the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Development of a machine learning model for river bedload
Claire Masteller
Colin Phillips
Abstract. Prediction of bedload sediment transport rates in rivers is a notoriously challenging problem due to inherent variability in river hydraulics and channel morphology. Machine learning offers a compelling approach to leverage the growing wealth of bedload transport observations towards the development of a data driven predictive model. We present an artificial neural network (ANN) model for predicting bedload transport rates informed by 8,117 measurements from 134 rivers. Inputs to the model were river discharge, flow width, bed slope, and four bed surface sediment sizes. A sensitivity analysis showed that all inputs to the ANN model contributed to a reasonable estimate of bedload flux. At individual sites, the ANN model was able to reproduce observed sediment rating curves with a variety of shapes and outperformed four standard bedload models. This ANN model has the potential to be broadly applied to predict bedload fluxes based on discharge and reach properties alone.
- Preprint
(1594 KB) -
Supplement
(1473 KB) - BibTeX
- EndNote
Hossein Hosseiny et al.
Status: final response (author comments only)
-
RC1: 'Comment on esurf-2022-23', Anonymous Referee #1, 23 May 2022
Dear Editor
Paper is interesting and well-written, but my main issue is regarding the lack of novelty. Authors just applied old model of ANN to bedload prediction. We have many new models like tree-based, rule based,... deep learning and so on. Due to lack of novelty, I have to reject the paper. Sorry for the decision.
Best
Citation: https://doi.org/10.5194/esurf-2022-23-RC1 -
RC2: 'Reply on RC1', Basil Gomez, 25 May 2022
- AC2: 'Reply on RC2', Claire Masteller, 24 Aug 2022
- AC1: 'Reply on RC1', Claire Masteller, 24 Aug 2022
-
RC2: 'Reply on RC1', Basil Gomez, 25 May 2022
-
CC1: 'Comment on esurf-2022-23', Xingyu Chen, 26 May 2022
The paper is well-written. A highligh of the study is the use of large dataset. The value of the study can be better addressed if the author better explore how the increase of the dataset size (from 500 to 8000+) can help improve the understanding of bedload transport. There must be a lot of information contained in the 8000+ bedload measurements that can be helpful in explaining the variability of bedload transport.
Citation: https://doi.org/10.5194/esurf-2022-23-CC1 -
AC3: 'Reply on CC1', Claire Masteller, 24 Aug 2022
Thank you for your comment and feedback.
We agree that the large database being used in this study provides a versatile model for bedload predictions. The predictions in this study are within the bounds of one order of magnitude (Fig.2), with a mean absolute error of 16.1 g/s/m . Previous study of Bhattacharya & Solomatine (2006) has shown that even with 407 observations, the machine learning model outperformed several empirical and physically based bedload models. However, they reported that the root mean squared error (RMSE) of the field-based ANN model was 68.4 x 10-5 m2/s (~1812.6 g/m/s; three orders of magnitude larger).
We would like to highlight that the novelty of this work is not solely based on using a large dataset. The proposed model in this study uniquely utilizes a number of easy-to-measure reach-scale variables along with discharge as the inputs to predict bedload flux, which in an of itself represents an advance.Â
Bhattacharya, B., Price, R. K., & Solomatine, D. P. (2007). Machine Learning Approach To Modeling Sediment Transport. Journal of Hydraulic Engineering, 133(4), 776–793. https://doi.org/10.1061/(ASCE)0733-9429(2007)133:4(440)
Citation: https://doi.org/10.5194/esurf-2022-23-AC3
-
AC3: 'Reply on CC1', Claire Masteller, 24 Aug 2022
-
RC3: 'Comment on esurf-2022-23', Anonymous Referee #3, 26 May 2022
The authors use an ANN to derive a nonlinear relationship between several river hydraulic variables and sediment transport rate. This machine learning technique and many others have been extensively tested in the last twenty years in many similar studies that the authors overlooked. Thus, I do not think that this study adds something new to the existing literature and I suggest rejection of the paper.
Citation: https://doi.org/10.5194/esurf-2022-23-RC3 - AC4: 'Reply on RC3', Claire Masteller, 24 Aug 2022
-
RC4: 'Comment on esurf-2022-23', Anonymous Referee #4, 07 Jul 2022
In this short and well-written manuscript, the authors present an ANN model for predicting bedload flux based on a published dataset. Machine learning is increasingly used for modeling and predicting natural dynamics, with known strengths and limitations. Bedload is perhaps one of the more challenging processes to model given its strong dependency on highly dynamic and local variables. A number of models have recently been published that attempt to predict bedload over large scales (continental and global; see below). This paper is therefore quite timely and adds to the broader communities' efforts to better predict fluvial dynamics. The following issues should be addressed before it is accepted for publication. These are not very major issues but will likely require additional analysis.
1. The observational dataset includes an unequal number of observations for each river - if the spatial variability is larger than the temporal variability this may lead to overfitting. The authors addressed that to a degree, but need to better discuss this issue. As it stands the model predicts temporal dynamics using observations from different rivers. ANN may be flexible enough to deal with this but, again, needs more discussion and maybe an additional analysis using some sort of average value for each site (regression may be more suitable in this case given the small sample size).
2. The removal of outliers is overall acceptable but can be very problematic when using a fluvial dataset as the 'extreme' values are often just the few large rivers in a dataset. The authors warn the reader to only use/interpret the results within the range of the variables but they should more carefully examine the outliers and try to include realistic observations and maximize the dataset (and thus model) representation of large rivers.
3. The metrics selected for representing the models' accuracy are reasonable but need some justification. Why MSE and not RMSE or PBIAS or R2?
4. The paper falls short in providing tools and guidelines for applying its outcomes. The paper's main outcome is to demonstrate the potential usefulness of ANN for modeling bedload flux. How can the reader use this knowledge moving forward? Will they have to develop their own ANN based on the dataset? How can it be used for other locations (as the authors suggested)? This is a common issue with ML modeling, but the authors can mitigate it with additional descriptions and tools (e.g. scripts).Â
5. The authors are encouraged to explore recently published papers such as:
Cohen, S., Syvitski, J., Ashely, T., Lammers, R., Fekete, B., & Li, H. Y. (2022). Spatial Trends and Drivers of Bedload and Suspended Sediment Fluxes in Global Rivers. Water Resources Research, e2021WR031583.
Gomez, B., & Soar, P. J. (2022). Bedload transport: beyond intractability. Royal Society Open Science, 9(3), 211932.ÂLammers, R. W., & Bledsoe, B. P. (2018). Parsimonious sediment transport equations based on Bagnold's stream power approach. Earth Surface Processes and Landforms, 43(1), 242-258.ÂLi, H. Y., Tan, Z., Ma, H., Zhu, Z., Abeshu, G. W., Zhu, S., ... & Leung, L. R. (2022). A new large-scale suspended sediment model and its application over the United States. Hydrology and Earth System Sciences, 26(3), 665-688.ÂTan, Z., Leung, L. R., Li, H. Y., & Cohen, S. (2022). Representing global soil erosion and sediment flux in Earth System Models. Journal of Advances in Modeling Earth Systems, 14(1).Citation: https://doi.org/10.5194/esurf-2022-23-RC4 - AC5: 'Reply on RC4', Claire Masteller, 24 Aug 2022
-
EC1: 'Comment on esurf-2022-23', Rebecca Hodge, 11 Jul 2022
This paper has received a range of reviewer comments. I would be happy to see a response from the authors indicting how they will address these reviews. The response needs to show how the work will be modified to address the comments and/or to provide a well justified rebuttal. I suggest that the authors focus most on the two sets of more substantial comments from Gomez and Referee #4. One theme is that the authors have not engaged with some recent literature that is of direct relevance to this work. The authors will need to consider the papers suggested by the reviewers and make it clear that their work has taken into account the ideas in those papers. Some of the other reviewers suggest that this work is not novel compared to existing literature, but unfortunately do not identify specific papers that the authors have missed. I encourage the authors to ensure that they have thoroughly searched the literature for relevant work. Some of the reviewers also question whether further analysis is necessary. The authors need to justify their choice of method, both in terms of the machine learning approach and in the selection and application of the bedload transport dataset.
Citation: https://doi.org/10.5194/esurf-2022-23-EC1 -
AC6: 'Reply on EC1', Claire Masteller, 24 Aug 2022
We thank the AE for their summary of the review. We have responded in detail to the reviews and comment individually above.Â
Citation: https://doi.org/10.5194/esurf-2022-23-AC6
-
AC6: 'Reply on EC1', Claire Masteller, 24 Aug 2022
Hossein Hosseiny et al.
Data sets
Bedload data set Hossein Hosseiny, Claire Masteller, Colin Phillips https://docs.google.com/spreadsheets/d/1TeGFcRfFqCaD-8keugCDIBftpl0Mxpz5/edit?usp=sharing&ouid=113425085155864679118&rtpof=true&sd=true
Hossein Hosseiny et al.
Viewed
HTML | XML | Total | Supplement | BibTeX | EndNote | |
---|---|---|---|---|---|---|
748 | 232 | 27 | 1,007 | 68 | 5 | 2 |
- HTML: 748
- PDF: 232
- XML: 27
- Total: 1,007
- Supplement: 68
- BibTeX: 5
- EndNote: 2
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1