Landslide hidden danger identification method and device based on unsupervised machine learning and time sequence InSAR

Abstract

The invention provides a landslide hidden danger identification method and device based on unsupervised machine learning and time sequence InSAR, and belongs to the technical field of landslide hidden danger identification. The method can divide a landslide area with a landslide hidden danger in a to-be-identified area, and further determine a single landslide sub-area in the landslide area; and the landslide causes of the single landslide sub-regions are obtained through analysis of the cause analysis model, so that prevention and control measures can be made and taken in time according to the single landslide sub-regions and the landslide causes thereof, and potential safety hazards existing in regions where landslides possibly occur are eliminated.

Description

Technical Field

[0001] The present invention relates to the technical field of landslide hazard identification, and particularly to a landslide hazard identification method and device based on unsupervised machine learning and time series InSAR. Background Art

[0002] A landslide phenomenon refers to a natural disaster in which soil, rock, or other debris on a slope slides downward as a whole along a certain sliding surface under the action of gravity. Monitoring and identifying landslide phenomena and taking timely prevention and control measures can reduce the disaster losses caused by landslide phenomena.

[0003] However, since the landslide phenomenon is a complex geological phenomenon, existing methods are difficult to locate the accurate location where the landslide occurs (for example, the area where the individual landslide is located) and determine the cause of the landslide, so that it is impossible to formulate and take prevention and control measures in time, leaving potential safety hazards in the areas where landslides may occur (also known as areas to be identified). Summary of the Invention

[0004] The present invention proposes a landslide hazard identification method and device based on unsupervised machine learning and time series InSAR, which can divide the landslide areas with potential landslide hazards in the areas to be identified, further determine the individual landslide sub-areas within the landslide areas, and then analyze the causes of the landslides in the individual landslide sub-areas through a cause analysis model. Furthermore, it is possible to formulate and take prevention and control measures in time according to the individual landslide sub-areas and their landslide causes, eliminating the potential safety hazards in the areas where landslides may occur.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] In a first aspect, the present invention provides a landslide hazard identification method based on unsupervised machine learning and temporal InSAR, including: processing the landslide displacements of multiple monitoring points in the area to be identified using the TS2Vec model to obtain multi-dimensional features of the landslide displacements of the multiple monitoring points; the landslide displacement of each monitoring point includes the landslide displacements of the monitoring point at multiple time points. Then, the multi-dimensional features are processed by the t-SNE algorithm for dimensionality reduction to obtain the landslide displacement features of the multiple monitoring points in the area to be identified. After clustering the landslide displacement features of the multiple monitoring points, the area where the cluster satisfying the landslide condition is located is determined as the landslide area in the area to be identified; the landslide condition is that the descending rate of the landslide displacement of the cluster in the horizontal direction satisfies the first descending condition, and the descending rate in the vertical direction satisfies the second descending condition; the landslide area is the area in the area to be identified where there are landslide hazards. And based on the landslide displacements and topographic data of multiple monitoring points in the landslide area, at least one individual landslide sub-area in the landslide area is determined. For each individual landslide sub-area, a cause analysis model is used to analyze the landslide data of the individual landslide sub-area to determine the cause of the landslide in the individual landslide sub-area; the individual landslide sub-area and the cause of the landslide in the individual landslide sub-area are the identification results of the landslide hazard; wherein, the landslide data of the individual landslide sub-area includes the landslide displacement and environmental information of the individual landslide sub-area; the cause analysis model is trained based on the constructed data training set; the data training set includes historical landslide data and the landslide causes determined based on the frequency domain features of the historical landslide data.

[0007] In the landslide hazard identification method based on unsupervised machine learning and temporal InSAR provided by the present invention, first, the TS2Vec model is used to process the landslide displacements of multiple monitoring points in the area to be identified to obtain multi-dimensional features of the landslide displacements of the multiple monitoring points, and then the t-SNE algorithm is used to reduce the dimensionality of the multi-dimensional features of the landslide displacements of the multiple monitoring points to obtain the landslide displacement features of the multiple monitoring points. The landslide area where there are landslide hazards in the area to be identified is determined according to the clustering result of the landslide displacement features of the multiple monitoring points. Based on the landslide displacements and topographic data of multiple monitoring points in the landslide area, at least one individual landslide sub-area in the landslide area is determined. Then, the cause analysis model is used to analyze and obtain the cause of the landslide in each individual landslide sub-area, and thus the identification result of the landslide hazard is obtained. From the above content, it can be seen that the above method can determine the individual landslide sub-areas and their causes of landslides, and then can timely formulate and take prevention and control measures according to the individual landslide sub-areas and their causes of landslides to eliminate the potential safety hazards in the areas where landslides may occur.

[0008] In an implementation of the first aspect, determining the cause of landslides based on the frequency-domain characteristics of historical landslide data includes: performing a fast Fourier transform on the landslide displacement characteristics of multiple monitoring points in the historical landslide data to obtain the landslide displacement frequency-domain characteristics of multiple monitoring points. Performing K-medoids clustering on the landslide displacement frequency-domain characteristics of multiple monitoring points to obtain multiple clusters. Performing periodic analysis on the landslide displacement frequency-domain characteristics of the cluster center of each cluster among the multiple clusters to determine the periodic characteristics of each cluster. Performing trend analysis on the landslide displacement frequency-domain characteristics of the cluster center of each cluster among the multiple clusters to determine the trend characteristics of each cluster. In the landslide cause analysis table, determine the landslide cause corresponding to the periodic characteristics, trend characteristics, and environmental information of each cluster as the landslide cause of the cluster.

[0009] In an implementation of the first aspect, the periodic characteristics are annual cycle, quarterly cycle, or aperiodic. The trend characteristics include trend intensity and significance index; the significance index is the P value; the trend intensity satisfies the following formula:

[0010] y = β0 + β1t + ε

[0011] where y represents the landslide displacement of the cluster center of the cluster, β0 represents the intercept, β1 represents the trend intensity, t represents time, and ε represents the residual.

[0012] In an implementation of the first aspect, in the landslide cause analysis table: when the environmental information includes periodic rainfall data, the periodic characteristic is an annual cycle, the magnitude of the trend characteristic is -0.2 mm / month, and the P value is less than 0.01, the corresponding landslide cause is rainfall. When the environmental information includes human activity data, the periodic characteristic is a quarterly cycle, the magnitude of the trend characteristic is 0.05 mm / month, and the P value is in the range of (0.01, 0.15), the corresponding landslide cause is human activity. When the environmental information includes earthquake data, the periodic characteristic is aperiodic, the magnitude of the trend characteristic is -0.5 mm / month, and the P value is less than 0.001, the corresponding landslide cause is earthquake.

[0013] In an implementation of the first aspect, the cause analysis model includes: a first input layer connected in sequence, multiple parallel InceptionTime networks, and a fully connected layer. Among them, the input ends of the multiple parallel InceptionTime networks are all connected to the input layer, and the output ends of the multiple parallel InceptionTime networks are all connected to the fully connected layer. Among the multiple parallel InceptionTime networks, each InceptionTime network includes, connected in sequence: a second input layer, a first residual module, a second residual module, and a global average pooling layer. Both the first residual module and the second residual module include, connected in residual connection in sequence: multiple Inception sub-modules and a multi-head latent attention layer; the residual connection is provided with learnable weights. The Inception sub-module includes, connected in sequence: a bottleneck layer, a parallel convolutional layer, a max pooling layer, and a fully connected layer. Among them, the parallel convolutional layer includes a 1×1 convolutional layer, a 3×3 convolutional layer, and a 5×5 convolutional layer. The input ends of the 3×3 convolutional layer, the 5×5 convolutional layer, and the max pooling layer are connected to the bottleneck layer. The input end of the 1×1 convolutional layer is connected to the input end of the Inception sub-module. The output ends of the 1×1 convolutional layer, the 3×3 convolutional layer, the 5×5 convolutional layer, and the max pooling layer are connected to the fully connected layer.

[0014] In the above implementation, since the Inception sub-module extracts local features through different convolutions, and the attention mechanism of the multi-head latent attention layer can capture long-range dependencies and form a complement, the cause analysis model can enhance the global dependence relationship of the time series features, avoid input noise interference, and reduce the error of the output result of the cause analysis model.

[0015] In an implementation of the first aspect, the first descent condition is that the range of the descent rate is [25, 50], and the second descent condition is that the range of the descent rate is [10, 25] within a time range of 1 to 3 months. Among them, the unit of the descent rate is mm / day.

[0016] Second aspect, the present invention provides a landslide hazard identification device based on unsupervised machine learning and temporal InSAR, including a multi-dimensional feature determination module, a landslide displacement feature determination module, a landslide area determination module, a single landslide sub-area determination module, and a landslide cause determination module. The multi-dimensional feature determination module is configured to process the landslide displacements of multiple monitoring points in the area to be identified using the TS2Vec model to obtain the multi-dimensional features of the landslide displacements of the multiple monitoring points; the landslide displacement of each monitoring point includes the landslide displacements of the monitoring point at multiple time points respectively. The landslide displacement feature determination module is configured to perform dimensionality reduction processing on the multi-dimensional features through the t-SNE algorithm to obtain the landslide displacement features of the multiple monitoring points in the area to be identified. The landslide area determination module is configured to, after clustering the landslide displacement features of the multiple monitoring points, determine the area where the cluster satisfying the landslide condition is located in the cluster result as the landslide area of the area to be identified; the landslide condition is that the downward rate of the landslide displacement in the horizontal direction of the cluster satisfies the first downward condition, and the downward rate in the vertical direction satisfies the second downward condition; the landslide area is the area with landslide hazards in the area to be identified. The single landslide sub-area determination module is configured to determine at least one single landslide sub-area within the landslide area according to the landslide displacements and terrain data of the multiple monitoring points in the landslide area. The landslide cause determination module is configured to, for each single landslide sub-area, analyze the landslide data of the single landslide sub-area using a cause analysis model to determine the landslide cause of the single landslide sub-area; the single landslide sub-area and the landslide cause of the single landslide sub-area are the identification results of the landslide hazards; wherein, the landslide data of the single landslide sub-area includes the landslide displacement and environmental information of the single landslide sub-area; the cause analysis model is trained based on the constructed data training set; the data training set includes historical landslide data and the landslide causes determined based on the frequency domain features of the historical landslide data.

[0017] Third aspect, the present invention provides an electronic device, including a processor and a memory coupled to the processor; the memory is used to store computer instructions, and when the electronic device runs, the processor executes the computer instructions stored in the memory so that the electronic device executes the method described in the first aspect or any one of its implementation manners above.

[0018] Fourth aspect, the present invention provides a computer-readable storage medium, including computer program instructions, and when the computer program instructions are executed by a computer, the computer is made to execute the method described in the first aspect or any one of its implementation manners above.

[0019] Fifth aspect, the present invention provides a computer program product, including computer program instructions, and when the computer program instructions run on a computer, the computer is made to execute the method described in the first aspect or any one of its implementation manners above.

[0020] For the technical effects corresponding to the second to fifth aspects and their possible implementation manners above, reference may be made to the description of the technical effects of the first aspect and its possible implementation manners above, which will not be elaborated here. Description of the Drawings

[0021] Figure 1 It is a schematic diagram of a landslide hazard identification method based on unsupervised machine learning and temporal InSAR provided by an embodiment of the present application;

[0022] Figure 2 It is a schematic diagram of the structures of the first residual module and the second residual module in the InceptionTime network provided by an embodiment of the present application;

[0023] Figure 3 It is a schematic diagram of the structure of the Inception sub-module provided by an embodiment of the present application;

[0024] Figure 4 It is a schematic diagram of the structure of a landslide hazard identification device based on unsupervised machine learning and temporal InSAR provided by an embodiment of the present application. Detailed Implementation Manner

[0025] In the description of the present invention, terms such as "first" and "second" in the description and claims of the present invention are used to distinguish different objects, rather than to describe a specific order of the objects.

[0026] In the embodiments of the present application, "and / or" represents the relationship between objects. For example, A and / or B may represent the following three situations: A exists alone, B exists alone, and A and B exist simultaneously.

[0027] In the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the embodiments of the present application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Exactly speaking, using words such as "exemplary" or "for example" is intended to present relevant concepts in a specific manner.

[0028] In the description of the present invention, unless otherwise specified, the meaning of "a plurality of" refers to two or more. For example, a plurality of monitoring points refers to two or more monitoring points.

[0029] The methods and devices provided by the embodiments of this application relate to landslide hazard identification, and can be used to locate individual landslide sub-areas with landslide hazards in the area to be identified, analyze the causes of landslides in the individual landslide sub-areas, and then formulate and take prevention and control measures in a timely manner based on the individual landslide sub-areas and their causes of landslides, so as to eliminate the potential safety hazards in the areas where landslides may occur. Among them, the above-mentioned landslide hazard refers to a landslide that is moving and has not been discovered.

[0030] It can be understood that the landslide area refers to the area with potential landslide hazards. The individual landslide sub-area refers to the area with potential individual landslide hazards. The cause of landslide refers to the reason for the landslide hazard. Among them, an individual landslide refers to a single landslide event that occurs on a relatively independent slope, usually with a clear scope and boundary, and its formation mechanism and movement characteristics are relatively independent.

[0031] Since it is difficult for existing methods to locate the exact location where a landslide occurs (for example, the area where an individual landslide is located) and determine the cause of the landslide, it is impossible to formulate and take prevention and control measures in a timely manner, resulting in potential safety hazards in the areas where landslides may occur (also known as the areas to be identified). The embodiments of this application provide a landslide hazard identification method and device based on machine learning, which can divide the landslide areas with landslide hazards in the area to be identified, further determine individual landslide sub-areas within the landslide areas, and then analyze the causes of landslides in the individual landslide sub-areas through a cause analysis model, and then can formulate and take prevention and control measures in a timely manner based on the individual landslide sub-areas and their causes of landslides, so as to eliminate the potential safety hazards in the areas where landslides may occur.

[0032] Exemplarily, the landslide hazard identification method based on machine learning provided by the embodiments of the present invention can be executed by an electronic device with processing functions. For example, the electronic device can be a computer, a server, etc. Taking the electronic device as a computer as an example, the hardware part of the computer can include: a processor, a memory, a network interface, a user interface, a communication bus, etc.

[0033] Among them, the processor is used to control the electronic device to execute relevant processing and calculation tasks. For example, determining multi-dimensional features, determining landslide displacement features, determining landslide areas, determining individual landslide sub-areas, and the causes of individual landslide sub-areas, etc. The processor can include a central processing unit (CPU) or other processors. The processor can be single-core or multi-core. For example, the processor can include multiple CPUs.

[0034] The memory is used to store computer instructions and related data. For example, it stores multi-dimensional features, landslide displacement features, landslide areas, individual landslide sub-areas, and the causes of individual landslide sub-areas, etc. The memory can be a random access memory (RAM), a read only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, or an optical memory, a magnetic disk storage medium, or any other magnetic storage device, or any other medium capable of storing program code or data that can be accessed by a computer. Optionally, the memory can be integrated within the processor, or the memory can be independent of the processor.

[0035] The network interface is used for the computer to communicate with other devices or communication networks. The network interface can be a transceiver with sending and receiving functions. Optionally, the network interface can include standard wired interfaces, wireless interfaces (such as WI-FI interfaces, Bluetooth interfaces, 5G interfaces).

[0036] The communication bus is used to implement connection and communication between various different components. For example, the above-mentioned processor, memory, network interface, and user interface can be interconnected through the communication bus.

[0037] The user interface can include a display screen, an input unit (such as a keyboard). Optionally, the user interface can also include standard wired interfaces, wireless interfaces.

[0038] Those skilled in the art can understand that the above computer may also include more or fewer components, or combine certain components, or have different component arrangements. The embodiments of the present application do not limit this.

[0039] Exemplarily, as Figure 1 shown, the landslide hazard identification method based on unsupervised machine learning and temporal InSAR provided by the embodiments of the present application includes S101 - S105.

[0040] S101. Use the TS2Vec model to process the landslide displacements of multiple monitoring points in the area to be identified, and obtain multi-dimensional features of the landslide displacements of the multiple monitoring points.

[0041] In the embodiments of the present application, the landslide displacement of each of the above-mentioned monitoring points includes the landslide displacements of the monitoring point at multiple time points. The landslide displacement includes the landslide displacement in the horizontal direction and the landslide displacement in the vertical direction. The landslide displacement of each of the above-mentioned monitoring points belongs to time series data.

[0042] Optionally, the landslide displacement can be surface deformation data (i.e., InSAR data) obtained by a synthetic aperture radar (SAR).

[0043] It is understandable that the above-mentioned TS2Vec model is a deep learning module for extracting general features from time series data. The above-mentioned TS2Vec model includes an input projection layer, a timestamp masking module, a dilated CNN module, and a global pooling layer connected in sequence. Among them, the dilated CNN module includes multi-scale convolutional kernels.

[0044] Specifically, the input content of the above-mentioned TS2Vec model is the landslide displacements of multiple monitoring points in the area to be identified, and the output content of the above-mentioned TS2Vec model is the multi-dimensional features of the landslide displacements of multiple monitoring points. The working process of the above-mentioned TS2Vec model will be described below.

[0045] Step 1: Combine the landslide displacement x in the horizontal direction and the landslide displacement y in the vertical direction among the landslide displacements of multiple monitoring points to obtain a vector k (the vector k includes the combined amount k of the landslide displacements at multiple time points, where i represents the i-th time point and t represents time), and input the vector k into the TS2Vec model. t and the landslide displacement y in the vertical direction t to obtain a vector k after combination t (the vector k t includes the combined amount k of the landslide displacements at multiple time points, where i represents the i-th time point and t represents time), and input the vector k i,t into the TS2Vec model. t

[0046] Step 2: The input projection layer of the TS2Vec model maps the combined amount k of the landslide displacements at each time point in the vector k to a high-dimensional latent vector z (for example, z is a 64-dimensional vector) to obtain the high-dimensional latent vectors z of the landslide displacements of multiple monitoring points. t in the combined amount k of the landslide displacements at each time point i,t to a high-dimensional latent vector z i,t (for example, z i,t is a 64-dimensional vector), to obtain the high-dimensional latent vectors z of the landslide displacements of multiple monitoring points. t

[0047] Step 3: Input the high-dimensional latent vector z into the timestamp masking module. The timestamp masking module randomly selects time steps (such as t = 5, 10, 15) and masks the z corresponding to the time steps in the high-dimensional latent vector z through a binary mask (the mask elements are sampled from the Bernoulli distribution) to obtain the masked high-dimensional latent vector z. t and masks the z corresponding to the time steps in the high-dimensional latent vector z t through a binary mask (the mask elements are sampled from the Bernoulli distribution) to obtain the masked high-dimensional latent vector z i,t to obtain the masked high-dimensional latent vector z. t

[0048] Step 4: Input the masked high-dimensional latent vector z into the dilated CNN module. The multi-scale convolutional kernels (such as 1x3, 3x3, 5x3) in the dilated CNN module extract the temporal features of the masked high-dimensional latent vector z in parallel. t and the multi-scale convolutional kernels (such as 1x3, 3x3, 5x3) in the dilated CNN module extract the temporal features of the masked high-dimensional latent vector z t in parallel.

[0049] Step 5: Input the masked high-dimensional latent vector z t ​​​The temporal features are input into the global pooling layer, and the global pooling layer concatenates the temporal features of the masked latent vector z t to obtain multi-dimensional features of landslide displacements at multiple monitoring points (for example, the multi-dimensional features are 64-dimensional vectors).

[0050] It should be understood that since the above TS2Vec model belongs to the prior art, the specific hierarchical structure and working process of the above TS2Vec model will not be further elaborated in the embodiments of the present application.

[0051] S102. Perform dimensionality reduction processing on the multi-dimensional features through the t-SNE algorithm to obtain landslide displacement features of multiple monitoring points in the area to be identified.

[0052] Among them, the above landslide displacement features refer to the change trends of landslide displacements at the monitoring points in the horizontal and vertical directions over time. Therefore, the above landslide displacement features can also be referred to as landslide displacement temporal features.

[0053] The above t-SNE (t-Distributed Stochastic Neighbor Embedding) algorithm is a dimensionality reduction algorithm. When the t-SNE algorithm maps high-dimensional data to a low-dimensional space, the local similarity between data points in the high-dimensional space is retained.

[0054] The working process of the above t-SNE algorithm is as follows.

[0055] Step 1. Calculate the similarity between data points in the high-dimensional space:

[0056] For each data point x in the multi-dimensional features i , calculate the similarity between the data point x i and other data points x in the multi-dimensional features j . The similarity p i between the data point x j and other data points x in the multi-dimensional features ij satisfies the following formula.

[0057]

[0058] Among them, p ij indicates the probability of the data point x j in the neighborhood of the data point x i ; |x i -x j | 2 represents the Euclidean distance between the data point x i and the data point x j ; σ i represents the width parameter of x i ; σ iDetermines the neighborhood size of the data point x i , σ i is usually set by a relative distance metric (such as the K-nearest neighbor method); n represents the number of data points in the multi-dimensional feature, |x i - x k | 2 is the Euclidean distance between the data point x i and the data point x k .

[0059] Step 2, Initialize the low-dimensional space: Randomly initialize the data points y in the low-dimensional space.

[0060] Step 3, Calculate the similarity between the data points in the low-dimensional space. The similarity q ij between the data points in the low-dimensional space satisfies the following formula.

[0061]

[0062] where y i represents the i-th data point in the low-dimensional space, y j represents the j-th data point in the low-dimensional space, y k represents the k-th data point in the low-dimensional space; |y i - y j | 2 is the Euclidean distance between the data points y i and y j , (1 + |y i - y j | 2 ) -1 is the Student's t-distribution (with 1 degree of freedom) between the data points y i and y j ; |y i - y k | 2 is the Euclidean distance between the data points y i and y j , (1 + |y i - y k | 2 ) -1 is the Student's t-distribution (with 1 degree of freedom) between the data points y i and y j .

[0063] Step 4, Optimize the representation of the multi-dimensional feature in the low-dimensional space:

[0064] Minimize the difference in probability distributions between the high-dimensional and low-dimensional spaces through the gradient descent method. Usually, the Kullback-Leibler divergence (KL divergence) is used as the loss function. The loss function DK(p||q) satisfies the following formula.

[0065]

[0066] Step 5: Repeat Steps 1 to 4 until the loss function DK(p||q) is minimized or reaches a set threshold, and then splice the data points in the low-dimensional space to obtain the landslide displacement characteristics (which can be 2D vectors) of multiple monitoring points.

[0067] S103: After clustering the landslide displacement characteristics of multiple monitoring points, determine the area where the cluster satisfying the landslide condition is located in the clustering result as the landslide area of the area to be identified.

[0068] In one implementation, the above-mentioned landslide condition is that the rate of decline of the landslide displacement of the cluster in the horizontal direction satisfies the first decline condition, and the rate of decline in the vertical direction satisfies the second decline condition. The landslide area is the area with potential landslide hazards in the area to be identified. The area where the cluster not satisfying the landslide condition is located in the above clustering result is the other area of the area to be identified. The other area can be a settlement area. The embodiments of the present application do not limit the specific geological conditions represented by the other area.

[0069] Optionally, the above S103 includes S1031 - S1033.

[0070] S1031: Cluster the landslide displacement characteristics of multiple monitoring points to obtain multiple clusters.

[0071] Exemplarily, the landslide displacement characteristics of multiple monitoring points can be clustered by the K-means algorithm or the DBSCAN algorithm, or can be clustered by other clustering methods. The embodiments of the present application do not limit the type of clustering method.

[0072] Exemplarily, the process of clustering the landslide displacement characteristics of multiple monitoring points using the K-means algorithm is as follows.

[0073] Step 1: Initialize the cluster centers of K clusters.

[0074] Randomly select the landslide displacement characteristics of K monitoring points from the landslide displacement characteristics of multiple monitoring points, and respectively use them as the cluster centers of K clusters. Use the landslide displacement characteristics of the multiple monitoring points other than the cluster centers of K clusters among the landslide displacement characteristics of multiple monitoring points as the remaining landslide displacement characteristics. The cluster center of each cluster after initialization satisfies the following formula.

[0075] C k =(ck1 , c k2 , …, c kD ), k = 1, 2, …, K

[0076] Among them, C k is the cluster center of the k-th cluster, and c k1 is the first element of C k , c k2 is the second element of C k , c kD is the D-th element of C k , and D is the dimension of C k .

[0077] Step 2: Assign the remaining landslide displacement features to K clusters.

[0078] Specifically, calculate the Euclidean distance between the landslide displacement feature of each monitoring point in the remaining landslide displacement features and the cluster centers of the K clusters, and assign the remaining landslide displacement features to the cluster with the smallest Euclidean distance.

[0079] Among them, the calculation formula of the Euclidean distance is as follows.

[0080]

[0081] Among them, d(X i , C k ) represents the Euclidean distance between the i-th remaining landslide displacement feature X i in the remaining landslide displacement features and the cluster center C k of the k-th cluster; x ij is the j-th element in X i , and c kj is the j-th element in C k .

[0082] Step 3: Update the cluster center of each cluster.

[0083] Calculate the mean value of all landslide displacement features in each of the K clusters respectively, and use this mean value as the new cluster center, and use the landslide displacement features of multiple monitoring points other than the cluster centers of the K clusters among the landslide displacement features of multiple monitoring points as the remaining landslide displacement features. The new cluster center satisfies the following formula.

[0084]

[0085] Among them, S k represents the set of all data points in the k-th cluster (i.e., the set of all landslide displacement features in the k-th cluster), and N k represents the number of data points in the k-th cluster (i.e., the number of landslide displacement features in the k-th cluster).

[0086] Step 4: Repeat Step 2 to Step 3 until the change of the clustering center is less than the set change threshold or the preset maximum number of iterations is reached.

[0087] It can be understood that during the process of clustering the landslide displacement characteristics of multiple monitoring points, the elbow method or the silhouette coefficient method can be used to limit the clustering results and the number of clusters obtained by clustering.

[0088] In the above elbow method, the clustering results and the number of clusters obtained by clustering are evaluated by the sum of squared errors (SSE). The sum of squared errors SSE satisfies the following formula.

[0089]

[0090] Among them, C i represents the i-th cluster, p represents the sample point in C i (that is, the landslide displacement characteristic of a monitoring point), and m i represents the centroid of C i .

[0091] In the above silhouette coefficient method, the clustering results and the number of clusters obtained by clustering are evaluated by the silhouette coefficient. The silhouette coefficient satisfies the following formula.

[0092]

[0093] Among them, S(i) represents the silhouette coefficient of the i-th sample point, b(i) represents the average distance from the i-th sample point to other sample points belonging to the same cluster, a(i) represents the minimum value of the average distance from the sample point to all samples in other clusters, and max{a(i), b(i)} represents the larger value of a(i) and b(i).

[0094] Since the above elbow method and the above silhouette coefficient method belong to common technical means in the technical field, the embodiments of the present application will not further elaborate on the specific implementation processes of the above two methods.

[0095] S1032: Obtain the centroid of each cluster.

[0096] The centroid of each cluster is the mean value of the landslide displacement characteristics of all monitoring points in the cluster.

[0097] S1033: Determine the area where the clusters whose centroid descent rates in the horizontal direction satisfy the first descent condition and the centroid descent rates in the vertical direction satisfy the second descent condition among the multiple clusters as the landslide area of the area to be identified.

[0098] The area where the above-mentioned cluster is located refers to the area where multiple monitoring points corresponding to multiple sample points included in the cluster are located.

[0099] In one implementation, the above-mentioned first descent condition is that the range of the descent rate is [25, 50], and the above-mentioned second descent condition is that the range of the descent rate is [10, 25] within a time range of 1 to 3 months, where the unit of the descent rate is mm / day.

[0100] S104. Determine at least one individual landslide sub-region within the landslide area according to the landslide displacement and terrain data of multiple monitoring points in the landslide area.

[0101] Optionally, the above-mentioned terrain data may include the longitude and latitude, regional boundary, and terrain features of multiple monitoring points in the landslide area; the terrain features may include slope and elevation data.

[0102] Exemplarily, the above-mentioned S104 includes S1041 - S1043.

[0103] S1041. Process the landslide displacement characteristics and terrain data of multiple monitoring points in the landslide area to obtain the displacement terrain vectors of multiple monitoring points in the landslide area.

[0104] Specifically, within the same coordinate system (such as the WGS84 coordinate system or the UTM coordinate system), align the landslide displacement characteristics and terrain data of multiple monitoring points in the landslide area so that the landslide displacement characteristics of multiple monitoring points in the landslide area correspond one-to-one with their geographical locations. Then splice the landslide displacement characteristics and terrain data of multiple monitoring points in the landslide area to obtain the displacement terrain vectors of multiple monitoring points in the landslide area.

[0105] S1042. Use the K-shape algorithm to cluster the displacement terrain vectors of multiple monitoring points in the landslide area to obtain multiple clusters.

[0106] The above-mentioned K-shape algorithm clusters the landslide displacement and terrain data of multiple monitoring points in the landslide area based on the Normalized Cross-Correlation (NCC) similarity.

[0107] The following describes the processing process of clustering the displacement terrain vectors of multiple monitoring points in the landslide area using the K-shape algorithm.

[0108] Step 1. Initialize the cluster centers.

[0109] Set the number of multiple clusters obtained by clustering the displacement terrain vectors of multiple monitoring points in the landslide area using the K-shape algorithm to be K, and initialize the cluster centers of each of the K clusters. Among them, the initialization method can randomly select the displacement terrain vectors of K monitoring points from the displacement terrain vectors of multiple monitoring points in the landslide area. Optionally, it can also be initialized by other methods (for example, selecting the most representative displacement terrain vector), which is not limited in the embodiments of the present application.

[0110] Step 2: Assign the displacement terrain vectors of multiple monitoring points in the landslide area to the K clusters.

[0111] For the displacement terrain vectors among the displacement terrain vectors of multiple monitoring points in the landslide area except for the cluster centers of the K clusters, calculate the normalized cross-correlation similarity between the displacement terrain vector and the cluster centers of the K clusters, and assign the displacement terrain vector to the cluster where the cluster center with the highest normalized cross-correlation similarity to the displacement terrain vector is located. The above-mentioned normalized cross-correlation similarity satisfies the following formula.

[0112]

[0113] Among them, NCC(X (i) ,C k ) represents the normalized cross-correlation similarity between the displacement terrain vector X (i) and the cluster center C k of the k-th cluster; X = (x1, x2,..., x T ), x1 represents the displacement terrain value corresponding to the first time point in X, x2 represents the displacement terrain value corresponding to the second time point in X, x T represents the displacement terrain value corresponding to the T-th time point in X; C k = (c1, c2,..., c T ), c1 represents the displacement terrain value corresponding to the first time point in C k , c2 represents the displacement terrain value corresponding to the second time point in C k , c T represents the displacement terrain value corresponding to the T-th time point in C k ; x t represents the displacement terrain value corresponding to the t-th time point in X, represents the mean value of the displacement terrain values in X, c t represents the displacement terrain value corresponding to the t-th time point in C k , represents the mean value of the displacement terrain values in C k .

[0114] Step 3: Update the cluster centers of the K clusters.

[0115] For each of the K clusters, the weighted average of the normalized cross - correlation similarities between all the points in the cluster (i.e., the displacement terrain vectors of multiple monitoring points included in the cluster) and the cluster center (i.e., the displacement terrain vector of a monitoring point serving as the cluster center) is used as the new cluster center. The new cluster center satisfies the following formula.

[0116]

[0117] Among them, S k represents the set of all data points in the k - th cluster (i.e., the set of displacement terrain vectors of multiple monitoring points included in C k ).

[0118] Step 4: Repeat Step 2 to Step 3 until the displacement terrain vectors of multiple monitoring points in the landslide area are all assigned to the K clusters.

[0119] S1043: Determine the area where each cluster is located as a single - landslide sub - area among the multiple clusters.

[0120] Among them, the area where each cluster is located refers to the area where the corresponding multiple monitoring points in the cluster are located.

[0121] S105: For each single - landslide sub - area, use the cause - analysis model to analyze the landslide data of the single - landslide sub - area to determine the cause of the landslide in the single - landslide sub - area.

[0122] Among them, the single - landslide sub - area and the cause of the landslide in the single - landslide sub - area are the identification results of landslide hazards. The landslide data of the single - landslide sub - area includes the landslide displacement and environmental information of the single - landslide sub - area.

[0123] Optionally, the above - mentioned environmental information may include remote - sensing image data, meteorological data, external - environment data, geological data, seismic data, and human - activity data. The meteorological data may include precipitation, temperature, wind speed, etc. The external - environment data may include geological structure and soil type, etc. The geological data may include slope, soil type, and rock type, etc. The seismic data may include the occurrence time, epicenter location, magnitude, and focal depth of the earthquake, etc. The human - activity data may include relevant data of events such as groundwater extraction.

[0124] Specifically, S105 is implemented by inputting the landslide data of the landslide sub - area into the cause - analysis model. After the cause - analysis model analyzes the landslide data of the landslide sub - area, it outputs the cause of the landslide in the landslide sub - area.

[0125] In one implementation, the above - mentioned cause - analysis model is trained based on the constructed data training set. The data training set includes historical landslide data and the causes of landslides determined based on the frequency - domain characteristics of the historical landslide data.

[0126] Optionally, during the process of determining the data training set, the causes of landslides are determined based on the frequency domain characteristics of historical landslide data, including the following steps 1 to 5.

[0127] Step 1: Perform a fast Fourier transform on the landslide displacement characteristics of multiple monitoring points in the historical landslide data to obtain the landslide displacement frequency domain characteristics of multiple monitoring points.

[0128] Step 2: Perform K-medoids clustering on the landslide displacement frequency domain characteristics of multiple monitoring points to obtain multiple clusters.

[0129] Step 3: Perform periodic analysis on the landslide displacement frequency domain characteristics of the cluster centers of each of the multiple clusters to determine the periodic characteristics of each cluster.

[0130] Exemplarily, the periodic characteristics can be annual cycle (12 months in a year), quarterly cycle (3 - 6 months), or no cycle. The periodic characteristics can also be daily cycle (1 day) or semi-daily cycle (12 hours). The trend characteristics include trend intensity and significance index. The significance index is the P value.

[0131] The trend intensity satisfies the following formula:

[0132] y = β0 + β1t + ε

[0133] where y represents the landslide displacement of the cluster center of the cluster, β0 represents the intercept, β1 represents the trend intensity, t represents time, and ε represents the residual.

[0134] In an application scenario, the correspondence between the periodic characteristics, driving factors, and landslide displacement is shown in Table 1 as follows.

[0135] Table 1: Correspondence table of periodic characteristics, driving factors, and landslide displacement

[0136]

[0137] Step 4: Perform trend analysis on the landslide displacement frequency domain characteristics of the cluster centers of each of the multiple clusters to determine the trend characteristics of each cluster.

[0138] Step 5: Determine the landslide cause corresponding to the periodic characteristics, trend characteristics, and environmental information of each cluster in the landslide cause analysis table as the landslide cause of the cluster.

[0139] In an implementation manner, the landslide cause analysis table is shown in Table 2 as follows.

[0140] Table 2: Landslide cause analysis table

[0141]

[0142] As can be seen from Table 2 above, in the landslide cause analysis table:

[0143] When the environmental information includes periodic rainfall data with an annual cycle as the periodic feature, a trend feature of -0.2 mm / month, and a P-value less than 0.01, the corresponding landslide cause is rainfall; then the landslide event corresponding to rainfall can be called a rainfall-induced landslide.

[0144] When the environmental information includes human activity data with a quarterly cycle as the periodic feature, a trend feature of 0.05 mm / month, and a P-value within the range of (0.01, 0.15), the corresponding landslide cause is human activity; then the landslide event corresponding to human activity can be called a human activity-induced landslide.

[0145] When the environmental information includes earthquake data with no cycle as the periodic feature, a trend feature of -0.5 mm / month, and a P-value less than 0.001, the corresponding landslide cause is earthquake, and the landslide event corresponding to earthquake can be called an earthquake-induced landslide.

[0146] Specifically, the above cause analysis model includes:

[0147] A first input layer of degree connection, multiple parallel InceptionTime networks, and a fully connected layer connected in sequence. Among them, the input ends of the multiple parallel InceptionTime networks are all connected to the input layer, and the output ends of the multiple parallel InceptionTime networks are all connected to the fully connected layer.

[0148] Among the multiple parallel InceptionTime networks, each InceptionTime network includes, connected in sequence: a second input layer, a first residual module, a second residual module, and a global average pooling layer.

[0149] As Figure 2 shown, both the first residual module and the second residual module include, connected by residual connection in sequence: multiple Inception sub-modules and a multi-head latent attention layer; the residual connection is provided with learnable weights.

[0150] Among them, the number of Inception sub-modules can be three, and the number of heads of the multi-head latent attention layer (Heads) can be four to eight.

[0151] For the multi-head latent attention layer, since the Inception sub-modules extract local features through different convolutions, and the attention mechanism of the multi-head latent attention layer can capture long-range dependencies and form a complementarity, enabling the cause analysis model to enhance the global dependency relationship of temporal features, avoid input noise interference, and reduce the error of the output result of the cause analysis model.

[0152] For the residual connection with learnable weights, in the traditional residual connection, the input is directly added to the output of the convolutional layer, and the formula is H(x) = F(x) + x, where F(x) is the result after the convolutional operation. The residual connection with learnable weights can automatically adjust the weights of the residual connection. For example, H(x) = α * F(x) + β * x, where α and β are learnable parameters or parameters dynamically adjusted according to the input. By introducing learnable scalar parameters (such as the above α and β) in the residual block, the weights can be automatically optimized through backpropagation.

[0153] As Figure 3 shown, the Inception sub-module includes, connected in sequence: a bottleneck layer, a parallel convolutional layer, a max pooling layer, and a fully connected layer. Among them, the parallel convolutional layer includes a 1×1 convolutional layer, a 3×3 convolutional layer, and a 5×5 convolutional layer. The input ends of the 3×3 convolutional layer, the 5×5 convolutional layer, and the max pooling layer are connected to the bottleneck layer. The input end of the 1×1 convolutional layer is connected to the input end of the Inception sub-module. The output ends of the 1×1 convolutional layer, the 3×3 convolutional layer, the 5×5 convolutional layer, and the max pooling layer are connected to the fully connected layer.

[0154] For the above-mentioned bottleneck layer, exemplarily, when the input is an M-dimensional multivariate time series, it first passes through the bottleneck layer in this module. The above-mentioned bottleneck layer uses m filters with a length of 1 and a stride of 1 to convert an M-dimensional signal into an m-dimensional (m << M) signal, aiming to reduce the dimension, reduce the model complexity at the same time, and alleviate the overfitting problem.

[0155] The training process of the above-mentioned cause analysis model is as follows.

[0156] Step 1, construct a deep learning model.

[0157] Specifically, define the number of layers (refer to the relevant description of the above-mentioned cause analysis model), activation function, loss function, and optimizer (such as the Adam optimizer) of the deep learning model.

[0158] In one application scenario, the loss function L of the above-mentioned cause analysis model satisfies the following formula.

[0159]

[0160] Among them, CS represents the number of landslide causes, y i represents the i-th landslide cause, and p_i represents the probability of the i-th landslide cause.

[0161] Step 2, train the deep learning model.

[0162] Step 2.1, forward propagation:

[0163] Use the historical landslide data in the training dataset as input data and input it into the deep learning model. The loss function calculates the error between the predicted value and the true value.

[0164] Step 2.1, Backpropagation:

[0165] Calculate the gradient of the loss function with respect to the model parameters and pass the gradient from the output layer to the input layer through the chain rule.

[0166] Step 2.3, Parameter Update: Use the optimizer to update the deep learning model parameters according to the gradient.

[0167] Step 2.4, Model Validation: Evaluate the performance (such as accuracy or precision) of the deep learning model through the validation dataset. Since the calculation formulas for accuracy and precision are common technical means in this technical field, they are not elaborated in this embodiment of the present application.

[0168] It can be understood that during the model validation process, if the deep learning model shows overfitting (i.e., the training set accuracy is very high, but the validation set accuracy is low), the deep learning model can be optimized by adding Dropout layers and applying Early Stopping and other methods.

[0169] Step 2.5, Iteration: Repeat Steps 2.1 to 2.4 until the deep learning model converges or reaches a predetermined number of training epochs, and use the trained deep learning model as the cause analysis model.

[0170] In summary, in the landslide hazard identification method based on unsupervised machine learning and temporal InSAR provided by the embodiments of the present application, first, the TS2Vec model is used to process the landslide displacements of multiple monitoring points in the area to be identified, obtaining multi-dimensional features of the landslide displacements of multiple monitoring points. Then, the t-SNE algorithm is used to reduce the dimensions of the multi-dimensional features of the landslide displacements of multiple monitoring points, obtaining the landslide displacement features of multiple monitoring points. Determine the landslide areas with potential landslide hazards in the area to be identified according to the clustering results of the landslide displacement features of multiple monitoring points. Based on the landslide displacements and terrain data of multiple monitoring points in the landslide area, determine at least one individual landslide sub-area in the landslide area. Then, use the cause analysis model to analyze and obtain the causes of landslides for each individual landslide sub-area. Thus, the identification result of the landslide hazard is obtained. From the above content, it can be seen that the above method can determine the individual landslide sub-areas and their causes of landslides, and thus can timely formulate and take prevention and control measures according to the individual landslide sub-areas and their causes of landslides, eliminating potential safety hazards in the areas where landslides may occur.

[0171] Correspondingly, the embodiments of the present application provide a landslide hazard identification device based on unsupervised machine learning and temporal InSAR, as Figure 4As shown in the figure, it includes a multi-dimensional feature determination module 501, a landslide displacement feature determination module 502, a landslide area determination module 503, a single landslide sub-area determination module 504, and a landslide cause determination module 505.

[0172] Among them, the multi-dimensional feature determination module 501 is used to process the landslide displacements of multiple monitoring points in the area to be identified by using the TS2Vec model, and obtain the multi-dimensional features of the landslide displacements of multiple monitoring points; the landslide displacement of each monitoring point includes the landslide displacements of the monitoring point at multiple time points respectively. For example, the multi-dimensional feature determination module 501 is used to implement S101 of the above-mentioned landslide hazard identification method based on machine learning.

[0173] The landslide displacement feature determination module 502 is used to perform dimensionality reduction processing on the multi-dimensional features through the t-SNE algorithm to obtain the landslide displacement features of multiple monitoring points in the area to be identified. For example, the landslide displacement feature determination module 502 is used to implement S102 of the above-mentioned landslide hazard identification method based on machine learning.

[0174] The landslide area determination module 503 is used to, after clustering the landslide displacement features of multiple monitoring points, determine the area where the cluster satisfying the landslide condition is located in the cluster result as the landslide area of the area to be identified; the landslide condition is that the descending rate of the landslide displacement of the cluster in the horizontal direction satisfies the first descending condition, and the descending rate in the vertical direction satisfies the second descending condition; the landslide area is the area with landslide hazards in the area to be identified. For example, the landslide area determination module 503 is used to implement S103 of the above-mentioned landslide hazard identification method based on machine learning.

[0175] The single landslide sub-area determination module 504 is used to determine at least one single landslide sub-area within the landslide area according to the landslide displacements and terrain data of multiple monitoring points in the landslide area. For example, the single landslide sub-area determination module 504 is used to implement S104 of the above-mentioned landslide hazard identification method based on machine learning.

[0176] The landslide cause determination module 505 is used to, for each single landslide sub-area, analyze the landslide data of the single landslide sub-area by using a cause analysis model to determine the landslide cause of the single landslide sub-area; the single landslide sub-area and the landslide cause of the single landslide sub-area are the identification results of the landslide hazard; among them, the landslide data of the single landslide sub-area includes the landslide displacement and environmental information of the single landslide sub-area; the cause analysis model is trained based on the constructed data training set; the data training set includes historical landslide data and landslide causes determined based on the frequency domain features of historical landslide data. For example, the landslide cause determination module 505 is used to implement S105 of the above-mentioned landslide hazard identification method based on machine learning.

[0177] Optionally, the single landslide sub-region determination module 504 is specifically configured to: process the landslide displacement characteristics and terrain data of multiple monitoring points in the landslide area to obtain the displacement terrain vectors of multiple monitoring points in the landslide area. Use the K-shape algorithm to cluster the displacement terrain vectors of multiple monitoring points in the landslide area to obtain multiple clusters. Determine the area where each cluster is located in the multiple clusters as a single landslide sub-region. For example, the single landslide sub-region determination module 504 is specifically configured to implement S1041-S1043 of the above-mentioned landslide hidden danger identification method based on machine learning.

[0178] Each module of the above-mentioned landslide hidden danger identification device based on unsupervised machine learning and temporal InSAR can also be used to execute other steps in the above method embodiments. All relevant contents involved in the above method embodiments can be cited in the function descriptions of the corresponding functional modules and will not be elaborated here.

[0179] The embodiment of the present application also provides an electronic device, including: a processor and a memory coupled to the processor; the memory is used to store computer instructions. When the electronic device runs, the processor executes the computer instructions stored in the memory so that the electronic device executes the method described in the above embodiment. Among them, the processor can implement the above-mentioned multi-dimensional feature determination module 501, landslide displacement feature determination module 502, landslide area determination module 503, single landslide sub-region determination module 504, and landslide cause determination module 505; the above-mentioned memory can also be used for multi-dimensional features, landslide displacement features, landslide areas, single landslide sub-regions, and the causes of single landslide sub-regions, etc.

[0180] The embodiment of the present application also provides a computer-readable storage medium, which includes a computer program. When the computer program runs on a computer, it is used to execute the method described in the above embodiment.

[0181] The embodiment of the present application also provides a computer program product, which includes computer program instructions. When the computer program instructions run on a computer, it is used to execute the method described in the above embodiment.

[0182] Each embodiment in this specification is described in a progressive manner. The same or similar parts among the embodiments can be referred to each other, and the key points of each embodiment are the differences from other embodiments.

[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A landslide hazard identification method based on unsupervised machine learning and temporal InSAR, characterized in that, Including: Using the TS2Vec model to process the landslide displacements of multiple monitoring points in the area to be identified, and obtaining multi-dimensional features of the landslide displacements of the multiple monitoring points; The landslide displacement of each monitoring point includes the landslide displacements of the monitoring point at multiple time points respectively; Performing dimensionality reduction processing on the multi-dimensional features through the t-SNE algorithm to obtain the landslide displacement features of multiple monitoring points in the area to be identified; After clustering the landslide displacement features of the multiple monitoring points, determining the area where the cluster satisfying the landslide condition is located in the cluster result as the landslide area of the area to be identified; the landslide condition is that the descending rate of the landslide displacement of the cluster in the horizontal direction satisfies the first descending condition, and the descending rate in the vertical direction satisfies the second descending condition; the landslide area is the area with landslide hazards in the area to be identified; According to the landslide displacements and topographic data of multiple monitoring points in the landslide area, determining at least one single landslide sub-area within the landslide area; For each single landslide sub-area, using a cause analysis model to analyze the landslide data of the single landslide sub-area to determine the landslide cause of the single landslide sub-area; the single landslide sub-area and the landslide cause of the single landslide sub-area are the identification results of the landslide hazards; wherein, the landslide data of the single landslide sub-area includes the landslide displacement and environmental information of the single landslide sub-area; the cause analysis model is trained based on a constructed data training set; the data training set includes historical landslide data and landslide causes determined based on the frequency domain features of the historical landslide data.

2. The method according to claim 1, characterized in that, Determining the landslide cause based on the frequency domain features of the historical landslide data includes: Performing a fast Fourier transform on the landslide displacement features of multiple monitoring points in the historical landslide data to obtain the landslide displacement frequency domain features of the multiple monitoring points; Performing K-medoids clustering on the landslide displacement frequency domain features of the multiple monitoring points to obtain multiple clusters; Performing periodic analysis on the landslide displacement frequency domain features of the cluster center of each cluster in the multiple clusters to determine the periodic features of each cluster; Performing trend analysis on the landslide displacement frequency domain features of the cluster center of each cluster in the multiple clusters to determine the trend features of each cluster; Determining the landslide cause corresponding to the periodic features, trend features, and environmental information of each cluster in the landslide cause analysis table as the landslide cause of the cluster.

3. The method according to claim 2, wherein The periodic feature is an annual cycle, a quarterly cycle, or no cycle; The trend feature includes a trend intensity and a significance index; the significance index is a P value; the trend intensity satisfies the following formula: y = β0 + β1t + ε where y represents the landslide displacement of the cluster center of the cluster, β0 represents the intercept, β1 represents the trend intensity, t represents time, and ε represents the residual.

4. The method according to claim 3, wherein In the landslide cause analysis table: When the environmental information includes periodic rainfall data with an annual cycle as the periodic feature, a trend feature magnitude of -0.2 mm / month, and a P-value less than 0.01, the corresponding landslide cause is rainfall; When the environmental information includes human activity data with a quarterly cycle as the periodic feature, a trend feature magnitude of 0.05 mm / month, and a P-value within the range of (0.01, 0.15), the corresponding landslide cause is human activity; When the environmental information includes earthquake data with no cycle as the periodic feature, a trend feature magnitude of -0.5 mm / month, and a P-value less than 0.001, the corresponding landslide cause is earthquake.

5. The method according to claim 1, wherein The cause analysis model includes: A first input layer of connection degree, multiple parallel InceptionTime networks, and a fully connected layer connected in sequence. Among them, the input ends of the multiple parallel InceptionTime networks are all connected to the input layer, and the output ends of the multiple parallel InceptionTime networks are all connected to the fully connected layer; Among the multiple parallel InceptionTime networks, each InceptionTime network includes, connected in sequence: a second input layer, a first residual module, a second residual module, and a global average pooling layer; Both the first residual module and the second residual module include, connected in residual connection in sequence: multiple Inception sub-modules and a multi-head latent attention layer; the residual connection is provided with learnable weights; The Inception sub-module includes, connected in sequence: a bottleneck layer, a parallel convolution layer, a max pooling layer, and a fully connected layer. Among them, the parallel convolution layer includes a 1×1 convolution layer, a 3×3 convolution layer, and a 5×5 convolution layer. The input ends of the 3×3 convolution layer, the 5×5 convolution layer, and the max pooling layer are connected to the bottleneck layer. The input end of the 1×1 convolution layer is connected to the input end of the Inception sub-module. The output ends of the 1×1 convolution layer, the 3×3 convolution layer, the 5×5 convolution layer, and the max pooling layer are connected to the fully connected layer.

6. The method according to claim 1, wherein, The first descent condition is that the range of the descent rate is [25, 50], and the second descent condition is that the range of the descent rate is [10, 25] within a time range of 1 to 3 months, where the unit of the descent rate is mm / day.

7. Landslide hazard identification device based on unsupervised machine learning and time-series InSAR, characterized in that, It includes a multi-dimensional feature determination module, a landslide displacement feature determination module, a landslide area determination module, a single landslide sub-area determination module, and a landslide cause determination module; The multi-dimensional feature determination module is used to process the landslide displacements of multiple monitoring points in the area to be identified by using the TS2Vec model to obtain the multi-dimensional features of the landslide displacements of the multiple monitoring points; the landslide displacement of each monitoring point includes the landslide displacements of the monitoring point at multiple time points respectively; The landslide displacement feature determination module is used to perform dimensionality reduction processing on the multi-dimensional features through the t-SNE algorithm to obtain the landslide displacement features of multiple monitoring points in the area to be identified; The landslide area determination module is configured to, after clustering the landslide displacement characteristics of the multiple monitoring points, determine the area where the cluster satisfying the landslide condition in the clustering result is located as the landslide area of the area to be identified; the landslide condition is that the descending rate of the landslide displacement of the cluster satisfies the first descending condition in the horizontal direction and the second descending condition in the vertical direction; the landslide area is the area with landslide hazards in the area to be identified; The single landslide sub-area determination module is configured to determine at least one single landslide sub-area within the landslide area according to the landslide displacement and terrain data of the multiple monitoring points in the landslide area; The landslide cause determination module is configured to, for each single landslide sub-area, analyze the landslide data of the single landslide sub-area by using a cause analysis model to determine the landslide cause of the single landslide sub-area; the single landslide sub-area and the landslide cause of the single landslide sub-area are the identification results of the landslide hazards; wherein, the landslide data of the single landslide sub-area includes the landslide displacement and environmental information of the single landslide sub-area; the cause analysis model is trained based on the constructed data training set; the data training set includes historical landslide data and the landslide causes determined based on the frequency domain characteristics of the historical landslide data.

8. An electronic device, characterized in that, It includes a processor and a memory coupled to the processor; the memory is used to store computer instructions, and when the electronic device runs, the processor executes the computer instructions stored in the memory so that the electronic device executes the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It includes computer program instructions, and when the computer program instructions are executed by a computer, the computer is made to execute the method according to any one of claims 1 to 6.

10. A computer program product, characterized in that, It includes computer program instructions, and when the computer program instructions run on a computer, the computer is made to execute the method according to any one of claims 1 to 6.

Images