstract occasionally drops that qualifier.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 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 Hapi hydrological forecasting model?
Hapi is a U-Net Swin Transformer V2 developed at Argonne National Laboratory that forecasts river discharge, surface runoff, snow water equivalent, and soil wetness across the contiguous United States at 0.05° (about 5 km) resolution for 24–72 hour lead times, using roughly 44 million parameters.
Does Hapi outperform GloFAS?
In the paper's 2024 test, Hapi beat the operational GloFAS forecast on flood-detection F1 at every threshold at 24 h and on most 48–72 h comparisons, and on median per-cell NSE. However, GloFAS kept better pooled RMSE/NSE/KGE at 48–72 h, and Hapi received perfect reanalysis weather forcing while GloFAS used real forecast forcing, so the comparison is not fully operational.
How fast is Hapi at inference?
A 72-hour, four-variable forecast covering the contiguous US takes on average 0.11 seconds with 1.69 GB of peak GPU memory on a single NVIDIA A100. Training the final model took about 830 A100-hours on 16 GPUs.
Is Hapi's code and data available?
The input datasets (ERA5-Land, GloFAS reanalysis and return-level fields, USGS gauges) are publicly available, but at the time of this review the paper's code repository section was still a placeholder stating the identifier would be inserted before publication, so code and weights were not yet accessible.
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.
- Extraction truncated several passages (part of §3 gated attention, portions of §4.4–4.6, and parts of Appendices F, G, N); some judgments rest on surrounding prose rather than the omitted material.
- Figures (maps, hydrographs, architecture diagram, performance diagram) could not be analyzed from text; figure-dependent claims were avoided or noted from captions only.
- Metadata lists 8 authors while the text lists 9 (including Tom Wall) and shows an unused Purdue University affiliation — likely an extraction or metadata artifact.
- Table 7 (ablation) and Table 8 (non-discharge variables) values are partially cited from prose because the extracted table bodies were truncated.
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.