Pith. sign in

REVIEW 3 major objections 5 minor 53 references

HydroGAT: Distributed Heterogeneous Graph Attention Transformer for Spatiotemporal Flood Prediction

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A graph that puts every land and river pixel into a node, with flow and catchment edges, predicts hourly discharge in two Midwest basins at Nash-Sutcliffe efficiency up to 0.97 and Kling-Gupta efficiency up to 0.96.

desk verdict Solid new GNN architecture for pixel-level flood forecasting, but the reported gains are undercut by a perfect-foresight rainfall input that baselines may not have gotten, and the KGE 'improvement' in the abstract contradicts the paper's own Table 2. read the letter →

arxiv 2509.02481 v1 pith:4X7RZ32C submitted 2025-09-02 cs.LG cs.DC

Aishwarya Sarkar , Autrin Hakimi , Xiaoqiong Chen , Hai Huang , Chaoqun Lu , Ibrahim Demir , Ali Jannesari This is my paper · ORCID
classification cs.LGcs.DC
keywords floodforecastinggraphneuralnetworksspatiotemporalpredictionriverdischargeattentionmechanismdistributedtrainingheterogeneousgraphshydrologicalrouting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a flood-forecasting model can keep every 4 km land and river pixel as a graph node—no coarse catchment polygons—and still train efficiently, and that this pixel-level topology improves hourly discharge prediction. HydroGAT couples a transformer that reads each node's rainfall and discharge history with two attention-based routing branches, one following steepest-descent flow directions and one following upstream–downstream catchment links, fused by a learnable weight. On the Cedar and Des Moines river basins over a 72-hour forecast horizon, it reports a Nash–Sutcliffe efficiency (NSE, a standard accuracy score where 1 is perfect) up to 0.97, a Kling–Gupta efficiency up to 0.96, and percent bias within ±5%, matching or exceeding five baseline architectures. The same attention weights double as interpretability: they show which upstream sub-basins and which past hours drive a forecast.

What carries the argument

The central object is the heterogeneous basin graph G=(V,E_F,E_C), where V is every 4 km land or river pixel, E_F is one outgoing steepest-descent flow edge per pixel, and E_C links upstream and downstream target stations across catchments. The carrying mechanism is gated attention-based message passing: a temporal transformer produces per-node embeddings; two GAT-based GRU branches route those embeddings along the two edge types; a learnable scalar α fuses the branches at target nodes; and a convolutional predictor combines the fused hidden state with forecasted rainfall. Graph attention is what turns static topography into dynamic, context-sensitive routing weights, and those weights are a

What would settle it

Take one basin and replace the attention-derived edge weights on catchment edges with randomized weights drawn from the same distribution; if the model's NSE barely changes, the learned influential-upstream mechanism is not carrying the claimed performance. Alternatively, recompute the flow edges from the original high-resolution elevation model instead of the 4 km resampling; a large performance drop would show that resampling artifacts, not the graph topology, supply the routing signal.

Watch

Extended reading notes

Core claim

HydroGAT's central claim is that the heterogeneous basin graph—every land and river pixel as a node, with local flow edges and long-range catchment edges—is the right substrate for spatiotemporal flood prediction. Message passing happens inside gated recurrent units whose linear layers are replaced by graph attention, so edge weights become time-varying, data-driven routing decisions; a scalar gate merges the two spatial branches at gauged target nodes. The authors argue this yields hourly discharge forecasts that match or beat five baseline architectures in two hydrologically different Midwestern basins, with ablation evidence that each component contributes: removing catchment edges drops

Load-bearing premise

The load-bearing premise is that the steepest-descent flow edges and the benchmark-derived catchment links faithfully represent how water actually moves through these basins, including tile drainage, road crossings, and small-catchment behavior; if either edge set misses real flow paths, message passing propagates wrong information and the reported accuracy and attention interpretability would not transfer.

Editorial extensions

If this is right

  • Full-resolution basin graphs become trainable: the distributed pipeline removes the memory pressure that previously forced pixel graphs to be collapsed into catchment polygons.
  • A single model can explain both the 'where' and 'when' of a flood forecast, since the same attention weights identify influential upstream nodes and influential past hours.
  • Long-range catchment edges are the main carrier of spatial information in smaller, sparser basins: removing them costs 15 NSE points in the Cedar basin and 5 points in the Des Moines basin.
  • Rainfall forecasts matter most where basins respond quickly; in slower basins a bad forecast is worse than no forecast, because the model trusts the input it was trained on.
  • The model can learn to ignore uninformative small catchments and concentrate on main tributaries, which is visible in the attention maps rather than inferred post hoc.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the attention maps reflect true hydrological influence—not just statistical correlation—they could be used to rank sub-basins for gauge placement or flood mitigation, a use the paper does not develop.
  • The fixed 24-hour causal attention window is a modeling choice that could be made adaptive; basins with long snowmelt lags or floodplain storage would be a natural test case.
  • Because labels exist only at gauged stations, the paper validates the graph at gauged nodes; a hold-one-gauge-out experiment would show whether land-node message passing actually generalizes to ungauged locations.
  • The clean separation of temporal and spatial attention suggests a template for other spatiotemporal forecasting problems where joint attention would be too expensive, not just hydrology.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes HydroGAT, a spatiotemporal graph model for hourly flood/discharge prediction at pixel-level resolution. It constructs a heterogeneous basin graph in which every land and river pixel is a node, connected by D8-derived flow edges and catchment-level upstream–downstream edges. The model combines a transformer-based temporal encoder with two GAT-based GRU spatial branches (flow and catchment), fuses them via a learnable scalar, and conditions the final prediction on forecasted rainfall. Training uses a DDP pipeline across multiple nodes/GPUs. Experiments on the Cedar River Basin (CRB) and Des Moines River Basin (DSMRB) report NSE up to 0.97, KGE up to 0.96, PBIAS within ±5%, and ablations supporting the architectural choices. The paper also presents attention-based interpretability and scalability up to 64 GPUs.

Significance. If the empirical results hold, the contribution is significant for high-resolution, topology-aware flood forecasting: most prior GNN hydrology models collapse pixels to catchments or gauge stations, whereas HydroGAT keeps full spatial resolution while making training feasible via distributed data parallelism. The paper is strong in that it provides public code, an ablation study, a sensitivity analysis of forecast noise, and attention maps tied to specific stations. However, the claimed KGE superiority is contradicted by the paper's own table, and the evaluation fairness with respect to future-rainfall input is not established. These issues must be resolved before the SOTA claim is credible.

major comments (3)
  1. [Abstract; §4.2; Table 2] The abstract claims 'improved KGE (up to 0.96)', but Table 2 shows HydroGAT achieves KGE=0.92 on DSMRB, lower than RGCN's 0.97, and on CRB it ties DCRNN at 0.96. The sentence in §4.2 that HydroGAT 'consistently outperforms or matches' the best baselines is therefore not supported. Please revise the claim and identify on which basin KGE is actually improved.
  2. [§3.4, Algorithm 1, §4.1.4] The Predictor ingests future precipitation P^{t+Δt}, obtained from Stage IV observations in the experiments. While §4.1.4 states baselines were 'adapted' to the same graph and pipeline, it never specifies whether DCRNN, GraphWaveNet, RGCN, GCRNN, and STGCN-WAVE also receive this forecasted rainfall. Without this information, the comparison may give HydroGAT an information advantage. The ablation in §4.4.4 shows the forecast branch has a modest effect, but that does not establish baseline fairness. Please state explicitly and, if needed, rerun baselines with the same forecast input.
  3. [§3.1.2, §4.1.1] The flow edges rely on D8 directions computed from a 1 arc-second DEM resampled to 4 km. At this resolution, flow paths may be distorted by resampling artifacts and by landscape features such as tile drainage and prairie potholes, which are described in §1 as important in Iowa. Since the graph defines the routing over which message passing occurs, errors in these edges can affect both predictions and the interpretability of attention maps. The paper should provide evidence that the 4 km D8 graph matches hydrological connectivity (e.g., comparison with higher-resolution flow directions or with observed upstream/downstream relations) or at least discuss this limitation.
minor comments (5)
  1. [General] The manuscript uses 'heterogenous' in several places (e.g., Contributions, abstract); should be 'heterogeneous.'
  2. [§4.1.2] Features and labels are log1p and min-max normalized; the paper does not state whether the reported NSE/KGE metrics are computed after inverse-transforming the predictions to discharge units. Please clarify.
  3. [§4.1.3] With batch size=8 and a distributed sequential sampler over 16 GPUs, clarify whether the batch size is per-GPU or global, and how the temporal chunks are partitioned.
  4. [Figure 2] The label 'TTemporalEmbeddings' appears to be a typo; also, the figure is small and the caption could better define the notation h_prev, h_next.
  5. [§4.5] The 120-hour horizon result is not clearly supported by the model definition in §3.1 (which specifies Δt=72) and the experiment setup in §4.1.3 (72-hour output). Describe how the 120-hour forecasts are generated (e.g., iterative/recursive prediction).

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; evaluation is out-of-sample and self-citations are non-load-bearing.

full rationale

HydroGAT is an empirical deep-learning paper. The central claim (NSE up to 0.97, KGE up to 0.96) is obtained by training on 2012-2015 data with validation in 2016 and testing on held-out 2017-2018 hourly discharge at gauge stations. No equation in Section 3 defines the target NSE/KGE in terms of a fitted parameter or of the model's own outputs; the predictor uses past precipitation/discharge plus a future-rainfall branch, but the ablation in Section 4.4.4 shows that removing the forecast branch causes at most a 1-point NSE drop in DSMRB and no change in CRB, so the headline performance is not forced by that input. Self-citations (WaterBench [6], pixel-graph idea [39], distributed pipeline [22], RGCN baseline [23]) are data or prior-architecture references, not an imported uniqueness theorem or ansatz; none is used to forbid alternatives or to define HydroGAT's output. The abstract's 'improved KGE (up to 0.96)' is difficult to reconcile with Table 2, where RGCN achieves KGE 0.97 on DSMRB while HydroGAT achieves 0.92, and the text's 'consistently outperforms or matches' is overstated; however, this is a factual/evaluation-validity issue, not a circular reduction. Likewise, using historical Stage IV rainfall as a 'forecast' input is a benchmark-realism concern, not circularity: it is an additional covariate, not a fitted target or a renamed version of the discharge labels. No circular step was found.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim depends on several hand-chosen hyperparameters and preprocessing choices. The graph construction and data imputation are domain assumptions from prior literature. No invented entities are introduced; the heterogeneous edges are modeling constructs, not new physical objects.

free parameters (5)
  • Input/output time window = 72 hours
    Hand-selected sequence length; all models and baselines use it; affects memory and context.
  • Learning rate = 0.01
    AdamW learning rate, fixed for all models; not tuned per model.
  • Hidden dimension = 32
    Hand-chosen model width.
  • Number of attention heads = 2
    Hand-chosen; the alpha fusion vector has one scalar per head.
  • Batch size = 8
    Set for DDP with 4 GPUs per machine.
assumptions (4)
  • domain assumption D8 flow directions from a 1 arc-second DEM resampled to 4km correctly represent surface water routing at the pixel scale.
    Used to build flow edges in Section 3.1.2; if incorrect, the spatial message passing is misdirected.
  • domain assumption Catchment upstream-downstream relationships from WaterBench are accurate for the gauge stations.
    Defines catchment edges in Section 3.1.2.
  • domain assumption Imputed discharge values (linear interpolation and regression on downstream flow) are reliable enough to serve as supervision.
    Two-step imputation in Section 4.1.1 fills missing labels; errors in imputation would bias training and evaluation.
  • domain assumption External real-time precipitation forecasts are causally valid and available at test time.
    The predictor uses forecasted rainfall as input in Section 3.4; the source of this forecast is unspecified, which is a weakness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HydroGAT: Distributed Heterogeneous Graph Attention Transformer for Spatiotemporal Flood Prediction." pith.science (2026). https://pith.science/paper/4X7RZ32C

@misc{pith2026250902481,
  author       = {Pith},
  title        = {Pith review of: HydroGAT: Distributed Heterogeneous Graph Attention Transformer for Spatiotemporal Flood Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4X7RZ32C}},
  note         = {Machine review of arXiv:2509.02481}
}
abstract

Accurate flood forecasting remains a challenge for water-resource management, as it demands modeling of local, time-varying runoff drivers (e.g., rainfall-induced peaks, baseflow trends) and complex spatial interactions across a river network. Traditional data-driven approaches, such as convolutional networks and sequence-based models, ignore topological information about the region. Graph Neural Networks (GNNs) propagate information exactly along the river network, which is ideal for learning hydrological routing. However, state-of-the-art GNN-based flood prediction models collapse pixels to coarse catchment polygons as the cost of training explodes with graph size and higher resolution. Furthermore, most existing methods treat spatial and temporal dependencies separately, either applying GNNs solely on spatial graphs or transformers purely on temporal sequences, thus failing to simultaneously capture spatiotemporal interactions critical for accurate flood prediction. We introduce a heterogenous basin graph where every land and river pixel is a node connected by physical hydrological flow directions and inter-catchment relationships. We propose HydroGAT, a spatiotemporal network that adaptively learns local temporal importance and the most influential upstream locations. Evaluated in two Midwestern US basins and across five baseline architectures, our model achieves higher NSE (up to 0.97), improved KGE (up to 0.96), and low bias (PBIAS within $\pm$5%) in hourly discharge prediction, while offering interpretable attention maps that reveal sparse, structured intercatchment influences. To support high-resolution basin-scale training, we develop a distributed data-parallel pipeline that scales efficiently up to 64 NVIDIA A100 GPUs on NERSC Perlmutter supercomputer, demonstrating up to 15x speedup across machines. Our code is available at https://github.com/swapp-lab/HydroGAT.

Figures

Figures reproduced from arXiv: 2509.02481 by the authors.

Figure 1
Figure 1. Heterogeneous graph construction. (a) Flow [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Architecture of HydroGAT. A transformer-based temporal module (blue) combined with two spatial modules (orange) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Proposed distributed training pipeline. Temporal [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: The locations of the Cedar River Basin (CRB) and [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Basin-level NSE of hourly discharge prediction [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: NSE/KGE score distributions for hourly discharge [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 9
Figure 9. Figure 9: NSE per target node vs. drainage area in Cedar Basin. [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: NSE computed per gauge station in Cedar River [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 11
Figure 11. Figure 11: Annual NSE per station in the Cedar River Basin: [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]
Figure 8
Figure 8. Figure 8: Cedar River basin graph with flow (grey) and catch [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 13
Figure 13. Figure 13: Sensitivity of model to Gaussian noise (STD=0.1- [PITH_FULL_IMAGE:figures/full_fig_p009_13.png]
Figure 12
Figure 12. Figure 12: Event-based precipitation-discharge correlation [PITH_FULL_IMAGE:figures/full_fig_p009_12.png]
Figure 16
Figure 16. Figure 16: Distribution of temporal attention across past in [PITH_FULL_IMAGE:figures/full_fig_p010_16.png]
Figure 17
Figure 17. Figure 17: Scalability of distributed training in Cedar River [PITH_FULL_IMAGE:figures/full_fig_p010_17.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 42 canonical work pages

  1. [1]

    Yazeed Alabbad, Enes Yildirim, and Ibrahim Demir. 2023. A web-based analytical urban flood damage and loss estimation framework. Environmental Modelling & Software 163 (2023), 105670

    work page 2023

  2. [2]

    Tao Bai and Pejman Tahmasebi. 2023. Graph neural network for groundwater level forecasting. Journal of Hydrology 616 (2023), 128792

    work page 2023

  3. [3]

    RA Bishop, J Joens, and J Zohrer. 1998. Iowa’s wetlands, present and future with a focus on prairie potholes. Journal of the Iowa Academy of Science: JIAS 105, 3 (1998), 89–93

    work page 1998

  4. [4]

    Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014)

    arXiv 2014

  5. [5]

    Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. 2016. Convolu- tional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems 29 (2016)

    work page 2016

  6. [6]

    Ibrahim Demir, Zhongrun Xiang, Bekir Demiray, and Muhammed Sit. 2022. Wa- terbench: a large-scale benchmark dataset for data-driven streamflow forecasting. Earth System Science Data Discussions 2022 (2022), 1–19

    work page 2022

  7. [7]

    Jianhua Du. 2011. NCEP/EMC 4KM Gridded Data (GRIB) Stage IV Data. https: //doi.org/10.5065/D6PG1QDD

    doi:10.5065/d6pg1qdd 2011

  8. [8]

    Olutoyin A Fashae, Adeyemi O Olusola, Ijeoma Ndubuisi, and Christopher God- win Udomboso. 2019. Comparing ANN and ARIMA model in predicting the discharge of River Opeki from 2010 to 2020. River research and applications 35, 2 (2019), 169–177

    work page 2019

Show all 53 references
  1. [9]

    Jun Feng, Haichao Sha, Yukai Ding, Le Yan, and Zhangheng Yu. 2022. Graph convolution based spatial-temporal attention LSTM model for flood forecasting. In 2022 International Joint Conference on Neural Networks (IJCNN) . IEEE, 1–8

    2022

  2. [10]

    Sepp Hochreiter. 1998. The vanishing gradient problem during learning recurrent neural nets and problem solutions. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems 6, 02 (1998), 107–116

    1998

  3. [11]

    Xiaowei Jia, Jacob Zwart, Jeffrey Sadler, Alison Appling, Samantha Oliver, Steven Markstrom, Jared Willard, Shaoming Xu, Michael Steinbach, Jordan Read, et al

  4. [12]

    Jiange Jiang, Zhan Liao, Yang Zhou, Hao Wang, and Qingqi Pei. 2022. A short- term flood prediction based on spatial deep learning network: A case study for Xi County, China. Journal of Hydrology 607 (04 2022), 127535. doi:10.1016/j.jhydrol. 2022.127535

    2022

  5. [13]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)

    2016 arXiv

  6. [14]

    Frederik Kratzert, Daniel Klotz, Martin Gauch, Christoph Klingler, Grey Nearing, and Sepp Hochreiter. 2021. Large-scale river network modeling using graph neural networks. In EGU General Assembly Conference Abstracts . EGU21–13375

    2021

  7. [15]

    Vikram Kumar, Selim Unal, Suraj Kumar Bhagat, and Tiyasha Tiyasha. 2024. A data-driven approach to river discharge forecasting in the Himalayan region: Insights from Aglar and Paligaad rivers. Results in Engineering 22 (2024), 102044

    2024

  8. [16]

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2018. Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting. In International Conference on Learning Representations (ICLR)

    2018

  9. [17]

    Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

    2017 arXiv

  10. [18]

    Grégoire Mialon, Dexiong Chen, Margot Selosse, and Julien Mairal. 2021. Graphit: Encoding graph structure in transformers.arXiv preprint arXiv:2106.05667 (2021)

    2021 arXiv

  11. [19]

    Hesham Mostafa. 2021. Sequential Aggregation and Rematerialization: Dis- tributed Full-batch Training of Graph Neural Networks on Large Graphs. arXiv

    2021

  12. [20]

    Daniel Asante Otchere, Tarek Omar Arbi Ganat, Raoof Gholami, and Syahrir Ridha. 2021. Application of supervised machine learning paradigms in the prediction of petroleum reservoir properties: Comparative analysis of ANN and SVM models. Journal of Petroleum Science and Engineer...

    2021

  13. [21]

    arXiv preprint arXiv:2111.06483 (2021)

    2021 arXiv

  14. [22]

    Aishwarya Sarkar, Jien Zhang, Chaoqun Lu, and Ali Jannesari. 2023. Domain- Aware Scalable Distributed Training for Geo-Spatiotemporal Data. In 2023 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) . 960–969. doi:10.1109/IPDPSW59300.2023.00159

    2023

  15. [23]

    Luana Ruiz, Fernando Gama, and Alejandro Ribeiro. 2019. Gated Graph Convo- lutional Recurrent Neural Networks. arXiv:1903.01888 [cs.LG] https://arxiv.org/ abs/1903.01888

    2019 arXiv

  16. [24]

    Muhammed Sit, Bekir Zahit Demiray, and Ibrahim Demir. 2023. Spatial downscal- ing of streamflow data with attention based spatio-temporal graph convolutional networks. EarthArxiv (2023)

    2023

  17. [25]

    Muhammed Sit, Bekir Demiray, and Ibrahim Demir. 2021. Short-term Hourly Streamflow Prediction with Graph Convolutional GRU Networks. arXiv:2107.07039 [cs.LG] https://arxiv.org/abs/2107.07039

    2021 arXiv

  18. [26]

    Michael Stalder, Firat Ozdemir, Artur Safin, Jonas Sukys, Damien Bouffard, and Fernando Perez-Cruz. 2021. Probabilistic modeling of lake surface water tem- perature using a Bayesian spatio-temporal graph convolutional neural network. arXiv preprint arXiv:2109.13235 (2021)

    2021 arXiv

  19. [27]

    Muhammed Sit, Bekir Z Demiray, Zhongrun Xiang, Gregory J Ewing, Yusuf Sermet, and Ibrahim Demir. 2020. A comprehensive review of deep learning applications in hydrology and water resources. Water Science and Technology 82, 12 (2020), 2635–2670

    2020

  20. [28]

    Alexander Y Sun, Peishi Jiang, Zong-Liang Yang, Yangxinyu Xie, and Xingyuan Chen. 2022. A graph neural network approach to basin-scale river network learning: The role of physics-based connectivity and data fusion. Hydrology and Earth System Sciences Discussions 2022 (2022), 1–35

    2022

  21. [29]

    Alexander Y Sun, Peishi Jiang, Maruti K Mudunuru, and Xingyuan Chen. 2021. Explore spatio-temporal learning of large sample hydrology using graph neural networks. Water Resources Research 57, 12 (2021), e2021WR030394

    2021

  22. [30]

    Prasanth Valayamkunnath, Michael Barlage, Fei Chen, David J Gochis, and Kristie J Franz. 2020. Mapping of 30-meter resolution tile-drained croplands using a geospatial modeling approach. Scientific data 7, 1 (2020), 257

    2020

  23. [31]

    United States Geological Survey. 2021. 3D Elevation Program 1 arc-second Digital Elevation Model. doi:10.5069/G98K778D

    2021 doi

  24. [32]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)

    2017 arXiv

  25. [33]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems 30 (2017)

    2017

  26. [34]

    M Waseem, N Mani, G Andiego, and M Usman. 2017. A review of criteria of fit for hydrological models. International Research Journal of Engineering and Technology (IRJET) 4, 11 (2017), 1765–1772

    2017

  27. [35]

    Chao Wang, Shijie Jiang, yi Zheng, Feng Han, Rohini Kumar, Oldrich Rakovec, and Siqi Li. 2024. Distributed Hydrological Modeling With Physics-Encoded Deep Learning: A General Framework and Its Application in the Amazon. Water Resources Research 60 (04 2024). doi:10.1029/2023WR036170

    2024 doi

  28. [36]

    Wei Ming Wong, Siva Kumar Subramaniam, Farah Shahnaz Feroz, Indra Devi Subramaniam, and Lew Ai Fen Rose. 2020. Flood prediction using ARIMA model in Sungai Melaka, Malaysia. International Journal 9, 4 (2020)

    2020

  29. [37]

    Jared D Willard, Charuleka Varadharajan, Xiaowei Jia, and Vipin Kumar. 2025. Time series predictions in unmonitored sites: A survey of machine learning techniques in water resources. Environmental Data Science 4 (2025), e7

    2025

  30. [38]

    Zhijian Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and Philip S Yu. 2019. Graph WaveNet for Deep Spatial-Temporal Graph Modeling. In IJCAI

    2019

  31. [39]

    Zhanghao Wu, Paras Jain, Matthew Wright, Azalia Mirhoseini, Joseph E Gonzalez, and Ion Stoica. 2021. Representing long-range context for graph neural networks with global attention. Advances in neural information processing systems 34 (2021), 13266–13279

    2021

  32. [40]

    Zhongrun Xiang and Ibrahim Demir. 2022. Fully distributed rainfall-runoff modeling using spatial-temporal graph neural network. EarthArxiv (2022)

    2022

  33. [41]

    Zhongrun Xiang and Ibrahim Demir. 2021. High-resolution rainfall-runoff mod- eling using graph neural network. arXiv preprint arXiv:2110.10833 (2021)

    2021 arXiv

  34. [42]

    Shuai Yang, Yueqin Zhang, and Zehua Zhang. 2023. Runoff prediction based on dynamic spatiotemporal graph neural network. Water 15, 13 (2023), 2463

    2023

  35. [43]

    Charlene Yang and Jack Deslippe. 2020. Accelerate science on perlmutter with nersc. Bulletin of the American Physical Society 65 (2020)

    2020

  36. [44]

    Enes Yildirim, Craig Just, and Ibrahim Demir. 2022. Flood risk assessment and quantification at the community and property level in the State of Iowa. Interna- tional journal of disaster risk reduction 77 (2022), 103106

    2022

  37. [45]

    Weichao Yang, Chuanxing Zheng, Xuelian Jiang, Hao Wang, Jijian Lian, De Hu, and Airong Zheng. 2023. Study on urban flood simulation based on a novel model of SWTM coupling D8 flow direction and backflow effect. Journal of Hydrology 621 (2023), 129608

    2023

  38. [46]

    Bing Yu, Haoteng Yin, and Zhanxing Zhu. 2017. Spatio-temporal graph con- volutional networks: A deep learning framework for traffic forecasting. arXiv preprint arXiv:1709.04875 (2017)

    2017 arXiv

  39. [47]

    Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie-Yan Liu. 2021. Do transformers really perform badly for graph representation? Advances in neural information processing systems 34 (2021), 28877–28888

    2021

  40. [48]

    Huafei Yu, Tinghua Ai, Min Yang, Lina Huang, and Jiaming Yuan. 2022. A recognition method for drainage patterns using a graph convolutional network. International Journal of Applied Earth Observation and Geoinformation 107 (2022), 102696

    2022

  41. [49]

    Cunjun Yu, Xiao Ma, Jiawei Ren, Haiyu Zhao, and Shuai Yi. 2020. Spatio-temporal graph transformer networks for pedestrian trajectory prediction. In European conference on computer vision . Springer, 507–523

    2020

  42. [50]

    Ariele Zanfei, Bruno M Brentan, Andrea Menapace, Maurizio Righetti, and Manuel Herrera. 2022. Graph convolutional recurrent neural networks for water demand forecasting. Water Resources Research 58, 7 (2022), e2022WR032299. SIGSPATIAL ’25, November 3–6, 2025, Minneapolis, MN, ...

    2022

  43. [51]

    Hao Yuan, Yajiong Liu, Yanfeng Zhang, Xin Ai, Qiange Wang, Chaoyi Chen, Yu Gu, and Ge Yu. 2023. Comprehensive evaluation of gnn training systems: A data management perspective. arXiv preprint arXiv:2311.13279 (2023)

    2023 arXiv

  44. [53]

    Qun Zhao, Yuelong Zhu, Kai Shu, Dingsheng Wan, Yufeng Yu, Xudong Zhou, and Huan Liu. 2020. Joint spatial and temporal modeling for hydrological prediction. Ieee Access 8 (2020), 78492–78503

    2020

  45. [2021]

    In Proceedings of the 2021 SIAM International Conference on Data Mining (SDM)

    Physics-guided recurrent graph model for predicting flow and temperature in river networks. In Proceedings of the 2021 SIAM International Conference on Data Mining (SDM). SIAM, 612–620

    2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.