Using Long Short-Term Memory networks to connect water table depth anomalies to precipitation anomalies over Europe

Ma, Yueling; Montzka, Carsten; Bayat, Bagher; Kollet, Stefan

doi:https://doi.org/10.5194/hess-25-3555-2021

Articles | Volume 25, issue 6

https://doi.org/10.5194/hess-25-3555-2021

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

https://doi.org/10.5194/hess-25-3555-2021

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

Articles | Volume 25, issue 6

Research article

|

23 Jun 2021

Research article |

| 23 Jun 2021

Using Long Short-Term Memory networks to connect water table depth anomalies to precipitation anomalies over Europe

Yueling Ma, Carsten Montzka, Bagher Bayat, and Stefan Kollet

Download

Final revised paper (published on 23 Jun 2021)
Preprint (discussion started on 24 Aug 2020)

Interactive discussion

Status: closed

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

- Printer-friendly version

- Supplement

RC1: 'comment hess-2020-382', Anonymous Referee #1, 21 Sep 2020
- AC1: 'Comment_responses_referee1_Ma_et_al', Yueling MA, 15 Nov 2020
RC2: 'Review of the manuscript 2020-382 “Using Long Short-Term Memory networks to connect water table depth anomalies to precipitation anomalies over Europe” by Ma et al.', Anonymous Referee #2, 19 Oct 2020
- AC2: 'Comment_responses_referee2_Ma_et_al', Yueling MA, 15 Nov 2020

Peer-review completion

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

ED: Publish subject to revisions (further review by editor and referees) (18 Nov 2020) by Zhongbo Yu

AR by Yueling Ma on behalf of the Authors (22 Dec 2020) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (28 Dec 2020) by Zhongbo Yu

RR by Anonymous Referee #1 (02 Jan 2021)

RR by Anonymous Referee #3 (09 Feb 2021)

Suggestions for revision or reasons for rejection

Suggestions for revision or reasons for rejection
The article presents the development of an artificial neural network that aims at connecting water table depth anomalies to precipitations anomalies. The topic is of interest. However, I have some doubts on the study.
The main one is that this neural network is constructed and assessed based on the simulation of a spatially distributed model. Therefore, there is no assessment on real observations. Moreover, it is not clear which is the advantage of using the neural network compares to the spatially distributed model. I guess it is the computational time, but, it is not clearly stated.
Additionally, only the water table depth anomaly is estimated. But what is the meaning of it? Can we expect similar situation in two places with a similar water table depth anomaly? A focus is made on drought, but, what is the general relationship between water table depth anomaly (wtda) and drought? ll these have to be explained and demonstrated
Special comments
Introduction: it is not clear how the water table depth anomaly is computed. Please explains and/or refers to equation 7… same for precipitation anomaly.
Section 2.1: the scheme presented as a complete subsurface water balance is in fact a rough simplification. There is only one aquifer layer, no connection with the rivers, no groundwater abstraction…
Section 2.2 I am not a neural network expert and would have appreciated a justification of the best suitability of LSTMs. Do other neural networks have memory?
Section 2.3 I’m not convinced of the added value of the wavelet transform.
Section 2.4 Figure 3 shows the mean water table depth simulated by the spatially distributed model on average on 20-year. From this figure, it appears clearly that rivers have a significant impact although they are not taken into account in scheme 1. Moreover, the range of value is huge, and this figure is very difficult to read.
Section 3:.1 Why do you focus on wtda>1.5 ? What do you expect it means? Aquifer level is not completely free to vary: a lower limit can be fixed by a bedrock, and the top surface (topography) can be the upper limit. Thus, I’m not sure that wtda>1.5 represent the same situation everywhere… The figure 5 is very pale, and it’s hard to see something…
Section 3.2 I’m not sure to be able to interpret figure 6. Indeed, we have no idea which percent of the domain is represented by each x-axis value. The pdf of R2 by region will have been more interesting
Figure 7 is also hard to read… Figure 8: percentage of the pixel that have a negative evapotranspiration. I guess there are not numerous pixels… How can we interpret this figure?
Section 3.3 again, not sure there is an added value of the wavelet transform..

Hide

ED: Publish subject to revisions (further review by editor and referees) (14 Feb 2021) by Zhongbo Yu

AR by Yueling Ma on behalf of the Authors (10 Mar 2021) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (11 Mar 2021) by Zhongbo Yu

RR by Anonymous Referee #1 (16 Mar 2021)

RR by Anonymous Referee #4 (16 May 2021)

Suggestions for revision or reasons for rejection

Recommendation: Major Revisions
In this manuscript, the authors tried to connect the water table depth anomalies to precipitation anomalies over Europe using Long Short-Term Memory networks based on the daily terrestrial simulations of the Terrestrial Systems Modeling Platform (TSMP) over Europe. The proposed method has the ability to reproduce the TSMP-G2A wtda maps. The authors also hypothesize that the proposed networks could simulate the high-frequency components of wtda, This manuscript is clear in construction and easy to follow. However, there are several concerns which need to be well addressed before the paper could be accepted.

Major Concerns:
1. What are the advantages of LSTM networks over TSMP, since TSMP could generate daily wtda time series. The authors need to give more justifications on LSTM networks.
2. The LSTM networks were trained based on the simulated outputs rather on the observed results. In my opinion, the authors should train the networks using observed data and compared the simulated results with other numerical models to demonstrate the capability of the proposed method.
3. The authors hypothesize that the proposed networks could simulate the high-frequency components of wtda by comparing the results from cross-wavelet transform analysis made on the outputs from both LSTM and TSMP. I am confused why the authors use cross-wavelet transform analysis to evaluate the performance, I don’t think the XWT is a good method for evaluation. The authors could provide more explanations. Meanwhile, the main concern may be the ability to simulate low frequency variations of wtd caused by extreme events such as long-term drought.

Minor coments:
1. Line 113-114: no confined aquifer?
2. For Figure 5, the authors may provide a anomaly map between wtda from TSMP-G2A and LSTM as it is really hard to see the difference.
3. The legend of Figure 6 should be put at right position.

Referee Report: PDF

Hide

ED: Publish subject to minor revisions (review by editor) (20 May 2021) by Zhongbo Yu

AR by Yueling Ma on behalf of the Authors (21 May 2021) Author's response Author's tracked changes Manuscript

ED: Publish as is (22 May 2021) by Zhongbo Yu

AR by Yueling Ma on behalf of the Authors (31 May 2021) Author's response Manuscript

Short summary

This study utilized spatiotemporally continuous precipitation anomaly (pr_a) and water table depth anomaly (wtd_a) data from integrated hydrologic simulation results over Europe in combination with Long Short-Term Memory (LSTM) networks to capture the time-varying and time-lagged relationship between pr_a and wtd_a in order to obtain reliable models to estimate wtda at the individual pixel level.