arxiv: 2604.25890 · v1 · submitted 2026-04-28 · ⚛️ physics.ao-ph

Recognition: unknown

Observation-Guided Neural Surrogate Learning for Scientific Simulation Emulation: A Single-Gauge Flood-Inundation Proof of Concept

Marzieh Alireza Mirhoseini

Authors on Pith no claims yet

Pith reviewed 2026-05-07 13:49 UTC · model grok-4.3

classification ⚛️ physics.ao-ph

keywords neural surrogateflood inundationsimulation emulationobservation-guided learningU-Nethydrodynamic modelingsingle gauge supervisionChicago flood

0 comments

The pith

A neural corrector trained at one gauge pixel reproduces full hydrodynamic flood maps with high fidelity on held-out events.

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

The paper establishes that a U-Net-ASPP model can refine a coarse ensemble Gaussian-process surrogate into an accurate emulation of LISFLOOD-FP flood-depth outputs by receiving supervision from only a single real gauge record at its mapped grid location. Simulation-derived losses are used everywhere except at that one pixel, where the datum-converted gauge depth serves as the training target. Across temporally held-out events the resulting emulator matches the simulator targets outside the gauge pixel to within 0.01 m mean absolute error and 0.99 R-squared, while remaining consistent with the gauge value at the constrained pixel. A reader would care because the result shows a practical way to inject sparse real-world observations into physics-based emulators without requiring dense measurement networks.

Core claim

The central claim is that an observation-guided neural surrogate, built from an EnsCGP coarse estimator followed by a U-Net-ASPP corrector, reproduces LISFLOOD-FP simulation targets with R-squared approximately 0.99 and mean absolute error below 0.01 m outside the gauge-constrained pixel across 2013-2019 temporally held-out events, while maintaining strong pointwise consistency with the converted Gauge L local depth target under the rolling-year protocol. The framework therefore demonstrates strong simulator-emulation agreement under single-site observation-guided correction.

What carries the argument

The EnsCGP ensemble-approximated Gaussian-process/local analogue surrogate that supplies a coarse depth field and uncertainty proxy, combined with the U-Net-ASPP neural corrector that refines the field using hybrid losses evaluated away from the single gauge pixel.

If this is right

Full inundation maps can be generated that remain consistent with both the underlying hydrodynamic simulator and the local gauge observation.
Sparse real observations can be incorporated into simulation emulators without needing dense sensor networks.
The approach generalizes temporally, as shown by performance on events held out by year under the rolling protocol.
The gauge constraint remains localized to one pixel while simulation fidelity is preserved elsewhere.

Where Pith is reading between the lines

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

The method could lower the cost of repeated flood simulations by replacing them with fast neural forward passes after initial training.
Extending the same single-site supervision idea to multiple gauges might improve spatial coverage in complex urban terrain.
Live gauge feeds could be fed directly into the corrector for near-real-time mapping, provided the mapping step remains unbiased.
The framework still emulates the simulator rather than predicting real-world inundation independently, so any simulator biases would carry through.

Load-bearing premise

The single gauge stage record can be mapped and datum-converted to a local water-depth target on the simulation grid without introducing systematic bias that the neural corrector then learns to exploit only at that pixel.

What would settle it

Running the emulator on a new held-out flood event and comparing its predicted depths at several independent locations away from the gauge against both a fresh full LISFLOOD-FP simulation and any available additional in-situ measurements; large systematic deviations at those locations would falsify the claim of faithful emulation.

Figures

Figures reproduced from arXiv: 2604.25890 by Marzieh Alireza Mirhoseini.

**Figure 1.** Figure 1: Study area and gauge network. (a) Twelve USGS river gauges (A–L) across the Chicago metropolitan region. (b) The 256 × 256 study crop centered on Gauge L used for high-resolution emulation. Background satellite imagery is from Google Earth and is shown only for geographic context; no model inputs, training targets, or quantitative results are derived from this imagery. Google Earth and any embedded datapr… view at source ↗

**Figure 2.** Figure 2: Maximum daily rainfall in 2019 and simulated flood inundation (27 April 2019). (a) Daily rainfall field on the wettest day of 2019. (b) Corresponding LISFLOOD-FP flood depth over the 256 × 256 crop; depths in meters. Gauge L location is marked in white circle. before being used in the surrogate-learning workflow. 2.2 Focus region around Gauge L view at source ↗

**Figure 3.** Figure 3: Two-stage observation-guided surrogate-emulation pipeline. Gridded rainfall is first mapped to a coarse flood-depth estimate and an uncertainty proxy by the EnsCGP local analogue surrogate. The U-Net–ASPP corrector refines this estimate using the coarse depth, uncertainty proxy, rainfall field, and coordinate channels. During training only, the Gauge L stage record is mapped to the simulation grid and conv… view at source ↗

**Figure 4.** Figure 4: Schematic of the refinement CNN architecture (U-Net+ASPP). The model takes a fivechannel input—the coarse EnsCGP flood-depth prediction, an analogue-ensemble uncertainty proxy, the rainfall field, and spatial x- and y-coordinate channels—and produces a high-resolution flood-depth map. The U-Net encoder (left) applies convolutional blocks and 2× pooling to downsample, while the decoder (right) uses upsampl… view at source ↗

**Figure 5.** Figure 5: Gauge-derived depth consistency. (a) Emulator-predicted local depth at Gauge L versus the converted gauge-derived local depth target for all test events (2013–2019). The dashed line denotes the 1:1 reference line; the rounded statistics are R2 = 1.00 and MAE ≈ 0.01 m, indicating strong pointwise agreement under the stated rolling-year protocol. (b) Absolute prediction error at Gauge L for each year (2013–2… view at source ↗

**Figure 6.** Figure 6: Per-event maximum error for 2015 floods. For each test event in 2015, the maximum absolute pixel error is plotted for the emulator (blue) and for the baseline coarse GP surrogate (EnsCGP, orange). Each point represents one flood event (date indicated along the horizontal axis). The gauge pixel is excluded from these error calculations. The emulator consistently yields much lower worst-case errors than the … view at source ↗

**Figure 7.** Figure 7: Simulation, emulator, and initial-guess inundation comparison in an extreme 2015 flood. Flood depth maps for the extreme event on 11 June 2015, comparing the LISFLOOD-FP simulation, the emulator, and the coarse EnsCGP surrogate (initial guess), with their respective error maps relative to the simulation. In this case, the EnsCGP surrogate severely underestimates the flood extent (producing insufficient inu… view at source ↗

**Figure 8.** Figure 8: Per-event maximum error for 2017 floods. Similar to view at source ↗

**Figure 9.** Figure 9: Simulation, emulator, and initial-guess inundation comparison in an extreme 2017 flood. Flood depth maps for the extreme event on 09 July 2017 from the LISFLOOD-FP simulation, the emulator, and the coarse EnsCGP surrogate, with their absolute error maps relative to the simulation. In this case, the EnsCGP surrogate substantially overestimates the flood extent, producing spurious inundation. The emulator su… view at source ↗

**Figure 10.** Figure 10: Emulator accuracy across held-out events (2013–2019). (a) Pixel-wise comparison of emulator-predicted vs. simulated flood depths, shown as a density scatter plot (log10 of pixel count). The dashed line indicates perfect agreement (1:1). The emulator achieves R2 ≈ 0.99 and overall MAE < 0.01 m, indicating strong agreement with the simulator under the stated evaluation protocol. (b) Per-event maximum absolu… view at source ↗

read the original abstract

We present an observation-guided neural surrogate-learning framework for scientific simulation emulation, demonstrated on urban flood-inundation mapping. The framework combines LISFLOOD-FP hydrodynamic simulations with a real Gauge L stage record that is mapped to the simulation grid and converted to a datum-consistent local water-depth target before being used as single-site supervision. Focusing on a 256 x 256 crop around Gauge L in the Chicago metropolitan area, the method first constructs an ensemble-approximated Gaussian-process/local analogue surrogate (EnsCGP) to obtain a coarse flood-depth estimate and an uncertainty proxy. A U-Net-ASPP neural corrector then refines the coarse map using only simulation-derived and geospatial inputs: EnsCGP depth, the uncertainty proxy, rainfall, and spatial coordinates. The converted gauge-derived local depth is used only as a pointwise training target at the mapped gauge pixel; simulation-based losses are evaluated away from that pixel. Across temporally held-out events from 2013-2019, the emulator closely reproduces LISFLOOD-FP simulation targets outside the gauge-constrained pixel, with R^2 approximately 0.99 and mean absolute error below 0.01 m, and shows strong pointwise consistency with the converted Gauge L local depth target under the stated rolling-year protocol. We interpret these results as strong simulator-emulation agreement with pointwise observation-guided correction, not as independent validation of real-world inundation accuracy or as a complete operational flood-forecasting system.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This is a clean but narrowly scoped proof-of-concept that gets a hybrid GP-plus-U-Net emulator to match LISFLOOD-FP outputs closely on held-out times with single-gauge guidance.

read the letter

Hey, the main takeaway is that this hybrid approach reproduces LISFLOOD-FP flood depths closely on held-out temporal events using only one gauge for supervision. They start with an ensemble Gaussian process to get a coarse depth map and uncertainty estimate, then feed that plus rainfall and coordinates into a U-Net with ASPP for refinement. The loss is mostly from the simulator outputs, but at the single mapped gauge pixel they use the converted stage record as target. On the 2013-2019 held-out set, they get R² around 0.99 and MAE under 0.01 m away from the gauge, with good match at the point too. The paper does well in being explicit that this demonstrates simulator emulation with pointwise correction, not real inundation accuracy. The setup with temporal hold-out and the rolling-year protocol for the gauge check is sensible. Soft spots are the tight scope—one gauge, one small crop—so scaling is untested. The abstract skips error bars and full details on the gauge-to-grid conversion, which matters for whether the model is learning something real or just fitting the mapping. There's inherent circularity from heavy simulator-based loss, though the single constraint and hold-out mitigate it somewhat. This is aimed at folks working on neural surrogates for physics simulations, especially in hydrology. A reader looking for examples of observation-guided emulation would find it useful. It deserves peer review as a well-executed limited demonstration with solid numbers, though reviewers would likely ask for more cases and uncertainty analysis.

Referee Report

2 major / 2 minor

Summary. The paper presents an observation-guided neural surrogate-learning framework for emulating LISFLOOD-FP hydrodynamic flood-inundation simulations. It first builds an ensemble-approximated Gaussian-process surrogate (EnsCGP) to produce coarse depth estimates and an uncertainty proxy, then applies a U-Net-ASPP neural corrector that refines the map using EnsCGP output, uncertainty, rainfall, and spatial coordinates. The only observation-based supervision is a single-point target at the mapped Gauge L pixel obtained by converting the stage record to local water depth; simulation-derived losses are used everywhere else. On temporally held-out events from 2013-2019 the emulator is reported to reproduce LISFLOOD-FP targets outside the gauge pixel with R² ≈ 0.99 and MAE < 0.01 m while remaining consistent with the gauge target.

Significance. If the quantitative results hold, the work supplies a concrete proof-of-concept for training neural emulators of physics-based simulators under sparse, single-site observational guidance. The explicit scoping to simulator-emulation agreement (rather than real-world inundation accuracy) and the use of temporal hold-out validation are appropriately cautious and strengthen the contribution. The approach could inform future data-assimilation strategies in hydrology, provided the gauge-mapping procedure is fully documented.

major comments (2)

[Abstract] Abstract: the central quantitative claim (R² ≈ 0.99 and MAE < 0.01 m on held-out events) is presented without error bars, standard deviations across events, or the exact number of events, making it difficult to judge the robustness of the emulation-agreement result.
[Framework description] Framework description: the gauge-to-grid mapping and datum-conversion procedure that produces the single-site local-depth target is not described in sufficient detail. Because this point supplies the only observation-derived supervision and the paper itself flags the risk that the corrector could exploit pixel-specific bias, the omission is load-bearing for evaluating the weakest assumption.

minor comments (2)

The abstract introduces the EnsCGP acronym after a long descriptive phrase; a parenthetical expansion on first use would improve immediate readability.
Consider adding a small table or supplementary figure that reports the per-event metrics (with standard deviations) rather than aggregate statements only; this would make the held-out performance easier to assess at a glance.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful to the referee for their thoughtful review and recommendation for minor revision. We have carefully considered the major comments and provide point-by-point responses below. We will make the suggested revisions to improve the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the central quantitative claim (R² ≈ 0.99 and MAE < 0.01 m on held-out events) is presented without error bars, standard deviations across events, or the exact number of events, making it difficult to judge the robustness of the emulation-agreement result.

Authors: We agree that including these statistics would strengthen the presentation of the results. In the revised manuscript, we will report the exact number of temporally held-out events from 2013-2019, along with the mean and standard deviation of the R² and MAE metrics across those events. If feasible, we will also include error bars or confidence intervals in the abstract or main text to better convey robustness. revision: yes
Referee: [Framework description] Framework description: the gauge-to-grid mapping and datum-conversion procedure that produces the single-site local-depth target is not described in sufficient detail. Because this point supplies the only observation-derived supervision and the paper itself flags the risk that the corrector could exploit pixel-specific bias, the omission is load-bearing for evaluating the weakest assumption.

Authors: We appreciate the referee highlighting this important detail. While the manuscript describes the mapping of the Gauge L stage record to the simulation grid and its conversion to a datum-consistent local water-depth target, we acknowledge that the procedure could benefit from greater elaboration to allow full evaluation of the supervision strategy. In the revision, we will expand this description, including specifics on the datum conversion, grid alignment, and any assumptions made, and we will explicitly discuss how this addresses or mitigates the risk of pixel-specific bias exploitation by the corrector. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper trains a U-Net corrector on LISFLOOD-FP simulation outputs (with losses evaluated away from the single gauge pixel) and reports generalization metrics on temporally held-out events. This is a standard supervised learning setup for emulator training; the reported R² ≈ 0.99 and MAE < 0.01 m on held-out data reflect successful fitting and generalization rather than any definitional or self-referential reduction. The gauge datum supplies only a pointwise constraint at one location and is not used to define the simulator targets themselves. No self-citation chains, ansatz smuggling, or uniqueness theorems are invoked to force the central result. The abstract explicitly scopes the claim to simulator-emulation agreement, avoiding any claim of independent real-world validation.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on the domain assumption that gauge stage can be converted to a consistent local depth target and on the standard assumption that neural-network training on simulation losses plus one point constraint yields a faithful emulator; no new physical entities are postulated.

free parameters (2)

U-Net-ASPP weights and hyperparameters
Trained parameters that define the corrector mapping; their values are fitted to the combined simulation and single-gauge targets.
EnsCGP ensemble size and kernel hyperparameters
Chosen to produce the coarse depth and uncertainty proxy used as network inputs.

axioms (2)

domain assumption The gauge stage record can be mapped to the simulation grid and converted to a datum-consistent local water-depth value without introducing unquantified bias.
Invoked when the converted gauge depth is used as the sole observation target at the mapped pixel.
domain assumption Simulation-based losses evaluated away from the gauge pixel are sufficient to enforce physical consistency in the neural corrector.
Underlies the claim that the emulator reproduces LISFLOOD-FP targets outside the constrained pixel.

pith-pipeline@v0.9.0 · 5577 in / 1727 out tokens · 63800 ms · 2026-05-07T13:49:40.798843+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

28 extracted references · 21 canonical work pages

[1]

Oliver E. J. Wing, Paul D. Bates, Christopher C. Sampson, Adam M. Smith, K. Ariel Johnson, and Timothy A. Erickson. Validation of a 30 m resolution flood hazard model of the conterminous United States.Water Resources Research, 53(9):7968–7986, 2017.https://doi.org/10.1002/2017WR020917

work page doi:10.1002/2017wr020917 2017
[2]

Global projections of river flood risk in a warmer world.Earth’s Future, 5(2):171–182, 2017.https://doi.org/10.1002/2016EF000485

Luigi Alfieri, Berny Bisselink, Francesco Dottori, Gustavo Naumann, Ad de Roo, Peter Salamon, Klaus Wyser, and Luc Feyen. Global projections of river flood risk in a warmer world.Earth’s Future, 5(2):171–182, 2017.https://doi.org/10.1002/2016EF000485

work page doi:10.1002/2016ef000485 2017
[3]

Rapid forecasting of urban flood inundation using multiple machine learning models.Natural Hazards, 108(2):2335–2356, 2021.https://doi.org/10.1007/s11069-021-04782-x

Jingming Hou, Nie Zhou, Guangzhao Chen, Miansong Huang, and Guangbi Bai. Rapid forecasting of urban flood inundation using multiple machine learning models.Natural Hazards, 108(2):2335–2356, 2021.https://doi.org/10.1007/s11069-021-04782-x

work page doi:10.1007/s11069-021-04782-x 2021
[4]

Zahura and Jonathan L

Faria T. Zahura and Jonathan L. Goodall. Predicting combined tidal and pluvial flood inundation using a machine learning surrogate model.Journal of Hydrology: Regional Studies, 41:101087, 2022. https://doi.org/10.1016/j.ejrh.2022.101087. 15

work page doi:10.1016/j.ejrh.2022.101087 2022
[5]

Xingyu Yan, Kui Xu, Wenqiang Feng, and Jing Chen. A rapid prediction model of urban flood in- undation in a high-risk area coupling machine learning and numerical simulation approaches.In- ternational Journal of Disaster Risk Science, 12(6):903–918, 2021.https://doi.org/10.1007/ s13753-021-00384-0

2021
[6]

Bates, Matthew S

Paul D. Bates, Matthew S. Horritt, and Timothy J. Fewtrell. A simple inertial formulation of the shallow water equations for efficient two-dimensional flood inundation modelling.Journal of Hydrology, 387(1–2):33–45, 2010.https://doi.org/10.1016/j.jhydrol.2010.03.027

work page doi:10.1016/j.jhydrol.2010.03.027 2010
[7]

Sharifian, Georges Kesserwani, Alovya A

Mohammad K. Sharifian, Georges Kesserwani, Alovya A. Chowdhury, Jeffrey Neal, and Paul Bates. LISFLOOD-FP 8.1: New GPU-accelerated solvers for faster fluvial/pluvial flood simulations.Geosci- entific Model Development, 16:2391–2413, 2023.https://doi.org/10.5194/gmd-16-2391-2023

work page doi:10.5194/gmd-16-2391-2023 2023
[8]

Surrogate modeling of joint flood risk across coastal watersheds

Benjamin Bass and Philip Bedient. Surrogate modeling of joint flood risk across coastal watersheds. Journal of Hydrology, 558:159–173, 2018.https://doi.org/10.1016/j.jhydrol.2018.01.014

work page doi:10.1016/j.jhydrol.2018.01.014 2018
[9]

Bo Liu, Yingbing Li, Minyuan Ma, and Bojun Mao. A comprehensive review of machine learning approaches for flood depth estimation.International Journal of Disaster Risk Science, 16(3):433–445, 2025.https://doi.org/10.1007/s13753-025-00639-0

work page doi:10.1007/s13753-025-00639-0 2025
[10]

Wang, Wenyan Wu, and Rory Nathan

Niels Fraehr, Quan J. Wang, Wenyan Wu, and Rory Nathan. Assessment of surrogate models for flood inundation: The physics-guided LSG model vs. state-of-the-art machine learning models.Water Research, 252:121202, 2024.https://doi.org/10.1016/j.watres.2024.121202

work page doi:10.1016/j.watres.2024.121202 2024
[11]

Wang, Wenyan Wu, and Rory Nathan

Niels Fraehr, Quan J. Wang, Wenyan Wu, and Rory Nathan. Upskilling low-fidelity hydrodynamic models of flood inundation through spatial analysis and Gaussian Process learning.Water Resources Research, 58(8):e2022WR032248, 2022.https://doi.org/10.1029/2022WR032248

work page doi:10.1029/2022wr032248 2022
[12]

2018.10.045

Maziar Raissi, Paris Perdikaris, and George E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational Physics, 378:686–707, 2019.https://doi.org/10.1016/j.jcp. 2018.10.045

work page doi:10.1016/j.jcp 2019
[13]

Physics-informed neural networks as surrogate models of hydrodynamic simulators.Science of the Total Environment, 912:168814, 2024

James Donnelly, Alireza Daneshkhah, and Soroush Abolfathi. Physics-informed neural networks as surrogate models of hydrodynamic simulators.Science of the Total Environment, 912:168814, 2024. https://doi.org/10.1016/j.scitotenv.2023.168814

work page doi:10.1016/j.scitotenv.2023.168814 2024
[14]

Bamber, Chaojun Ouyang, and Xiao Xiang Zhu

Qingsong Xu, Yilei Shi, Jonathan L. Bamber, Chaojun Ouyang, and Xiao Xiang Zhu. Large-scale flood modeling and forecasting with FloodCast.Water Research, 264:122162, 2024.https://doi.org/10. 1016/j.watres.2024.122162

work page arXiv 2024
[15]

Learning surrogate rainfall-driven inundation models with few data, 2024

Marzieh Alireza Mirhoseini. Learning surrogate rainfall-driven inundation models with few data, 2024. arXiv:2411.19323;https://doi.org/10.48550/arXiv.2411.19323

work page doi:10.48550/arxiv.2411.19323 2024
[16]

Learning surrogate extreme rainfall-driven inundation models with few data

Marzieh Alireza Mirhoseini, Anamitra Saha, and Sai Ravela. Learning surrogate extreme rainfall-driven inundation models with few data. InAGU Fall Meeting Abstracts, pages H21E–04, 2024. AGU24, Washington, DC;https://ui.adsabs.harvard.edu/abs/2024AGUFMH21E...04A/abstract

2024
[17]

U-Net: Convolutional Networks for Biomedical Image Segmentation, pp.\ 234–241

Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. InMedical Image Computing and Computer-Assisted Intervention – MICCAI 2015, pages 234–241, 2015.https://doi.org/10.1007/978-3-319-24574-4_28

work page doi:10.1007/978-3-319-24574-4_28 2015
[18]

Encoder- decoder with atrous separable convolution for semantic image segmentation

Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder- decoder with atrous separable convolution for semantic image segmentation. InProceedings of the European Conference on Computer Vision, pages 801–818, 2018.https://doi.org/10.1007/ 978-3-030-01234-2_49. 16

2018
[19]

Farr, Paul A

Tom G. Farr, Paul A. Rosen, Edward Caro, Robert Crippen, Riley Duren, Scott Hensley, Michael Kobrick, Mimi Paller, Ernesto Rodriguez, Ladislav Roth, David Seal, Scott Shaffer, Joanne Shimada, Jeffrey Umland, Marian Werner, Michael Oskin, Douglas Burbank, and Douglas Alsdorf. The shuttle radar topography mission.Reviews of Geophysics, 45:RG2004, 2007.https...

2007
[20]

Conterminous United States land cover change patterns 2001–2016 from the 2016 National Land Cover Database

Collin Homer, Jon Dewitz, Suming Jin, George Xian, Catherine Costello, Patrick Danielson, Leila Gass, Michelle Funk, James Wickham, Stephen Stehman, Roger Auch, and Kurt Riitters. Conterminous United States land cover change patterns 2001–2016 from the 2016 National Land Cover Database. ISPRS Journal of Photogrammetry and Remote Sensing, 162:184–199, 2020...

2001
[21]

M. M. Thornton, R. Shrestha, Y. Wei, P. E. Thornton, S.-C. Kao, and B. E. Wilson. Daymet: Daily surface weather data on a 1-km grid for North America, version 4 r1, 2022.https://doi.org/10. 3334/ORNLDAAC/2129

2022
[22]

Geological Survey

U.S. Geological Survey. USGS Water Data for the Nation: U.S. Geological Survey National Water Information System database, 2026. Accessed 2026-04-26;https://doi.org/10.5066/F7P55KJN

work page doi:10.5066/f7p55kjn 2026
[23]

Hydrofunctions 0.2.4 documentation: Python tools for downloading and working with USGS NWIS hydrology data, 2023.https://hydrofunctions.readthedocs.io/

Michael Roberge. Hydrofunctions 0.2.4 documentation: Python tools for downloading and working with USGS NWIS hydrology data, 2023.https://hydrofunctions.readthedocs.io/

2023
[24]

Fast ensemble smoothing.Ocean Dynamics, 57(2):123–134, 2007

Sai Ravela and Dennis McLaughlin. Fast ensemble smoothing.Ocean Dynamics, 57(2):123–134, 2007. https://doi.org/10.1007/s10236-006-0098-6

work page doi:10.1007/s10236-006-0098-6 2007
[25]

A realtime observatory for laboratory simulation of planetary flows.Experiments in Fluids, 48(5):915–925, 2010.https: //doi.org/10.1007/s00348-009-0752-0

Sai Ravela, John Marshall, Chris Hill, Andrew Wong, and Scott Stransky. A realtime observatory for laboratory simulation of planetary flows.Experiments in Fluids, 48(5):915–925, 2010.https: //doi.org/10.1007/s00348-009-0752-0

work page doi:10.1007/s00348-009-0752-0 2010
[26]

An intriguing failing of convolutional neural networks and the CoordConv solution

Rosanne Liu, Joel Lehman, Piero Molino, Felipe Petroski Such, Eric Frank, Alex Sergeev, and Jason Yosinski. An intriguing failing of convolutional neural networks and the CoordConv solution. InAd- vances in Neural Information Processing Systems, volume 31, 2018.https://proceedings.neurips. cc/paper/2018/hash/60106888f8977b71e1f15db7bc9a88d1-Abstract.html

work page arXiv 2018
[27]

Rapid urban flood inun- dation forecasting using a physics-informed deep learning approach.Journal of Hydrology, 643:131998, 2024.https://doi.org/10.1016/j.jhydrol.2024.131998

Fang Yang, Wu Ding, Jianshi Zhao, Lixiang Song, Dawen Yang, and Xudong Li. Rapid urban flood inun- dation forecasting using a physics-informed deep learning approach.Journal of Hydrology, 643:131998, 2024.https://doi.org/10.1016/j.jhydrol.2024.131998

work page doi:10.1016/j.jhydrol.2024.131998 2024
[28]

Hamed Farahmand, Yuanchang Xu, and Ali Mostafavi. A spatial–temporal graph deep learning model for urban flood nowcasting leveraging heterogeneous community features.Scientific Reports, 13:6768, 2023.https://doi.org/10.1038/s41598-023-32548-x. 17

work page doi:10.1038/s41598-023-32548-x 2023