Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 69 of the 69 references it cites were found in a registry or an index; none of the 34 sources its claims cite is checked for support yet.

The Geometry of Flow: Advancing Predictions of River Geometry With Multi‐Model Machine Learning

Open access

Abstract

Hydraulic geometry parameters describing river hydrogeomorphic relationships are critical for determining a channel's capacity to convey water and sediment which is important for flood forecasting. Although well‐established, power‐law hydraulic geometry curves have been widely used to understand riverine systems and mapping flooding inundation worldwide for the past 70 years, we have become increasingly aware of their limitations. In the present study, we have moved beyond these traditional power‐law relationships, testing the ability of machine‐learning models to provide improved predictions of river width and depth. For this work, we have used an unprecedentedly large river measurement data set (HYDRoSWOT) as well as a suite of watershed predictor data to develop novel data‐driven approaches to better estimate river geometries over the contiguous United States (CONUS). Our Random Forest, XGBoost, and neural network models out‐performed the traditional, regionalized power law‐based hydraulic geometry equations for both width and depth, providing R‐squared values of as high as 0.75 for width and as high as 0.67 for depth, compared with R‐squared values of 0.45 for width and 0.18 for depth from the regional hydraulic geometry equations. Our results also show diverse performance outcomes across stream orders and geographical regions for the different machine‐learning models, demonstrating the value of using multi‐model approaches to maximize the predictability of river geometry. The developed models have been used to create the newly publicly available STREAM‐geo data set, which provides river width, depth, width/depth ratio, and river and stream surface area (%RSSA) for nearly 2.7 million NHDPlus stream reaches across the contiguous US.

Overview

CoChat Summary

TL;DR

This study tests multi-model machine learning to predict river width and depth across the contiguous US using HYDRoSWOT observations and 17 watershed predictors, finding that ML models outperform traditional regional power-law hydraulic geometry. It releases STREAM-geo predictions for nearly 2.7 million NHDPlus stream reaches.

Key findings

  • Random Forest, XGBoost, and neural network (MLP) models outperformed regional power-law equations for median river width and depth. Width: RF R² = 0.75, XGBoost R² = 0.73, MLP R² = 0.73; depth: RF R² = 0.67, XGBoost R² = 0.67, MLP R² = 0.66. Regional power-law equations gave R² = 0.45 for width and R² = 0.18 for depth; multiple linear regression was worse for width (R² = 0.33).
  • Median river discharge was the most important predictor in the Random Forest model (RFI width = 0.67, RFI depth = 0.78), with catchment area second-most important for width. Simple regressions showed discharge explained width R² = 0.61 and depth R² = 0.57; catchment area explained width R² = 0.36 and depth R² = 0.25; upstream dams explained width R² = 0.21 and depth R² = 0.16, though dam RFI was only 0.01.
  • Performance varied by stream order and ecoregion. Power-law models overall underpredicted width by about 24% and depth by about 22%; ML models had overall under/overprediction of 7% or less for width and 13% or less for depth. For width, power-law performed best at 4th- and 5th-order streams (PBIAS = 8.9% and 4.6%) but underpredicted higher-order widths by 50% or more. For depth, power-law overpredicted 4th-order streams (median PBIAS = 8.5%) but underpredicted 1st-order (22.5%) and 9th-order (54.23%) streams. For 1st-order width, median PBIAS was RF = 22.6%, XGBoost = 30.0%, MLP = 28.9%. All ML width models performed well in Southeastern US plains (ER 8.3, NSE > 0.84), while RF vs. MLP differences reversed by region and by width vs. depth.
  • STREAM-geo provides predictions for 2,610,904 of 2,691,339 NHDPlus reaches, with remaining reaches marked NaN. It includes river width, depth, width/depth ratio, and %RSSA. Across CONUS, streams and rivers cover nearly 75,000 km², or 0.93% of the continental US. Median width/depth ratio ranged from 13.0 for 1st-order streams to 48.4 for 10th-order streams, plateauing above 8th order. More than 96.9% of US stream reaches are narrower than 50 m.

Methods

  • Data: HYDRoSWOT USGS ADCP observations (220,000+ records, >5,000 stations, 1940s–2014). Cleaning removed per-site width/depth values above the 95th or below the 5th percentile, NaNs, and sites with fewer than 6 data points; median width and depth were computed per site. Final training data included 5,269 reaches for width and 2,678 reaches for depth. Seventeen stream and catchment predictors were used, including climate, streamflow, NHDPlus reach metadata, bed D50, land use, vegetation, soils, population density, and dams; HYDRoSWOT locations were matched to the closest NHDPlus reach within a 500-m buffer.
  • Models: regional power-law equations for 20 Hydrologic Landscape Regions, multiple linear regression, Random Forest, XGBoost, and multilayer perceptron (MLP). Data were split 75:25 training/testing. Evaluation metrics were RMSE, PBIAS, R², and NSE. Hyperparameters were optimized with GridSearchCV and 5-fold cross-validation; the MLP used 100 hyperparameter sets, Adam optimization, MSE loss, and early stopping. Random Forest predictor importance was quantified by permutation-based MSE reduction.

Limitations

  • The paper states that HYDRoSWOT sampling is uneven, with more intensive sampling in the eastern US and more width than depth data; training data disproportionately represent 4th- and 5th-order streams while 1st-order streams are under-sampled. Depth predictions were less accurate than width predictions, multiple linear regression performed poorly for width, and reasons for ecoregion differences were not fully understood. Bedrock data are unavailable at national scale, so slope and elevation act as proxies. SWOT observes only rivers wider than 100 m, potentially as small as 50 m, and has limited temporal sampling, while >96.9% of US reaches are narrower than 50 m.
  • Evident from the text: the models predict median width and depth under median-flow conditions, not full distributions or event-based dynamics. Performance varies substantially by stream order and region, so a single model may be unreliable, especially for low-order streams.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1029/2023wr036733

Cites

69 of 69 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Observed river depth and width data across the contiguous US. (a) Schematic of river width and depth definitions, as obtained from HYDRoSWOT; (b, c): the spatial distribution of measured sites and the number of records per site for river width and depth; (d, e): frequency plots of river width and depth.
Figure 1 Observed river depth and width data across the contiguous US. (a) Schematic of river width and depth definitions, as obtained from HYDRoSWOT; (b, c): the spatial distribution of measured sites and the number of records per site for river width and depth; (d, e): frequency plots of river width and depth.CC-BY-NC-ND · From the PDF
Figure 2 Modeling approaches predicting stream channel width and depth. Note that while traditional approaches use only watershed area and streamflow (discharge) to predict channel width and depth, our data-driven, machine-learning approaches use 17 different predictors.
Figure 2 Modeling approaches predicting stream channel width and depth. Note that while traditional approaches use only watershed area and streamflow (discharge) to predict channel width and depth, our data-driven, machine-learning approaches use 17 different predictors.CC-BY-NC-ND · From the PDF
Image 3
CC-BY-NC-ND · From the PDF
Figure 3 Relative predictor importance of variables used in the machine-learning models, as determined from the random forest model. Predictor importance is shown for both (a) median river width and (b) median river depth.
Figure 3 Relative predictor importance of variables used in the machine-learning models, as determined from the random forest model. Predictor importance is shown for both (a) median river width and (b) median river depth.CC-BY-NC-ND · From the PDF
Image 5
CC-BY-NC-ND · From the PDF
Figure 4 Model performance results for the river width and depth predictions: (a) NSE, (b) R 2 , and (c) PBIAS.
Figure 4 Model performance results for the river width and depth predictions: (a) NSE, (b) R 2 , and (c) PBIAS.CC-BY-NC-ND · From the PDF
Image 7
CC-BY-NC-ND · From the PDF
Figure 5 Variations in model performance as a function of stream order using the traditional power-law approach (a, b) as well as the random forest (c, d), XGBoost (e, f), and neural network (MLP) (g, h) models based on PBIAS values.
Figure 5 Variations in model performance as a function of stream order using the traditional power-law approach (a, b) as well as the random forest (c, d), XGBoost (e, f), and neural network (MLP) (g, h) models based on PBIAS values.CC-BY-NC-ND · From the PDF
Image 9
CC-BY-NC-ND · From the PDF
Figure 6 Stream order distributions. Panel (a) shows the distribution of stream orders for all available width data. Panel (b) shows the distribution for all available depth data. Panel (c) shows the distribution of stream orders for all rivers and streams across the contiguous U.S. Note that the sampling data represents an over-sampling of middle-order streams, while 1st-order streams are under-sampled.
Figure 6 Stream order distributions. Panel (a) shows the distribution of stream orders for all available width data. Panel (b) shows the distribution for all available depth data. Panel (c) shows the distribution of stream orders for all rivers and streams across the contiguous U.S. Note that the sampling data represents an over-sampling of middle-order streams, while 1st-order streams are under-sampled.CC-BY-NC-ND · From the PDF
Image 11
CC-BY-NC-ND · From the PDF
Figure 7 Spatially varying model performance to predict channel width and depth. Figure panels show Nash-Sutcliffe Efficiency values for the four different modeling approaches: (a, b) power law; (c, d) Random Forest; (e, f) XGBoost; (g, h) MLP; (i) Zone II North American Ecoregions (ERs).
Figure 7 Spatially varying model performance to predict channel width and depth. Figure panels show Nash-Sutcliffe Efficiency values for the four different modeling approaches: (a, b) power law; (c, d) Random Forest; (e, f) XGBoost; (g, h) MLP; (i) Zone II North American Ecoregions (ERs).CC-BY-NC-ND · From the PDF
Image 13
CC-BY-NC-ND · From the PDF
Figure 8 Differences in performance between the random forest and multilayer perceptron models. The maps show the difference in NSE values, ΔNSE, between the two models for (a) width and (b) depth. Darker blue represents superior performance for the multilayer perceptron (MLP) models, darker red represents superior performance for the random forest (RF) models, and white or light colors suggest little or no difference in model performance.
Figure 8 Differences in performance between the random forest and multilayer perceptron models. The maps show the difference in NSE values, ΔNSE, between the two models for (a) width and (b) depth. Darker blue represents superior performance for the multilayer perceptron (MLP) models, darker red represents superior performance for the random forest (RF) models, and white or light colors suggest little or no difference in model performance.CC-BY-NC-ND · From the PDF
Image 15
CC-BY-NC-ND · From the PDF
Figure 9 STREAM-geo modeled stream geometry results for the contiguous US. The U.S. maps show the median river (a) width and (d) depth at the scale of individual stream reaches. (b, e) River width and depth distributions across the CONUS. (c, f) River width and depth distributions across stream orders.
Figure 9 STREAM-geo modeled stream geometry results for the contiguous US. The U.S. maps show the median river (a) width and (d) depth at the scale of individual stream reaches. (b, e) River width and depth distributions across the CONUS. (c, f) River width and depth distributions across stream orders.CC-BY-NC-ND · From the PDF
Figure 10 STREAM-geo river width/depth ratios and percent river surface area (%RSSA) values. The U.S. maps show (a) the river width/depth ratios at the scale of individual NHDPlus stream reaches and (d) %RSSA values at the scale of HUC8watersheds. (b) River width/depth distributions across the CONUS. (c) River width/depth ratios across stream orders. (e) %RSSA distributions across HUC8 watersheds.
Figure 10 STREAM-geo river width/depth ratios and percent river surface area (%RSSA) values. The U.S. maps show (a) the river width/depth ratios at the scale of individual NHDPlus stream reaches and (d) %RSSA values at the scale of HUC8watersheds. (b) River width/depth distributions across the CONUS. (c) River width/depth ratios across stream orders. (e) %RSSA distributions across HUC8 watersheds.CC-BY-NC-ND · From the PDF

Verification

Claims the paper makes

0 Valid · 0 Not valid · 19 Not checked · 64 Own results and interpretations

35 more claims

Create a free account to see all 83 claims and to verify the 26 references still open.

See all claims — free account
  • It may state two separate claims as one. 33
  • Probably not a substantial claim. 14

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Research works printed but not linked: 2

  1. Chollet F., 2021, Deep learning with Python
  2. 10.48550/arxiv.2301.04027 · arXiv · doi:10.48550/arxiv.2301.04027

Cited by

See all 23

Metadata

Authors
Chang, S. Y., Ghahremani, Z., Manuel, L., Erfani, S. M. H., Shen, C., Cohen, S., Van Meter, K. J., Pierce, J. L., Meselhe, E. A., & Goharian, E.
Affiliations
  • Pennsylvania State University (US)
  • Boise State University (US)
  • Tulane University (US)
  • University of South Carolina (US)
  • University of Alabama (US)
  • Department of Geography Pennsylvania State University University Park PA USA
Corresponding author
  • Shen, Chaopeng
  • Cohen, Sagy
Venue
Water Resources Research
Published
2024-10
Volume
60
Issue
10
ISSN
  • 0043-1397 (ISSN-L)
  • 1944-7973
OpenAlex
W4403686990
Cited by
23
References
69
Citations by year
2026: 8 · 2025: 14 · 2024: 1
FWCI
4.98
Type
journal article
Language
en
Open access
Locations
  • Water Resources Research (publishedVersion, cc-by-nc-nd)
  • arXiv (Cornell University) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • Consortium of Universities for the Advancement of Hydrologic Science
  • National Oceanic and Atmospheric Administration (A22-0307-S003)
  • National Water Center, United Arab Emirates University
Topics
  • Physical Sciences › Earth and Planetary Sciences › Earth-Surface Processes › Hydrology and Sediment Transport Processes
  • Physical Sciences › Environmental Science › Global and Planetary Change › Flood Risk Assessment and Management
  • Physical Sciences › Environmental Science › Water Science and Technology › Hydrology and Watershed Management Studies
Keywords
river geometry, hydraulic geometry, river width, river depth, power-law relationship, multi-model ensemble, random forest, XGBoost, artificial neural networks, flood inundation, contiguous United States

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-19, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-19 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.