the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 23 May 2024
Analysis of surface deformation prediction in high mountain canyon areas based on time-series InSAR technology and improved Elman neural network
Abstract. To address the issues of over-reliance on deformation data and model singularity in existing surface deformation prediction methods in high mountain canyon areas, this study proposes the improvement of Elman neural network using cuckoo search algorithm and grey wolf optimization algorithm (CS-Elman and GWO-Elman) from the perspective of multi-temporal and multi-factor analysis. Firstly, surface deformation in the study area is monitored using SBAS-InSAR and PS-InSAR techniques. Then, the optimal evaluation factors are determined from 13 evaluation factors including digital elevation model (DEM) and slope using grey correlation analysis and correlation matrix analysis in SPSSAU software. These optimal factors, combined with surface deformation monitoring values obtained from InSAR technology, are used to construct CS-Elman and GWO-Elman prediction models from a multi-factor and multi-temporal perspective. Finally, the optimal prediction model is determined through comparative experiments and its prediction performance is validated. Results indicate: (1) SBAS-InSAR and PS-InSAR techniques exhibit a high correlation coefficient (R2=0.85) between monitored radar line of sight (LOS) deformation rates, demonstrating the feasibility of joint analysis of the two techniques. (2) The CS-Elman model has a smaller absolute error range compared to the GWO-Elman model. The optimal convergence iteration number, mean square error, mean absolute error (MAE) and mean absolute percentage error (MAPE) of the CS-Elman model are 3 iterations, 0.020 mm/a, 1.620 mm/a and 21.500 %, respectively, which are all superior to the GWO-Elman model. This indicates that the Elman network optimized by the CS algorithm exhibits better performance and higher accuracy in predicting surface deformation in high mountain canyon areas. (3) Comparative analysis with SVM, LSTM and PSO-BP models, as well as prediction of temporal deformation trends at deformation points, validate the advantages and effectiveness of the CS-Elman model in surface deformation prediction. This method can serve as an effective means for long-term deformation prediction in high mountain canyon areas.
- Preprint
(5015 KB) - Metadata XML
-
Supplement
(13 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2024-1220', Anonymous Referee #1, 25 Jun 2024
Chen et al have presented neural network based methods to estimate vertical deformations using (fairly) readily available geospatial features such as DEM, vegetation, rainfall, soil type, etc. The ground truth data are calculated based on two InSAR techniques. This study and similar methods are highly relevant for predicting future deformations in order to inform risk mitigation decisions related to landslides and debris flow.
However, I have several concerns in the authors’ manuscript which make it difficult to validate their methodologies, and to relate it with their stated objectives. I have described my concerns below starting with the major concerns:
- While the authors’ stated goal for their methodology is predicting deformations, they have only demonstrated the prediction at 2 points in their study region. Moreover the section on future predictions does not provide sufficient detail about the spatial and temporal evolution of the prediction quality. As a result, it is unclear how the authors’ approach can be implemented in practice and be used for real-world applications. I would suggest focusing the study more on future prediction of deformations, and providing spatial and temporal evolution based on the training set. It will be important to identify how the length of the training set impacts the quality of future predictions. Additionally, since the authors have demonstrated a high dependence of deformations on rainfall seasons, it will be necessary to determine for how many seasons can the future deformations be reliably predicted.
- The authors have used 970 points for training and 30 for testing. This distribution is not sufficient to validate their results. A typical distribution of test set is 20-30% of the data, especially for small datasets containing 1000 total points. Moreover, the authors’ stated goal is future prediction and not spatial prediction on unobserved points. Therefore, the test set must comprise of future observations in addition to withholding spatial points. The test set must also include geographically withheld spatial points in addition to randomly selected ones, to validate their methodology on previously unseen proximate regions. As a result of these deficiencies, the authors' methodology cannot be fully validated. I would suggest adding more information about selection of training data, including spatial and temporal distributions, and a more representative distribution of the test set that matches the authors’ stated objectives.
- One of the major limitations of the authors’ current methodology is that training the model requires the availability on InSAR data in the region of interest. However, if the InSAR data is already available, it is unclear why a prediction model is necessary to identify deformations across spatial points in a study area. In this scenario, a prediction model will only be useful for future temporal predictions. Similar to the above points, I would suggest focussing the implementation and results of their methodologies on temporal predictions instead of spatial estimates, on a sufficiently large test set.
- The authors have used a neural network model, but combined it with optimization methods like CS and GWO. However, it is unclear from the manuscript why these optimization methods are needed, as a neural network model can be trained directly using back-propagation.
- The authors have not presented documentation for how the deformation predictions may be used for disaster risk mitigation for landslides and debris flow. It would be helpful to include at least some information in the discussion/conclusions section for translating deformation predictions to disaster likelihood.
- Line 56 - Please define “grey models”.
- Line 72 - Please provide some examples of “certain limitations”.
- Line 131 - Please provide references for TWI and SPI, and their brief descriptions.
- Line 133 - Please describe grey relational analysis.
- Table 1 - The table seems to have already been described in lines 139-141, and is unnecessary.
- Eq 1-3 - Please define k.
- Eq 1-3 - It looks like x, y, and u are vectors/matrices, but are not represented as such. Please use vector notation by changing to boldface.
- Section 2.3 - It is not clear why the CS and GWO algorithms are needed. The weights of the neural network should be able to be calculated using back-propagation. In case of any issues observed from overfitting (which is what local minima would imply), typical neural network regularization techniques, like pooling layers, l1, l2 regularization, quantization, etc. can be employed.
- Line 242 - Please include the magnitude scale, date, and reference for the earthquake.
- Line 245 - Please provide a reference for the statement.
- Line 246 - The statement is unclear.
- Figure 4 - Please include the date on which Google collected the satellite image.
- Line 259 - Please provide a reference for Sentinel-1A data.
- Table 2 - From the text, it looks like column 3 represents the resolution, not the scale. Please update accordingly.
- Table 2 - Please clarify whether 1.07 m is both the horizontal and azimuthal resolution of Google Satellite imagery.
- Figure 5 - The units of rainfall (mm/a) are not clear.
- Table 3 - What is the reason for classification of continuous variables like DEM, slope, etc.? Please provide references for classifications of continuous variables if they are based on other studies. Please clarify if classifications for non-continuos variables, like soil type, lithology, etc. are obtained as is from the data sources. Please clarify how these classifications were used (and if they were used) in the authors’ models.
- Line 289 - How can the techniques select their own master images? Please clarify the process of selecting the master images.
- Figures 6, 7 - The units mm.a^{-1} are unclear.
- Figures 6, 7 - Please provide higher resolution figures as they are currently not legible.
- Figures 6,7 - What are the vertical resolutions of SBAS- and PS-InSAR methods? The data represented up to 3 decimal points in mm, indicates a resolution of 10^{-3} mm.
- Figures 6,7 - Red outlined boxes for debris flow gully could not be located in the figures. Are debris flow gullies represented by red lines? If that is the case, please change the box in the legend to a line.
- Line 316 - The authors’ statement that the deformation rate at point Q is relatively flat during the rainy season is supported in the first season, but does not seem to be supported in the second season from Figure 8.
- Line 327 - Please quantify “several”. Why weren’t all points selected on a uniform grid, or x number of points selected randomly from the study area?
- Table 5 - As a suggestion, please change Table 5 to a figure, with the correlation values also represented on a color scale, to easily identify variables with high and low correlations.
- Table 6 - Please describe the functions as their proprietary MATLAB names are insufficient for understanding the model. I would also suggest removing Table 6 and providing the information in the text description only.
- Table 7 - Please provide a reasoning for selecting each of the parameters of the model.
- Line 364 - Since the lowest resolution of the selected 9 features is 30m, please describe how the points were generated randomly across the study region. Were they generated from a random 30x30m grid, or were they distributed continuously in the study region? If they were distributed continuously, how were the feature values obtained from their native resolution to a continuous distribution across the study region?
- Figure 10, 11 - The x-axis is not clear. What does the number of samples imply? If this represents the sample number from the 30 samples in the test set, I would suggest changing x-axis label to “sample number in test set”.
- Figure 12 - The following legend labels are introduced for the first time in the figure and have not been described elsewhere - validation, best, goal.
- Figure 12 - How is the criteria for convergence determined? From the figure, it appears that all models converge in less than 5 epochs. Additionally, I suggest keeping the same x-axis scale and stacking the 3 subfigures vertically for easier comparisons.
- Line 421 - Why have the number of points reduced from 1000 to 507 for this comparison?
- Line 445 - Please quantify “period”.
- Line 453 - The authors state that the InSAR accuracy is +- 10mm. However, throughout the manuscript, they have represented InSAR deformations at a resolution of 10^{-3} mm. Please clarify.
Citation: https://doi.org/10.5194/egusphere-2024-1220-RC1 -
CC1: 'Reply on RC1', Kuayue Chen, 24 Sep 2024
Thank you for your detailed feedback and suggestions on my article. I will respond to each piece of feedback from the reviewers one by one for your review. Additionally, I have revised the manuscript, and all revisions are highlighted in red font.
Question #1. While the authors’ stated goal for their methodology is predicting deformations, they have only demonstrated the prediction at 2 points in their study region. Moreover the section on future predictions does not provide sufficient detail about the spatial and temporal evolution of the prediction quality. As a result, it is unclear how the authors’ approach can be implemented in practice and be used for real-world applications. I would suggest focusing the study more on future prediction of deformations, and providing spatial and temporal evolution based on the training set. It will be important to identify how the length of the training set impacts the quality of future predictions. Additionally, since the authors have demonstrated a high dependence of deformations on rainfall seasons, it will be necessary to determine for how many seasons can the future deformations be reliably predicted.
Answer: Thank you for your kind advice. This study focuses on improving the Elman neural network using the Cuckoo Search (CS) algorithm and Grey Wolf Optimization (GWO) algorithm. The evaluation factors such as DEM are used as inputs, and the annual average deformation rate is used as the output. First, spatial prediction is conducted to predict the annual average deformation rate at unknown points, and the accuracy of the model is validated by comparing it with the InSAR-monitored annual average deformation rate. The results show that the CS-Elman model has higher prediction accuracy. In practical applications, once the DEM and other evaluation factors for a certain area are obtained, the network model can be used to predict the spatial deformation rate in that area. Based on the surface deformation rate, appropriate measures can be taken to reduce the damage caused by geological disasters. Moreover, using the network model for surface deformation prediction can reduce the time cost associated with InSAR technology. Then, time-series predictions are made for typical deformation points. The first 26 periods are used as training samples, and the deformation trends of the subsequent 4 periods are predicted. The model's accuracy is verified by comparing the predictions with the InSAR-monitored time-series deformation. Based on review suggestions, we will add the following experiments: In Section 4.2.3, we will add future deformation predictions for typical deformation points, using time-series deformation data from 2022-2023 to predict the deformation trend for 2024. We will also analyze how many months of future deformation can be reliably predicted and predict the future deformation trends for the entire study area.
Question #2. The authors have used 970 points for training and 30 for testing. This distribution is not sufficient to validate their results. A typical distribution of test set is 20-30% of the data, especially for small datasets containing 1000 total points. Moreover, the authors' stated goal is future prediction and not spatial prediction on unobserved points. Therefore, the test set must comprise of future observations in addition to withholding spatial points. The test set must also include geographically withheld spatial points in addition to randomly selected ones, to validate their methodology on previously unseen proximate regions. As a result of these deficiencies, the authors' methodology cannot be fully validated. I would suggest adding more information about selection of training data, including spatial and temporal distributions, and a more representative distribution of the test set that matches the authors' stated objectives.
Answer: Thank you for your valuable opinion. This study includes both future time predictions and spatial predictions for unknown points. Based on the review suggestions, we have added experiments for future predictions of typical deformation points and spatial predictions for the entire study area.
Question #3. One of the major limitations of the authors' current methodology is that training the model requires the availability on InSAR data in the region of interest. However, if the InSAR data is already available, it is unclear why a prediction model is necessary to identify deformations across spatial points in a study area. In this scenario, a prediction model will only be useful for future temporal predictions. Similar to the above points, I would suggest focussing the implementation and results of their methodologies on temporal predictions instead of spatial estimates, on a sufficiently large test set.
Answer: Thank you for your careful review. This study uses a prediction model to forecast deformation at spatial points within the study area because InSAR technology requires a large amount of image data for acquisition, storage, preprocessing, and a series of operations that consume significant processing time. By using the prediction model, time costs can be reduced. Therefore, the prediction model is used to forecast deformation at unknown spatial points, and the results are compared with InSAR monitoring data to verify the model's accuracy. This demonstrates that the prediction model can replace InSAR technology by using only 13 evaluation factors such as DEM to predict the deformation of the area. Based on the review comments, we have added experiments for future predictions of typical deformation points.
Question #4. The authors have used a neural network model, but combined it with optimization methods like CS and GWO. However, it is unclear from the manuscript why these optimization methods are needed, as a neural network model can be trained directly using back-propagation.
Answer: Thank you for your valuable opinion. In lines 94-94 and 217-230 of the text, the reasons for using the CS and GWO algorithms to optimize the Elman neural network are explained: "However, determining the weights, thresholds and learning rates of the Elman neural network is often challenging, requiring optimization." and "Specifically: (1) The initial values of neural network weights have a significant impact on the training process. If the initial weights are not properly chosen, the network may get stuck in a local optimum or experience slow convergence. The search for the best weight combination in high-dimensional space is complex, and efficient optimization algorithms are needed to find the global optimum. (2) Thresholds in neural networks play a crucial role in adjusting the output. Like weights, the choice of thresholds affects network performance and convergence speed. Thresholds also need to be optimized to ensure that the network performs well on training data and can effectively generalize to new data. (3) Learning rate determines the step size for updating network weights. If the learning rate is too large, the training process may oscillate or fail to converge. Conversely, if the learning rate is too small, the training process will be slow. Dynamic adjustments to the learning rate may be required during different stages of training, making the choice of an appropriate learning rate and adjustment strategy critical. Optimization algorithms, such as genetic algorithms (GA), particle swarm optimization (PSO), grey wolf optimization (GWO), and cuckoo search (CS), can more effectively search for the optimal combination of weights, thresholds, and learning rates, thus improving the overall performance of the network. The optimization process can also help prevent overfitting and enhance the model's generalization ability, ensuring better performance on unseen data.".
Question #5. The authors have not presented documentation for how the deformation predictions may be used for disaster risk mitigation for landslides and debris flow. It would be helpful to include at least some information in the discussion/conclusions section for translating deformation predictions to disaster likelihood.
Answer: Thank you for your suggestion. Based on the review comments, we have added the following to the conclusion section: "The method in this paper validates that the CS-Elman model has a small error compared to InSAR monitoring technology and is suitable for short-term monitoring. Therefore, in practical applications, local evaluation factors such as DEM can be used as inputs to the CS-Elman model, with the annual average deformation rate as the output. Based on the IUGS/WGL (1995) geological hazard intensity classification criteria, with 2 mm/a and 16 mm/a as thresholds, regions where the predicted deformation rate exceeds 16 mm/a should be identified. Protective measures should be taken to prevent geological disasters in these areas."
IUGS / WGL, International Union of Geological Science Working Group on Landslides, 1995, A suggested method for describing the rate of movement of a landslide. IAEG Bulletin, 1995, 52, pp. 75-78.
Question #6. Line 56 - Please define "grey models".
Answer: Thanks for your kind suggestion. Based on the review comments, we have added the definition of the "grey model" in the text: "Among them, the grey model is a forecasting method used for small samples and uncertain systems, mainly applied to solve prediction problems with incomplete information and sparse data.".
Question #7. Line 72 - Please provide some examples of "certain limitations".
Answer: Thank you for your kind advice. Based on the review comments, we have added examples of "certain limitations" in the text: "If a model is based solely on deformation data, its lack of multi-factor analysis capability will limit its prediction accuracy and comprehensiveness. Deformation data typically reflect slow changes, making it difficult to predict sudden disasters such as debris flows and landslides.".
Question #8. Line 131 - Please provide references for TWI and SPI, and their brief descriptions.
Answer: Thank you very much for your careful review. Based on the review comments, we have added references and brief explanations for TWI and SPI in the text: "In which, TWI (Sörensen, Zinko, and Seibert 2006) is an index that reflects the influence of terrain on water accumulation and distribution, primarily used to describe the potential moisture conditions of an area. SPI (Parker and Davey 2023) is based on the combination of slope and flow, and is mainly used to analyze the erosion and scouring strength of water flow on the surface.".
Sörensen, R., Zinko, U., and Seibert, J.: On the calculation of the topographic wetness index: evaluation of different methods based on field observations, Hydrol. Earth Syst. Sci., 10, 101-112, https://doi.org/10.5194/hess-10-101-2006, 2006.
Parker, C., and Davey, J.: Stream power indices correspond poorly with observations of alluvial river channel adjustment, Earth Surf. Proc. Land., 48, 1290-1304, https://doi.org/10.1002/esp.5550, 2023.
Question #9. Line 133 - Please describe grey relational analysis.
Answer: Thank you for your kind advice. Based on the review comments, we have added the concept of "grey relational analysis" in the text: "Grey relational analysis is a method within grey system theory that aims to reveal the relationships between different factors or variables by analyzing their similarity. The basic concept is that if two sequences have more similar trends in variation, their degree of correlation is higher.".
Question #10. Table 1 - The table seems to have already been described in lines 139-141, and is unnecessary.
Answer: Thank you for your careful review. Based on the review comments, we have made reductions to the description of R in the text.
Question #11. Eq 1-3 - Please define k.
Answer: Thank you for your careful review. According to the revision comments, we have defined k as "k represents any given moment.".
Question #12. Eq 1-3 - It looks like x, y, and u are vectors/matrices, but are not represented as such. Please use vector notation by changing to boldface.
Answer: Thank you for your careful review. According to the revision comments, we have changed the vectors x, y, u, b1 and, b2 to bold.
Question #13. Section 2.3 - It is not clear why the CS and GWO algorithms are needed. The weights of the neural network should be able to be calculated using back-propagation. In case of any issues observed from overfitting (which is what local minima would imply), typical neural network regularization techniques, like pooling layers, l1, l2 regularization, quantization, etc. can be employed.
Answer: Thank you for your attention and questions. The issue has been addressed in the fourth question.
Question #14. Line 242 - Please include the magnitude scale, date, and reference for the earthquake.
Answer: Thank you very much for your valuable feedback. According to the revision comments, we have added the earthquake magnitude, time, and references in Section 3.1: "The high mountain canyon area of the Minjiang River Basin is located in Wenchuan County, Sichuan Province, China. It is a typical high-risk area for geological disasters, particularly frequent debris flows and landslides. This basin was affected by the Ms8.0 earthquake on May 12, 2008 (Dong et al. 2008).".
Dong, S. W., Zhang, Y. Q., Wu, Z. H., Yang, N., Ma, Y. S., Shi, W., Chen, Z. L., Long, C. X., and An, M. J.: Surface Rupture and Co‐seismic Displacement Produced by the Ms 8.0 Wenchuan Earthquake of May 12th, 2008, Sichuan, China: Eastwards Growth of the Qinghai‐Tibet Plateau, Acta Geol. Sin-Engl., 82, 938-948, https://doi.org/10.1111/j.1755-6724.2008.tb00649.x, 2008.
Question #15. Line 245 - Please provide a reference for the statement.
Answer: Thank you for your careful review. According to the revision comments, we have added the reference for rainfall-induced landslide debris flows in Section 3.1: "Zhou, W., Tang, C., Van Asch, T. W., and Zhou, C. H.: Rainfall-triggering response patterns of post-seismic debris flows in the Wenchuan earthquake area. Nat. Hazards, 70: 1417-1435, https://doi.org/10.1007/s11069-013-0883-8, 2014.".
Question #16. Line 246 - The statement is unclear.
Answer: Thank you for your attentiveness. According to the revision comments, we have changed "This study selects the high mountain canyons on both sides of the Minjiang River and the Zagunao River as the study area." to "This study selects the mountainous canyon areas along both sides of the Minjiang River and Zagunao River as the study area.".
Question #17. Figure 4 - Please include the date on which Google collected the satellite image.
Answer: Thank you for your kind advice. According to the revision comments, we have added the date of Google satellite image collection in Figures 4 and 5: "Image capture date: January 14, 2024".
Question #18. Line 259 - Please provide a reference for Sentinel-1A data.
Answer: Thank you for your careful review. According to the revision comments, we have added the reference for Sentinel-1A data: (Shankar et al. 2023).
Shankar, H., Chauhan, P., Singh, D., Bhandari, R., Bhatt, C. M., Roy, A., Kannaujiya, S., and Singh, R. P. (2023). Multi-temporal InSAR and Sentinel-1 for assessing land surface movement of Joshimath town, India. Geomat. Nat. Haz. risk, 14, 2253972, https://doi.org/10.1080/19475705.2023.2253972, 2023.
Question #19. Table 2 - From the text, it looks like column 3 represents the resolution, not the scale. Please update accordingly.
Answer: Thank you for your valuable opinion. According to the revision comments, we corrected the third column of Table 2 to "Spatial resolution.".
Question #20. Table 2 - Please clarify whether 1.07 m is both the horizontal and azimuthal resolution of Google Satellite imagery.
Answer: Thank you for your kind advice. In Table 2, "1.07 m" refers to the pixel resolution of Google Earth.
Question #21. Figure 5 - The units of rainfall (mm/a) are not clear.
Answer: Thank you for your careful review. The unit "mm/a" in Figure 5 stands for millimeters per year.
Question #22. Table 3 - What is the reason for classification of continuous variables like DEM, slope, etc.? Please provide references for classifications of continuous variables if they are based on other studies. Please clarify if classifications for non-continuos variables, like soil type, lithology, etc. are obtained as is from the data sources. Please clarify how these classifications were used (and if they were used) in the authors' models.
Answer: Thank you for your valuable opinion. According to the revision suggestions, we do not use the classification of continuous variables in the model, so we delete the original Table 3, update the table numbering in the subsequent text, and revise Figure 5. The non-continuous variables (such as soil type, lithology, and vegetation) are downloaded from the corresponding websites (as shown in "Data availability") and processed using ArcGIS software.
Question #23. Line 289 - How can the techniques select their own master images? Please clarify the process of selecting the master images.
Answer: Thank you for your careful review. According to the revision suggestions, we add the process of selecting the master image in Section 4.1:"Selecting the primary image in InSAR technology is crucial, and the following three factors should be considered: (1) Temporal stability: The primary image should be chosen at a suitable time point within the monitoring period, typically selecting an image from the middle of the period. This ensures that subsequent images have shorter temporal baselines relative to the primary image, reducing atmospheric effects and temporal decorrelation issues. (2) Surface features: The surface features in the image should be stable. Avoid selecting an image from dates when large-scale surface changes occurred, such as after major natural disasters (e.g., earthquakes, floods, landslides), as they can negatively affect the interferogram. (3) Data quality: The quality of the image is a key factor in selecting the primary image. Ensure that the chosen image has no significant noise, data loss, or other disturbances. Taking into account the aforementioned factors, the two time-series InSAR techniques select April 1, 2023 and December 26, 2022 as the master images, respectively.".
Question #24. Figures 6, 7 - The units mm.a^{-1} are unclear.
Answer: Thank you for your kind advice. The notation "mm.a^{-1}" in Figures 6 and 7 has been changed to "mm/a" to represent millimeters per year.
Question #25. Figures 6, 7 - Please provide higher resolution figures as they are currently not legible.
Answer: Thank you for your careful review. Based on the revision suggestions, Figures 6 and 7 have been updated in the text.
Question #26. Figures 6, 7 - What are the vertical resolutions of SBAS-InSAR and PS-InSAR methods? The data represented up to 3 decimal points in mm, indicates a resolution of 10^{-3} mm.
Answer: Thanks for your kind suggestion. The SBAS-InSAR and PS-InSAR techniques primarily focus on monitoring surface deformation and coherence analysis, rather than the vertical resolution of specific points on the surface. The spatial resolution of the Sentinel-1A data used is 5 m × 20 m (range × azimuth). The data represent the line-of-sight deformation rates monitored by both techniques, with the unit being millimeters per year.
Question #27. Figures 6,7 - Red outlined boxes for debris flow gully could not be located in the figures. Are debris flow gullies represented by red lines? If that is the case, please change the box in the legend to a line.
Answer: Thank you for your careful review. In Figures 6 and 7, debris flows are represented by red outlines, and the red box in the legend has been changed to a red line.
Question #28. Line 316 - The authors' statement that the deformation rate at point Q is relatively flat during the rainy season is supported in the first season, but does not seem to be supported in the second season from Figure 8.
Answer: Thank you for your kind advice. The deformation at point Q is relatively stable during the rainy season (May to September) compared to the changing trend in the non-rainy season (October to the following April). Taking SBAS-InSAR as an example, from October 2022 to April 2023, the maximum deformation at point Q changed from -47.21 mm/a to -135.21 mm/a, a change of 88 mm/a. From May 2023 to September 2023, the maximum deformation at point Q changed from -137.41 mm/a to -180.46 mm/a, a change of 43.05 mm/a. Therefore, this indicates that the deformation trend at point Q is relatively stable during the rainy season.
Question #29. Line 327 - Please quantify "several". Why weren't all points selected on a uniform grid, or x number of points selected randomly from the study area?
Answer: Thank you very much for your careful review. According to the revision suggestions, we changed "Several points with the same name were selected in the study area" to "300 points with the same name were randomly selected in the study area."
Question #30. Table 5 - As a suggestion, please change Table 5 to a figure, with the correlation values also represented on a color scale, to easily identify variables with high and low correlations.
Answer: Thanks for your kind suggestion. According to the revision suggestions, we changed the original Table 5 to a graphical representation of the relationships among the factors.
Question #31. Table 6 - Please describe the functions as their proprietary MATLAB names are insufficient for understanding the model. I would also suggest removing Table 6 and providing the information in the text description only.
Answer: Thanks for your kind suggestion. According to the revision suggestions, we deleted the original Table 6 and described the functions used in the model in the text: "The chosen training function is trainlm, and tanh and purelin serve as the activation functions for the hidden and output layers, respectively. Trainlm trains feedforward neural networks using the Levenberg-Marquardt algorithm, known for fast convergence, robustness, and excellent performance in handling small to medium-sized problems. Tanh is a hyperbolic tangent function commonly applied in hidden layer nodes to enable the network to learn and approximate nonlinear functions, while purelin is a linear activation function typically employed in the output layer for regression tasks.".
Question #32. Table 7 - Please provide a reasoning for selecting each of the parameters of the model.
Answer: Thank you for your kind advice. The reasons for each parameter setting in the model are as follows: (1) Training iterations set to 1000: A sufficient number of training iterations ensures the model fully learns from the training set, avoiding underfitting. However, too many iterations may lead to overfitting and wasted computational resources. Setting 1000 iterations strikes a balance between learning effectiveness and computational cost. (2) Learning rate: The learning rate determines the step size for each parameter update. A commonly used learning rate of 0.01 ensures the model parameters converge stably. A rate too high may cause oscillations, while too low a rate slows down the convergence process. (3) Minimum error for training goal set to 10-5: This ensures high precision for the model. Achieving such a low error indicates the model fits the training data very well. (4) Display frequency: Controls how often results are displayed during training. Displaying results every 25 iterations allows users to monitor training progress without interrupting the process too frequently. (5) Momentum factor: Used to accelerate the convergence of the gradient descent algorithm. A momentum factor of 0.01 smooths the parameter update process, preventing interference from local minima. (6) Maximum number of generations in evolutionary algorithms: Limits the runtime and computational cost of the algorithm. 50 generations are usually sufficient to find a solution close to the global optimum without significantly increasing computation time. (7) Initial population size in the Cuckoo Search algorithm: Influences the diversity and convergence speed of the algorithm. A population size of 10 provides enough diversity without incurring excessive computational overhead. (8) Probability in the Cuckoo Search algorithm: Determines the frequency of individual replacement in the population. A probability of 0.25 balances the need for exploring new solutions and exploiting the current best solution. (9) Number of wolves in the Grey Wolf Optimization algorithm: Affects the optimization effectiveness. 20 wolves provide sufficient diversity and exploration capability without excessively increasing computational costs. (10) Search range for wolves: Determines the search space of the algorithm. A range of 30 ensures that the algorithm has enough space to explore the global optimum without unnecessarily increasing the computational complexity by having too large a search space.
Question #33. Line 364 - Since the lowest resolution of the selected 9 features is 30m, please describe how the points were generated randomly across the study region. Were they generated from a random 30x30m grid, or were they distributed continuously in the study region? If they were distributed continuously, how were the feature values obtained from their native resolution to a continuous distribution across the study region?
Answer: Thank you very much for your careful review. The random points are randomly generated using the "Create Random Points" tool in ArcGIS software.
Question #34. Figure 10, 11 - The x-axis is not clear. What does the number of samples imply? If this represents the sample number from the 30 samples in the test set, I would suggest changing x-axis label to "sample number in test set".
Answer: Thank you for your attention and questions. In Figures 10 and 11, the x-axis represents the number of samples in the test set. According to the revision suggestions, we changed the x-axis of Figures 10 and 11 to "The number of samples in the test set."
Question #35. Figure 12 - The following legend labels are introduced for the first time in the figure and have not been described elsewhere - validation, best, goal.
Answer: Thank you for your kind advice. According to the revision suggestions, we added descriptions of validation, best, and goal in the text: "In this figure, validation refers to the model's performance on the validation set during training. The validation set is used to evaluate the model's generalization ability and to avoid overfitting. It shows how the error (such as mean squared error) on the validation data changes as training progresses. Best indicates the model's optimal state during training, which is the moment when the lowest error on the validation set is achieved. It represents the model parameters that perform best on the validation set in the current training. Goal refers to the target error for training, which is typically a threshold set by the user. When the model's error drops to this target value, training can be terminated early, indicating that the desired outcome has been achieved.".
Question #36. Figure 12 - How is the criteria for convergence determined? From the figure, it appears that all models converge in less than 5 epochs. Additionally, I suggest keeping the same x-axis scale and stacking the 3 subfigures vertically for easier comparisons.
Answer: Thanks for your kind suggestion. When the model reaches the minimum error of the training goal, the model converges.
Question #37. Line 421 - Why have the number of points reduced from 1000 to 507 for this comparison?
Answer: Thank you for your careful review. Both 1000 and 507 are randomly selected data points. The use of different data sets in the two comparative experiments is intended to reduce the randomness of the experimental results and ensure a more comprehensive and reliable evaluation of the model's performance. By conducting experiments on different data sets, the chance of obtaining results due to specific data distributions is minimized, providing a more accurate reflection of the model's actual generalization ability.
Question #38. Line 445 - Please quantify "period".
Answer: Thank you for your careful review. A period of 24 days.
Question #39. Line 453 - The authors state that the InSAR accuracy is +- 10mm. However, throughout the manuscript, they have represented InSAR deformations at a resolution of 10^{-3} mm. Please clarify.
Answer: Thank you for your kind advice. "InSAR deformation accuracy is ±10 mm," which means that the error between the surface deformation monitored by InSAR technology and the actual surface deformation is within ±10 mm. The absolute error between the model predictions and the InSAR measurements is within 5 mm, indicating that the model's prediction accuracy is relatively high. Here, ±10 mm is a shorthand for ±10.000 mm.
-
AC1: 'Reply on RC1', Baoyun Wang, 29 Sep 2024
Thank you for your detailed feedback and suggestions on my article. I will respond to each piece of feedback from the reviewers one by one for your review. Additionally, I have revised the manuscript, and all revisions are highlighted in red font.
Question #1. While the authors’ stated goal for their methodology is predicting deformations, they have only demonstrated the prediction at 2 points in their study region. Moreover the section on future predictions does not provide sufficient detail about the spatial and temporal evolution of the prediction quality. As a result, it is unclear how the authors’ approach can be implemented in practice and be used for real-world applications. I would suggest focusing the study more on future prediction of deformations, and providing spatial and temporal evolution based on the training set. It will be important to identify how the length of the training set impacts the quality of future predictions. Additionally, since the authors have demonstrated a high dependence of deformations on rainfall seasons, it will be necessary to determine for how many seasons can the future deformations be reliably predicted.
Answer: Thank you for your kind advice. This study focuses on improving the Elman neural network using the Cuckoo Search (CS) algorithm and Grey Wolf Optimization (GWO) algorithm. The evaluation factors such as DEM are used as inputs, and the annual average deformation rate is used as the output. First, spatial prediction is conducted to predict the annual average deformation rate at unknown points, and the accuracy of the model is validated by comparing it with the InSAR-monitored annual average deformation rate. The results show that the CS-Elman model has higher prediction accuracy. In practical applications, once the DEM and other evaluation factors for a certain area are obtained, the network model can be used to predict the spatial deformation rate in that area. Based on the surface deformation rate, appropriate measures can be taken to reduce the damage caused by geological disasters. Moreover, using the network model for surface deformation prediction can reduce the time cost associated with InSAR technology. Then, time-series predictions are made for typical deformation points. The first 26 periods are used as training samples, and the deformation trends of the subsequent 4 periods are predicted. The model's accuracy is verified by comparing the predictions with the InSAR-monitored time-series deformation. Based on review suggestions, we will add the following experiments: In Section 4.2.3, we will add future deformation predictions for typical deformation points, using time-series deformation data from 2022-2023 to predict the deformation trend for 2024. We will also analyze how many months of future deformation can be reliably predicted and predict the future deformation trends for the entire study area.
Question #2. The authors have used 970 points for training and 30 for testing. This distribution is not sufficient to validate their results. A typical distribution of test set is 20-30% of the data, especially for small datasets containing 1000 total points. Moreover, the authors' stated goal is future prediction and not spatial prediction on unobserved points. Therefore, the test set must comprise of future observations in addition to withholding spatial points. The test set must also include geographically withheld spatial points in addition to randomly selected ones, to validate their methodology on previously unseen proximate regions. As a result of these deficiencies, the authors' methodology cannot be fully validated. I would suggest adding more information about selection of training data, including spatial and temporal distributions, and a more representative distribution of the test set that matches the authors' stated objectives.
Answer: Thank you for your valuable opinion. This study includes both future time predictions and spatial predictions for unknown points. Based on the review suggestions, we have added experiments for future predictions of typical deformation points and spatial predictions for the entire study area.
Question #3. One of the major limitations of the authors' current methodology is that training the model requires the availability on InSAR data in the region of interest. However, if the InSAR data is already available, it is unclear why a prediction model is necessary to identify deformations across spatial points in a study area. In this scenario, a prediction model will only be useful for future temporal predictions. Similar to the above points, I would suggest focussing the implementation and results of their methodologies on temporal predictions instead of spatial estimates, on a sufficiently large test set.
Answer: Thank you for your careful review. This study uses a prediction model to forecast deformation at spatial points within the study area because InSAR technology requires a large amount of image data for acquisition, storage, preprocessing, and a series of operations that consume significant processing time. By using the prediction model, time costs can be reduced. Therefore, the prediction model is used to forecast deformation at unknown spatial points, and the results are compared with InSAR monitoring data to verify the model's accuracy. This demonstrates that the prediction model can replace InSAR technology by using only 13 evaluation factors such as DEM to predict the deformation of the area. Based on the review comments, we have added experiments for future predictions of typical deformation points.
Question #4. The authors have used a neural network model, but combined it with optimization methods like CS and GWO. However, it is unclear from the manuscript why these optimization methods are needed, as a neural network model can be trained directly using back-propagation.
Answer: Thank you for your valuable opinion. In lines 94-94 and 217-230 of the text, the reasons for using the CS and GWO algorithms to optimize the Elman neural network are explained: "However, determining the weights, thresholds and learning rates of the Elman neural network is often challenging, requiring optimization." and "Specifically: (1) The initial values of neural network weights have a significant impact on the training process. If the initial weights are not properly chosen, the network may get stuck in a local optimum or experience slow convergence. The search for the best weight combination in high-dimensional space is complex, and efficient optimization algorithms are needed to find the global optimum. (2) Thresholds in neural networks play a crucial role in adjusting the output. Like weights, the choice of thresholds affects network performance and convergence speed. Thresholds also need to be optimized to ensure that the network performs well on training data and can effectively generalize to new data. (3) Learning rate determines the step size for updating network weights. If the learning rate is too large, the training process may oscillate or fail to converge. Conversely, if the learning rate is too small, the training process will be slow. Dynamic adjustments to the learning rate may be required during different stages of training, making the choice of an appropriate learning rate and adjustment strategy critical. Optimization algorithms, such as genetic algorithms (GA), particle swarm optimization (PSO), grey wolf optimization (GWO), and cuckoo search (CS), can more effectively search for the optimal combination of weights, thresholds, and learning rates, thus improving the overall performance of the network. The optimization process can also help prevent overfitting and enhance the model's generalization ability, ensuring better performance on unseen data.".
Question #5. The authors have not presented documentation for how the deformation predictions may be used for disaster risk mitigation for landslides and debris flow. It would be helpful to include at least some information in the discussion/conclusions section for translating deformation predictions to disaster likelihood.
Answer: Thank you for your suggestion. Based on the review comments, we have added the following to the conclusion section: "The method in this paper validates that the CS-Elman model has a small error compared to InSAR monitoring technology and is suitable for short-term monitoring. Therefore, in practical applications, local evaluation factors such as DEM can be used as inputs to the CS-Elman model, with the annual average deformation rate as the output. Based on the IUGS/WGL (1995) geological hazard intensity classification criteria, with 2 mm/a and 16 mm/a as thresholds, regions where the predicted deformation rate exceeds 16 mm/a should be identified. Protective measures should be taken to prevent geological disasters in these areas."
IUGS / WGL, International Union of Geological Science Working Group on Landslides, 1995, A suggested method for describing the rate of movement of a landslide. IAEG Bulletin, 1995, 52, pp. 75-78.
Question #6. Line 56 - Please define "grey models".
Answer: Thanks for your kind suggestion. Based on the review comments, we have added the definition of the "grey model" in the text: "Among them, the grey model is a forecasting method used for small samples and uncertain systems, mainly applied to solve prediction problems with incomplete information and sparse data.".
Question #7. Line 72 - Please provide some examples of "certain limitations".
Answer: Thank you for your kind advice. Based on the review comments, we have added examples of "certain limitations" in the text: "If a model is based solely on deformation data, its lack of multi-factor analysis capability will limit its prediction accuracy and comprehensiveness. Deformation data typically reflect slow changes, making it difficult to predict sudden disasters such as debris flows and landslides.".
Question #8. Line 131 - Please provide references for TWI and SPI, and their brief descriptions.
Answer: Thank you very much for your careful review. Based on the review comments, we have added references and brief explanations for TWI and SPI in the text: "In which, TWI (Sörensen, Zinko, and Seibert 2006) is an index that reflects the influence of terrain on water accumulation and distribution, primarily used to describe the potential moisture conditions of an area. SPI (Parker and Davey 2023) is based on the combination of slope and flow, and is mainly used to analyze the erosion and scouring strength of water flow on the surface.".
Sörensen, R., Zinko, U., and Seibert, J.: On the calculation of the topographic wetness index: evaluation of different methods based on field observations, Hydrol. Earth Syst. Sci., 10, 101-112, https://doi.org/10.5194/hess-10-101-2006, 2006.
Parker, C., and Davey, J.: Stream power indices correspond poorly with observations of alluvial river channel adjustment, Earth Surf. Proc. Land., 48, 1290-1304, https://doi.org/10.1002/esp.5550, 2023.
Question #9. Line 133 - Please describe grey relational analysis.
Answer: Thank you for your kind advice. Based on the review comments, we have added the concept of "grey relational analysis" in the text: "Grey relational analysis is a method within grey system theory that aims to reveal the relationships between different factors or variables by analyzing their similarity. The basic concept is that if two sequences have more similar trends in variation, their degree of correlation is higher.".
Question #10. Table 1 - The table seems to have already been described in lines 139-141, and is unnecessary.
Answer: Thank you for your careful review. Based on the review comments, we have made reductions to the description of R in the text.
Question #11. Eq 1-3 - Please define k.
Answer: Thank you for your careful review. According to the revision comments, we have defined k as "k represents any given moment.".
Question #12. Eq 1-3 - It looks like x, y, and u are vectors/matrices, but are not represented as such. Please use vector notation by changing to boldface.
Answer: Thank you for your careful review. According to the revision comments, we have changed the vectors x, y, u, b1 and, b2 to bold.
Question #13. Section 2.3 - It is not clear why the CS and GWO algorithms are needed. The weights of the neural network should be able to be calculated using back-propagation. In case of any issues observed from overfitting (which is what local minima would imply), typical neural network regularization techniques, like pooling layers, l1, l2 regularization, quantization, etc. can be employed.
Answer: Thank you for your attention and questions. The issue has been addressed in the fourth question.
Question #14. Line 242 - Please include the magnitude scale, date, and reference for the earthquake.
Answer: Thank you very much for your valuable feedback. According to the revision comments, we have added the earthquake magnitude, time, and references in Section 3.1: "The high mountain canyon area of the Minjiang River Basin is located in Wenchuan County, Sichuan Province, China. It is a typical high-risk area for geological disasters, particularly frequent debris flows and landslides. This basin was affected by the Ms8.0 earthquake on May 12, 2008 (Dong et al. 2008).".
Dong, S. W., Zhang, Y. Q., Wu, Z. H., Yang, N., Ma, Y. S., Shi, W., Chen, Z. L., Long, C. X., and An, M. J.: Surface Rupture and Co‐seismic Displacement Produced by the Ms8.0 Wenchuan Earthquake of May 12th, 2008, Sichuan, China: Eastwards Growth of the Qinghai‐Tibet Plateau, Acta Geol. Sin-Engl., 82, 938-948, https://doi.org/10.1111/j.1755-6724.2008.tb00649.x, 2008.
Question #15. Line 245 - Please provide a reference for the statement.
Answer: Thank you for your careful review. According to the revision comments, we have added the reference for rainfall-induced landslide debris flows in Section 3.1: "Zhou, W., Tang, C., Van Asch, T. W., and Zhou, C. H.: Rainfall-triggering response patterns of post-seismic debris flows in the Wenchuan earthquake area. Nat. Hazards, 70: 1417-1435, https://doi.org/10.1007/s11069-013-0883-8, 2014.".
Question #16. Line 246 - The statement is unclear.
Answer: Thank you for your attentiveness. According to the revision comments, we have changed "This study selects the high mountain canyons on both sides of the Minjiang River and the Zagunao River as the study area." to "This study selects the mountainous canyon areas along both sides of the Minjiang River and Zagunao River as the study area.".
Question #17. Figure 4 - Please include the date on which Google collected the satellite image.
Answer: Thank you for your kind advice. According to the revision comments, we have added the date of Google satellite image collection in Figures 4 and 5: "Image capture date: January 14, 2024".
Question #18. Line 259 - Please provide a reference for Sentinel-1A data.
Answer: Thank you for your careful review. According to the revision comments, we have added the reference for Sentinel-1A data: (Shankar et al. 2023).
Shankar, H., Chauhan, P., Singh, D., Bhandari, R., Bhatt, C. M., Roy, A., Kannaujiya, S., and Singh, R. P. (2023). Multi-temporal InSAR and Sentinel-1 for assessing land surface movement of Joshimath town, India. Geomat. Nat. Haz. risk, 14, 2253972, https://doi.org/10.1080/19475705.2023.2253972, 2023.
Question #19. Table 2 - From the text, it looks like column 3 represents the resolution, not the scale. Please update accordingly.
Answer: Thank you for your valuable opinion. According to the revision comments, we corrected the third column of Table 2 to "Spatial resolution.".
Question #20. Table 2 - Please clarify whether 1.07 m is both the horizontal and azimuthal resolution of Google Satellite imagery.
Answer: Thank you for your kind advice. In Table 2, "1.07 m" refers to the pixel resolution of Google Earth.
Question #21. Figure 5 - The units of rainfall (mm/a) are not clear.
Answer: Thank you for your careful review. The unit "mm/a" in Figure 5 stands for millimeters per year.
Question #22. Table 3 - What is the reason for classification of continuous variables like DEM, slope, etc.? Please provide references for classifications of continuous variables if they are based on other studies. Please clarify if classifications for non-continuos variables, like soil type, lithology, etc. are obtained as is from the data sources. Please clarify how these classifications were used (and if they were used) in the authors' models.
Answer: Thank you for your valuable opinion. According to the revision suggestions, we do not use the classification of continuous variables in the model, so we delete the original Table 3, update the table numbering in the subsequent text, and revise Figure 5. The non-continuous variables (such as soil type, lithology, and vegetation) are downloaded from the corresponding websites (as shown in "Data availability") and processed using ArcGIS software.
Question #23. Line 289 - How can the techniques select their own master images? Please clarify the process of selecting the master images.
Answer: Thank you for your careful review. According to the revision suggestions, we add the process of selecting the master image in Section 4.1:"Selecting the primary image in InSAR technology is crucial, and the following three factors should be considered: (1) Temporal stability: The primary image should be chosen at a suitable time point within the monitoring period, typically selecting an image from the middle of the period. This ensures that subsequent images have shorter temporal baselines relative to the primary image, reducing atmospheric effects and temporal decorrelation issues. (2) Surface features: The surface features in the image should be stable. Avoid selecting an image from dates when large-scale surface changes occurred, such as after major natural disasters (e.g., earthquakes, floods, landslides), as they can negatively affect the interferogram. (3) Data quality: The quality of the image is a key factor in selecting the primary image. Ensure that the chosen image has no significant noise, data loss, or other disturbances. Taking into account the aforementioned factors, the two time-series InSAR techniques select April 1, 2023 and December 26, 2022 as the master images, respectively.".
Question #24. Figures 6, 7 - The units mm.a^{-1} are unclear.
Answer: Thank you for your kind advice. The notation "mm.a^{-1}" in Figures 6 and 7 has been changed to "mm/a" to represent millimeters per year.
Question #25. Figures 6, 7 - Please provide higher resolution figures as they are currently not legible.
Answer: Thank you for your careful review. Based on the revision suggestions, Figures 6 and 7 have been updated in the text.
Question #26. Figures 6, 7 - What are the vertical resolutions of SBAS-InSAR and PS-InSAR methods? The data represented up to 3 decimal points in mm, indicates a resolution of 10^{-3} mm.
Answer: Thanks for your kind suggestion. The SBAS-InSAR and PS-InSAR techniques primarily focus on monitoring surface deformation and coherence analysis, rather than the vertical resolution of specific points on the surface. The spatial resolution of the Sentinel-1A data used is 5 m × 20 m (range × azimuth). The data represent the line-of-sight deformation rates monitored by both techniques, with the unit being millimeters per year.
Question #27. Figures 6,7 - Red outlined boxes for debris flow gully could not be located in the figures. Are debris flow gullies represented by red lines? If that is the case, please change the box in the legend to a line.
Answer: Thank you for your careful review. In Figures 6 and 7, debris flows are represented by red outlines, and the red box in the legend has been changed to a red line.
Question #28. Line 316 - The authors' statement that the deformation rate at point Q is relatively flat during the rainy season is supported in the first season, but does not seem to be supported in the second season from Figure 8.
Answer: Thank you for your kind advice. The deformation at point Q is relatively stable during the rainy season (May to September) compared to the changing trend in the non-rainy season (October to the following April). Taking SBAS-InSAR as an example, from October 2022 to April 2023, the maximum deformation at point Q changed from -47.21 mm/a to -135.21 mm/a, a change of 88 mm/a. From May 2023 to September 2023, the maximum deformation at point Q changed from -137.41 mm/a to -180.46 mm/a, a change of 43.05 mm/a. Therefore, this indicates that the deformation trend at point Q is relatively stable during the rainy season.
Question #29. Line 327 - Please quantify "several". Why weren't all points selected on a uniform grid, or x number of points selected randomly from the study area?
Answer: Thank you very much for your careful review. According to the revision suggestions, we changed "Several points with the same name were selected in the study area" to "300 points with the same name were randomly selected in the study area."
Question #30. Table 5 - As a suggestion, please change Table 5 to a figure, with the correlation values also represented on a color scale, to easily identify variables with high and low correlations.
Answer: Thanks for your kind suggestion. According to the revision suggestions, we changed the original Table 5 to a graphical representation of the relationships among the factors.
Question #31. Table 6 - Please describe the functions as their proprietary MATLAB names are insufficient for understanding the model. I would also suggest removing Table 6 and providing the information in the text description only.
Answer: Thanks for your kind suggestion. According to the revision suggestions, we deleted the original Table 6 and described the functions used in the model in the text: "The chosen training function is trainlm, and tanh and purelin serve as the activation functions for the hidden and output layers, respectively. Trainlm trains feedforward neural networks using the Levenberg-Marquardt algorithm, known for fast convergence, robustness, and excellent performance in handling small to medium-sized problems. Tanh is a hyperbolic tangent function commonly applied in hidden layer nodes to enable the network to learn and approximate nonlinear functions, while purelin is a linear activation function typically employed in the output layer for regression tasks.".
Question #32. Table 7 - Please provide a reasoning for selecting each of the parameters of the model.
Answer: Thank you for your kind advice. The reasons for each parameter setting in the model are as follows: (1) Training iterations set to 1000: A sufficient number of training iterations ensures the model fully learns from the training set, avoiding underfitting. However, too many iterations may lead to overfitting and wasted computational resources. Setting 1000 iterations strikes a balance between learning effectiveness and computational cost. (2) Learning rate: The learning rate determines the step size for each parameter update. A commonly used learning rate of 0.01 ensures the model parameters converge stably. A rate too high may cause oscillations, while too low a rate slows down the convergence process. (3) Minimum error for training goal set to 10-5: This ensures high precision for the model. Achieving such a low error indicates the model fits the training data very well. (4) Display frequency: Controls how often results are displayed during training. Displaying results every 25 iterations allows users to monitor training progress without interrupting the process too frequently. (5) Momentum factor: Used to accelerate the convergence of the gradient descent algorithm. A momentum factor of 0.01 smooths the parameter update process, preventing interference from local minima. (6) Maximum number of generations in evolutionary algorithms: Limits the runtime and computational cost of the algorithm. 50 generations are usually sufficient to find a solution close to the global optimum without significantly increasing computation time. (7) Initial population size in the Cuckoo Search algorithm: Influences the diversity and convergence speed of the algorithm. A population size of 10 provides enough diversity without incurring excessive computational overhead. (8) Probability in the Cuckoo Search algorithm: Determines the frequency of individual replacement in the population. A probability of 0.25 balances the need for exploring new solutions and exploiting the current best solution. (9) Number of wolves in the Grey Wolf Optimization algorithm: Affects the optimization effectiveness. 20 wolves provide sufficient diversity and exploration capability without excessively increasing computational costs. (10) Search range for wolves: Determines the search space of the algorithm. A range of 30 ensures that the algorithm has enough space to explore the global optimum without unnecessarily increasing the computational complexity by having too large a search space.
Question #33. Line 364 - Since the lowest resolution of the selected 9 features is 30m, please describe how the points were generated randomly across the study region. Were they generated from a random 30x30m grid, or were they distributed continuously in the study region? If they were distributed continuously, how were the feature values obtained from their native resolution to a continuous distribution across the study region?
Answer: Thank you very much for your careful review. The random points are randomly generated using the "Create Random Points" tool in ArcGIS software.
Question #34. Figure 10, 11 - The x-axis is not clear. What does the number of samples imply? If this represents the sample number from the 30 samples in the test set, I would suggest changing x-axis label to "sample number in test set".
Answer: Thank you for your attention and questions. In Figures 10 and 11, the x-axis represents the number of samples in the test set. According to the revision suggestions, we changed the x-axis of Figures 10 and 11 to "The number of samples in the test set."
Question #35. Figure 12 - The following legend labels are introduced for the first time in the figure and have not been described elsewhere - validation, best, goal.
Answer: Thank you for your kind advice. According to the revision suggestions, we added descriptions of validation, best, and goal in the text: "In this figure, validation refers to the model's performance on the validation set during training. The validation set is used to evaluate the model's generalization ability and to avoid overfitting. It shows how the error (such as mean squared error) on the validation data changes as training progresses. Best indicates the model's optimal state during training, which is the moment when the lowest error on the validation set is achieved. It represents the model parameters that perform best on the validation set in the current training. Goal refers to the target error for training, which is typically a threshold set by the user. When the model's error drops to this target value, training can be terminated early, indicating that the desired outcome has been achieved.".
Question #36. Figure 12 - How is the criteria for convergence determined? From the figure, it appears that all models converge in less than 5 epochs. Additionally, I suggest keeping the same x-axis scale and stacking the 3 subfigures vertically for easier comparisons.
Answer: Thanks for your kind suggestion. When the model reaches the minimum error of the training goal, the model converges.
Question #37. Line 421 - Why have the number of points reduced from 1000 to 507 for this comparison?
Answer: Thank you for your careful review. Both 1000 and 507 are randomly selected data points. The use of different data sets in the two comparative experiments is intended to reduce the randomness of the experimental results and ensure a more comprehensive and reliable evaluation of the model's performance. By conducting experiments on different data sets, the chance of obtaining results due to specific data distributions is minimized, providing a more accurate reflection of the model's actual generalization ability.
Question #38. Line 445 - Please quantify "period".
Answer: Thank you for your careful review. A period of 24 days.
Question #39. Line 453 - The authors state that the InSAR accuracy is +- 10mm. However, throughout the manuscript, they have represented InSAR deformations at a resolution of 10^{-3} mm. Please clarify.
Answer: Thank you for your kind advice. "InSAR deformation accuracy is ±10 mm," which means that the error between the surface deformation monitored by InSAR technology and the actual surface deformation is within ±10 mm. The absolute error between the model predictions and the InSAR measurements is within 5 mm, indicating that the model's prediction accuracy is relatively high. Here, ±10 mm is a shorthand for ±10.000 mm.
Citation: https://doi.org/10.5194/egusphere-2024-1220-AC1
-
RC2: 'Comment on egusphere-2024-1220', Mohammad Amin Khalili, 10 Sep 2024
Summary of the Article
The article focuses on surface deformation prediction in high mountain canyon areas using time-series InSAR technology combined with an improved Elman neural network model. The authors propose two optimization methods for the Elman network: Cuckoo Search (CS) and Grey Wolf Optimization (GWO). The study uses SBAS-InSAR and PS-InSAR technologies to monitor surface deformation and incorporates 13 evaluation factors related to terrain and environmental conditions. These factors, such as DEM, slope, and rainfall, are filtered through grey relational and correlation analysis to select optimal inputs for the Elman models. The CS-Elman model outperforms the GWO-Elman model in terms of mean square error, mean absolute error (MAE), and mean absolute percentage error (MAPE). Results show the CS-Elman model to be more effective and accurate in predicting surface deformation, especially when compared to traditional models like SVM and LSTM.
Reviewer’s Comments (Major Revision Required)
- Abstract
- The abstract is informative, but it needs refinement for clarity and conciseness. For example, mentioning the specific results of MAE and MAPE for other methods such as LSTM and SVM etc. would enhance understanding.
- The abstract should also briefly mention the significance of the problem, i.e., why improving surface deformation predictions in high mountain canyon areas is important for hazard prevention (the motivation of doing this study like three or four lines of introduction section).
- This type of writing for Abstract is not common ‘Results indicates’ near 10 lines.
- Introduction
- The introduction adequately establishes the context but is verbose and could benefit from restructuring for clarity. Sentences like “China is a country prone to geological disasters” could be simplified.
- The literature review is extensive, but it lacks a critical discussion of the limitations of the current approaches. More emphasis should be placed on how the proposed model fills these gaps in the literature.
- The problem statement should be more focused. While the paper acknowledges the non-linearity of surface deformation in high mountain canyon areas, the rationale for choosing CS and GWO as optimization techniques should be more explicitly linked to these challenges.
- The introduction presentation is generally unsuitable, comprehensive, and in order. First, you start from InSAR and jump to prediction, and then you come back, bring references from SBAS and PSI, and go to prediction again. It is not very readable and does not convey useful information. The paragraphs should be in order and use a mix of old and new references. Use various and new references; for example, Khalili et al. 2023 and Sun et al. 2024 have up-to-date studies with high accuracy on landslide prediction by InSAR and Hybrid AI model.
- Methodology
- The methodology is detailed but lacks clarity in several places. The flowchart in Figure 1 is helpful, but more explanation is needed to connect the steps logically.
- The choice of 13 factors is justified through grey relational and correlation analysis, but the explanation of these techniques is somewhat technical for a general audience. Simplifying or expanding explanations for readers unfamiliar with these statistical methods would be beneficial. Also, open the part related to lines 133 to 139 and explain more.
- If I am not mistaken you want to use cumulative deformation and predict it, however this type of data is temporal and the conditioning factors are spatial, how did you address it?
- The explanation of the Elman neural network, CS, and GWO algorithms is overly technical without enough context on why these models were chosen practically. The mathematical models are fine for specialists but could be overwhelming for broader audiences. Consider providing a more intuitive explanation of how CS and GWO optimize the Elman network.
- As the authors mentioned the primary references in detail, there are many studies in this field that have fundamentally explained these SBAS and PSI methods. It is enough to include parameters that are effective in your processing in the methodology section. In addition to the expression of its software. Additional explanations are not necessary.
- Case Study and data sets
- The study area section lacks essential explanations, such as about the geological setting. Also, a figure should be given for these explanations. The phenomenon under investigation and prediction is landslides, so a landslide inventory map with full explanations can be provided in this section. We are looking for a reliable prediction, and this detail is important for us in interpreting the results.
- Results
- The results section provides quantitative evidence of the model’s performance, but it could be better structured. For example, the comparison of SBAS-InSAR and PS-InSAR data should be presented with a clearer connection to the model’s predictions.
- Figures 6 and 7 are visually effective, but the discussion around them lacks depth. It would be helpful to explicitly link these figures to the overall goals of the study. For instance, explaining how the deformation trends in these figures correlate with predictions made by the CS-Elman model.
- Cross-validation results (Figure 9) are important but are not discussed thoroughly. Why is the correlation coefficient of 0.85 significant? Does this meet a specific threshold for accuracy in similar studies?
- The discussion of error metrics like MAE and MAPE (Figure 11 and 12) is technical, and the benefits of the CS-Elman model should be articulated more clearly. The practical implications of reducing the error by this margin should be highlighted.
- Major point:
- you were working and processing on Descending orbit, the words of Subsidence and uplift are correct? (Lines 292-293). Also, lines 286 to 287 must go to the methodology section.
- The processing techniques are different and have different details, but both inputs are the same. How do you evaluate the results obtained from SBAS with the results obtained from PSI? If the S1A images have errors or problems, it is for both. You need another source of data to do this evaluation. This validation seems to be useless and does not stamp approval or rejection on SBAS outputs.
- The result needs to be visualized. These statistical graphs are not enough.
- Comparison with Other Models
- The paper compares CS-Elman with GWO-Elman, but the comparison with other models like SVM, LSTM, and PSO-BP is relatively weak. There needs to be a deeper discussion of how these models perform in different scenarios (e.g., different types of terrain). Also, there are some studies that out perform these type of traditional DLMs, may be for evaluation you need to compare with GCN-LSTM, CNN-LSTM, and other hybrid model.
- The section should address why the CS-Elman model outperforms traditional models and what specific features make it better suited for surface deformation prediction. Simply mentioning that the CS-Elman model is superior without a clear analysis of the causes of this superiority is insufficient.
- Figures and Tables
- Many figures (e.g., Figures 6-12) are effective but require more thorough descriptions in the text. For example, the deformation maps (Figures 6 and 7) should be more directly linked to the points discussed in the results section.
- Table 5 shows correlation coefficients, but the discussion is overly technical. A clearer explanation of why certain factors was excluded would improve readability.
- All your Figures need the coordinate system, correct all the figures and check them.
- Discussion and Conclusion
- The discussion lacks depth and should more clearly articulate the implications of the findings. For instance, while the paper shows that CS-Elman is superior in prediction accuracy, the practical implications for geological hazard monitoring in high mountain canyon areas should be emphasized.
- The conclusion is weak. It should restate the significance of the findings more strongly and suggest future research directions based on the limitations of the study. For example, the study might explore additional influencing factors or further improve the model by hybridizing CS with other optimization techniques.
- Language and Grammar
- The paper contains some grammatical errors and awkward phrasing. For example, phrases like "the Elman network optimized by CS algorithm exhibits better performance" could be reworded to "the CS-optimized Elman network performs better."
- The tone is quite technical and might be difficult for some readers. Simplifying the language in certain sections would improve accessibility.
- Additional Points
- Limitations: The limitations of the study are not discussed enough. For example, the issue of vegetation interference is briefly mentioned but not explored in terms of how it could be mitigated in future research.
- Reproducibility: The steps for constructing the CS-Elman and GWO-Elman models are well detailed, but more information on the data preprocessing (especially normalization techniques) is needed to ensure that others can replicate the study.
Final Recommendation: Major Revision
The article is well-researched and tackles a significant problem in surface deformation prediction. However, the manuscript needs substantial revision in terms of clarity, depth of discussion, and comparison with other models. The paper should also better connect its findings to the broader implications for geological hazard monitoring.
-
CC2: 'Reply on RC2', Kuayue Chen, 24 Sep 2024
Thank you for your detailed feedback and suggestions on my article. I will respond to each piece of feedback from the reviewers one by one for your review. Additionally, I have revised the manuscript, and all revisions are highlighted in red font.
Question #1. Abstract
(1) The abstract is informative, but it needs refinement for clarity and conciseness. For example, mentioning the specific results of MAE and MAPE for other methods such as LSTM and SVM etc. would enhance understanding.
Answer: Thank you for your kind advice. Based on the revision comments, we have simplified the results section of the abstract and mentioned the specific results of the CS-Elman model compared to other models: "(3) By comparing with models such as SVM, the CS-Elman model has the smallest error range (0.071–1.843 mm/a) and better accuracy (MAE 0.818 mm/a, MAPE 9.353%).".
(2) The abstract should also briefly mention the significance of the problem, i.e., why improving surface deformation predictions in high mountain canyon areas is important for hazard prevention (the motivation of doing this study like three or four lines of introduction section).
Answer: Thank you for your careful review. According to the revision suggestions, we add the statement "Mountain canyon areas often have complex terrain and unstable geological conditions, making them prone to disasters such as landslides and debris flows. Surface deformation is an early sign of these disasters, and accurate prediction of deformation can help detect potential hazards in advance, reducing the likelihood of disaster occurrence." in the abstract to highlight the importance of predicting surface deformation in high mountain canyon areas.
(3) This type of writing for Abstract is not common 'Results indicates' near 10 lines.
Answer: Thank you for your valuable opinion. According to the revision suggestions, the results section of the abstract is simplified: "Results indicate: (1) SBAS-InSAR and PS-InSAR techniques have a strong correlation (R2 = 0.85) in monitored deformation rates, confirming their potential for joint analysis. (2) The CS-Elman model outperforms the GWO-Elman model, achieving a lower error range, faster convergence (3 iterations), and better metrics (MSE: 0.020 mm/a, MAE: 1.620 mm/a, MAPE: 21.500%). (3) Compared to SVM, LSTM, and PSO-BP models, the CS-Elman model has the smallest error range (0.071–1.843 mm/a) and better accuracy (MAE 0.818 mm/a, MAPE 9.353%). (4) The CS-Elman model excels in short-term predictions but is less effective for long-term forecasting. It predicts a maximum surface uplift of 120.913 mm, offering insights for disaster prevention.".
Question #2. Introduction
(1) The introduction adequately establishes the context but is verbose and could benefit from restructuring for clarity. Sentences like "China is a country prone to geological disasters" could be simplified.
Answer: Thank you for your careful review. According to the revision suggestions, we simplified the language in the introduction. For example, we changed "China is a country prone to geological disasters, and various types of geological disasters have caused enormous losses to the lives and property security of its people." to "China is prone to geological disasters, which have caused significant losses to lives and property security of its people." We also simplified "Landslides, debris flows and collapses are major geological disasters characterized by strong concealment, significant hazards and high suddenness, and they are widely distributed in mountainous areas and canyons in China." to "Landslides, debris flows, and collapses, common in mountainous and canyon areas, are highly hazardous due to their suddenness and hidden nature".
(2) The literature review is extensive, but it lacks a critical discussion of the limitations of the current approaches. More emphasis should be placed on how the proposed model fills these gaps in the literature. (3) The problem statement should be more focused. While the paper acknowledges the non-linearity of surface deformation in high mountain canyon areas, the rationale for choosing CS and GWO as optimization techniques should be more explicitly linked to these challenges.
Answer: Thank you for your suggestion. According to the revision suggestions, we added a discussion of the limitations of current methods in the introduction, as well as how the proposed model addresses these issues: "Although the aforementioned methods can effectively combine InSAR technology and neural network models to monitor and predict surface deformation information, they all have certain drawbacks and limitations. Some scholars (Teng, Wang, and Jiang 2022; Yang et al. 2022; Ye et al. 2022) propose prediction models that overly rely on deformation data, considering fewer other factors that may trigger disasters. They only utilize deformation data as the input and output layers for prediction, which presents certain shortcomings. If a model is based solely on deformation data, its lack of multi-factor analysis capability will limit its prediction accuracy and comprehensiveness. Deformation data typically reflect slow changes, making it difficult to predict sudden disasters such as debris flows and landslides. Additionally, other scholars (Wang et al. 2019; Radman, Akhoondzadeh, and Hosseiny 2021) construct prediction models from the perspective of influencing factors, but the prediction accuracy of these models is relatively low, primarily due to insufficient optimization of the weights and thresholds of the network model. To enhance model performance, it is necessary to introduce optimization algorithms, such as Cuckoo Search (CS) and Grey Wolf Optimization (GWO), to better optimize network parameters.".
(4) The introduction presentation is generally unsuitable, comprehensive, and in order. First, you start from InSAR and jump to prediction, and then you come back, bring references from SBAS and PSI, and go to prediction again. It is not very readable and does not convey useful information. The paragraphs should be in order and use a mix of old and new references. Use various and new references; for example, Khalili et al. 2023 and Sun et al. 2024 have up-to-date studies with high accuracy on landslide prediction by InSAR and Hybrid AI model.
Answer: Thanks for your kind suggestion. The first paragraph of the introduction explains the importance of monitoring and predicting surface deformation in mountain canyon areas. The second paragraph first discusses InSAR monitoring and then transitions to prediction, citing references that combine InSAR technology with neural network models for monitoring and predicting surface deformation, ultimately summarizing the shortcomings of the aforementioned scholars. The third paragraph introduces the content of this paper based on these shortcomings: geological disasters occur due to multiple factors, so this study uses multiple factors as inputs to the network model. Since the weights and thresholds of the network model significantly influence its performance, it introduces optimization algorithms, such as Cuckoo Search (CS) and Grey Wolf Optimization (GWO), to optimize the parameters of the neural network. The fourth paragraph provides a summary.
Question #3. Methodology
(1) The methodology is detailed but lacks clarity in several places. The flowchart in Figure 1 is helpful, but more explanation is needed to connect the steps logically.
Answer: Thank you for your kind advice. According to the revision suggestions, we described Figure 1 as follows: "This study first uses SBAS-InSAR and PS-InSAR techniques to obtain surface deformation information in the study area. Then, the optimal evaluation factors are selected based on correlation analysis. Next, the optimal evaluation factors are used as inputs for the CS-Elman and GWO-Elman models, with the annual average deformation rate as the output. The output results are compared with the annual average deformation rate obtained from InSAR technology to verify the model's accuracy. Finally, the optimal model is applied for future time series deformation prediction. The overall technical process is shown in Figure 1.".
(2) The choice of 13 factors is justified through grey relational and correlation analysis, but the explanation of these techniques is somewhat technical for a general audience. Simplifying or expanding explanations for readers unfamiliar with these statistical methods would be beneficial. Also, open the part related to lines 133 to 139 and explain more.
Answer: Thank you very much for your careful review. According to the revision suggestions, we added the concepts of grey relational analysis and correlation analysis in the text: "Grey relational analysis is a method within grey system theory that aims to reveal the relationships between different factors or variables by analyzing their similarity. The basic concept is that if two sequences have more similar trends in variation, their degree of correlation is higher.", "Bivariate correlation analysis is a statistical method used to measure the strength and direction of the relationship between two variables. By calculating the correlation coefficient, the linear relationship between the two variables can be quantified. The results can help determine whether there is a positive correlation (both variables increase or decrease together), a negative correlation (one variable increases while the other decreases), or no correlation (no obvious relationship between the two variables).".
(3) If I am not mistaken you want to use cumulative deformation and predict it, however this type of data is temporal and the conditioning factors are spatial, how did you address it?
Answer: Thank you for your careful review. For time-series predictions, the input is the deformation time series; for predicting the deformation rate at unknown spatial points, the input consists of evaluation factors such as DEM.
(4) The explanation of the Elman neural network, CS, and GWO algorithms is overly technical without enough context on why these models were chosen practically. The mathematical models are fine for specialists but could be overwhelming for broader audiences. Consider providing a more intuitive explanation of how CS and GWO optimize the Elman network.
Answer: Thank you for your attention and questions. Based on the revisions, we have added the following in Section 2.3.4: "(1) Choice of Elman Neural Network: The Elman neural network is a neural network model particularly suited for handling time-series data due to its internal feedback mechanism, which allows it to capture the temporal dependencies in data. This makes it especially effective for tasks with time-series characteristics, such as surface deformation prediction. (2) Why Optimization is Needed: Although the Elman neural network can handle time-series data, its training process faces challenges, especially with complex datasets. The network's weights and thresholds can be difficult to adjust to their optimal state, resulting in low prediction accuracy or slow training. Therefore, an optimization algorithm is needed to improve the model's performance. (3) CS Optimization Algorithm: CS is an optimization algorithm inspired by the brood parasitism behavior of cuckoos in nature, using a random search method to find the optimal solution. In neural network training, CS helps optimize the network's weights and thresholds, thereby improving the model's prediction accuracy. GWO Optimization Algorithm: GWO simulates the hunting behavior of wolf packs, introducing leader wolves and follower wolves in the search mechanism to gradually approach the optimal solution. The GWO algorithm is used to adjust the parameters of the Elman neural network, enabling it to find a better-performing solution in a shorter time.".
(5) As the authors mentioned the primary references in detail, there are many studies in this field that have fundamentally explained these SBAS and PSI methods. It is enough to include parameters that are effective in your processing in the methodology section. In addition to the expression of its software. Additional explanations are not necessary.
Answer: Thank you very much for your valuable feedback. According to the revision suggestions, we simplify the methodology section of SBAS-InSAR and PS-InSAR techniques in Section 2.1.
Question #4. Case Study and data sets
The study area section lacks essential explanations, such as about the geological setting. Also, a figure should be given for these explanations. The phenomenon under investigation and prediction is landslides, so a landslide inventory map with full explanations can be provided in this section. We are looking for a reliable prediction, and this detail is important for us in interpreting the results.
Answer: Thank you for your attention and questions. According to the revisions, we have enriched the introduction of the geological environment of the study area in Section 3.1 by adding the following: "The high mountain canyon area of the Minjiang River Basin is located in Wenchuan County, Sichuan Province, China. It is a typical high-risk area for geological disasters, particularly frequent debris flows and landslides.", "The terrain in this area is steep, with a maximum elevation difference of 4185 meters. This extreme elevation variation makes the mountains prone to gravity-induced landslides and debris flows. Due to the low vegetation cover and concentrated rainfall during the rainy season, severe soil erosion occurs in the area. The soil in certain localities is loose, especially after being disturbed by seismic activity, making the soil layers even more unstable and highly susceptible to landslides and the formation of debris flows during heavy rainfall.", "Through field investigations of geological disasters, historical data, and visual interpretation of remote sensing images, a total of 25 debris flow occurrences were identified within the study area. Their specific distribution is shown in Figure 4.".
Question #5. Results
(1) The results section provides quantitative evidence of the model's performance, but it could be better structured. For example, the comparison of SBAS-InSAR and PS-InSAR data should be presented with a clearer connection to the model's predictions.
Answer: Thank you very much for your careful review and valuable corrections. Due to the lack of precise measurement data such as leveling and GPS, SBAS-InSAR and PS-InSAR monitoring results are used for cross-validation to verify the reliability of the time-series InSAR experiment results and the feasibility of the InSAR technique. Then, evaluation factors such as DEM are input into the neural network, with the annual average deformation rate as the output. The output results are compared with those monitored by the InSAR technique to verify the accuracy of the model.
(2) Figures 6 and 7 are visually effective, but the discussion around them lacks depth. It would be helpful to explicitly link these figures to the overall goals of the study. For instance, explaining how the deformation trends in these figures correlate with predictions made by the CS-Elman model.
Answer: Thank you for your kind advice. Using SBAS-InSAR and PS-InSAR techniques, the deformation trends shown in Figures 6 and 7 are obtained to monitor the deformation in the study area. These trends can be used to analyze the deformation from 2022 to 2023 in the study area. Time-series analysis of typical deformation points, combined with rainfall data, provides a basis for the neural network model to predict deformation in 2024. The effectiveness of the predicted deformation trends can be judged by comparing them with the deformation trends monitored by InSAR technology.
(3) Cross-validation results (Figure 9) are important but are not discussed thoroughly. Why is the correlation coefficient of 0.85 significant? Does this meet a specific threshold for accuracy in similar studies?
Answer: Thank you very much for your careful review and valuable corrections. The cross-validation results demonstrate the consistency of surface deformation monitoring between the two InSAR techniques, confirming the reliability of InSAR in monitoring surface deformation. The reference (Liu et al. 2023) reports a significant correlation of 0.81 when using SBAS-InSAR and PS-InSAR for cross-validation. Therefore, the correlation coefficient of 0.85 in this study also indicates a significant correlation.
Liu, H., Xu, X. Y., Chen, M., Chen, F. L., Ding, R. L., and Liu, F.: Time-series InSAR-based dynamic remote sensing monitoring of the Great Wall of the Ming Dynasty in Qinhuangdao, Remote Sensing for Natural Resources, 35, 202-211, https://doi.cnki.net/10.6046/zrzyyg.2021163, 2023.
(4) The discussion of error metrics like MAE and MAPE (Figure 11 and 12) is technical, and the benefits of the CS-Elman model should be articulated more clearly. The practical implications of reducing the error by this margin should be highlighted.
Answer: Thank you for your suggestion. According to the revision suggestions, we add the practical significance of reducing errors: "Therefore, for unknown areas of surface deformation, the CS-Elman model predicts surface deformation more effectively by utilizing evaluation factors such as DEM, potentially reducing the time and technical costs associated with InSAR technology.".
Major point:
(5) you were working and processing on Descending orbit, the words of Subsidence and uplift are correct? (Lines 292-293). Also, lines 286 to 287 must go to the methodology section.
Answer: Thank you for your valuable opinion. The terms "uplift" and "subsidence" are used correctly, as surface deformation includes both uplift and subsidence. According to the revision suggestions, we move the original statement from Section 4.1, "The study utilizes the SBAS-InSAR and PS-InSAR techniques available in the Sarscape 5.6.2 module of ENVI 5.6.2 software to process Sentinel-1A images and extract deformation information in the study area." to the methods section in 2.1.
(6) The processing techniques are different and have different details, but both inputs are the same. How do you evaluate the results obtained from SBAS with the results obtained from PSI? If the S1A images have errors or problems, it is for both. You need another source of data to do this evaluation. This validation seems to be useless and does not stamp approval or rejection on SBAS outputs.
Answer: Thank you for your kind advice. Due to the lack of precise measurement data such as leveling and GPS, and considering that many scholars have already verified the reliability of Sentinel-1A data (Ye et al. 2022; Li, Dai, and Zheng 2024), this study employs SBAS-InSAR and PS-InSAR techniques for cross-validation based on Sentinel-1A data. Compared to monitoring deformation with a single InSAR technique, utilizing multiple time-series InSAR techniques results in higher monitoring accuracy and more reliable outcomes.
Ye, Y. C., Yan, C. D., Luo, X. X., Zhang, R. F., and Yuan, G. J.: Analysis of ground subsidence along Zhengzhou metro based on time series InSAR, National Remote Sensing Bulletin, 26, 1342-1353, https://doi.org/10.11834/jrs.20211246, 2022.
Li, H. R., Dai, S. L., and Zheng, J. X.: Subsidence prediction of high-fill areas based on InSAR monitoring data and the PSO-SVR model, The Chinese Journal of Geological Hazard and Control, 35, 127-136, https://doi.cnki.net/10.16031/j.cnki.issn.1003-8035.202210005, 2024.
(7) The result needs to be visualized. These statistical graphs are not enough.
Answer: Thank you for your careful review. According to the revision suggestions, we have added visual results of the predictions in Section 4.2.3.
Question #6. Comparison with Other Models
(1) The paper compares CS-Elman with GWO-Elman, but the comparison with other models like SVM, LSTM, and PSO-BP is relatively weak. There needs to be a deeper discussion of how these models perform in different scenarios (e.g., different types of terrain). Also, there are some studies that out perform these type of traditional DLMs, may be for evaluation you need to compare with GCN-LSTM, CNN-LSTM, and other hybrid model.
Answer: Thank you for your valuable opinion. Based on the revision comments, we have added experiments on the GCN-LSTM (Chen et al. 2023), CNN-LSTM (Kim and Cho 2019), and BO-XGboost (Qiu et al. 2022) models in section 4.2.2.
Chen, H. H., Zhu, M. Y., Hu, X., Wang, J. R., Sun, Y., and Yang, J. D.: Research on short-term load forecasting of new-type power system based on GCN-LSTM considering multiple influencing factors, Energy Rep., 9, 1022-1031, https://doi.org/10.1016/j.egyr.2023.05.048, 2023.
Kim, T. Y., Cho, S. B.: Predicting residential energy consumption using CNN-LSTM neural networks, Energy, 182, 72-81, https://doi.org/10.1016/j.energy.2019.05.230, 2019.
Qiu, Y. G., Zhou, J., Khandelwal, M., Yang, H. T., Yang, P. X., and Li, C. Q.: Performance evaluation of hybrid WOA-XGBoost, GWO-XGBoost and BO-XGBoost models to predict blast-induced ground vibration, Eng. Comput-germany, 38, 4145-4162, https://doi.org/10.1007/s00366-021-01393-9, 2022.
(2) The section should address why the CS-Elman model outperforms traditional models and what specific features make it better suited for surface deformation prediction. Simply mentioning that the CS-Elman model is superior without a clear analysis of the causes of this superiority is insufficient.
Answer: Thank you for your valuable opinion. According to the revision suggestions, we have added the reasons why the CS-Elman model outperforms traditional models in Section 4.2.2: "The CS-Elman model is superior to traditional models and is more suitable for surface deformation prediction for the following reasons: (1) Model architecture advantage: The CS-Elman model combines the Cuckoo Search (CS) algorithm with the Elman neural network. The CS algorithm’s global search capability optimizes the Elman network’s parameters, thereby improving the model’s convergence speed and prediction accuracy. (2) Nonlinear modeling capability: Unlike traditional linear models, the CS-Elman model captures nonlinear relationships more effectively. Surface deformation is often influenced by multiple complex factors, and the deep learning structure of the CS-Elman model enables it to handle these nonlinear features effectively. (3) Adaptability and flexibility: The CS-Elman model is highly adaptable and can adjust its parameters based on different input features, such as DEM, rainfall, and seismic activity. This allows the model to maintain high prediction performance under various geographical and climatic conditions. (4) Comprehensive factor evaluation: The model can consider multiple evaluation factors simultaneously, providing more comprehensive prediction results. In contrast to traditional models that rely on a single or limited number of variables, CS-Elman excels in multi-factor analysis. (5) Reliability of validation results: Experimental results show that the CS-Elman model outperforms traditional models in terms of accuracy and stability. By comparing with InSAR monitoring results, it effectively reduces prediction errors.".
Question #7. Figures and Tables
(1) Many figures (e.g., Figures 6-12) are effective but require more thorough descriptions in the text. For example, the deformation maps (Figures 6 and 7) should be more directly linked to the points discussed in the results section.
Answer: Thank you for your careful review. In Figures 6 and 7, the typical deformation points P and Q are used in the deformation prediction, where the time-series data from 2022-2023 monitored by InSAR is utilized to predict the deformation of points P and Q in 2024.
(2) Table 5 shows correlation coefficients, but the discussion is overly technical. A clearer explanation of why certain factors was excluded would improve readability.
Answer: Thank you for your suggestion. In Section 2.2 of the methods part, we explain why it is important to exclude highly correlated factors: "Surface deformation is influenced by multiple factors, which exhibit diversity and complexity. Moreover, certain factors may exhibit correlations, and high correlations can lead to model complexity and reduced operational speed. Therefore, conducting correlation analysis among various factors is crucial. By analyzing the correlations among factors, it is possible to exclude highly correlated factors, which is significant for model establishment and surface deformation monitoring.".
(3) All your Figures need the coordinate system, correct all the figures and check them.
Answer: Thanks for your kind suggestion. According to the revision suggestions, we have added coordinate systems to Figures 5, 6, 7, and 17.
Question #8. Discussion and Conclusion
(1) The discussion lacks depth and should more clearly articulate the implications of the findings. For instance, while the paper shows that CS-Elman is superior in prediction accuracy, the practical implications for geological hazard monitoring in high mountain canyon areas should be emphasized.
Answer: Thank you for your kind advice. According to the revision suggestions, In the conclusion section, we added the practical significance of monitoring geological disasters in alpine canyon areas: "The method in this paper validates that the CS-Elman model has a small error compared to InSAR monitoring technology and is suitable for short-term monitoring. Therefore, in practical applications, local evaluation factors such as DEM can be used as inputs to the CS-Elman model, with the annual average deformation rate as the output. Based on the IUGS/WGL (1995) geological hazard intensity classification criteria, with 2 mm/a and 16 mm/a as thresholds, regions where the predicted deformation rate exceeds 16 mm/a should be identified. Protective measures should be taken to prevent geological disasters in these areas.".
IUGS / WGL, International Union of Geological Science Working Group on Landslides, 1995, A suggested method for describing the rate of movement of a landslide. IAEG Bulletin, 1995, 52, pp. 75-78.
(2) The conclusion is weak. It should restate the significance of the findings more strongly and suggest future research directions based on the limitations of the study. For example, the study might explore additional influencing factors or further improve the model by hybridizing CS with other optimization techniques.
Answer: Thank you very much for your careful review. According to the revisions, we have added future research directions in the conclusion:"Additionally, we will consider factors such as distance to rivers, faults, and roads as evaluation factors in the future. We will also improve optimization algorithms like CS by incorporating long short-term memory networks (LSTM), autoregressive integrated moving average models (ARIMA), and exponential smoothing models to predict future deformation.".
Question #9. Language and Grammar
(1) The paper contains some grammatical errors and awkward phrasing. For example, phrases like "the Elman network optimized by CS algorithm exhibits better performance" could be reworded to "the CS-optimized Elman network performs better."(2) The tone is quite technical and might be difficult for some readers. Simplifying the language in certain sections would improve accessibility.
Answer: Thank you very much for your careful review. According to the revision suggestions, we simplified the sentences, for example, changing "the Elman network optimized by CS algorithm exhibits better performance" to "the CS-optimized Elman network performs better".
Question #10. Additional Points
(1) Limitations: The limitations of the study are not discussed enough. For example, the issue of vegetation interference is briefly mentioned but not explored in terms of how it could be mitigated in future research.
Answer: Thank you for your careful review. According to the revisions, we have added the following in the conclusion: "To address the issue of vegetation affecting the accuracy of deformation monitoring, the following measures will be taken in the future to mitigate vegetation interference: (1) Data preprocessing: Use image processing techniques, such as denoising and image segmentation, to identify and remove areas affected by vegetation interference. (2) Choosing the right timing: Collect data during different seasons of vegetation growth to select periods with minimal interference. (3) Multi-source data fusion: Combine different data sources, such as optical remote sensing and LiDAR, to improve monitoring accuracy.".
(2) Reproducibility: The steps for constructing the CS-Elman and GWO-Elman models are well detailed, but more information on the data preprocessing (especially normalization techniques) is needed to ensure that others can replicate the study.
Answer: Thank you very much for your valuable feedback. According to the revision suggestions, we have added information on normalization techniques in Step 1.
-
AC2: 'Reply on RC2', Baoyun Wang, 29 Sep 2024
Thank you for your detailed feedback and suggestions on my article. I will respond to each piece of feedback from the reviewers one by one for your review. Additionally, I have revised the manuscript, and all revisions are highlighted in red font.
Question #1. Abstract
(1) The abstract is informative, but it needs refinement for clarity and conciseness. For example, mentioning the specific results of MAE and MAPE for other methods such as LSTM and SVM etc. would enhance understanding.
Answer: Thank you for your kind advice. Based on the revision comments, we have simplified the results section of the abstract and mentioned the specific results of the CS-Elman model compared to other models: "(3) By comparing with models such as SVM, the CS-Elman model has the smallest error range (0.071–1.843 mm/a) and better accuracy (MAE 0.818 mm/a, MAPE 9.353%).".
(2) The abstract should also briefly mention the significance of the problem, i.e., why improving surface deformation predictions in high mountain canyon areas is important for hazard prevention (the motivation of doing this study like three or four lines of introduction section).
Answer: Thank you for your careful review. According to the revision suggestions, we add the statement "Mountain canyon areas often have complex terrain and unstable geological conditions, making them prone to disasters such as landslides and debris flows. Surface deformation is an early sign of these disasters, and accurate prediction of deformation can help detect potential hazards in advance, reducing the likelihood of disaster occurrence." in the abstract to highlight the importance of predicting surface deformation in high mountain canyon areas.
(3) This type of writing for Abstract is not common 'Results indicates' near 10 lines.
Answer: Thank you for your valuable opinion. According to the revision suggestions, the results section of the abstract is simplified: "Results indicate: (1) SBAS-InSAR and PS-InSAR techniques have a strong correlation (R2 = 0.85) in monitored deformation rates, confirming their potential for joint analysis. (2) The CS-Elman model outperforms the GWO-Elman model, achieving a lower error range, faster convergence (3 iterations), and better metrics (MSE: 0.020 mm/a, MAE: 1.620 mm/a, MAPE: 21.500%). (3) Compared to SVM, LSTM, and PSO-BP models, the CS-Elman model has the smallest error range (0.071–1.843 mm/a) and better accuracy (MAE 0.818 mm/a, MAPE 9.353%). (4) The CS-Elman model excels in short-term predictions but is less effective for long-term forecasting. It predicts a maximum surface uplift of 120.913 mm, offering insights for disaster prevention.".
Question #2. Introduction
(1) The introduction adequately establishes the context but is verbose and could benefit from restructuring for clarity. Sentences like "China is a country prone to geological disasters" could be simplified.
Answer: Thank you for your careful review. According to the revision suggestions, we simplified the language in the introduction. For example, we changed "China is a country prone to geological disasters, and various types of geological disasters have caused enormous losses to the lives and property security of its people." to "China is prone to geological disasters, which have caused significant losses to lives and property security of its people." We also simplified "Landslides, debris flows and collapses are major geological disasters characterized by strong concealment, significant hazards and high suddenness, and they are widely distributed in mountainous areas and canyons in China." to "Landslides, debris flows, and collapses, common in mountainous and canyon areas, are highly hazardous due to their suddenness and hidden nature".
(2) The literature review is extensive, but it lacks a critical discussion of the limitations of the current approaches. More emphasis should be placed on how the proposed model fills these gaps in the literature. (3) The problem statement should be more focused. While the paper acknowledges the non-linearity of surface deformation in high mountain canyon areas, the rationale for choosing CS and GWO as optimization techniques should be more explicitly linked to these challenges.
Answer: Thank you for your suggestion. According to the revision suggestions, we added a discussion of the limitations of current methods in the introduction, as well as how the proposed model addresses these issues: "Although the aforementioned methods can effectively combine InSAR technology and neural network models to monitor and predict surface deformation information, they all have certain drawbacks and limitations. Some scholars (Teng, Wang, and Jiang 2022; Yang et al. 2022; Ye et al. 2022) propose prediction models that overly rely on deformation data, considering fewer other factors that may trigger disasters. They only utilize deformation data as the input and output layers for prediction, which presents certain shortcomings. If a model is based solely on deformation data, its lack of multi-factor analysis capability will limit its prediction accuracy and comprehensiveness. Deformation data typically reflect slow changes, making it difficult to predict sudden disasters such as debris flows and landslides. Additionally, other scholars (Wang et al. 2019; Radman, Akhoondzadeh, and Hosseiny 2021) construct prediction models from the perspective of influencing factors, but the prediction accuracy of these models is relatively low, primarily due to insufficient optimization of the weights and thresholds of the network model. To enhance model performance, it is necessary to introduce optimization algorithms, such as Cuckoo Search (CS) and Grey Wolf Optimization (GWO), to better optimize network parameters.".
(4) The introduction presentation is generally unsuitable, comprehensive, and in order. First, you start from InSAR and jump to prediction, and then you come back, bring references from SBAS and PSI, and go to prediction again. It is not very readable and does not convey useful information. The paragraphs should be in order and use a mix of old and new references. Use various and new references; for example, Khalili et al. 2023 and Sun et al. 2024 have up-to-date studies with high accuracy on landslide prediction by InSAR and Hybrid AI model.
Answer: Thanks for your kind suggestion. The first paragraph of the introduction explains the importance of monitoring and predicting surface deformation in mountain canyon areas. The second paragraph first discusses InSAR monitoring and then transitions to prediction, citing references that combine InSAR technology with neural network models for monitoring and predicting surface deformation, ultimately summarizing the shortcomings of the aforementioned scholars. The third paragraph introduces the content of this paper based on these shortcomings: geological disasters occur due to multiple factors, so this study uses multiple factors as inputs to the network model. Since the weights and thresholds of the network model significantly influence its performance, it introduces optimization algorithms, such as Cuckoo Search (CS) and Grey Wolf Optimization (GWO), to optimize the parameters of the neural network. The fourth paragraph provides a summary.
Question #3. Methodology
(1) The methodology is detailed but lacks clarity in several places. The flowchart in Figure 1 is helpful, but more explanation is needed to connect the steps logically.
Answer: Thank you for your kind advice. According to the revision suggestions, we described Figure 1 as follows: "This study first uses SBAS-InSAR and PS-InSAR techniques to obtain surface deformation information in the study area. Then, the optimal evaluation factors are selected based on correlation analysis. Next, the optimal evaluation factors are used as inputs for the CS-Elman and GWO-Elman models, with the annual average deformation rate as the output. The output results are compared with the annual average deformation rate obtained from InSAR technology to verify the model's accuracy. Finally, the optimal model is applied for future time series deformation prediction. The overall technical process is shown in Figure 1.".
(2) The choice of 13 factors is justified through grey relational and correlation analysis, but the explanation of these techniques is somewhat technical for a general audience. Simplifying or expanding explanations for readers unfamiliar with these statistical methods would be beneficial. Also, open the part related to lines 133 to 139 and explain more.
Answer: Thank you very much for your careful review. According to the revision suggestions, we added the concepts of grey relational analysis and correlation analysis in the text: "Grey relational analysis is a method within grey system theory that aims to reveal the relationships between different factors or variables by analyzing their similarity. The basic concept is that if two sequences have more similar trends in variation, their degree of correlation is higher.", "Bivariate correlation analysis is a statistical method used to measure the strength and direction of the relationship between two variables. By calculating the correlation coefficient, the linear relationship between the two variables can be quantified. The results can help determine whether there is a positive correlation (both variables increase or decrease together), a negative correlation (one variable increases while the other decreases), or no correlation (no obvious relationship between the two variables).".
(3) If I am not mistaken you want to use cumulative deformation and predict it, however this type of data is temporal and the conditioning factors are spatial, how did you address it?
Answer: Thank you for your careful review. For time-series predictions, the input is the deformation time series; for predicting the deformation rate at unknown spatial points, the input consists of evaluation factors such as DEM.
(4) The explanation of the Elman neural network, CS, and GWO algorithms is overly technical without enough context on why these models were chosen practically. The mathematical models are fine for specialists but could be overwhelming for broader audiences. Consider providing a more intuitive explanation of how CS and GWO optimize the Elman network.
Answer: Thank you for your attention and questions. Based on the revisions, we have added the following in Section 2.3.4: "(1) Choice of Elman Neural Network: The Elman neural network is a neural network model particularly suited for handling time-series data due to its internal feedback mechanism, which allows it to capture the temporal dependencies in data. This makes it especially effective for tasks with time-series characteristics, such as surface deformation prediction. (2) Why Optimization is Needed: Although the Elman neural network can handle time-series data, its training process faces challenges, especially with complex datasets. The network's weights and thresholds can be difficult to adjust to their optimal state, resulting in low prediction accuracy or slow training. Therefore, an optimization algorithm is needed to improve the model's performance. (3) CS Optimization Algorithm: CS is an optimization algorithm inspired by the brood parasitism behavior of cuckoos in nature, using a random search method to find the optimal solution. In neural network training, CS helps optimize the network's weights and thresholds, thereby improving the model's prediction accuracy. GWO Optimization Algorithm: GWO simulates the hunting behavior of wolf packs, introducing leader wolves and follower wolves in the search mechanism to gradually approach the optimal solution. The GWO algorithm is used to adjust the parameters of the Elman neural network, enabling it to find a better-performing solution in a shorter time.".
(5) As the authors mentioned the primary references in detail, there are many studies in this field that have fundamentally explained these SBAS and PSI methods. It is enough to include parameters that are effective in your processing in the methodology section. In addition to the expression of its software. Additional explanations are not necessary.
Answer: Thank you very much for your valuable feedback. According to the revision suggestions, we simplify the methodology section of SBAS-InSAR and PS-InSAR techniques in Section 2.1.
Question #4. Case Study and data sets
The study area section lacks essential explanations, such as about the geological setting. Also, a figure should be given for these explanations. The phenomenon under investigation and prediction is landslides, so a landslide inventory map with full explanations can be provided in this section. We are looking for a reliable prediction, and this detail is important for us in interpreting the results.
Answer: Thank you for your attention and questions. According to the revisions, we have enriched the introduction of the geological environment of the study area in Section 3.1 by adding the following: "The high mountain canyon area of the Minjiang River Basin is located in Wenchuan County, Sichuan Province, China. It is a typical high-risk area for geological disasters, particularly frequent debris flows and landslides.", "The terrain in this area is steep, with a maximum elevation difference of 4185 meters. This extreme elevation variation makes the mountains prone to gravity-induced landslides and debris flows. Due to the low vegetation cover and concentrated rainfall during the rainy season, severe soil erosion occurs in the area. The soil in certain localities is loose, especially after being disturbed by seismic activity, making the soil layers even more unstable and highly susceptible to landslides and the formation of debris flows during heavy rainfall.", "Through field investigations of geological disasters, historical data, and visual interpretation of remote sensing images, a total of 25 debris flow occurrences were identified within the study area. Their specific distribution is shown in Figure 4.".
Question #5. Results
(1) The results section provides quantitative evidence of the model's performance, but it could be better structured. For example, the comparison of SBAS-InSAR and PS-InSAR data should be presented with a clearer connection to the model's predictions.
Answer: Thank you very much for your careful review and valuable corrections. Due to the lack of precise measurement data such as leveling and GPS, SBAS-InSAR and PS-InSAR monitoring results are used for cross-validation to verify the reliability of the time-series InSAR experiment results and the feasibility of the InSAR technique. Then, evaluation factors such as DEM are input into the neural network, with the annual average deformation rate as the output. The output results are compared with those monitored by the InSAR technique to verify the accuracy of the model.
(2) Figures 6 and 7 are visually effective, but the discussion around them lacks depth. It would be helpful to explicitly link these figures to the overall goals of the study. For instance, explaining how the deformation trends in these figures correlate with predictions made by the CS-Elman model.
Answer: Thank you for your kind advice. Using SBAS-InSAR and PS-InSAR techniques, the deformation trends shown in Figures 6 and 7 are obtained to monitor the deformation in the study area. These trends can be used to analyze the deformation from 2022 to 2023 in the study area. Time-series analysis of typical deformation points, combined with rainfall data, provides a basis for the neural network model to predict deformation in 2024. The effectiveness of the predicted deformation trends can be judged by comparing them with the deformation trends monitored by InSAR technology.
(3) Cross-validation results (Figure 9) are important but are not discussed thoroughly. Why is the correlation coefficient of 0.85 significant? Does this meet a specific threshold for accuracy in similar studies?
Answer: Thank you very much for your careful review and valuable corrections. The cross-validation results demonstrate the consistency of surface deformation monitoring between the two InSAR techniques, confirming the reliability of InSAR in monitoring surface deformation. The reference (Liu et al. 2023) reports a significant correlation of 0.81 when using SBAS-InSAR and PS-InSAR for cross-validation. Therefore, the correlation coefficient of 0.85 in this study also indicates a significant correlation.
Liu, H., Xu, X. Y., Chen, M., Chen, F. L., Ding, R. L., and Liu, F.: Time-series InSAR-based dynamic remote sensing monitoring of the Great Wall of the Ming Dynasty in Qinhuangdao, Remote Sensing for Natural Resources, 35, 202-211, https://doi.cnki.net/10.6046/zrzyyg.2021163, 2023.
(4) The discussion of error metrics like MAE and MAPE (Figure 11 and 12) is technical, and the benefits of the CS-Elman model should be articulated more clearly. The practical implications of reducing the error by this margin should be highlighted.
Answer: Thank you for your suggestion. According to the revision suggestions, we add the practical significance of reducing errors: "Therefore, for unknown areas of surface deformation, the CS-Elman model predicts surface deformation more effectively by utilizing evaluation factors such as DEM, potentially reducing the time and technical costs associated with InSAR technology.".
Major point:
(5) you were working and processing on Descending orbit, the words of Subsidence and uplift are correct? (Lines 292-293). Also, lines 286 to 287 must go to the methodology section.
Answer: Thank you for your valuable opinion. The terms "uplift" and "subsidence" are used correctly, as surface deformation includes both uplift and subsidence. According to the revision suggestions, we move the original statement from Section 4.1, "The study utilizes the SBAS-InSAR and PS-InSAR techniques available in the Sarscape 5.6.2 module of ENVI 5.6.2 software to process Sentinel-1A images and extract deformation information in the study area." to the methods section in 2.1.
(6) The processing techniques are different and have different details, but both inputs are the same. How do you evaluate the results obtained from SBAS with the results obtained from PSI? If the S1A images have errors or problems, it is for both. You need another source of data to do this evaluation. This validation seems to be useless and does not stamp approval or rejection on SBAS outputs.
Answer: Thank you for your kind advice. Due to the lack of precise measurement data such as leveling and GPS, and considering that many scholars have already verified the reliability of Sentinel-1A data (Ye et al. 2022; Li, Dai, and Zheng 2024), this study employs SBAS-InSAR and PS-InSAR techniques for cross-validation based on Sentinel-1A data. Compared to monitoring deformation with a single InSAR technique, utilizing multiple time-series InSAR techniques results in higher monitoring accuracy and more reliable outcomes.
Ye, Y. C., Yan, C. D., Luo, X. X., Zhang, R. F., and Yuan, G. J.: Analysis of ground subsidence along Zhengzhou metro based on time series InSAR, National Remote Sensing Bulletin, 26, 1342-1353, https://doi.org/10.11834/jrs.20211246, 2022.
Li, H. R., Dai, S. L., and Zheng, J. X.: Subsidence prediction of high-fill areas based on InSAR monitoring data and the PSO-SVR model, The Chinese Journal of Geological Hazard and Control, 35, 127-136, https://doi.cnki.net/10.16031/j.cnki.issn.1003-8035.202210005, 2024.
(7) The result needs to be visualized. These statistical graphs are not enough.
Answer: Thank you for your careful review. According to the revision suggestions, we have added visual results of the predictions in Section 4.2.3.
Question #6. Comparison with Other Models
(1) The paper compares CS-Elman with GWO-Elman, but the comparison with other models like SVM, LSTM, and PSO-BP is relatively weak. There needs to be a deeper discussion of how these models perform in different scenarios (e.g., different types of terrain). Also, there are some studies that out perform these type of traditional DLMs, may be for evaluation you need to compare with GCN-LSTM, CNN-LSTM, and other hybrid model.
Answer: Thank you for your valuable opinion. Based on the revision comments, we have added experiments on the GCN-LSTM (Chen et al. 2023), CNN-LSTM (Kim and Cho 2019), and BO-XGboost (Qiu et al. 2022) models in section 4.2.2.
Chen, H. H., Zhu, M. Y., Hu, X., Wang, J. R., Sun, Y., and Yang, J. D.: Research on short-term load forecasting of new-type power system based on GCN-LSTM considering multiple influencing factors, Energy Rep., 9, 1022-1031, https://doi.org/10.1016/j.egyr.2023.05.048, 2023.
Kim, T. Y., Cho, S. B.: Predicting residential energy consumption using CNN-LSTM neural networks, Energy, 182, 72-81, https://doi.org/10.1016/j.energy.2019.05.230, 2019.
Qiu, Y. G., Zhou, J., Khandelwal, M., Yang, H. T., Yang, P. X., and Li, C. Q.: Performance evaluation of hybrid WOA-XGBoost, GWO-XGBoost and BO-XGBoost models to predict blast-induced ground vibration, Eng. Comput-germany, 38, 4145-4162, https://doi.org/10.1007/s00366-021-01393-9, 2022.
(2) The section should address why the CS-Elman model outperforms traditional models and what specific features make it better suited for surface deformation prediction. Simply mentioning that the CS-Elman model is superior without a clear analysis of the causes of this superiority is insufficient.
Answer: Thank you for your valuable opinion. According to the revision suggestions, we have added the reasons why the CS-Elman model outperforms traditional models in Section 4.2.2: "The CS-Elman model is superior to traditional models and is more suitable for surface deformation prediction for the following reasons: (1) Model architecture advantage: The CS-Elman model combines the Cuckoo Search (CS) algorithm with the Elman neural network. The CS algorithm’s global search capability optimizes the Elman network’s parameters, thereby improving the model’s convergence speed and prediction accuracy. (2) Nonlinear modeling capability: Unlike traditional linear models, the CS-Elman model captures nonlinear relationships more effectively. Surface deformation is often influenced by multiple complex factors, and the deep learning structure of the CS-Elman model enables it to handle these nonlinear features effectively. (3) Adaptability and flexibility: The CS-Elman model is highly adaptable and can adjust its parameters based on different input features, such as DEM, rainfall, and seismic activity. This allows the model to maintain high prediction performance under various geographical and climatic conditions. (4) Comprehensive factor evaluation: The model can consider multiple evaluation factors simultaneously, providing more comprehensive prediction results. In contrast to traditional models that rely on a single or limited number of variables, CS-Elman excels in multi-factor analysis. (5) Reliability of validation results: Experimental results show that the CS-Elman model outperforms traditional models in terms of accuracy and stability. By comparing with InSAR monitoring results, it effectively reduces prediction errors.".
Question #7. Figures and Tables
(1) Many figures (e.g., Figures 6-12) are effective but require more thorough descriptions in the text. For example, the deformation maps (Figures 6 and 7) should be more directly linked to the points discussed in the results section.
Answer: Thank you for your careful review. In Figures 6 and 7, the typical deformation points P and Q are used in the deformation prediction, where the time-series data from 2022-2023 monitored by InSAR is utilized to predict the deformation of points P and Q in 2024.
(2) Table 5 shows correlation coefficients, but the discussion is overly technical. A clearer explanation of why certain factors was excluded would improve readability.
Answer: Thank you for your suggestion. In Section 2.2 of the methods part, we explain why it is important to exclude highly correlated factors: "Surface deformation is influenced by multiple factors, which exhibit diversity and complexity. Moreover, certain factors may exhibit correlations, and high correlations can lead to model complexity and reduced operational speed. Therefore, conducting correlation analysis among various factors is crucial. By analyzing the correlations among factors, it is possible to exclude highly correlated factors, which is significant for model establishment and surface deformation monitoring.".
(3) All your Figures need the coordinate system, correct all the figures and check them.
Answer: Thanks for your kind suggestion. According to the revision suggestions, we have added coordinate systems to Figures 5, 6, 7, and 17.
Question #8. Discussion and Conclusion
(1) The discussion lacks depth and should more clearly articulate the implications of the findings. For instance, while the paper shows that CS-Elman is superior in prediction accuracy, the practical implications for geological hazard monitoring in high mountain canyon areas should be emphasized.
Answer: Thank you for your kind advice. According to the revision suggestions, In the conclusion section, we added the practical significance of monitoring geological disasters in alpine canyon areas: "The method in this paper validates that the CS-Elman model has a small error compared to InSAR monitoring technology and is suitable for short-term monitoring. Therefore, in practical applications, local evaluation factors such as DEM can be used as inputs to the CS-Elman model, with the annual average deformation rate as the output. Based on the IUGS/WGL (1995) geological hazard intensity classification criteria, with 2 mm/a and 16 mm/a as thresholds, regions where the predicted deformation rate exceeds 16 mm/a should be identified. Protective measures should be taken to prevent geological disasters in these areas.".
IUGS / WGL, International Union of Geological Science Working Group on Landslides, 1995, A suggested method for describing the rate of movement of a landslide. IAEG Bulletin, 1995, 52, pp. 75-78.
(2) The conclusion is weak. It should restate the significance of the findings more strongly and suggest future research directions based on the limitations of the study. For example, the study might explore additional influencing factors or further improve the model by hybridizing CS with other optimization techniques.
Answer: Thank you very much for your careful review. According to the revisions, we have added future research directions in the conclusion: "Additionally, we will consider factors such as distance to rivers, faults, and roads as evaluation factors in the future. We will also improve optimization algorithms like CS by incorporating long short-term memory networks (LSTM), autoregressive integrated moving average models (ARIMA), and exponential smoothing models to predict future deformation.".
Question #9. Language and Grammar
(1) The paper contains some grammatical errors and awkward phrasing. For example, phrases like "the Elman network optimized by CS algorithm exhibits better performance" could be reworded to "the CS-optimized Elman network performs better."(2) The tone is quite technical and might be difficult for some readers. Simplifying the language in certain sections would improve accessibility.
Answer: Thank you very much for your careful review. According to the revision suggestions, we simplified the sentences, for example, changing "the Elman network optimized by CS algorithm exhibits better performance" to "the CS-optimized Elman network performs better".
Question #10. Additional Points
(1) Limitations: The limitations of the study are not discussed enough. For example, the issue of vegetation interference is briefly mentioned but not explored in terms of how it could be mitigated in future research.
Answer: Thank you for your careful review. According to the revisions, we have added the following in the conclusion: "To address the issue of vegetation affecting the accuracy of deformation monitoring, the following measures will be taken in the future to mitigate vegetation interference: (1) Data preprocessing: Use image processing techniques, such as denoising and image segmentation, to identify and remove areas affected by vegetation interference. (2) Choosing the right timing: Collect data during different seasons of vegetation growth to select periods with minimal interference. (3) Multi-source data fusion: Combine different data sources, such as optical remote sensing and LiDAR, to improve monitoring accuracy.".
(2) Reproducibility: The steps for constructing the CS-Elman and GWO-Elman models are well detailed, but more information on the data preprocessing (especially normalization techniques) is needed to ensure that others can replicate the study.
Answer: Thank you very much for your valuable feedback. According to the revision suggestions, we have added information on normalization techniques in Step 1.
Citation: https://doi.org/10.5194/egusphere-2024-1220-AC2
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 260 | 57 | 391 | 708 | 27 | 14 | 15 |
- HTML: 260
- PDF: 57
- XML: 391
- Total: 708
- Supplement: 27
- BibTeX: 14
- EndNote: 15
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1