A landslide deformation prediction method based on space-time topology features

Abstract

This invention belongs to the field of landslide deformation prediction technology, and specifically to a landslide deformation prediction method based on spatiotemporal topological features. The method includes the following steps: S1: SBAS-InSAR processing is performed on the acquired SAR image set to obtain temporal deformation data of the landslide area. Simultaneously, rainfall data and soil moisture content data are acquired to form influencing factor data. S2: The temporal deformation data of the landslide area is decomposed into trend-term deformation and periodic-term deformation through temporal decomposition. Dataset 1 is constructed based on the trend-term deformation, and Dataset 2 is constructed based on the periodic-term deformation and influencing factor data. S3: Dataset 1 and Dataset 2 are respectively input into the landslide deformation prediction model based on spatiotemporal topological features to obtain predicted values ​​for Dataset 1 and Dataset 2. The predicted results are then summed to obtain the final predicted landslide deformation value. This invention can capture the correlation between influencing factors and landslide deformation, improving the accuracy of landslide deformation prediction. Compared with the traditional GRU model, it has smaller errors and higher accuracy.

Description

Technical Field

[0001] This invention relates to the field of landslide deformation prediction technology, specifically a landslide deformation prediction method based on spatiotemporal topological features. Background Technology

[0002] Landslides are serious geological disasters that can directly cause casualties, damage to houses, disruption of transportation networks, and even blockage of rivers, triggering numerous secondary disasters such as mudslides. my country has many areas prone to landslides, and the landslides that occur every year have a huge impact on human life and property, causing significant casualties and damaging infrastructure, agricultural land, and housing.

[0003] In recent years, temporal InSAR technology has gradually developed and matured in the field of deformation monitoring, and researchers at home and abroad have carried out research on combining InSAR technology with deep learning to achieve deformation prediction. However, due to the influence of complex mechanisms and numerous variable factors on landslide deformation sequences, their deformation trends have complex uncertainties, making prediction a complex problem. Methods for landslide deformation prediction based on InSAR technology also have limited accuracy. Therefore, we propose a landslide deformation prediction method based on spatiotemporal topological features to solve the above problems. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a landslide deformation prediction method based on spatiotemporal topological features, which solves the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention specifically adopts the following technical solution: A landslide deformation prediction method based on spatiotemporal topological features includes the following steps: S1: Acquire SAR image set, precise orbit data, rainfall data and soil moisture data of the landslide area. Based on the precise orbit data, perform SBAS-InSAR processing on the acquired SAR image set to obtain temporal deformation data of the landslide area. At the same time, rainfall data and soil moisture data constitute influencing factor data. S2: Decompose the temporal deformation data of the landslide area into trend deformation and periodic deformation through time series decomposition. Construct dataset one based on trend deformation and dataset two based on periodic deformation and influencing factor data. S3: Input Dataset 1 and Dataset 2 into the landslide deformation prediction model based on spatiotemporal topological features to obtain the predicted values ​​of Dataset 1 and Dataset 2, and add the predicted results to obtain the final predicted value of landslide deformation.

[0006] Furthermore, in S1, the SAR image set is processed using SBAS-InSAR technology. Specifically, the first scene in the SAR image set is selected as the main image. The time baseline threshold is set to 180 days and the spatial baseline threshold is set to 200 meters. High coherence points are determined. Differential interferometry is performed on the high coherence points to obtain the differential interferometric phase of each point. After Goldstein filtering, phase unwrapping is performed to obtain the deformation phase, atmospheric phase, elevation residual phase and noise. Then, the atmospheric phase, elevation residual phase and noise are removed by filtering to obtain the deformation phase. Finally, the landslide area time-series deformation data is obtained by SBAS inversion.

[0007] Furthermore, in step S2, the temporal deformation data of the landslide area is decomposed into trend-term deformation and periodic-term deformation through time-series decomposition. Specifically, the original deformation time-series signal is represented as the sum of several sub-components with different frequency characteristics. , in the formula, This represents the k-th sub-component, where K is the preset decomposition level; and under the constraint of complete signal reconstruction: Next, frequency domain compactness constraints are introduced for each sub-component, and the following optimization objective is constructed: ,in, Let represent the center frequency of the k-th sub-component. Then, by introducing constraint adjustment parameters and consistency control variables, the above optimization problem is transformed into an iteratively solvable form, and an alternating optimization strategy is adopted for each sub-component. and its center frequency The system is updated so that each sub-component gradually converges to different frequency ranges in the frequency domain; when the convergence condition is met... Stop iterating when the time comes; Finally, based on the frequency distribution characteristics of each sub-component, the low-frequency dominant component is extracted as the trend term deformation, and the high-frequency or periodic significant component is classified as the periodic term deformation.

[0008] Furthermore, in S2, both dataset one and dataset two are divided into training and test sets in a 7:3 ratio.

[0009] Furthermore, the landslide deformation prediction model based on spatiotemporal topological features includes: a prediction module based on historical time-series dependencies and a multi-scale spatiotemporal coupling prediction module.

[0010] Furthermore, the prediction module based on historical time-series dependencies includes: performing a stability test on dataset one; if the test indicates non-stationarity, then performing a first-order difference on dataset one. A stationary sequence is obtained, where, This represents a d-order difference operator. Given a stationary sequence, if stationarity is tested, a recursive relationship between the stationary sequence and historical observations is constructed, and an error correction term is introduced, resulting in the following prediction expression: ; in, and For the parameters to be estimated, The random perturbation term is represented by p and q, which represent the order of historical dependence and the order of error correction, respectively. The parameters are then estimated by minimizing the error between the predicted and observed values ​​to complete the model fitting. Finally, recursive predictions are made for future times, and the original scale is restored through inverse difference operations. This allows us to obtain the predicted value of the trend term deformation.

[0011] Furthermore, the multi-scale spatiotemporal coupled prediction module includes: processing the rainfall data in dataset two. Soil moisture content Landslide deformation periodic term As a multimodal input, standardization is performed: first, the mean and standard deviation are calculated separately. Standardization ; Based on the spatial relationships of landslide monitoring points, an adjacency matrix A is constructed, and a graph structure is defined. , where V is the set of monitoring points and E is the spatial relationship between monitoring points; The contribution weights of different modalities in the time dimension are calculated using a multimodal attention mechanism, and the multimodal temporal feature matrix is ​​obtained by weighted fusion. The fused features are input into the spatiotemporal graph convolution module, which uses spatial graph convolution to extract spatial topological features between monitoring points and uses temporal convolution to extract dynamic evolution features of long time series. Bundle Input fully connected layer prediction: Output The loss function is ,in As a hyperparameter, it outputs the predicted values ​​of the periodic deformation at each monitoring point at future times.

[0012] Furthermore, the multimodal attention mechanism uses standardized rainfall data... Soil moisture content and landslide deformation periodic term Each modality is mapped to a high-dimensional feature space to obtain its initial feature vector. The contribution weights of each mode at time t are calculated using an attention mechanism. Its mathematical expression is: ; in, The relevance scoring function for modal features is used; the weighted summation of each modal feature is performed to obtain the fused multimodal temporal feature matrix. The expression is: ; The matrix It serves as the input to subsequent spatiotemporal graph convolutional layers for further extraction of spatial topological features.

[0013] Furthermore, the expression for the spatiotemporal graph convolution operator in the spatiotemporal graph convolution module is: in, It is an adjacency matrix that includes self-connections. For degree matrix, The input features are those obtained after multimodal fusion. These are learnable weight parameters. The PReLU activation function has the following mathematical expression: ; in, It is a constant. It is input. This is the output.

[0014] Furthermore, the mathematical expression for the LeakyReLU activation function is: ; Where a is a constant and x is the input.

[0015] Compared with existing technologies, this invention provides a landslide deformation prediction method based on spatiotemporal topological features, which has the following advantages: Traditional algorithms do not distinguish between periodic deformation and trend deformation related to the internal structure of the slope when predicting landslide deformation, and they also fail to consider that the occurrence of landslides is often related to a variety of external factors. This invention distinguishes the trend and periodic terms of landslide deformation through time series decomposition, and combines rainfall and soil moisture data to predict landslide deformation, thereby improving the accuracy of landslide deformation prediction. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method steps of the present invention; Figure 2 This is a diagram showing the time and space baselines of the present invention; Figure 3 This is a schematic diagram of the prediction module structure based on historical time-series dependencies according to the present invention; Figure 4 This is a schematic diagram of the multi-scale spatiotemporal coupling prediction module structure of the present invention; Figure 5This is a schematic diagram of the multimodal attention fusion module structure in the multi-scale spatiotemporal coupling prediction module of the present invention; Figure 6 This is a schematic diagram of the spatiotemporal graph convolution module in the multi-scale spatiotemporal coupling prediction module of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0018] like Figure 1 As shown in the figure, an embodiment of the present invention proposes a landslide deformation prediction method based on spatiotemporal topological features, which includes the following steps: S1: Acquire SAR image set, precise orbit data, rainfall data, and soil moisture content data of the landslide area. Based on the precise orbit data, perform SBAS-InSAR processing on the acquired SAR image set to obtain the temporal deformation data of the landslide area. At the same time, the rainfall data and soil moisture content data constitute the influencing factor data. Among them, the daily rainfall data, the time range and location are consistent with the landslide deformation monitoring area, and the unit is mm. The soil moisture content data is in cm3 / cm3.

[0019] Select the first scene in the SAR image set as the master image, such as Figure 2 The time baseline threshold was set to 180 days and the spatial baseline threshold to 200 meters. High coherence points were determined, and differential interferometry was performed on the high coherence points to obtain the differential interferometric phase of each point. After Goldstein filtering, phase unwrapping was performed to obtain the deformation phase, atmospheric phase, elevation residual phase, and noise. Then, the atmospheric phase, elevation residual phase, and noise were removed by filtering to obtain the deformation phase. Finally, the landslide area time-series deformation data was obtained by SBAS inversion.

[0020] S2: The temporal deformation data of the landslide area is decomposed into trend deformation and periodic deformation through time series decomposition. Dataset 1 is constructed based on the trend deformation, and Dataset 2 is constructed based on the periodic deformation and the influencing factor data. Dataset 1 and Dataset 2 are both divided into training set and test set in a 7:3 ratio.

[0021] The original deformation time-series signal is represented as the sum of several sub-components with different frequency characteristics: , in the formula, This represents the k-th sub-component, where K is the preset decomposition level; and under the constraint of complete signal reconstruction: Next, frequency domain compactness constraints are introduced for each sub-component, and the following optimization objective is constructed: ,in, Let represent the center frequency of the k-th sub-component. Then, by introducing constraint adjustment parameters and consistency control variables, the above optimization problem is transformed into an iteratively solvable form, and an alternating optimization strategy is adopted for each sub-component. and its center frequency The system is updated so that each sub-component gradually converges to different frequency ranges in the frequency domain; when the convergence condition is met... Stop iterating when the time comes; Finally, based on the frequency distribution characteristics of each sub-component, the low-frequency dominant component is extracted as the trend term deformation, and the high-frequency or periodic significant component is classified as the periodic term deformation.

[0022] S3: Input Dataset 1 and Dataset 2 into the landslide deformation prediction model based on spatiotemporal topological features to obtain the predicted values ​​of Dataset 1 and Dataset 2, and add the predicted results to obtain the final predicted value of landslide deformation.

[0023] The landslide deformation prediction model based on spatiotemporal topological features includes a prediction module based on historical time series dependencies and a multi-scale spatiotemporal coupling prediction module.

[0024] Input dataset 1 into Figure 3 In the prediction module based on historical time-series dependencies shown, a stability test is performed on dataset 1. If the test indicates non-stationarity, then first-order differencing is performed on dataset 1. A stationary sequence is obtained, where, This represents a d-order difference operator. Given a stationary sequence, if stationarity is tested, a recursive relationship between the stationary sequence and historical observations is constructed, and an error correction term is introduced, resulting in the following prediction expression: ; in, and For the parameters to be estimated, The random perturbation term is represented by p and q, which represent the order of historical dependence and the order of error correction, respectively. The parameters are then estimated by minimizing the error between the predicted and observed values ​​to complete the model fitting. Finally, recursive predictions are made for future times, and the original scale is restored through inverse difference operations. This allows us to obtain the predicted value of the trend term deformation.

[0025] Simultaneously, dataset two is input into the multi-scale spatiotemporal coupling prediction module, the module structure of which is as follows: Figure 4 As shown, firstly, the rainfall data in dataset two... Soil moisture content Landslide deformation periodic term As a multimodal input, standardization is performed: first, the mean and standard deviation are calculated separately. Standardization ; Based on the spatial relationships of landslide monitoring points, an adjacency matrix A is constructed, and a graph structure is defined. , where V is the set of monitoring points and E is the spatial relationship between monitoring points; like Figure 5 As shown, the standardized rainfall amount Soil moisture content and landslide deformation periodic term Each modality is mapped to a high-dimensional feature space to obtain its initial feature vector. The contribution weights of each mode at time t are calculated using an attention mechanism. Its mathematical expression is: ; in, The relevance scoring function for modal features is used; the weighted summation of each modal feature is performed to obtain the fused multimodal temporal feature matrix. The expression is: ; Input the fused features Figure 6 The spatiotemporal graph convolution module shown extracts spatial topological features between monitoring points using spatial graph convolution and extracts dynamic evolution features of long-term series using temporal convolution. The expression for the spatiotemporal graph convolution operator is: in, It is an adjacency matrix that includes self-connections. For degree matrix, The input features are those obtained after multimodal fusion. These are learnable weight parameters. The PReLU activation function has the following mathematical expression: ; in, It is a constant. It is input. This is the output.

[0026] Bundle Input fully connected layer prediction: Output The loss function is ,in For hyperparameters, the mathematical expression of the LeakyReLU activation function is: ; Where 'a' is a constant and 'x' is the input. The final output is the predicted periodic deformation value for each monitoring point at future times.

[0027] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A landslide deformation prediction method based on spatiotemporal topological features, characterized in that: Includes the following steps: S1: Acquire SAR image set, precise orbit data, rainfall data and soil moisture data of the landslide area. Based on the precise orbit data, perform SBAS-InSAR processing on the acquired SAR image set to obtain temporal deformation data of the landslide area. At the same time, rainfall data and soil moisture data constitute influencing factor data. S2: Decompose the temporal deformation data of the landslide area into trend deformation and periodic deformation through time series decomposition. Construct dataset one based on trend deformation and dataset two based on periodic deformation and influencing factor data. S3: Input Dataset 1 and Dataset 2 into the landslide deformation prediction model based on spatiotemporal topological features to obtain the predicted values ​​of Dataset 1 and Dataset 2, and add the predicted results to obtain the final predicted value of landslide deformation.

2. The landslide deformation prediction method based on spatiotemporal topological features according to claim 1, characterized in that: In S1, SBAS-InSAR technology is used to process the SAR image set. Specifically, the first scene in the SAR image set is selected as the main image. The time baseline threshold is set to 180 days and the spatial baseline threshold is set to 200 meters. High coherence points are determined. Differential interferometry is performed on the high coherence points to obtain the differential interferometric phase of each point. After Goldstein filtering, phase unwrapping is performed to obtain the deformation phase, atmospheric phase, elevation residual phase and noise. Then, the atmospheric phase, elevation residual phase and noise are removed by filtering to obtain the deformation phase. Finally, the landslide area time-series deformation data is obtained by SBAS inversion.

3. The landslide deformation prediction method based on spatiotemporal topological features according to claim 1, characterized in that: In step S2, the time-series deformation data of the landslide area is decomposed into trend-term deformation and periodic-term deformation through time-series decomposition. Specifically, the original deformation time-series signal is represented as the sum of several sub-components with different frequency characteristics. , in the formula, This represents the k-th sub-component, where K is the preset decomposition level; and under the constraint of complete signal reconstruction: Next, frequency domain compactness constraints are introduced for each sub-component, and the following optimization objective is constructed: ,in, Let represent the center frequency of the k-th sub-component. Then, by introducing constraint adjustment parameters and consistency control variables, the above optimization problem is transformed into an iteratively solvable form, and an alternating optimization strategy is adopted for each sub-component. and its center frequency The system is updated so that each sub-component gradually converges to different frequency ranges in the frequency domain; when the convergence condition is met... Stop iterating when the time comes; Finally, based on the frequency distribution characteristics of each sub-component, the low-frequency dominant component is extracted as the trend term deformation, and the high-frequency or periodic significant component is classified as the periodic term deformation.

4. The landslide deformation prediction method based on spatiotemporal topological features according to claim 1, characterized in that: In S2, both dataset 1 and dataset 2 are divided into training and test sets in a 7:3 ratio.

5. The landslide deformation prediction method based on spatiotemporal topological features according to claim 1, characterized in that: The landslide deformation prediction model based on spatiotemporal topological features in S3 includes: a prediction module based on historical time series dependencies and a multi-scale spatiotemporal coupling prediction module.

6. The landslide deformation prediction method based on spatiotemporal topological features according to claim 5, characterized in that: The prediction module based on historical time-series dependencies includes: performing a stability test on dataset 1; if the test indicates non-stationarity, then performing a first-order difference on dataset 1. A stationary sequence is obtained, where, This represents a d-order difference operator. Given a stationary sequence, if stationarity is tested, a recursive relationship between the stationary sequence and historical observations is constructed, and an error correction term is introduced, resulting in the following prediction expression: ,in, and For the parameters to be estimated, The random perturbation term is represented by p and q, which represent the order of historical dependence and the order of error correction, respectively. The parameters are then estimated by minimizing the error between the predicted and observed values ​​to complete the model fitting. Finally, recursive predictions are made for future times, and the original scale is restored through inverse difference operations. This allows us to obtain the predicted value of the trend term deformation.

7. The landslide deformation prediction method based on spatiotemporal topological features according to claim 5, characterized in that: The multi-scale spatiotemporal coupled prediction module includes: processing the rainfall data in dataset two. Soil moisture content Landslide deformation periodic term As a multimodal input, standardization is performed: first, the mean and standard deviation are calculated separately. Standardization ; Based on the spatial relationships of landslide monitoring points, an adjacency matrix A is constructed, and a graph structure is defined. , where V is the set of monitoring points and E is the spatial relationship between monitoring points; The contribution weights of different modalities in the time dimension are calculated using a multimodal attention mechanism, and the multimodal temporal feature matrix is ​​obtained by weighted fusion. The fused features are input into the spatiotemporal graph convolution module, which uses spatial graph convolution to extract spatial topological features between monitoring points and uses temporal convolution to extract dynamic evolution features of long time series. Bundle Input fully connected layer prediction: Output The loss function is ,in As a hyperparameter, it outputs the predicted values ​​of the periodic deformation at each monitoring point at future times.

8. The landslide deformation prediction method based on spatiotemporal topological features according to claim 7, characterized in that: The multimodal attention mechanism uses standardized rainfall data. Soil moisture content and landslide deformation periodic term Each modality is mapped to a high-dimensional feature space to obtain its initial feature vector. ; The contribution weights of each mode at time t are calculated using an attention mechanism. Its mathematical expression is: ； in, The relevance scoring function for modal features is used; the weighted summation of each modal feature is performed to obtain the fused multimodal temporal feature matrix. The expression is: ； The matrix It serves as the input to subsequent spatiotemporal graph convolutional layers for further extraction of spatial topological features.

9. The landslide deformation prediction method based on spatiotemporal topological features according to claim 7, characterized in that: The expression for the spatiotemporal graph convolution operator in the spatiotemporal graph convolution module is: in, It is an adjacency matrix that includes self-connections. For degree matrix, The input features are those obtained after multimodal fusion. These are learnable weight parameters. The PReLU activation function has the following mathematical expression: ； in, It is a constant. It is input. This is the output.

10. A landslide deformation prediction method based on spatiotemporal topological features according to claim 7, characterized in that: The mathematical expression for the LeakyReLU activation function is: ； Where a is a constant and x is the input.

Images