DACLnet: A Dual-Attention-Mechanism CNN-LSTM Network for the Accurate Prediction of Nonlinear InSAR Deformation

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a dual attention mechanism CNN-LSTM network (DACLnet) for accurately predicting nonlinear InSAR surface deformation. The paper describes the construction and methodology of DACLnet and its application in areas of excessive groundwater extraction in the Turpan Basin. The authors claim that this model performs excellently in predicting nonlinear surface deformation. However, the paper needs improvement in structure, data processing methods, and certain analysis results. The following are detailed suggestions for revisions.

 

1. The abstract should be more concise and clear, avoiding repetitive information. It is recommended to succinctly describe the research methods, key results, and application prospects. The introduction section needs to systematically introduce the background and importance of nonlinear surface deformation, clearly state the deficiencies of current research, and highlight the innovations of this paper.

 

2. When describing the DACLnet model structure, the working principles of the dual attention mechanism should be explained in more detail, accompanied by clearer diagrams. In the data processing section, the improvements of the IPTA-InSAR technique should be described in more detail, and the specific steps and reasons for data preprocessing should be explained.

 

3. The proportion of training and test sets should be clearly stated, and the basis for selecting this proportion should be explained. In the network training section, the basis for selecting all key parameters should be listed, and the reasons for selecting these parameters and their impact on the model's performance should be explained.

 

4. It is recommended to re-examine the results section. The results section should systematically display the predictive performance of the DACLnet model, providing more quantitative metrics such as MAE, RMSE, and MAPE, and comparing them in detail with other models. Charts should have detailed captions explaining the specific meaning and data sources of each figure. In the discussion section, the advantages and limitations of the DACLnet model should be analyzed more deeply, discussing the applicability of the model in different scenarios, and suggesting future research directions.

 

5. In the past decade, InSAR technology has rapidly developed and has become one of the main technical means for surface deformation monitoring, with relevant theories and applications becoming very mature. Commercial or open-source software like ENVI SARscape already provides relatively mature and reliable processing modules, and the future development of InSAR technology relies more on the advancement of SAR satellite resolution. Recent research rarely performs correlation analysis and more often discusses surface subsidence detected by InSAR. However, the authors used ascending and descending track data (AT41F135, DT121F449) to obtain deformation in the study area. From Figure 4 and Figure 5, the monitoring results are basically consistent. The ascending and descending track data have differences in heading, incidence, observation time, and influences from the ionosphere and troposphere, which theoretically should result in significant differences in deformation monitoring. Therefore, it is suggested to retain only the results of either the ascending or descending track and remove the correlation analysis (e.g., Figure 5). The focus of the paper should be on the introduction of data processing (explaining the preprocessing software and main processes used for the time series) and strengthening the analysis and interpretation of deformation.

 

6. The paper briefly analyzes the cumulative subsidence of P1 and P2; however, overall, the length of the paper is short (with only 9 figures and 3 tables). The intermediate process data figures for point target LSTM training, testing, and prediction are not clearly provided in the paper. It is recommended to enhance the content of the paper, increase the analysis, and even consider making the code public to enhance the credibility of the paper. Additionally, research on deep learning for InSAR deformation prediction has been increasing in recent years. To avoid homogenization, it is suggested to add analysis of time series deformation, study area subsidence conditions, and consider the impact of factors such as hydraulic engineering, groundwater extraction, surface water evaporation, and salt accumulation on deformation cycles, and even consider establishing related models.

Comments on the Quality of English Language

Moderate editing of English language required

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This paper examine the spatiotemporal deformation based on IPTA-InSAR method in Turpan Basin and utilized InSAR derived result to predict nonlinear deformation based on DACLnet. Generally, the manuscript was well organized and clearly stated. Following are some aspects or weakness that the paper did not stated very clearly and should be strenthened:

1.     Were the 574,662 deformation time series selected from the ascending deformation result or the descending result or both?

2.     Since the non-linear deformation in the research area has the periodic and seasonal characteriscis, if the training set was generated randomly from the initial 70% of the 5748 deformation data, how can the training set retain the periodic and seasonal characteriscis of the original data? The predicted deformation was only spanning a very short time period (from Feb. 27 2020 to April 27, 2020). It is hard to see that the prediction captured the spatio-temporal characteristics of the non-linear deformation, let alone the periodic and seasonal changes of the deformation.

3.     The correlation between the observed and DACLnet simulated deformaion varied in different time spans according to Figure 7, yet the paper did not analyze or discuss this inconsistency.

4.     As we know that deformation derived from one InSAR set can only capture deformation occurred in LOS direction, while the actuall deformation might occur in three dimensions (vertical, horizontal and north-south directions). Therefore, the prediction result might not reflect the real deformation. It is recommended to integrate ascending and descending InSAR results, as well as the GNSS data to train the model and capture the actual deformation characteristics in the rearch area.

Comments on the Quality of English Language

The English Language could be polished for publicaiton. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This study developed a CNN-LSTM network to predict nonlinear InSAR deformation. The proposed model is validated by comparing it with other ML models using a case study. Generally speaking, this work is well-written and organized. However, before recommending publication, the following issues should be clarified:

1. What is nonlinear InSAR deformation? It should be clearly explained in the introduction section.

2. There are too many abbreviations, and a list is recommended.

3. More details about the theory on InSAR data detecting surface deformation are recommended to be introduced.

4. What are the input parameters of the other ML methods used in this study? What is the difference compared with the proposed model?

5. The effects of earthquakes and underground mining on ground deformation should also be mentioned in the introduction section. The related works, titled “An energy‐frequency parameter for earthquake ground motion intensity measure” and “Main frequency band of blast vibration signal based on wavelet packet transform” are recommended for citation.

6. Figure 7: Could the authors provide longer prediction values? While there is no recorded data as a control group, it could provide more details about the performance of the model in prediction.

7. The limitations of the proposed method are suggested to be explained in the discussion section instead of the final paragraph of the conclusion. The applicability of the proposed model could also be discussed in the discussion section.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Moderate editing of English language required

Comments on the Quality of English Language

Moderate editing of English language required

Author Response

Thanks for your comments and suggestions on our manuscript. We have addressed the concern about the English language by utilizing one of the recommended editing services listed at https://www.mdpi.com/authors/english. We believe these revisions have significantly improved the clarity and readability of the manuscript. Thank you once again for your time and efforts in evaluating our manuscript.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have adequately addressed my concerns. 

Author Response

Thanks for your positive comment on our revised manuscript. We appreciate your acknowledgment of our efforts.