the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
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: open (until 10 Jul 2024)
-
RC1: 'Comment on egusphere-2024-1220', Anonymous Referee #1, 25 Jun 2024
reply
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
Viewed
HTML | XML | Total | Supplement | BibTeX | EndNote | |
---|---|---|---|---|---|---|
116 | 29 | 13 | 158 | 13 | 4 | 5 |
- HTML: 116
- PDF: 29
- XML: 13
- Total: 158
- Supplement: 13
- BibTeX: 4
- EndNote: 5
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1