An unsupervised deep clustering co-seismic landslide identification method, device and equipment

Abstract

The application provides an aseismic landslide identification method, device and equipment of unsupervised deep clustering, and relates to the technical field of graphic image processing, and comprises the following steps: acquiring a multi-source data set, preprocessing the multi-source data set, converting it into a preset image format, inputting multispectral remote sensing images in the multi-source data set into a remote sensing feature extraction branch for feature reconstruction to obtain first-class features, and inputting normalized vegetation index NDVI, terrain data, geographic data, geological maps and seismic data in the multi-source data set into an implicit knowledge feature extraction branch to obtain second-class features; the second-class features are implicit knowledge; under the supervision of a soft label, the first-class features and the second-class features are fused and clustered, and the corresponding aseismic landslide category of a to-be-measured area is determined based on a clustering result. By constructing implicit knowledge features, the mechanism of landslide disasters can be implicitly captured and expressed in the model through node features and graph structures, and the accuracy of identification is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of graphic image processing, and in particular to a coseismic landslide identification method, device and equipment based on unsupervised deep clustering. BACKGROUND

[0002] The identification of coseismic landslides has traditionally relied on field measurement methods such as field surveys to collect basic information about slope instability and surface damage. Such field surveys are usually carried out along rivers and roads, but if these roads or rivers are blocked or damaged due to an earthquake or coseismic landslide, disaster investigation work will be severely affected. This traditional method often fails to observe all coseismic landslides, especially those occurring in inaccessible areas. With the continuous progress of remote sensing technology, the investigation of coseismic landslides has been significantly improved, benefiting from various types of remote sensing data, including multispectral, hyperspectral remote sensing and synthetic aperture radar data. These remote sensing images are crucial for defining the boundaries of coseismic landslides and determining their geographic location. Nevertheless, the process of manually interpreting remote sensing images is still subjective and time-consuming, which limits the efficiency of post-earthquake landslide identification. Therefore, improving the automation and objectivity of remote sensing image interpretation is of great significance for quickly and effectively identifying coseismic landslides.

[0003] Currently, unsupervised feature landslide identification has made important progress, and a large number of excellent results have emerged. However, there are still limitations as follows: for the task of identifying coseismic landslides, it is not known in advance how many ground object categories are in the image, or which features correspond to coseismic landslides. It is worth studying how to identify coseismic landslides in the case where the number of image categories is unknown and it is not known which category belongs to coseismic landslides. In addition, the method using a knowledge graph requires a large amount of complex geological data and deep domain expertise, which puts high demands on the knowledge base of researchers and the technology used for knowledge graph representation and modeling. SUMMARY

[0004] The present application aims to solve the problem of current unsupervised feature landslide identification. The present application provides a coseismic landslide identification method, device and equipment based on unsupervised deep clustering.

[0005] The technical solution of the present application embodiment is as follows:

[0006] The present application provides a coseismic landslide identification method based on unsupervised deep clustering, comprising:

[0007] obtaining a multi-source data set; the multi-source data set includes multispectral remote sensing images, normalized vegetation index NDVI, terrain data, geographic data, geological maps and seismic data related to the area to be measured;

[0008] Preprocess the multi-source data set, and convert the multi-source data set into a preset image format;

[0009] input the multispectral remote sensing image in the preset image format into a remote sensing feature extraction branch to perform feature reconstruction, to obtain first-class features, and input the normalized vegetation index NDVI, the terrain data, the geographic data, the geological map, and the seismic data in the preset image format into an implicit knowledge feature extraction branch, to obtain second-class features; the second-class features are implicit knowledge;

[0010] Under soft label supervision, fuse and cluster the first-class features and the second-class features to obtain a clustering result, and determine the coseismic landslide category corresponding to the to-be-measured area based on the clustering result.

[0011] Optionally, the preprocessing of the multi-source data set and the conversion of the multi-source data set into a preset image format comprise:

[0012] convert the multi-source data set into an image format, and perform superimposition processing to obtain a multi-band image;

[0013] crop the multi-band image into a block with a size of 128x128.

[0014] Optionally, the remote sensing feature extraction branch comprises a self-encoder, and the input of the multispectral remote sensing image in the preset image format into the remote sensing feature extraction branch to perform feature reconstruction to obtain first-class features comprises:

[0015] input the cropped multi-band image into the self-encoder, reconstruct the image by using the self-encoder, and continuously iterate to minimize the reconstruction error until the first-class features are obtained.

[0016] Optionally, the structure principle of the self-encoder comprises:

[0017] :

[0018] wherein, is an encoder function, g is a decoder function, is an input multi-band image, is an output first-class feature.

[0019] Optionally, the input of the normalized vegetation index NDVI, the terrain data, the geographic data, the geological map, and the seismic data in the preset image format into the implicit knowledge feature extraction branch to obtain second-class features comprises:

[0020] Creating a node feature matrix based on the normalized vegetation index NDVI, the terrain data, the geographic data, the geological map, and the image corresponding to the seismic data; each node of the node feature matrix corresponds to a pixel point, and the node feature corresponds to the pixel value of the pixel point;

[0021] Constructing an adjacency matrix of the node feature matrix using an 8-connected neighborhood;

[0022] The adjacency matrix is ​​used to perform a graph convolution operation to obtain the second type of features.

[0023] Optionally, the fusing and clustering the first category features and the second category features under soft label supervision to obtain a clustering result includes:

[0024] The difference between the soft label distribution and the target distribution is quantified using the divergence KL, and the first type of features and the second type of features are fused and clustered to obtain a clustering result;

[0025] Given a feature map :

[0026] ;

[0027] Given cluster centers:

[0028] ;

[0029] ;

[0030] This gives the soft label distribution of the probability that each feature point belongs to each cluster:

[0031] ;

[0032] in, represents the eigenvector of point i, represents the initial cluster center j, Represents the i-th sample point With the jth cluster center The square of the Euclidean distance between is the updated cluster center, is the scaling parameter.

[0033] Optionally, fusing and clustering the first category features and the second category features under soft label supervision to obtain a clustering result includes:

[0034] Under the supervision of soft labels, clustering is completed after fusing the first category features and the second category features;

[0035] The optimal number of clusters is determined by hierarchical clustering, and the cluster centers are updated based on the optimal number.

[0036] The second aspect of the embodiment of the application provides an unsupervised deep clustering co-seismic landslide identification device, comprising a data acquisition module, a preprocessing module, a feature reconstruction module and a fusion clustering module, wherein,

[0037] The data acquisition module is configured to acquire a multi-source data set; the multi-source data set comprises multispectral remote sensing images, normalized vegetation index NDVI, terrain data, geographic data, geological maps and seismic data related to a to-be-measured region;

[0038] The preprocessing module is configured to preprocess the multi-source data set and convert the multi-source data set into a preset image format;

[0039] The feature reconstruction module is configured to input the multispectral remote sensing images in the preset image format to a remote sensing feature extraction branch for feature reconstruction to obtain first-type features, and input the normalized vegetation index NDVI, the terrain data, the geographic data, the geological maps and the seismic data in the preset image format to an implicit knowledge feature extraction branch to obtain second-type features; the second-type features are implicit knowledge;

[0040] The fusion clustering module is configured to fuse and cluster the first-type features and the second-type features under soft label supervision to obtain a clustering result, and determine a co-seismic landslide category corresponding to the to-be-measured region based on the clustering result.

[0041] The third aspect of the embodiment of the application provides an electronic device comprising a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the unsupervised deep clustering co-seismic landslide identification method of the first aspect.

[0042] The fourth aspect of the embodiment of the application provides a computer readable storage medium having a computer program stored thereon, and the program, when executed by a processor, implements the steps of the method of the first aspect.

[0043] Compared with the prior art, the technical solution provided by the application has the following beneficial effects:

[0044] The application provides an aseismic landslide identification method, device and equipment based on unsupervised deep clustering. The method comprises the following steps: acquiring a multi-source data set, respectively reconstructing features by using a remote sensing feature extraction branch to obtain first type features, and using an implicit knowledge feature extraction branch to obtain second type features, fusing and clustering the first type features and the second type features under soft label supervision to obtain a clustering result, and determining an aseismic landslide category corresponding to a to-be-detected area based on the clustering result. By constructing the implicit knowledge features, the mechanism of the landslide disaster can be implicitly captured and expressed in the model through the node features and the graph structure, and the accuracy of the identification is improved. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 A flowchart of an aseismic landslide identification method based on unsupervised deep clustering provided by the application;

[0046] Figure 2 Results of aseismic landslide identification in Hokkaido by using six different methods;

[0047] Figure 3 Results of aseismic landslide identification in Haiti by using six different methods;

[0048] Figure 4 Results of aseismic landslide identification in Luding by using six different methods;

[0049] Figure 5 A structural diagram of an aseismic landslide identification device based on unsupervised deep clustering provided by the application;

[0050] Figure 6 A structural diagram of an electronic device provided by the application. DETAILED DESCRIPTION

[0051] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. It should be understood, however, that the description is merely for illustration, and is not intended to limit the scope of the present application. In the following detailed description of embodiments of the present application, numerous specific details are set forth in order to provide a thorough understanding of the embodiments. However, it will be apparent to one skilled in the art that one or more embodiments can be practiced without these specific details. In other instances, well-known structures and functions have not been described in detail in order to avoid obscuring aspects of the application.

[0052] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the present application. As used herein, the term "includes" and "comprising" and the like means the inclusion of the stated features, steps, operations, and / or components, but not the exclusion of one or more other features, steps, operations, or components.

[0053] All terms used herein, including technical and scientific terms, have the meanings commonly understood by one of ordinary skill in the art, unless otherwise defined. It should be further understood that the terms used herein should be interpreted as having a meaning that is consistent with the understanding of those terms by those skilled in the relevant art and that the terms should not be interpreted in an overly rigid or overly formal manner unless clearly defined.

[0054] Some of the diagrams illustrated in the drawings are block diagrams and / or flowcharts. It should be understood that some of the blocks in the block diagrams and / or flowcharts, or combinations thereof, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, so that the instructions executed by the processor can create a means for implementing the functions / operations specified in the block diagrams and / or flowcharts.

[0055] In some embodiments, please refer to Figure 1 , Figure 1 A flowchart of an unsupervised deep clustering co-seismic landslide identification method provided by an embodiment of the present application; the unsupervised deep clustering co-seismic landslide identification method provided by the embodiment of the present application comprises:

[0056] S110, acquiring a multi-source data set; the multi-source data set comprises multispectral remote sensing images, normalized vegetation index NDVI, terrain data, geographic data, a geological map, and seismic data related to a region to be measured.

[0057] In the embodiment, the multispectral remote sensing images can be Planet remote sensing images, the normalized vegetation index NDVI is generated by calculating remote sensing images, the terrain data can be used to obtain elevation, slope, and slope direction, the geographic data can be used to obtain road and water system vectors, in addition, the road and water system vectors are converted into distance from road distance and distance from water system distance through Euclidean distance, the lithology data can be obtained through the geological map, and the seismic data can adopt the seismic data published by the United States Geological Survey, and the intensity and PGA (Peak Ground Acceleration, peak ground acceleration) can be extracted through the data.

[0058] S120, pre-processing the multi-source data set, and converting the multi-source data set into a preset image format;

[0059] In some embodiments, S120, pre-processing the multi-source data set, and converting the multi-source data set into a preset image format, comprises:

[0060] Converting the multi-source data set into an image format and performing superimposition processing to obtain a multi-band image;

[0061] The multi-band image is cropped into a block with a size of 128x128.

[0062] In this embodiment, the multi-source data set is converted into a preset size image format, standardized, and unified in size, and then a plurality of images are superimposed together to generate a multi-band image, and the multi-band image is cropped to adapt to subsequent processing by the autoencoder. Here, multi-threading or parallel processing can be used to speed up the data loading and cropping process.

[0063] In S130, the multi-spectral remote sensing image in the preset image format is input to the remote sensing feature extraction branch for feature reconstruction to obtain the first type of features, and the normalized vegetation index NDVI, terrain data, geographic data, geological map and seismic data in the preset image format are input to the implicit knowledge feature extraction branch to obtain the second type of features. The second type of features are implicit knowledge.

[0064] In this embodiment, the multi-spectral remote sensing image is input to the remote sensing feature extraction branch for feature reconstruction, and the factor data composed of NDVI, seismic, terrain data, geographic data, geological map and seismic data is input to the implicit knowledge feature extraction branch, so that the mechanism of landslide disaster is converted into implicit knowledge, which is helpful for more accurately identifying and predicting landslide areas. In some embodiments, the remote sensing feature extraction branch includes an autoencoder, S130, the multi-spectral remote sensing image in the preset image format is input to the remote sensing feature extraction branch for feature reconstruction to obtain the first type of features, including:

[0065] The cropped multi-band image is input to the autoencoder, and the autoencoder is used to reconstruct the image, and the reconstruction error is iteratively minimized until the first type of features is obtained.

[0066] The autoencoder AE is a popular deep neural network architecture suitable for unsupervised learning tasks. In the context of landslide identification, AE helps to learn features from landslide images without labeled data. The network can reconstruct the input image, thereby automatically extracting and generating the most representative features of landslides. The parameter settings of the autoencoder AE are shown in Table 1.

[0067]

[0068] In some embodiments, the structure and principle of the autoencoder include:

[0069] :

[0070] wherein, is the encoder function, g is the decoder function, is the input multi-band image, is the output first type of features.

[0071] In some embodiments, S130, the normalized vegetation index NDVI of the preset image format, terrain data, geographic data, geological maps and seismic data are input into an implicit knowledge feature extraction branch to obtain second type features, including:

[0072] A node feature matrix is created based on the normalized vegetation index NDVI, terrain data, geographic data, geological maps and seismic data corresponding to the image; each node of the node feature matrix corresponds to a pixel point, and the node feature corresponds to the pixel value of the pixel point;

[0073] An adjacency matrix of the node feature matrix is constructed using an 8-connected neighborhood;

[0074] Graph convolution operation is performed using the adjacency matrix to obtain the second type features.

[0075] The present embodiment reduces false positives and improves accuracy by converting image data into graph data and constructing implicit knowledge to constrain landslide detection.

[0076] In one example, for each image, it is flattened into a node feature matrix, where each pixel is a node, and its feature is the value of the pixel. The shape of the node feature matrix is (HxW, C). The node feature matrix is specifically:

[0077] ;

[0078] Next, an adjacency matrix is constructed using an 8-connected neighborhood, where each pixel is connected to its 8 adjacent pixels. The adjacency matrix satisfies:

[0079] ;

[0080] Then, the node feature matrix is normalized:

[0081] ;

[0082] where, is the mean value, is the standard deviation of the feature. The normalized feature and the adjacency matrix are converted into a tensor. Finally, a linear transformation and a ReLU activation node feature matrix are applied, and graph convolution operation is performed using the adjacency matrix, and then the process is repeated. The general formula for graph convolution operation using the adjacency matrix is as follows:

[0083] ;

[0084] Through this process, the influencing factors affecting co-seismic landslide occurrence can be converted into graph data, and GNN can be used to maintain the spatial structure of the image and learn implicit knowledge, which helps to accurately identify landslides by combining the output of the remote sensing branch.

[0085] S140 , under the supervision of soft labels, fusing and clustering the first type of features and the second type of features to obtain a clustering result, and determining the coseismic landslide category corresponding to the test area based on the clustering result.

[0086] Soft labeling refers to assigning a probability distribution to each sample in the training data instead of a single hard label, which can reflect the confidence or ambiguity of the sample belonging to different categories.

[0087] In some embodiments, S140, under the supervision of soft labels, fusing and clustering the first category features and the second category features to obtain clustering results includes:

[0088] The divergence KL is used to quantify the difference between the soft label distribution and the target distribution, and the first and second category features are fused and clustered to obtain the clustering results.

[0089] In one example, given a feature map :

[0090] ;

[0091] Given cluster centers:

[0092] ;

[0093] ;

[0094] From this, we can get the soft label distribution of the probability that each feature point belongs to each cluster:

[0095] ;

[0096] in, represents the eigenvector of point i, represents the initial cluster center j (c initial centroids are obtained by minibatch K-means clustering on the original multi-source dataset), Represents the i-th sample point With the jth cluster center The square of the Euclidean distance between is the updated cluster center, is the scaling parameter.

[0097] The target distribution is used to enhance the confidence of cluster assignment and guide features closer to the centroid of their respective clusters. The principle can be expressed as:

[0098] ;

[0099] in, .

[0100] In this embodiment, the Kullback-Leibler (KL) divergence is used to quantify the difference between the soft label distribution and the target distribution. The specific principle is as follows:

[0101] ;

[0102] Since the number of classes in the image is unknown, hierarchical clustering is selected to adaptively determine the number of clusters. At the end of each epoch, the feature map is extracted from the model, and the optimal cluster number is determined by the following formula:

[0103] ;

[0104] Then, based on the obtained optimal cluster number, the cluster center of the model is dynamically updated:

[0105] ;

[0106] wherein, is a link matrix generated by the link function , and represents the second column of , that is, the index of the cluster.

[0107] In some embodiments, S140, under the supervision of the soft label, the first type of feature and the second type of feature are fused and clustered to obtain a clustering result, including:

[0108] After the first type of feature and the second type of feature are fused under the supervision of the soft label, the clustering is completed;

[0109] The optimal number of clusters is determined by hierarchical clustering, and the cluster center is updated based on the optimal number.

[0110] In this embodiment, the clustering result can be optimized in combination with the loss function. Here, the total loss is composed of two parts: image reconstruction loss and Kullback-Leibler supervised adaptive clustering loss.

[0111] ;

[0112] wherein, is a predefined hyperparameter used to balance the relative importance of image reconstruction loss and clustering loss.

[0113] The image reconstruction loss is composed of two parts: the loss of the remote sensing branch and the output loss of the entire reconstruction module, both of which use the Huber loss function. The self-training loss uses the KL divergence loss to measure the divergence between the predicted distribution and the target distribution.

[0114] ;

[0115] ;

[0116] ;

[0117] In one example, the data of the 2021 Mw6.6 Hokkaido earthquake, the 2021 Ms7.2 Haiti earthquake and the 2022 Ms Luding earthquake are used for verification. Compared with methods such as K-means, Mini batch K-means, FCM, FSECSGL and CAE, the coseismic landslide identification method of unsupervised deep clustering provided in the application can obtain the highest OA, mIoU and F1 values.

[0118] Table 2. Accuracy comparison of different methods on three data sets

[0119]

[0120] Please refer to Figures 2-4 , Figures 2-4 are the coseismic landslide identification results in the Hokkaido, Haiti and Luding earthquake regions respectively. Among them, Figure 2 are the coseismic landslide identification results in Hokkaido using six different methods. (a) is the label, and (b)-(g) are the identification results of K-means, Mini batch K-means, FCM, FSECSGL, CAE and the method of the application respectively. Figure 3 are the coseismic landslide identification results in Haiti using six different methods. (a) is the label, and (b)-(g) are the identification results of K-means, Mini batch K-means, FCM, FSECSGL, CAE and the method of the application respectively. Figure 4 are the coseismic landslide identification results in Luding using six different methods. (a) is the label, and (b)-(g) are the identification results of K-means, Mini batch K-means, FCM, FSECSGL, CAE and the method of the application respectively.

[0121] In an optional embodiment, the application provides an unsupervised deep clustering co-seismic landslide identification method, which is implemented by an unsupervised network model. The unsupervised network model is composed of a remote sensing feature extraction branch, an implicit knowledge feature extraction branch, and a Kullback-Leibler supervised adaptive clustering module. The remote sensing feature extraction branch adopts an autoencoder structure with a skip connection to automatically extract and learn key landslide features from remote sensing images, effectively representing co-seismic landslides through image reconstruction. The implicit knowledge feature extraction branch integrates seismic, geological, geographical, and topographical features into graph data, extracts features through a graph convolutional neural network to capture implicit knowledge related to co-seismic landslides. The Kullback-Leibler supervised adaptive clustering module generates a new feature map by synthesizing the features of the two branches. From this feature map, the target distribution and soft labels are calculated, which are then used to supervise the clustering process. By combining a hierarchical clustering method to autonomously determine the optimal number of clusters and update the cluster centers, the dependence on human intervention is minimized, thereby improving the automation level of the model.

[0122] In the embodiment of the application, the multi-source data set is obtained, converted into a preset image format, and then feature reconstruction is performed using the remote sensing feature extraction branch to obtain the first type of features, and the implicit knowledge feature extraction branch to obtain the second type of features. Then, under the supervision of soft labels, the first type of features and the second type of features are fused and clustered to obtain a clustering result, and the co-seismic landslide class corresponding to the to-be-measured area is determined based on the clustering result. By constructing implicit knowledge features, the mechanism of landslide disasters can be implicitly captured and expressed in the model through node features and graph structures, improving the accuracy of identification.

[0123] In some embodiments, referring to Figure 5 , Figure 5 A structural schematic diagram of an unsupervised deep clustering co-seismic landslide identification device is provided in the embodiment of the application. The unsupervised deep clustering co-seismic landslide identification device 500 comprises a data acquisition module 510, a preprocessing module 520, a feature reconstruction module 530, and a fusion and clustering module 540, wherein

[0124] The data acquisition module 510 is configured to acquire a multi-source data set. The multi-source data set comprises multispectral remote sensing images, normalized vegetation index NDVI, topographic data, geographical data, geological maps, and seismic data related to the to-be-measured area.

[0125] The preprocessing module 520 is configured to preprocess the multi-source data set and convert the multi-source data set into a preset image format.

[0126] The feature reconstruction module 530 is configured to input the multispectral remote sensing image in the preset image format to the remote sensing feature extraction branch for feature reconstruction to obtain first-class features, and input normalized vegetation index NDVI, terrain data, geographical data, geological maps and seismic data in the preset image format to the implicit knowledge feature extraction branch to obtain second-class features; the second-class features are implicit knowledge.

[0127] The fusion and clustering module 540 is configured to fuse and cluster the first-class features and the second-class features under soft label supervision to obtain a clustering result, and determine a co-seismic landslide category corresponding to the to-be-measured area based on the clustering result.

[0128] In some embodiments, the preprocessing module 520 is specifically configured to:

[0129] convert the multi-source data set into an image format and perform superimposition processing to obtain a multi-band image;

[0130] crop the multi-band image into a block with a size of 128x128.

[0131] In some embodiments, the remote sensing feature extraction branch includes a self-encoder, and the feature reconstruction module 530 is specifically configured to:

[0132] input the cropped multi-band image into the self-encoder, reconstruct the image by using the self-encoder, and continuously iterate to minimize the reconstruction error until the first-class features are obtained.

[0133] In some embodiments, the structure principle of the self-encoder includes:

[0134]

[0135] wherein, is an encoder function, g is a decoder function, is the input multi-band image, is the output first-class features.

[0136] In some embodiments, the feature reconstruction module 530 is specifically configured to:

[0137] create a node feature matrix based on images corresponding to the normalized vegetation index NDVI, the terrain data, the geographical data, the geological maps and the seismic data; each node of the node feature matrix corresponds to a pixel point, and a node feature corresponds to a pixel value of the pixel point;

[0138] construct an adjacency matrix of the node feature matrix by using an 8-connected neighborhood;

[0139] perform graph convolution operation by using the adjacency matrix to obtain the second-class features.

[0140] ​In some embodiments, the fusion clustering module 540 is specifically configured to:

[0141] The divergence KL is used to quantify the difference between the soft label distribution and the target distribution, and the first and second category features are fused and clustered to obtain the clustering results;

[0142] Given a feature map :

[0143] ;

[0144] Given cluster centers:

[0145] ;

[0146] ;

[0147] This gives the soft label distribution of the probability that each feature point belongs to each cluster:

[0148] ;

[0149] in, represents the eigenvector of point i, represents the initial cluster center j, Represents the i-th sample point With the jth cluster center The square of the Euclidean distance between is the updated cluster center, is the scaling parameter.

[0150] In some embodiments, the fusion clustering module 540 is specifically configured to:

[0151] Under the supervision of soft labels, clustering is completed by fusing the first category features and the second category features;

[0152] The optimal number of clusters is determined through hierarchical clustering, and the cluster centers are updated based on the optimal number.

[0153] The unsupervised deep clustering co-seismic landslide identification device provided in the embodiment of the present application can implement each process in the embodiment corresponding to the above-mentioned unsupervised deep clustering co-seismic landslide identification method. To avoid repetition, they are not described here.

[0154] It should be noted that the unsupervised deep clustering co-seismic landslide identification device provided in the embodiment of the present application and the unsupervised deep clustering co-seismic landslide identification method provided in the embodiment of the present application are based on the same application concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned unsupervised deep clustering co-seismic landslide identification method, and the repeated parts will not be repeated.

[0155] In some embodiments, refer to Figure 6 , Figure 6 A structural schematic diagram of an electronic device is provided in some embodiments of the present application. An electronic device 600 provided in an embodiment of the present application includes a processor 610 and a memory 620; the memory 620 has a computer program stored therein, wherein the computer program, when executed by the processor, implements the above-mentioned method for identifying an aseismic landslide by unsupervised deep clustering.

[0156] Specifically, the processor 610 may, for example, include a general-purpose microprocessor, an instruction set processor, and / or a related chipset and / or a special-purpose microprocessor (such as an application-specific integrated circuit (ASIC)), etc. The processor 610 can also include on-board memory for cache use. The processor 610 can be a single processing unit or a plurality of processing units for performing different actions of the method flow according to an embodiment of the present application.

[0157] The memory 620 may, for example, be any medium capable of containing, storing, communicating, propagating or transmitting instructions. For example, the memory 620 may include, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared or semiconductor system, device, apparatus or propagation medium. Specific examples of the memory 620 include a magnetic storage device such as a magnetic tape or a hard disk drive (HDD); an optical storage device such as a compact disc (CD-ROM); also a random access memory (RAM) or a flash memory; and / or a wired / wireless communication link.

[0158] The present application also provides a computer readable medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned method for identifying an aseismic landslide by unsupervised deep clustering. The computer readable medium can be included in the device / apparatus / system described in the above embodiments; or can exist separately and not be assembled into the device / apparatus / system. The above computer readable medium carries one or more programs, which, when executed, implement the method according to an embodiment of the present application.

[0159] According to the embodiments of the present application, the computer readable medium can be a computer readable signal medium or a computer readable storage medium or any combination thereof. The computer readable storage medium can be, for example but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or any suitable combination thereof. More specific examples of the computer readable storage medium can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this application, the computer readable storage medium can be any tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus or device. In this application, the computer readable signal medium can include a computer readable program code transmitted in baseband or as part of a carrier wave in a propagated signal, in which the computer readable program code contains a program for use by or in connection with an instruction execution system, apparatus or device. The propagated signal can take any suitable form, including but not limited to an electromagnetic signal, an optical signal or any suitable combination thereof. The computer readable signal medium can also be any computer readable medium that can be used to carry or store a program for use by or in connection with an instruction execution system, apparatus or device, except for the computer readable storage media described above. The program code contained in the computer readable media can be transmitted in any suitable form, including but not limited to, radio frequency, electrical, optical, acoustic or any suitable combination thereof.

[0160] Those skilled in the art will appreciate that features recited in the various embodiments and / or claims of this application can be combined and / or interchanged, even if this is not explicitly stated in the application. In particular, the features of the various embodiments and / or claims of this application can be combined and / or interchanged, without departing from the spirit and teachings of this application. All combinations and / or interchanges of the features of the various embodiments and / or claims of this application are expressly envisioned and are within the scope of this application. Accordingly, the scope of the application should not be limited to the above-described embodiments, but should be determined by the appended claims and their equivalents.

Claims

1. A co-seismic landslide identification method based on unsupervised deep clustering, characterized in that: include: Acquiring a multi-source data set; the multi-source data set includes multispectral remote sensing images, normalized difference vegetation index (NDVI), topographic data, geographic data, geological maps, and seismic data related to the area to be measured; Preprocessing the multi-source dataset to convert the multi-source dataset into a preset image format; Inputting the multispectral remote sensing image in a preset image format into a remote sensing feature extraction branch for feature reconstruction to obtain first-category features, and inputting the Normalized Difference Vegetation Index (NDVI), the terrain data, the geographic data, the geological map, and the seismic data in a preset image format into an implicit knowledge feature extraction branch to obtain second-category features; the second-category features are implicit knowledge; The normalized difference vegetation index (NDVI), the terrain data, the geographic data, the geological map, and the seismic data are input into the implicit knowledge feature extraction branch to obtain the second type of features, including: Creating a node feature matrix based on the normalized vegetation index NDVI, the terrain data, the geographic data, the geological map, and the image corresponding to the seismic data; each node of the node feature matrix corresponds to a pixel point, and the node feature corresponds to the pixel value of the pixel point; Constructing an adjacency matrix of the node feature matrix using an 8-connected neighborhood; Performing graph convolution operation using the adjacency matrix to obtain the second type of features; Under the supervision of soft labels, the first type of features and the second type of features are fused and clustered to obtain a clustering result, and the co-seismic landslide category corresponding to the area to be measured is determined based on the clustering result.

2. The unsupervised deep clustering coseismic landslide identification method according to claim 1, characterized in that: The preprocessing of the multi-source dataset to convert the multi-source dataset into a preset image format includes: Converting the multi-source data sets into an image format and performing overlay processing to obtain a multi-band image; The multi-band image is cropped into blocks of 128×128 size.

3. The unsupervised deep clustering coseismic landslide identification method according to claim 2, characterized in that: The remote sensing feature extraction branch includes an autoencoder, and the multispectral remote sensing image is input into the remote sensing feature extraction branch for feature reconstruction to obtain a first type of feature, including: The cropped multi-band image is input into the autoencoder, and the image is reconstructed using the autoencoder / decoder, and the reconstruction error is continuously minimized iteratively until the first type of features are obtained.

4. The unsupervised deep clustering coseismic landslide identification method according to claim 3, characterized in that: The structural principles of the autoencoder include: f:χ→κ Where f is the encoder function, g is the decoder function, X is the input multi-band image, is the first type of feature output.

5. The unsupervised deep clustering coseismic landslide identification method according to claim 1, characterized in that: The step of fusing and clustering the first category features and the second category features under soft label supervision to obtain a clustering result includes: The difference between the soft label distribution and the target distribution is quantified using the divergence KL, and the first type of features and the second type of features are fused and clustered to obtain a clustering result; Given a feature map Z: Given cluster centers: This gives the soft label distribution of the probability that each feature point belongs to each cluster: Among them, Z i represents the eigenvector of point i, μ j represents the initial cluster center j, ‖Z i -μ j ‖ 2 Represents the i-th sample point x i and the jth cluster center μ j The square of the Euclidean distance between is the updated cluster center, and v is the scaling parameter.

6. The unsupervised deep clustering coseismic landslide identification method according to claim 1, characterized in that: The step of fusing and clustering the first category features and the second category features under soft label supervision to obtain a clustering result includes: Under the supervision of soft labels, clustering is completed after fusing the first category features and the second category features; The optimal number of clusters is determined through hierarchical clustering, and the cluster centers are updated based on the optimal number.

7. An unsupervised deep clustering coseismic landslide identification device, characterized in that: include: Data acquisition module, preprocessing module, feature reconstruction module and fusion clustering module, among which, The data acquisition module is configured to acquire a multi-source data set; the multi-source data set includes multispectral remote sensing images, normalized difference vegetation index (NDVI), topographic data, geographic data, geological maps, and seismic data related to the area to be measured; The preprocessing module is configured to preprocess the multi-source dataset and convert the multi-source dataset into a preset image format; The feature reconstruction module is configured to input the multispectral remote sensing image in a preset image format into a remote sensing feature extraction branch for feature reconstruction to obtain first-category features, and input the Normalized Difference Vegetation Index (NDVI) in a preset image format, the terrain data, the geographic data, the geological map, and the seismic data into an implicit knowledge feature extraction branch to obtain second-category features; the second-category features are implicit knowledge; The normalized difference vegetation index (NDVI), the terrain data, the geographic data, the geological map, and the seismic data are input into the implicit knowledge feature extraction branch to obtain the second type of features, including: Creating a node feature matrix based on the normalized vegetation index NDVI, the terrain data, the geographic data, the geological map, and the image corresponding to the seismic data; each node of the node feature matrix corresponds to a pixel point, and the node feature corresponds to the pixel value of the pixel point; Constructing an adjacency matrix of the node feature matrix using an 8-connected neighborhood; Performing graph convolution operation using the adjacency matrix to obtain the second type of features; The fusion clustering module is configured to fuse and cluster the first type of features and the second type of features under soft label supervision to obtain a clustering result, and determine the coseismic landslide category corresponding to the test area based on the clustering result.

8. An electronic device comprising a processor and a memory; the memory stores a computer program, wherein: When the computer program is executed by the processor, the computer program implements the unsupervised deep clustering co-seismic landslide identification method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images