Large-scale landslide deformation prediction method based on insar inversion and multiple influence factors

Abstract

The application discloses a large-range landslide deformation prediction method based on InSAR inversion and multiple influence factors, comprising: performing InSAR inversion processing on SAR image data to obtain time-series deformation data; clustering the time-series deformation data to obtain time-series deformation data of several categories; decomposing the time-series deformation data of each category into a periodic term deformation sequence and a trend term deformation sequence; determining an influence factor significantly related to the periodic term deformation; respectively establishing LSTM models of the time-series deformation data of each category to realize prediction of the deformation of each category; adding the periodic term deformation prediction value and the trend term deformation prediction value of each category to obtain a deformation amount prediction result of each category; and merging the deformation amount prediction results of each category to obtain a large-range landslide deformation prediction result. The application can effectively predict large-range landslide deformation, and solves the problems of small prediction range and high cost in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of geological disaster prediction, in particular to a large-range landslide deformation prediction method based on InSAR inversion and multiple influence factors. BACKGROUND

[0002] Landslide is a complex geological evolution process triggered by internal and external factors such as geological structure and rainfall. Frequent and widely distributed landslides and their chain disasters have a serious impact on regional water resources development, new urbanization construction, railway and highway traffic trunk line engineering construction and operation. Landslide deformation prediction, as an effective means to realize landslide disaster prediction, is one of the basic works of landslide disaster prevention, and the related theories and methods have also developed rapidly and been applied in recent years.

[0003] The invention patent with publication number CN112270400A takes into account the influence of landslide deformation influence factors on landslides, and uses a deep learning LSTM (Long Short-Term Memory) network to establish a model, considering the dynamic characteristics of landslide evolution, and improves the prediction accuracy. However, this scheme uses single-point monitoring data obtained by traditional landslide monitoring as the main data source. This method is limited by the small monitoring range and high cost of traditional landslide monitoring, and the single-point monitoring / prediction result cannot comprehensively reflect the overall deformation characteristics of the landslide.

[0004] The invention patent with publication number CN113251947A uses InSAR technology to obtain ground deformation time series results, and then combines a deep learning LSTM network to establish a model to predict ground deformation. However, on the one hand, this method does not classify the InSAR ground deformation inversion results, and cannot consider different deformation trends in different regions, which may affect the prediction effect of deformation; on the other hand, this method does not consider the influence of landslide deformation influence factors, and is not suitable for landslide areas that are greatly affected by dynamic influence factors. SUMMARY

[0005] Therefore, the purpose of the present application is to overcome the defects in the prior art, and to provide a large-range landslide deformation prediction method based on InSAR inversion and multiple influence factors, which can effectively predict large-range landslide deformation and solve the problems of small prediction range and high cost in the prior art.

[0006] The large-range landslide deformation prediction method based on InSAR inversion and multiple influence factors of the present application comprises the following steps:

[0007] S1. Collect SAR image data of the target area, and perform InSAR inversion processing on the SAR image data to obtain time series deformation data;

[0008] S2. Cluster the time series deformation data to obtain time series deformation data of several categories;

[0009] S3. Decompose the time series deformation data of each category into a periodic term deformation sequence and a trend term deformation sequence;

[0010] S4. Extract several influence factors, calculate the correlation degree between each influence factor and the periodic term deformation sequence of each category, select the influence factors significantly related to the periodic term deformation, and take the significantly related influence factors as target influence factors;

[0011] S5. Establish an LSTM model for each category of time series deformation data, specifically including:

[0012] For the trend term deformation of the same category, the trend term deformation is taken as a variable input to establish a trend term single-variable LSTM model;

[0013] For the periodic term deformation of the same category, the periodic term deformation and the target influence factor are taken as variable inputs to establish a periodic term multivariate LSTM model;

[0014] S6. Network model training is performed on the single-variable LSTM model and the multivariate LSTM model in step S5 to obtain a trained LSTM model, and the trained LSTM model is used to predict the deformation of each category to obtain periodic term deformation prediction values and trend term deformation prediction values of each category;

[0015] S7. The periodic term deformation prediction values and the trend term deformation prediction values of each category are added to obtain deformation variable prediction results of each category, and the deformation variable prediction results of each category are combined to obtain large-scale landslide deformation prediction results.

[0016] Further, the SAR image data is subjected to InSAR inversion processing, specifically including:

[0017] S11. Preprocessing of SAR image data; the preprocessing includes data import and data cropping;

[0018] S12. Select a super master image from the SAR image, and register all other images to the super master image;

[0019] S13. Obtain the basic data for SBAS inversion estimation through baseline estimation, connection graph generation, differential interference, interferogram filtering, and phase unwrapping processing;

[0020] S14. Remove residual terrain errors and residual atmospheric phases from the basic data, and convert the deformation information into units to obtain a deformation result; the deformation result includes a deformation rate and a deformation time sequence;

[0021] S15. Convert the deformation result to a geographic coordinate system to obtain a plurality of high-coherence points of the target area, each of which has time-series deformation data of a plurality of time periods.

[0022] Further, the time-series deformation data is subjected to clustering processing, specifically including:

[0023] S21. Randomly determine N data in all time-series deformation data as clustering centers;

[0024] S22. Calculate the cosine similarity value CS of each time-series deformation data with the N clustering centers respectively, and assign the data to the cluster with the highest CS value;

[0025] S23. Calculate the CS value of each data in each cluster with other data in the cluster, and redefine the data with the highest average CS value as the clustering center;

[0026] S24. Repeat steps S22-S23 for iteration until the data in each cluster no longer changes or the iteration reaches K times;

[0027] S25. Calculate the CS value of each clustering center data, if the CS value is greater than A, add 1 to the category number N, and re-cluster; if the CS value is less than B, reduce 1 to the category number N, and re-cluster; after clustering, output the data of each category respectively.

[0028] Further, the cosine similarity value CS is calculated according to the following formula:

[0029]

[0030] wherein, and represent two different time-series data, x0, …, x N-1 and y0, …, y N-1 represent the deformation of N time series respectively.

[0031] Further, in step S3, the decomposition uses wavelet decomposition, the decomposition function of the wavelet is Daubechies, and the decomposition level of the wavelet is four.

[0032] Further, the plurality of influence factors include rainfall, reservoir water level elevation, and reservoir water level change.

[0033] Further, the correlation degree between each influence factor and the periodic term deformation sequence of each category is calculated according to the following steps:

[0034] S41. Take the periodic term deformation sequence as Take the n influence factor sequences as And the periodic term deformation sequence With the impact factor sequence Perform normalization processing;

[0035] S42. Calculate the absolute difference |x0(t)-x| between corresponding elements of the periodic deformation sequence and the influence factor sequence. i (t)|;where x0(t) is the periodic deformation sequence The value of x at time t i (t) is the impact factor sequence The value at time t;

[0036] S43. Calculate the correlation coefficient ξ 0i (t):

[0037]

[0038] Where, ξ 0i (t) is a periodic term deformation sequence With the impact factor sequence The correlation coefficient at time t, This represents the minimum absolute difference between two sequences. The maximum absolute difference between the two sequences is represented by ρ, where ρ is the resolution coefficient.

[0039] S44. Calculate the correlation degree r 0i :

[0040]

[0041] Where n is the length of the time sequence.

[0042] The beneficial effects of this invention are as follows: This invention discloses a large-scale landslide deformation prediction method based on InSAR inversion and multiple influencing factors. Using InSAR inversion results as the data source, it solves the problems of small scope and high cost associated with existing technologies that utilize traditional landslide monitoring data. Compared to existing technologies, it can obtain large-scale deformation prediction results, better reflecting the deformation characteristics of the entire landslide area. By clustering the surface deformation results obtained from InSAR inversion and predicting the clustered results separately, it solves the problem of low prediction accuracy caused by directly using InSAR surface deformation results for prediction in existing technologies. By decomposing the clustered surface deformation results and incorporating dynamic influencing factors of landslide deformation such as rainfall and reservoir water level, the InSAR inversion results are made suitable for landslide prediction, making the prediction more accurate. Attached Figure Description

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0044] Figure 1This is a schematic diagram of the prediction method of the present invention;

[0045] Figure 2 This is a schematic diagram of the time-series deformation decomposition results of a highly coherent point in Category 1 of the present invention;

[0046] Figure 3 This is a schematic diagram of the LSTM cell unit structure of the present invention;

[0047] Figure 4 This is a schematic diagram comparing the predicted deformation results with the actual deformation of selected points in Category 1 of the present invention.

[0048] Figure 5 This is a schematic diagram of the large-scale landslide deformation prediction results of the present invention. Detailed Implementation

[0049] The present invention will be further described below with reference to the accompanying drawings, as shown in the figures:

[0050] The large-scale landslide deformation prediction method based on InSAR inversion and multiple influencing factors of the present invention includes the following steps:

[0051] S1. Collect SAR image data of the target area and perform InSAR inversion processing on the SAR image data to obtain time-series deformation data;

[0052] S2. Cluster the time-series deformation data to obtain several categories of time-series deformation data;

[0053] S3. Decompose the time-series deformation data of each category into periodic deformation sequences and trend deformation sequences;

[0054] S4. Extract several influencing factors, calculate the correlation between each influencing factor and the periodic term deformation sequence of each category, select the influencing factors that are significantly related to the periodic term deformation, and take the significantly related influencing factors as the target influencing factors;

[0055] S5. Establish LSTM models for each category of time-series deformation data, specifically including:

[0056] For trend item deformation of the same category, the trend item deformation is used as a variable input to build a univariate LSTM model of the trend item.

[0057] For the same category of periodic term deformation, the periodic term deformation and the target influencing factor are used as variable inputs to establish a periodic term multivariate LSTM model;

[0058] S6. Network model training is performed on the single-variable LSTM model and the multi-variable LSTM model in step S5 to obtain trained LSTM models, and each type of deformation is predicted using the trained LSTM models to obtain each type of periodic term deformation prediction value and trend term deformation prediction value;

[0059] S7. The periodic term deformation prediction value and the trend term deformation prediction value of each type are added to obtain the deformation variable prediction result of each type, and the deformation variable prediction results of each type are combined to obtain a large-scale landslide deformation prediction result.

[0060] In this embodiment, with the development of InSAR (Interferometric Synthetic Aperture Radar), it has been widely used in ground deformation monitoring. The inversion result of InSAR can be used as a reliable data source for landslide deformation prediction.

[0061] A certain study area is selected as a target area, and SBAS-InSAR technology is used to process 30 scenes of time series Sentinel-1 SAR images of the study area using Envi software, and the time interval of each image is 12 days, and the process is as follows:

[0062] S11. Preprocessing of SAR image data; the preprocessing includes data import and data cropping; the data cropping specifically crops the study area range;

[0063] S12. Select a SAR image with good quality as a super master image, and register all other images to the super master image to realize image registration;

[0064] S13. Through baseline estimation, connection graph generation, differential interference, interferogram filtering and phase unwrapping processing, the basic data for SBAS inversion estimation is obtained;

[0065] S14. Remove residual terrain errors and residual atmospheric phases from the basic data by SBAS inversion estimation, convert the deformation information to units, and obtain the deformation result; the deformation result includes deformation rate and deformation time series;

[0066] S15. Convert the deformation result to the geographic coordinate system to obtain 2765 high-coherence points in the target area, each high-coherence point having 30 periods of time series deformation data.

[0067] In this embodiment, the cosine similarity is used as a similarity measure to cluster the time series deformation result data obtained by InSAR technology to define different deformation regions. The specific steps are as follows:

[0068] S21. Set the number of cluster categories as N, and randomly determine N data as cluster centers in all time series deformation data;

[0069] S22. Calculate the cosine similarity CS of each time series deformation data with N cluster centers respectively, and allocate the data to the cluster with the highest CS value; all data are finally divided into N categories;

[0070] S23. Calculate the CS value of each data in each cluster with other data in the cluster, and redefine the data with the highest average CS value as the cluster center;

[0071] S24. Repeat steps S22-S23 for iteration until the data in each cluster no longer changes or the iteration reaches K times; the value of K is 100;

[0072] S25. Calculate the CS value between each cluster center data, if the CS value is greater than A, it is considered that the categories are relatively similar and need further classification, then the number of categories N is increased by 1 and re-clustered; if the CS value is less than B, it is considered that the categories are too different and need to be appropriately reduced, then the number of categories N is reduced by 1 and re-clustered; after clustering, output the data of each category respectively. Wherein, the value of A is 0.8, and the value of B is 0.6.

[0073] In this embodiment, the cosine similarity CS is calculated according to the following formula:

[0074]

[0075] Wherein, And represent two different time series data, x0,…,x N-1 and y0,…,y N-1 represent N time series deformation;

[0076] The above formula calculates the similarity between two data, which only represents the correlation between the two data; regarding the determination of the cluster center, in addition to the initial random determination in S21, in the loop of S23, each data is calculated with other data in the cluster to determine the cluster center with the highest average similarity;

[0077] According to the different deformation trends, 2765 high-coherence points obtained by InSAR ground deformation inversion are clustered and processed, and finally three different categories of deformation are obtained through iterative calculation, category 1 has 846 data, category 2 has 1355 data, and category 3 has 564 data.

[0078] In this embodiment, in step S3, the time series deformation of each data in each category is decomposed into a periodic term deformation sequence and a trend term deformation sequence:

[0079] S(t) = X(t) + Y(t);

[0080] Where S(t) is the time series deformation sequence, X(t) is the periodic term deformation sequence, and Y(t) is the trend term deformation sequence.

[0081] The time series deformation of each category is decomposed into a periodic term time series deformation and a trend term time series deformation. Wavelet decomposition calculation is performed using Matlab software, which requires selection of appropriate wavelet functions and decomposition levels. After multiple experiments, Daubechies wavelet is used for four-layer decomposition in this example, which can obtain good decomposition effect for the time series deformation of the study area, Figure 2 is the time series deformation decomposition result of a certain high coherence point in category 1.

[0082] In this embodiment, in step S4, multiple dimensions of influence factors such as rainfall and reservoir water level are extracted from relevant monitoring data. In this example, five influence factors are selected, including 12-day rainfall, 24-day rainfall, reservoir water level elevation, 12-day reservoir water level change, and 24-day reservoir water level change. After data processing, the influence factors are analyzed for grey correlation degree with the periodic term deformation of each category. Taking category 1 as an example, the correlation analysis result is shown in Table 1, which shows that the selected influence factors are closely related to the periodic term deformation.

[0083] Table 1

[0084]

[0085] In this embodiment, the correlation degree between each influence factor and the periodic term deformation sequence of each category is calculated according to the following steps:

[0086] S41. The periodic term deformation sequence is taken as The n influence factor sequences are taken as The periodic term deformation sequence is normalized with the influence factor sequence

[0087] S42. The absolute difference |x0(t)-x i (t)| of corresponding elements of the periodic term deformation sequence and the influence factor sequence is calculated; where x0(t) is the value of the periodic term deformation sequence at time t, and x i (t) is the value of the influence factor sequence at time t.

[0088] S43. The correlation coefficient ξ​​​0i (t):

[0089]

[0090] wherein, ξ 0i (t) is a periodic term deformation sequence and an impact factor sequence the correlation coefficient at time t, denotes the minimum absolute difference between two sequences, denotes the maximum absolute difference between two sequences, and ρ is a resolution coefficient, the ρ value interval is (0, 1), and 0.5 is usually taken;

[0091] S44. Calculate the correlation degree r 0i :

[0092]

[0093] wherein, n is the length of the time sequence;

[0094] The correlation degree obtained by the above calculation is compared with a set threshold value. If the correlation degree of the impact factor and the periodic term deformation sequence is greater than the set threshold value, the impact factor and the periodic term deformation are significantly correlated, otherwise, the impact factor and the periodic term deformation are not significantly correlated; wherein, the set threshold value is 0.6.

[0095] In this embodiment, in step S5, different categories of deformations are obtained by clustering, and models are established respectively to realize fine prediction.

[0096] Using the Keras framework, TensorFlow as the back end, and python coding language, the periodic term and the trend term deformation obtained by decomposition of the same category are used to construct a double-LSTM model for prediction, that is, only the trend term deformation is used as a variable input to establish a trend term single-variable LSTM model to predict the trend term deformation; the periodic term deformation and the target impact factor are used as variable inputs to establish a periodic term multivariable LSTM model to predict the periodic term deformation.

[0097] The structure of the LSTM cell unit is as shown in Figure 3 The corresponding algorithm of the main structure is as follows:

[0098] i t = σ (W xi x t + W hi h t-1 + W ci c t-1 + b i )

[0099] f t = σ (W xfx t +W hf h t-1 +W cf c t-1 +b f )

[0100] c t =f t c t-1 +i t tanh(W xc x t +W hc h t-1 +b c )

[0101] O t =σ(W xo x t +W ho h t-1 +W co c t +b o )

[0102] h t =O t tanh(c t )

[0103] In the formula, i, f, c, o respectively represent the input gate, the forget gate, the cell state and the output gate; x t represents the input at the t time, h t represents the hidden state output of the corresponding cell unit; W and b are respectively the corresponding weight coefficient matrix and the bias term; and sigma and tanh are respectively the sigmoid and the hyperbolic tangent activation function.

[0104] In the embodiment, the univariate LSTM model and the multivariate LSTM model are subjected to network model training, the input length and the output length are determined, the training set, the verification set and the test set are divided, the training set is used for model training, the verification set is used for hyperparameter fine-tuning in the model training process, the test set is used for model testing, and finally the periodic term and the trend term prediction value of each category are obtained.

[0105] Specifically, the input length is 24 and the output length is 1, that is, the deformation of the previous 24 periods is used to make a single-step prediction of the future 1 period. Taking category 1 as an example, the training set is the time series of the deformations of 846 high-coherence points from period 1 to period 25, and periods 1-24 are used as samples and period 25 is used as a label. Among them, 70% is selected as the training set and 30% is selected as the validation set; the test set is the deformation of 846 high-coherence points from period 2 to period 26. According to the grid search method, the parameters are optimized, and in the univariate LSTM model of the trend term, the number of LSTM layers is set to 2 and the number of hidden neurons is set to 32; in the multivariate LSTM model of the periodic term, the number of LSTM layers is set to 3 and the number of hidden neurons is set to 64.

[0106] The trained model is used to directly predict the test set, and the periodic term deformation and the trend term deformation of the 26th period are predicted. The predicted values of the periodic term deformation and the trend term deformation of each category are added to obtain the deformation prediction results of each high-coherence point in each category. As shown in Figure 4 , the average absolute error of each point in the figure is 1.871 mm, and the maximum point error is 3.755 mm. Among them, the average absolute error is: , where x i and are the true value and the predicted value respectively, and N is the number of samples.

[0107] By combining the deformation prediction values of each category and visualizing the output, the large-scale landslide deformation prediction results can be obtained as shown in Figure 5 .

[0108] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A method for predicting large-scale landslide deformation based on InSAR inversion and multiple influencing factors, characterized in that: Includes the following steps: S1. Collect SAR image data of the target area and perform InSAR inversion processing on the SAR image data to obtain time-series deformation data; S2. Cluster the time-series deformation data to obtain several categories of time-series deformation data; S3. Decompose the time-series deformation data of each category into periodic deformation sequences and trend deformation sequences; S4. Extract several influencing factors, calculate the correlation between each influencing factor and the periodic term deformation sequence of each category, select the influencing factors that are significantly related to the periodic term deformation, and take the significantly related influencing factors as the target influencing factors; The correlation between each influencing factor and the periodic deformation sequence of each category is calculated according to the following steps: S41. Take the periodic term deformation sequence as... Take the n impact factor sequences as and the periodic term deformation sequence With the impact factor sequence Perform normalization processing; S42. Calculate the absolute difference between corresponding elements of the periodic term deformation sequence and the influence factor sequence. ;in, For periodic term deformation sequence exist The value at time, Impact factor sequence exist The value at time; S43. Calculate the correlation coefficient : ； in, For periodic term deformation sequence With the impact factor sequence The correlation coefficient at time t, This represents the minimum absolute difference between two sequences. This represents the maximum absolute difference between two sequences. The resolution coefficient; S44. Calculate the correlation degree : ； Where n is the length of the time series sequence; S5. Establish LSTM models for each category of time-series deformation data, specifically including: For trend item deformation of the same category, the trend item deformation is used as a variable input to build a univariate LSTM model of the trend item. For the same category of periodic term deformation, the periodic term deformation and the target influencing factor are used as variable inputs to establish a periodic term multivariate LSTM model; S6. Train the network model of the univariate LSTM model and the multivariate LSTM model in step S5 to obtain the trained LSTM model, and use the trained LSTM model to predict the deformation of each category to obtain the predicted values ​​of the periodic deformation and the trend deformation of each category. S7. Add the predicted deformation values ​​of the periodic terms and the predicted deformation values ​​of the trend terms for each category to obtain the predicted deformation results for each category. Then, merge the predicted deformation results for each category to obtain the predicted deformation results for the large-scale landslide.

2. The method for large-scale landslide deformation prediction based on InSAR inversion and multiple influencing factors according to claim 1, characterized in that: InSAR inversion processing of SAR image data specifically includes: S11. Preprocess the SAR image data; the preprocessing includes data import and data cropping; S12. Select a super master image from the SAR images and register all other images to the super master image; S13. The basic data for SBAS inversion estimation is obtained through baseline estimation, connection diagram generation, differential interferometry, interferogram filtering, and phase unwrapping. S14. Remove residual topographic errors and residual atmospheric phase from the basic data, perform unit conversion on the deformation information, and obtain the deformation results; the deformation results include deformation rate and deformation time series; S15. Transform the deformation results into a geographic coordinate system to obtain several highly coherent points in the target area. Each highly coherent point has time-series deformation data for several time periods.

3. The method for large-scale landslide deformation prediction based on InSAR inversion and multiple influencing factors according to claim 1, characterized in that: Clustering of time-series deformation data includes: S21. Randomly select N data points from all time-series deformation data as cluster centers; S22. Calculate the cosine similarity value CS between each time-series deformation data and N cluster centers, and assign the data to the cluster with the highest CS value; S23. Calculate the CS value of each data in each cluster with other data in the same cluster, and redefine the data with the highest average CS value with the other data as the cluster center; S24. Repeat steps S22-S23 to iterate until the data in each cluster no longer changes or the iteration reaches K times; S25. Calculate the CS value of each cluster center data. If the CS value is greater than A, increment the number of clusters N by 1 and re-cluster. If the CS value is less than B, decrement the number of clusters N by 1 and re-cluster. Output the data of each category after clustering is completed.

4. The method for large-scale landslide deformation prediction based on InSAR inversion and multiple influencing factors according to claim 3, characterized in that: Calculate the cosine similarity value CS using the following formula: ； in, , , and These represent two different time series data. as well as These represent the deformations of N time series.

5. The method for large-scale landslide deformation prediction based on InSAR inversion and multiple influencing factors according to claim 1, characterized in that: In step S3, the decomposition is performed using wavelet decomposition, the decomposition function of the wavelet is Daubechies, and the number of wavelet decomposition layers is four.

6. The method for large-scale landslide deformation prediction based on InSAR inversion and multiple influencing factors according to claim 1, characterized in that: The aforementioned influencing factors include rainfall, reservoir water level elevation, and reservoir water level change.

Images