Groundwater level forecasting with artificial neural networks:  a comparison of long short-term memory (LSTM),  convolutional neural networks (CNNs), and non-linear  autoregressive networks with exogenous input (NARX)

Wunsch, Andreas; Liesch, Tanja; Broda, Stefan

doi:https://doi.org/10.5194/hess-25-1671-2021

Articles | Volume 25, issue 3

https://doi.org/10.5194/hess-25-1671-2021

© Author(s) 2021. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/hess-25-1671-2021

© Author(s) 2021. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 25, issue 3

Research article

|

01 Apr 2021

Research article |

| 01 Apr 2021

Groundwater level forecasting with artificial neural networks: a comparison of long short-term memory (LSTM), convolutional neural networks (CNNs), and non-linear autoregressive networks with exogenous input (NARX)

Andreas Wunsch, Tanja Liesch, and Stefan Broda

Download

Final revised paper (published on 01 Apr 2021)
Supplement to the final revised paper
Preprint (discussion started on 23 Nov 2020)
Supplement to the preprint

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'The Rabbit and the Turtle: Or, an LSTM, a CNN and a NARX.', Anonymous Referee #1, 23 Dec 2020
- AC1: 'Reply on RC1', Andreas Wunsch, 15 Jan 2021
RC2: 'Groundwater Level Forecasting with Artificial Neural Networks: A Comparison of LSTM, CNN and NARX', Anonymous Referee #2, 24 Dec 2020
- AC2: 'Reply on RC2', Andreas Wunsch, 15 Jan 2021

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

ED: Reconsider after major revisions (further review by editor and referees) (31 Jan 2021) by Mauro Giudici

AR by Andreas Wunsch on behalf of the Authors (01 Feb 2021) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (01 Feb 2021) by Mauro Giudici

RR by Anonymous Referee #2 (16 Feb 2021)

RR by Anonymous Referee #1 (01 Mar 2021)

Suggestions for revision or reasons for rejection

Dear authors,

I compliment you for the adaption of the manuscript. I think it has indeed become a better overall contribution. I do still have some comments remaining, which are listed below:

L51. Sentence reads a bit strange due to the double usage of well. A possible adaptation might be something like: “Overall, these studies show that both, LSTMs and CNNs, are very well suited to forecast groundwater …”
L111ff. Still would recommend to use “sequence to one” here and throughout the manuscript.
L169f/L.189/L.198. I would write “squared Pearson’s correlation coefficient” here instead of the lower case r. I am aware that it is not always the case, but usually the capitalization of the R is used to differentiate between the computation of the correlation between a predicant and a predictor (lower case r) and the correlation between an estimate and a observation (capitalized R).
L.198. I would recommend the statement about the coefficient of determination stronger and more general (in its current version the statement about the simple linear model is not precise, since the estimations are not derived ). My recommendation would be something like the following: “In our case, we use the squared Pearson correlation coefficient R^2 as a general coefficient of determination, since it compares the linear relation between simulated and observed GWLs.”
L.259. I would not undersell the finding here. Maybe replace “not surprising” with “expected”.
L.341-L.344. I do not understand these sentences. Especially the last one. I would recommend rewriting them, since it is not clear what exactly is computed “at once”. Similarly, if the difference in behavior is caused by the different modelling approach, then why not choose the same setup for the models? If I understand the model descriptions correctly, I see no reason why this was not possible.
L.401. “Models” is used twice in “If accuracy is especially important, one should stick with NARX models models”.
L. 418f. The two sentences are formulated in a strange manner since it does read as if CNNs in general are not Deep Learning approaches. I would recommend slightly restructuring the sentences.
L.424. This is a strong conclusion. Please try to keep it no matter what.

Hide

ED: Publish subject to technical corrections (02 Mar 2021) by Mauro Giudici

AR by Andreas Wunsch on behalf of the Authors (03 Mar 2021) Manuscript