Predicting Surface Subsidence in Northern Huainan Based on a Hybrid LSTM–Transformer Model

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Summary: The paper proposes a hybrid LSTM–Transformer architecture for predicting mining-induced subsidence using SBAS-InSAR data. The paper describes the significance of the problem and proposes an automated solution to forecast subsidence rates. Some evidence is provided to demonstrate efficiency of the solution.

Some things that can be improved:

• In Section 3, LSTM and attention are described in unnecessary detail. These are well-known concept, and authors could refer instead to the original papers instead.
• In Section 4, the input data used for training/analysis is not described clearly. The section would benefit from a clear description of a dataset (# of data points used for training, dimensionality, size, illustration), including what a single data point looks like exactly.
• Section 4: the baseline is not clearly described. SBAS-InSAR is both a data processing technique and a baseline - how is it used?
• Sections 3 and 4 lack an explanation of the model architecture choice. There are simpler architectures that are known to be efficient on time series data, e.g. temporal CNNs, gradient boosting that the authors didn't seem to consider.
• Sections 3 and 4 would benefit from an ablation study backing up the choice of the model architecture (related to the previous point). Did the authors consider parts of the ensemble separately (LSTM only and transformer only)?
• Section 4.3. The choice of evaluation dataset needs some explanation. Why only 8 points were considered? How the points were chosen?

Overall, the problem looks relevant and the proposed solution seems to be efficient but the content lack some key details crucial for understanding design choices made by the authors.

Author Response

Thank you very much for your valuable comments. Please see the attachment for our responses.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

• The authors standardize the time step to 3‑day intervals and extend to a “360‑period” series. Could you spell out the exact resampling/interpolation scheme, any smoothing applied, and how the added autocorrelation was handled in the forecast task? Also, please report results using the native Sentinel‑1 sampling (no resampling) and confirm that the evaluation is strictly out‑of‑sample (e.g., rolling‑origin).

 

• I might be misreading, but the phrase “the final 7 periods, from January 6, 2022, to December 21, 2024” seems to span most of the study window. Please clarify the exact train/val/test indices or dates and adopt a forward‑only (chronological) split. A small table with start/end dates for each split would resolve the ambiguity immediately.

 

• SBAS‑InSAR is the monitoring source, not a predictor. Its “RMSE” is undefined without an external ground truth (leveling/GNSS). What the paper needs is the forecast error of your model versus held‑out SBAS observations (and, if available, against in‑situ data). I suggest removing “SBAS RMSE” from model comparisons and rewriting the section accordingly.

 

• Two different maxima appear for the same zone (e.g., −85.70 vs −83.94 mm/yr), and the urban maximum varies across sections. Please select the correct values, identify the figure/table that establishes them, and propagate the same numbers through Results, captions, and Conclusions so the story stays coherent.

 

• I was unable to locate several SBAS specifics required for reproduction: interferogram network criteria (temporal/spatial baselines), coherence mask threshold, filtering and unwrapping method, reference area/pixel, atmospheric correction strategy (including stratified troposphere), DEM source, and residual-topography handling, as well as LOS→vertical assumptions. Please add these to the Research Methodology so an independent group can reproduce the time series.

 

• The manuscript references mining / urban / decommissioned zones but does not document how these were derived. Please describe the features used, the clustering or classification algorithm, thresholds/hyperparameters, and any manual post‑editing. A small validation against independent land‑use or mining records (even if qualitative) would strengthen credibility.

 

• For the hybrid LSTM‑Transformer, could you report the complete configuration (g., depths, hidden sizes, attention dims/heads, dropout, positional encoding, optimizer/learning‑rate schedule, weight decay)? Also, please explain why different input windows were used by zone (e.g., 200 for mining vs 50 elsewhere) and provide a brief sensitivity check that performance is not an artifact of window length or epoch count.

Author Response

Thank you very much for your valuable comments. Please see the attachment for our responses.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

The study addresses an important global issue: predicting subsidence in mining areas. The integration of SBAS-InSAR with a hybrid LSTM-Transformer model is an interesting and promising approach. However, I recommend considering the following:

Introduction

• Strengthen the research gap statement. Currently the introduction does a good job describing the techniques (InSAR, SBAS, LSTM), but it doesn't make clear enough what specific aspect previous studies failed to address.
• Include more recent international references.
• Emphasize the global practical relevance.

Methodology

• Provide more detailed justification for the model parameters.
• If possible, add sensitivity tests or cross-validation.

Results

• If possible, present comparative tables or figures with previous studies.
• Don't just report subsidence values, explain what they mean for decision-making in mining cities.

Discussion and Conclusions

• Link the findings to real-world applications.
• Write clearer conclusions that focus on the contributions and limitations.

Author Response

Thank you very much for your valuable comments. Please see the attachment for our responses.

Author Response File: Author Response.docx

Reviewer 4 Report

Comments and Suggestions for Authors

Review of the manuscript titled “Predicting Surface Subsidence in Northern Huainan Based on a Hybrid LSTM-Transformer Model”, submitted to Applied Science.

- Avoid abbreviations in the title and in the keywords.

- Line 23: it is -57.42 mm/yr? In “The average subsidence … were -57.42 mm/yr, -5.37 mm/yr, …”.

- In the introduction, a review of literature is required to highlight and discuss some of the existing researches that deal with the application of InSAR, LSTM… 
- Line 295: delete the title “Continuation of Table 3. The parameters of the LSTM-Transformer model”. If a part of the table occurs on the next page, it is a continuation of the table and should not have an independent title.
- The last paragraph of your introduction should be devoted to highlighting only the objectives and not the results.
- In the introduction, I suggest discussing the utility of the geophysical methods as potential tools for subsidence detection. Here, I suggest some references for some geophysical methods (electric: https://doi.org/10.1155/2019/2565430 ; seismic: https://doi.org/10.62593/2090-2468.1044) 
- I suggest adding some real pictures showing the surface subsidence in the study area.
- You should show some examples of the Sentinel-1A images in the research data section.
- The formula and equations must be cited within the text. For example, the formula 1 (line 122) was presented suddenly without being mentioned before.
- Line 129: Is it C ̃_t or C ̃t?
- Line 165: What does the abbreviation “lstm_hidden_dim” mean?
- Please highlight the software used to execute the different operations.
- Figure 2 is not interpreted. The content and the different steps should be explained.
- Line 225: Why do you name this subheading as “Deformation Rate Analysis”? It is a subsidence rate analysis.
- Line 245: specify the figure number in “… as shown in the figure,…”.
- Figure 3: Use a distinctive color to highlight the 8 research points (A, B, C…). 
- Line 266: According to the table, the subsidence rate does not range from -50 mm/yr to -70 mm/yr. Provide the exact values.
- I ask if the process used allows delineating all the subsidence areas in the northern Huainan City.
- Table 3 and Figure 5 are not interpreted nor discussed. This point is very essential. It is evident to present data without being interpreted.
- Upon incorporating the recommended modifications, this paper is very important and has the potential to provide a significant contribution to the body of scientific literature.

Author Response

Thank you very much for your valuable comments. Please see the attachment for our responses.

Author Response File: Author Response.docx

Reviewer 5 Report

Comments and Suggestions for Authors

This paper explores a hybrid LSTM-Transformer model for use in monitoring and early warning of surface subsidence hazards. The paper is quite interesting, is relevant to the scope of the journal, and can be published after addressing a number of issues.The idea behind the paper is to use forgetting and recollection in recurrent neural networks (upper left in the diagram). The overall novelty is likely the combination of their existing model with forgetting. Overall, there is an improvement in prediction on the presented data, but it's not entirely clear whether this is due to the model or not.

1. The diagram in Figure 2 is described rather superficially and briefly. It is difficult to understand the full algorithm cycle.
2.  The points in Figure 3 (A, B, etc.) blend into the background, and their positions are difficult to understand. 
3. What is the actual size of the data used to train the model? 
4. What is the "time" axis in Figure 5? What units are used? What are the numbers? It would also be useful to show how the original model works, without the improvements proposed in the paper (see Figure 5).
5. The data is quite small (360 dataset?), so authors divided it into sections and analyzed them. It's not entirely clear from the neural network itself how repeatable this neural network is. It's also unclear what type of operator is used to link the time data. Typically, local information is loaded at each time step, but here we use separate time and spatial steps (perhaps these should be combined). 

Author Response

Thank you very much for your valuable comments. Please see the attachment for our responses.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript was revised accordingly.

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript shows several important improvements over the previous version. The introduction better defines the research gap and relies on more current references. The methodology is clearer, the results are better understood, and the conclusions are linked to real-world applications. Overall, the article is more solid and consistent.

Reviewer 4 Report

Comments and Suggestions for Authors

Dear Author, we would like to congratulate you and your team for the revisions made, which have improved the quality of the manuscript. 
Regards