Pith. sign in

REVIEW 3 major objections 4 minor 48 references

This paper claims that keeping each station's local forecast as a stable anchor and adding only a state- and lead-dependent bounded correction drawn from neighboring observations improves prediction of sustained high-water plateaus without

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-04 21:05 UTC pith:JW2VI75K

load-bearing objection A solid, inventive anchored-forecasting paper whose main mechanism claim is well supported, but the episode-metric protocol and two weak graph baselines need work before the headline plateau gains can be fully trusted. the 3 major comments →

arxiv 2608.01775 v1 pith:JW2VI75K submitted 2026-08-03 cs.LG cs.AI

Multi-Source Dynamic Graph Learning for Compound-Flood Forecasting in Managed Coastal Systems

Liangjun You , Min Wu , Orlando Woods , Dongsheng Luo This is my paper
classification cs.LG cs.AI
keywords compound floodinghigh-water forecastingdynamic graph learningbounded residual correctionspatiotemporal forecastingmanaged coastal systemsepisode evaluationwater-level forecasting
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that flood-relevant high-water plateaus in a managed coastal system can be forecast better by treating a station's own recent history as an anchor and letting observations from other stations—rain, groundwater, neighboring water levels, pumps, and gates—add only a small, state-conditioned correction. The central claim is that this anchored decomposition, with a bounded residual, improves detection, onset timing, peak magnitude, and duration of sustained high-water episodes while keeping routine water-level accuracy at least as good as strong baselines. The authors argue that global errors like MAE and MSE hide how poorly models reproduce the shape of high-water events, so they evaluate episodes directly through temporal alignment. A sympathetic reader would care because the target phenomenon—prolonged high water—is the intermediate state that precedes flood impact and drives early-warning decisions in managed basins.

Core claim

The discovery is a forecasting architecture whose output is decomposed as a local anchor plus a correction. The anchor is produced only from the target station's surface-water history, while the correction is drawn from the full heterogeneous network of rain, groundwater, neighboring water levels, pumps, and gates. The correction is computed through a dynamic graph that re-weights which stations matter for each forecast window, a regime gate that decides how much correction the current hydrometeorological and operational state supports, and a lead-dependent budget that bounds the maximum departure. The bound guarantees that network information can revise but never replace the local forecast.

What carries the argument

The load-bearing identity is the anchored decomposition of each forecast into a local trajectory plus a bounded residual: the forecast equals the anchor A plus the correction C. The correction is defined as C = g * beta * tanh(d), where d is the graph decoder's raw proposal, g is a regime gate between 0 and 1 computed from multi-source summary statistics, and beta is a lead-dependent budget that increases with forecast distance. The identity guarantees the magnitude of C is at most beta, so cross-station information is strictly confined to a controlled departure. The dynamic adjacency is built from node states typed by station kind and geographic distance, so the graph's structure is re-deri

Load-bearing premise

The load-bearing premise is that the episode evaluation protocol—training-derived percentiles, three-exceedance minimum, six-hour merge, 24-hour issue stride, and greedy one-to-one overlap matching—is a fair measure of sustained high-water fidelity; if those construction choices interact with the model's smooth bounded corrections, the headline plateau-reliability gain could be an artifact of the metric.

What would settle it

Re-run the S7 three-day evaluation with the episode construction changed—merge gap at 0 versus 12 hours, minimum duration at 1 versus 6 hours, issue stride at 12 versus 48 hours, and matching by largest overlap versus earliest onset—and check whether the proposed model's Episode F1 and Duration MAE advantage over the patching-transformer baseline persists at every setting. If any setting flips the ranking, the plateau-reliability claim is tied to the evaluation protocol rather than to the forecast itself.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, flood early-warning systems can use cross-station data without sacrificing local forecast stability, because the anchor keeps the baseline trajectory explicit and the correction is bounded.
  • The separation between anchor and correction makes the contribution of distributed observations inspectable: operators can see when a predicted departure is supported by the wider network and when it is purely local.
  • The method is transportable to any managed basin with heterogeneous gage types (rain, groundwater, levels, control operations) and needs no event labels during training, since only observed water levels are used in the loss.
  • Because the bound grows with lead time, the model is designed to be more conservative about cross-site departures at short horizons and more willing to revise at long horizons, matching the uncertainty profile of multi-day forecasts.
  • Adopting episode-level evaluation alongside global errors would give a more honest view of high-water fidelity for operational models than MSE alone.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The episode protocol's merge gap, minimum duration, and matching rule are untested; if a six-hour merge gap or greedy matching interacts with smooth bounded corrections, the plateau-reliability advantage could be partly a metric artifact rather than a hydrological gain.
  • The same anchored-correction principle could be tested on river networks with reservoir releases as the 'management' signal, where the gate-and-decode split would make human operations visible as an explicit correction term.
  • The bounded residual can be read as an uncertainty statement: the lead-dependent budget is a worst-case envelope, so the model could be extended to output interval forecasts by treating the budget as a calibration parameter.
  • A testable extension with direct bearing on the claim: train the same architecture with the correction bound removed and compare the distribution of predicted episode durations; if durations inflate without the bound, the hypothesis that bounded corrections preserve plateau shape gains support.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. This paper proposes an anchored dynamic-graph forecaster for hourly surface-water stage forecasting in the managed South Florida coastal system. The forecast at each issue time is decomposed into a local temporal anchor A_t (a channel-wise PatchTST operating on the target WATER history only) and a network-supported bounded correction C_t (Eq. 1, Eq. 9). The correction uses a GRU over heterogeneous stations, a state-dependent learned directed graph, a multi-source regime gate, and a lead-dependent trust-region budget. Training uses masked MSE only; high-water episodes are defined post hoc from training-period quantile thresholds and matched one-to-one within forecast windows. The paper reports lower full-record MAE/MSE, higher Episode F1, and lower onset/peak/duration errors than eight baselines across four horizons, three chronological splits, and three spatial partitions, and supports the mechanism claims with on/off ablations and threshold-q sensitivity.

Significance. If the comparative ranking is robust, the paper makes a useful architectural and evaluation contribution to hydrologic forecasting in managed coastal systems. The explicit information boundary between a local temporal reference and a bounded, regime-gated, cross-station correction is a clean way to incorporate distributed observations without destabilizing local persistence. The paper's positive features include publicly released data/code, a derivation of the trust-region guarantee (Eq. 10), a training objective that does not use event labels, and evaluation that goes beyond global error metrics to event-scale onset, peak, and duration. The principal fragility is that the central 'sustained high-water plateau' claim is measured only through an episode-construction protocol whose non-threshold components are fixed without sensitivity analysis; the reported gains could therefore in part reflect the protocol rather than genuine process fidelity.

major comments (3)
  1. [§4.2, Appendix B] The headline claim about sustained high-water plateau forecasting is evaluated only through the episode protocol of §4.2/Appendix B: training-period quantile thresholds (Eq. 11), minimum 3 exceedance hours, 6-hour merge gap, 24-hour issue stride, and greedy one-to-one overlap matching. The paper varies only the threshold q (Figure 3, Table 11); none of the other construction choices is sensitivity-tested. This is load-bearing because the proposed correction (Eq. 9) is a smooth, bounded, regime-gated residual that changes threshold-crossing behavior and temporal smoothness relative to the baselines. A forecast hovering just above the threshold can produce longer merged episodes or match more easily, inflating Duration MAE and Episode F1 without a real improvement in plateau fidelity. Please add a sensitivity analysis over reasonable alternatives for the minimum duration, merge gap, issue
  2. [Table 7 / Appendix D] Table 7 lists "Random seeds 3", and all main tables report mean ± standard deviation, but Appendix D explicitly states that the standard deviations summarize variation over the three official spatial partitions, not variation from random model initialization. No seed-level statistics are reported. This makes the uncertainty estimates and the phrase "Random seeds 3" misleading. Please report seed-level means/standard deviations for the main comparisons (or at least verify that the rankings are stable across seeds), and clarify in the table captions that the reported intervals are over spatial partitions only. With only three spatial partitions and no seed variance, the claim of consistent wins is not yet statistically supported.
  3. [Tables 8–10 vs. §5.2] Section 5.2 states that the proposed model achieves the lowest Onset MAE throughout the evaluated horizons, but the split-level detail tables contain clear counterexamples. In Table 8 (S5), Graph WaveNet has Onset MAE 0.962 h at 3D versus 1.491 h for Ours, and 3.093 h at 7D versus 3.431 h; in Table 9 (S6), FourierGNN has 3.707 h at 7D versus 4.252 h; in Table 10 (S7), MTGNN has 2.996 h at 7D versus 4.680 h. Similar reversals occur for Duration MAE (e.g., Table 8 3D: PatchTST 8.667 h vs. Ours 9.328 h). Because the paper's central claim is a consistency result over splits, the main text should either report split-level results alongside the aggregate, or qualify the claim to make clear that the lowest-on-average result does not hold at every chronological split. Please address this discrepancy explicitly.
minor comments (4)
  1. [Algorithm 1 / Eq. (9)] Algorithm 1 describes the correction as computing (alpha_t,i, b_t,i) and then clipping b_t,i to [-alpha_t,i, alpha_t,i], while §3.2 defines the correction as C_t,i,l = g_t,i,l * beta_l * tanh(d_t,i,l). The relationship between alpha_t,i and (g_t,i,l, beta_l, tanh) is not stated. Please align the algorithm's notation with the equations.
  2. [Figure 4] The y-axis of Figure 4 is described only as "relative performance degradation". Please define the metric (e.g., percentage loss relative to the full model for each of the six metrics) in the caption or in the text.
  3. [Table 2] The row for Ours in Table 2 shows eight numbers after the first four F1 entries; the caption is clear, but a visual separator or column grouping would reduce the risk of misreading the F1 and Onset blocks.
  4. [Table 10] The S7 MSE values for MTGNN and Graph WaveNet are much larger than their S5/S6 values (e.g., MTGNN 1D MSE 15.993×10^-2 vs. 1.994×10^-2 in S5). This suggests possible pathological forecasts in one partition; please add a sentence explaining whether these are due to a known data issue or a genuine model failure, since otherwise readers may question the baseline setup.

Circularity Check

0 steps flagged

No significant circularity: the high-water improvements are emergent from MSE training and evaluated on held-out data.

full rationale

The paper's derivation chain is self-contained. The training objective (Algorithm 1, line 11) is masked MSE only; event labels and episode metrics are explicitly excluded from training ('Event labels and episode metrics do not enter training' and 'Event labels and episode metrics are used only after forecasts are issued, for evaluation'). The anchored decomposition Y_t = A_t + C_t (Eq. 1) and the bounded correction (Eq. 9) are modeling definitions, not fitted targets. The headline claim about sustained high-water plateaus is evaluated on held-out test splits with thresholds estimated only from training-period quantiles (Eq. 11), and the model is compared against eight external baselines. The only author-overlapping citation is SF2Bench [42], co-authored by Dongsheng Luo, but it is a public benchmark with fixed chronological and spatial splits, publicly available data and code, and no event labels in training; it therefore provides independent, reproducible evidence rather than load-bearing self-citation. The skeptical concern that episode-protocol construction choices (minimum duration, merge gap, issue stride, one-to-one matching) are not sensitivity-tested is a validity or robustness issue, not circularity: no construction rule is equivalent to the model's output or objective, and the paper explicitly acknowledges the high-water definition is a reproducible proxy rather than a regulatory flood measure. No circular step can be exhibited from the paper's own equations or citations.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central empirical claim is benchmark-based rather than derived, so the ledger is dominated by evaluation-definition choices and dataset assumptions rather than physical axioms. The most load-bearing items are the author-defined episode protocol (which defines the plateau-reliability target) and the single-10-epoch training schedule assumption. No invented entities; the only hand-fitted constants are trust-region and hyperparameter choices.

free parameters (5)
  • trust-region budget endpoints beta_min, beta_max
    Chosen by hand; bound the maximum allowed graph correction |C_{t,i,l}| <= beta_l (Eqs. 9-10). Values are not stated in the text, and no sensitivity analysis is reported beyond the on/off ablation (Fig. 4).
  • dynamic neighborhood size K = 20
    Hand-chosen; selects the top-K non-self sources in Eq. (6). Not swept.
  • regime statistic window r = 24 h
    Hand-chosen; the mean-absolute regime statistic mu_{t,c} is computed over the most recent r hours. Not swept.
  • episode construction parameters = q=0.95, min duration 3 h, merge gap 6 h, issue stride 24 h
    Define the evaluation target metric. q is swept (Fig. 3, Table 11); the other three are fixed without sensitivity checks. These choices directly shape Episode F1, Onset, Peak, and Duration MAE.
  • lookback L and training schedule = L=48 h, epochs=10, lr=1e-3, batch=64
    Common configuration applied to all models (Table 7); the 10-epoch schedule may disadvantage complex baselines (see red flags).
axioms (4)
  • domain assumption Sustained high surface-water stage is a flood-relevant intermediate state consistent with compound-flood hazard when accompanied by coastal or hydrologic forcing
    Section 1 frames local sustained high stage as the 'flood-relevant intermediate hydrologic state' the whole evaluation targets; Section 7 concedes it is not a regulatory flood stage, inundation extent, or damage measure.
  • ad hoc to paper Training-period quantile exceedance runs (merged with a 6 h gap, min 3 h) capture the high-water plateaus relevant to early warning
    Section 4.2 and Appendix B define episodes exclusively this way; the plateau-reliability claim is evaluated only through this protocol, and matching parameters are not sensitivity-tested.
  • domain assumption SF2Bench station coverage and observation masks provide a sufficiently complete, unbiased picture of the managed system within each split and partition
    Appendix B describes evolving coverage (Table 4) and partitions with zero PUMP stations (Table 5); the model must work with missing types, and systematic sensor bias would propagate into the learned corrections.
  • standard math AdamW training with 10 epochs and validation-MSE checkpoint selection yields converged, comparable models across architectures
    Section 4 and Table 7 set one schedule for all models; the extreme MTGNN and Graph WaveNet errors suggest this axiom plausibly fails for those baselines.

reviewed 2026-08-04 · how reviews work

0 comments
read the original abstract

Compound flooding in managed coastal systems is influenced by hydrological conditions and water-management activity observed across multiple monitoring stations. Current forecasting models can capture temporal dependencies with low average errors, but global error metrics may conceal poor reproduction of prolonged high-water plateaus that are relevant to flood early warning. Because hydrometeorological and operational signals are distributed across heterogeneous gages, single-site records do not fully represent high-water dynamics. Nevertheless, unconstrained fusion of cross-site signals can degrade the stability of local temporal forecasts. This work proposes an anchored forecasting framework that incorporates cross-site information through state- and lead-dependent bounded residual corrections. A multi-source regime representation constructed from hydrometeorological and operational observations adaptively calibrates inter-site relationships and correction scales, enabling targeted cross-site adjustment while preserving the local temporal forecast as a stable anchor. Beyond conventional global error statistics, we evaluate event-scale high-water characteristics through the temporal alignment of forecasted and observed high-water processes. Experiments demonstrate that selectively integrating multi-station dynamic conditions improves the prediction reliability of sustained high-water plateaus while maintaining high accuracy during routine hydrological conditions, supporting flood early warning and water-management decision support.

Figures

Figures reproduced from arXiv: 2608.01775 by Dongsheng Luo, Liangjun You, Min Wu, Orlando Woods.

Figure 1
Figure 1. Figure 1: Distributed observations across a high-water [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Managed coastal monitoring and water-control system in South Florida, where rainfall, groundwater, surface water, [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sensitivity of episode-level performance to the station-specific high-water threshold for the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Relative performance degradation under architecture and observation-source ablations for the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 19 canonical work pages · 2 internal anchors

  1. [1]

    Lei Bai, Lina Yao, Can Li, Xianzhi Wang, and Can Wang. 2020. Adaptive Graph Convolutional Recurrent Network for Traffic Forecasting. InAdvances in Neural Information Processing Systems, Vol. 33. 17804–17815. arXiv:2007.02842 [cs.LG] https://arxiv.org/abs/2007.02842

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  2. [2]

    Vousdoukas, Giuseppe Zappa, Kevin Hodges, Theodore G

    Emanuele Bevacqua, Michalis I. Vousdoukas, Giuseppe Zappa, Kevin Hodges, Theodore G. Shepherd, Douglas Maraun, Lorenzo Mentaschi, and Luc Feyen

    work page

  3. [3]

    Frame, Frederik Kratzert, Daniel Klotz, Martin Gauch, Guy Shalev, Oren Gilon, Logan M

    Jonathan M. Frame, Frederik Kratzert, Daniel Klotz, Martin Gauch, Guy Shalev, Oren Gilon, Logan M. Qualls, Hoshin V. Gupta, and Grey S. Nearing. 2022. Deep Learning Rainfall–Runoff Predictions of Extreme Events.Hydrology and Earth System Sciences26, 13 (2022), 3377–3392. doi:10.5194/hess-26-3377-2022

    work page doi:10.5194/hess-26-3377-2022 2022

  4. [4]

    Shengnan Guo, Youfang Lin, Ning Feng, Chao Song, and Huaiyu Wan. 2019. Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting.Proceedings of the AAAI Conference on Artificial Intelligence33, 1 (2019), 922–929. doi:10.1609/aaai.v33i01.3301922

    doi:10.1609/aaai.v33i01.3301922 2019

  5. [5]

    Gupta, Harald Kling, Koray K

    Hoshin V. Gupta, Harald Kling, Koray K. Yilmaz, and Guillermo F. Martinez

    work page

  6. [6]

    Yukiko Hirabayashi, Roobavannan Mahendran, Sujan Koirala, Lisako Konoshima, Dai Yamazaki, Satoshi Watanabe, Hyungjun Kim, and Shinjiro Kanae. 2013. Global Flood Risk under Climate Change.Nature Climate Change3, 9 (2013), 816–821. doi:10.1038/nclimate1911

    doi:10.1038/nclimate1911 2013

  7. [7]

    Robert Jane, Luis Cadavid, Jayantha Obeysekera, and Thomas Wahl. 2020. Multi- variate Statistical Modelling of the Drivers of Compound Flood Events in South Florida.Natural Hazards and Earth System Sciences20, 10 (2020), 2681–2699. doi:10.5194/nhess-20-2681-2020

    doi:10.5194/nhess-20-2681-2020 2020

  8. [8]

    Sheo Yon Jhin, Seojin Kim, and Noseong Park. 2024. Addressing Prediction Delays in Time Series Forecasting: A Continuous GRU Approach with Derivative Regularization. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1234–1245. doi:10.1145/3637528.3671969

    arXiv 2024

  9. [9]

    Faghmous, Michael Steinbach, Arindam Banerjee, Auroop Ganguly, Shashi Shekhar, Nagiza Samatova, and Vipin Kumar

    Anuj Karpatne, Gowtham Atluri, James H. Faghmous, Michael Steinbach, Arindam Banerjee, Auroop Ganguly, Shashi Shekhar, Nagiza Samatova, and Vipin Kumar. 2017. Theory-Guided Data Science: A New Paradigm for Scientific Discovery from Data.IEEE Transactions on Knowledge and Data Engineering29, 10 (2017), 2318–2331. doi:10.1109/TKDE.2017.2720168

    work page arXiv 2017

  10. [10]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. InInternational Conference on Learning Represen- tations. arXiv:1609.02907 [cs.LG] https://openreview.net/forum?id=SJU4ayYgl

    Pith/arXiv arXiv 2017

  11. [11]

    Nikolas Kirschstein and Yixuan Sun. 2024. The Merit of River Network Topology for Neural Flood Forecasting. InProceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235). 24713– 24725. https://proceedings.mlr.press/v235/kirschstein24a.html

    work page 2024

  12. [12]

    Daniel Klotz, Frederik Kratzert, Martin Gauch, Alden Keefe Sampson, Johannes Brandstetter, Günter Klambauer, Sepp Hochreiter, and Grey Nearing. 2022. Uncer- tainty Estimation with Deep Learning for Rainfall–Runoff Modeling.Hydrology and Earth System Sciences26, 7 (2022), 1673–1693. doi:10.5194/hess-26-1673-2022

    work page doi:10.5194/hess-26-1673-2022 2022

  13. [13]

    Frederik Kratzert, Daniel Klotz, Claire Brenner, Karsten Schulz, and Mathew Herrnegger. 2018. Rainfall–Runoff Modelling Using Long Short-Term Memory (LSTM) Networks.Hydrology and Earth System Sciences22, 11 (2018), 6005–6022. doi:10.5194/hess-22-6005-2018

    doi:10.5194/hess-22-6005-2018 2018

  14. [14]

    Sampson, Sepp Hochreiter, and Grey S

    Frederik Kratzert, Daniel Klotz, Mathew Herrnegger, Alden K. Sampson, Sepp Hochreiter, and Grey S. Nearing. 2019. Toward Improved Predictions in Ungauged Basins: Exploiting the Power of Machine Learning.Water Resources Research55, 12 (2019), 11344–11354. doi:10.1029/2019WR026065

    doi:10.1029/2019wr026065 2019

  15. [15]

    Guokun Lai, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu. 2018. Modeling Long- and Short-Term Temporal Patterns with Deep Neural Networks. InThe 41st International ACM SIGIR Conference on Research and Development in Information Retrieval. 95–104. doi:10.1145/3209978.3210006

    arXiv 2018

  16. [16]

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2018. Diffusion Convo- lutional Recurrent Neural Network: Data-Driven Traffic Forecasting. InIn- ternational Conference on Learning Representations. arXiv:1707.01926 [cs.LG] https://openreview.net/forum?id=SJiHXGWAZ

    Pith/arXiv arXiv 2018

  17. [18]

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. 2024. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting. InInternational Conference on Learning Representations. arXiv:2310.06625 [cs.LG] https://openreview.net/forum?id=JePfAI8fah

    Pith/arXiv arXiv 2024

  18. [19]

    Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. 2024. AutoTimes: Autoregressive Time Series Forecasters via Large Language Models. InAdvances in Neural Information Processing Systems, Vol. 37. 122154–122184. arXiv:2402.02370 [cs.LG] https://openreview.net/forum?id=FOvZztnp1H

    Pith/arXiv arXiv 2024

  19. [20]

    Amir Mosavi, Pinar Ozturk, and Kwok-Wing Chau. 2018. Flood Prediction Using Machine Learning Models: Literature Review.Water10, 11 (2018), 1536. doi:10.3390/w10111536

    work page doi:10.3390/w10111536 2018

  20. [21]

    Grey Nearing, Deborah Cohen, Vusumuzi Dube, Martin Gauch, Oren Gilon, Shaun Harrigan, Avinatan Hassidim, Daniel Klotz, Frederik Kratzert, Asher Metzger, Sella Nevo, Florian Pappenberger, Christel Prudhomme, Guy Shalev, Shlomo Shenzis, Tadele Yednkachw Tekalign, Dana Weitzner, and Yossi Matias. 2024. Global Prediction of Extreme Floods in Ungauged Watershe...

    doi:10.1038/s41586-024-07145-1 2024

  21. [22]

    Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

    Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. A Time Series Is Worth 64 Words: Long-Term Forecasting with Transformers. In International Conference on Learning Representations. arXiv:2211.14730 [cs.LG] https://openreview.net/forum?id=Jbdc0vTOcol

    Pith/arXiv arXiv 2023

  22. [23]

    Jayantha Obeysekera, Michelle Irizarry, Joseph Park, Jenifer Barnes, and Tibebe Dessalegne. 2011. Climate Change and Its Implications for Water Resources Man- agement in South Florida.Stochastic Environmental Research and Risk Assessment 25, 4 (2011), 495–516. doi:10.1007/s00477-010-0418-8

    work page doi:10.1007/s00477-010-0418-8 2011

  23. [24]

    Oreshkin, Dmitri Carpov, Nicolas Chapados, and Yoshua Bengio

    Boris N. Oreshkin, Dmitri Carpov, Nicolas Chapados, and Yoshua Bengio

    work page

  24. [25]

    Read, Xiaowei Jia, Jared Willard, Alison P

    Jordan S. Read, Xiaowei Jia, Jared Willard, Alison P. Appling, Jacob A. Zwart, Samantha K. Oliver, Anuj Karpatne, Gretchen J. A. Hansen, Paul C. Hanson, William Watkins, Michael Steinbach, and Vipin Kumar. 2019. Process-Guided Deep Learning Predictions of Lake Water Temperature.Water Resources Research 55, 11 (2019), 9173–9190. doi:10.1029/2019WR024922

    work page doi:10.1029/2019wr024922 2019

  25. [26]

    David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. 2020. DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks. International Journal of Forecasting36, 3 (2020), 1181–1191. doi:10.1016/j.ijforecast. 2019.07.001

    doi:10.1016/j.ijforecast 2020

  26. [27]

    InInternational Conference on Learning Representations

    N-BEATS: Neural Basis Expansion Analysis for Interpretable Time Series Forecasting. InInternational Conference on Learning Representations. arXiv:1905.10437 [cs.LG] https://openreview.net/forum?id=r1ecqn4YwB

    Pith/arXiv arXiv 1905

  27. [28]

    Chaopeng Shen. 2018. A Transdisciplinary Review of Deep Learning Research and Its Relevance for Water Resources Scientists.Water Resources Research54, 11 (2018), 8558–8593. doi:10.1029/2018WR022643

    work page doi:10.1029/2018wr022643 2018

  28. [29]

    Shuman, Sunil K

    David I. Shuman, Sunil K. Narang, Pascal Frossard, Antonio Ortega, and Pierre Vandergheynst. 2013. The Emerging Field of Signal Processing on Graphs: Extend- ing High-Dimensional Data Analysis to Networks and Other Irregular Domains. IEEE Signal Processing Magazine30, 3 (2013), 83–98. doi:10.1109/MSP.2012.2235192

    work page arXiv 2013

  29. [30]

    Laio Oriel Seman, Stefano Frizzo Stefenon, Kin-Choong Yow, Leandro dos San- tos Coelho, and Viviana Cocco Mariani. 2026. Fourier-Enhanced Sequence- to-Sequence Latent Graph Neural Networks for Multi-Node Spatiotemporal Forecasting in a Hydroelectric Reservoir.Engineering Applications of Artificial Intelligence167 (2026), 113939. doi:10.1016/j.engappai.2026.113939

    work page arXiv 2026

  30. [31]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. InInternational Conference on Learning Representations. arXiv:1710.10903 [stat.ML] https: //openreview.net/forum?id=rJXMpikCZ

    Pith/arXiv arXiv 2018

  31. [32]

    Meyers, and Mark E

    Thomas Wahl, Shaleen Jain, Jens Bender, Steven D. Meyers, and Mark E. Luther

    work page

  32. [33]

    Chao Song, Youfang Lin, Shengnan Guo, and Huaiyu Wan. 2020. Spatial-Temporal Synchronous Graph Convolutional Networks: A New Framework for Spatial- Temporal Network Data Forecasting.Proceedings of the AAAI Conference on Artificial Intelligence34, 1 (2020), 914–921. doi:10.1609/aaai.v34i01.5438

    doi:10.1609/aaai.v34i01.5438 2020

  33. [34]

    Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

  34. [35]

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Ming- sheng Long. 2023. TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis. InInternational Conference on Learning Representations. arXiv:2210.02186 [cs.LG] https://openreview.net/forum?id=ju_Uqw384Oq

    Pith/arXiv arXiv 2023

  35. [36]

    Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, Xiaojun Chang, and Chengqi Zhang. 2020. Connecting the Dots: Multivariate Time Series Forecasting with Graph Neural Networks. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 753–763. doi:10.1145/3394486. 3403118

    doi:10.1145/3394486 2020

  36. [37]

    Ward, Anaïs Couasnon, Dirk Eilander, Ivan D

    Philip J. Ward, Anaïs Couasnon, Dirk Eilander, Ivan D. Haigh, Alistair Hendry, Sanne Muis, Ted I. E. Veldkamp, Hessel C. Winsemius, and Thomas Wahl. 2018. Dependence between High Sea-Level and High River Discharge Increases Flood Hazard in Global Deltas and Estuaries.Environmental Research Letters13, 8 (2018), 084012. doi:10.1088/1748-9326/aad400

    work page doi:10.1088/1748-9326/aad400 2018

  37. [38]

    Kun Yi, Qi Zhang, Wei Fan, Hui He, Liang Hu, Pengyang Wang, Ning An, Longbing Cao, and Zhendong Niu. 2023. FourierGNN: Rethinking Multivariate Time Series Forecasting from a Pure Graph Perspective. In Advances in Neural Information Processing Systems, Vol. 36. 69638–69660. arXiv:2311.06190 [cs.LG] https://proceedings.neurips.cc/paper_files/paper/2023/ has...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  38. [39]

    Bing Yu, Haoteng Yin, and Zhanxing Zhu. 2018. Spatio-Temporal Graph Con- volutional Networks: A Deep Learning Framework for Traffic Forecasting. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence. 3634–3640. doi:10.24963/ijcai.2018/505

    doi:10.24963/ijcai.2018/505 2018

  39. [40]

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. Are Transformers Effective for Time Series Forecasting?Proceedings of the AAAI Conference on Artificial Intelligence37, 9 (2023), 11121–11128. doi:10.1609/aaai.v37i9.26317

    doi:10.1609/aaai.v37i9.26317 2023

  40. [41]

    Chuanpan Zheng, Xiaoliang Fan, Cheng Wang, and Jianzhong Qi. 2020. GMAN: A Graph Multi-Attention Network for Traffic Prediction.Proceedings of the AAAI Conference on Artificial Intelligence34, 1 (2020), 1234–1241. doi:10.1609/aaai. v34i01.5477

    doi:10.1609/aaai 2020

  41. [42]

    Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. 2019. Graph WaveNet for Deep Spatial-Temporal Graph Modeling. InProceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence. 1907–1913. doi:10.24963/ijcai.2019/264 KDD ’27, August 1–5, 2027, San Jose, CA, USA Liangjun You, Min Wu, Orlando Woods, and Dongsheng Luo

    doi:10.24963/ijcai.2019/264 2019

  42. [43]

    Haoyi Zhou, Shang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting.Proceedings of the AAAI Conference on Artificial Intelli- gence35, 12 (2021), 11106–11115. doi:10.1609/aaai.v35i12.17325

    doi:10.1609/aaai.v35i12.17325 2021

  43. [44]

    Horton, Bart van den Hurk, Amir AghaKouchak, Aglaé Jézéquel, Miguel D

    Jakob Zscheischler, Olivia Martius, Seth Westra, Emanuele Bevacqua, Colin Raymond, Radley M. Horton, Bart van den Hurk, Amir AghaKouchak, Aglaé Jézéquel, Miguel D. Mahecha, Douglas Maraun, Alexandre M. Ramos, Nina N. Ridder, Wim Thiery, and Edoardo Vignotto. 2020. A Typology of Compound Weather and Climate Events.Nature Reviews Earth & Environment1, 7 (20...

    doi:10.1038/s43017-020-0060-z 2020

  44. [47]

    Xu Zheng, Chaohao Lin, Sipeng Chen, Zhuomin Chen, Jimeng Shi, Jayantha Obey- sekera, Jingchao Ni, Wei Cheng, Jason Liu, and Dongsheng Luo. 2026. Uncovering Insights of Compound Flooding with Data-Driven AI. Accepted to the SIGKDD 2026 AI for Science Track. arXiv:2506.04281 [cs.LG]. arXiv:2506.04281 [cs.LG] https://arxiv.org/abs/2506.04281

    Pith/arXiv arXiv 2026

  45. [2009]

    doi:10.1016/j.jhydrol.2009.08.003

    Decomposition of the Mean Squared Error and NSE Performance Criteria: Implications for Improving Hydrological Modelling.Journal of Hydrology377, 1–2 (2009), 80–91. doi:10.1016/j.jhydrol.2009.08.003

    doi:10.1016/j.jhydrol.2009.08.003 2009

  46. [2015]

    doi:10.1038/ nclimate2736

    Increasing Risk of Compound Flooding from Storm Surge and Rainfall for Major US Cities.Nature Climate Change5, 12 (2015), 1093–1097. doi:10.1038/ nclimate2736

    work page 2015

  47. [2020]

    doi:10.1038/s43247-020-00044-z

    More Meteorological Events That Drive Compound Coastal Flooding Are Projected under Climate Change.Communications Earth & Environment1, 1 (2020), 47. doi:10.1038/s43247-020-00044-z

    work page doi:10.1038/s43247-020-00044-z 2020

  48. [2023]

    Surveys55, 4, Article 66 (2023), 37 pages

    Integrating Scientific Knowledge with Machine Learning for Engineering and Environmental Systems.Comput. Surveys55, 4, Article 66 (2023), 37 pages. doi:10.1145/3514228

    doi:10.1145/3514228 2023

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.