A Landslide Area Detection Method Based on Dual-Channel Image Convolution

Abstract

This invention provides a landslide area detection method based on dual-channel graph convolution, applicable to landslide area detection in remote sensing images. The specific steps are as follows: First, a fully convolutional network is used to extract deep-level features from the landslide area image in the remote sensing image; triple information of the graph structure is constructed from the deep-level features, including nodes, edges, and distances between nodes; adjacency matrix, degree matrix, feature matrix, and weight matrix are obtained; the adjacency matrix, degree matrix, feature matrix, and weight matrix are input into a dual-channel graph convolution to obtain new feature information; the new feature information is upsampled to obtain upsampled feature information; the upsampled feature information is mapped through dimensionality reduction and softmax function operations to output the final landslide area detection result in the remote sensing image. This invention can effectively improve the automation and intelligence level of landslide area detection in remote sensing images and can meet the actual on-orbit requirements of satellites for detecting landslide areas in remote sensing images.

Description

Technical Field

[0001] This invention relates to the detection of landslide areas based on remote sensing images, specifically an automatic detection method for landslide areas in remote sensing images based on dual-channel image convolution. Background Technology

[0002] A landslide is a geological phenomenon in which rocks and soil move when they can no longer withstand their own weight and external forces. Landslides are one of the major geological disasters in my country, causing significant impacts on people's lives and livelihoods. When a landslide occurs, it can lead to traffic disruptions, river flow interruptions, reduced agricultural production, and threats to life and property. Therefore, landslide detection is of great value. Landslide detection involves determining the geographical location and shape of a landslide. Through landslide detection, quantitative statistical information can be generated, and a database can be established, providing an important means for the prevention and control of landslide disasters. Simultaneously, analyzing the spatial relationship between landslides and their influencing factors can create landslide sensitivity maps, providing a reference for landslide prevention.

[0003] Current landslide detection methods are generally classified into three categories: field detection, machine learning, and deep learning. Field detection requires extensive manual investigation and mapping of the landslide-affected area, followed by data archiving. This approach is time-consuming, labor-intensive, incomplete, and heavily influenced by human factors and geographical environment, making it difficult to conduct rapid and large-scale accurate detection and assessment of affected areas in real time. Machine learning technology uses influencing factors to predict landslides; however, these factors vary across different regions, data collection is challenging, and machine learning accuracy is not high.

[0004] With the development of optical remote sensing technology and the advancement of various platforms and equipment such as aerospace, high-resolution, multispectral remote sensing images have provided ample and reliable data for mapping large-area landslide areas, offering new technical means for landslide identification and detection, and rapid post-disaster surveys. Accurate extraction of landslide information based on remote sensing imagery and topographic data such as DEMs, seeking reliable data from Gaofen, Sentinel, and Landsat series satellites, and utilizing deep learning technology for comprehensive analysis to accurately extract landslide extents have become the main methods and means. Currently, using convolutional neural networks to identify landslides and map landslide distribution using optical imagery and topographic features yields excellent results. High-resolution optical remote sensing images can reflect the optical characteristics of landslides; therefore, much current research and practical work utilizes high-resolution remote sensing images to train models for landslide detection and identification. However, when using fully convolutional networks to extract features from medical images, the limited receptive field can lead to the loss of local information about diseased tissues, resulting in reduced segmentation accuracy, excessive computational complexity, and slow processing speed. Summary of the Invention

[0005] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide a landslide area detection method based on dual-channel graph convolution.

[0006] The technical solution of the present invention is:

[0007] This invention discloses a landslide area detection method based on dual-channel graph convolution, comprising the following steps:

[0008] S1: Use a fully convolutional network to extract deep features from remote sensing images of landslide areas;

[0009] S2: Use graph construction methods to construct the nodes, edges and distances between nodes of the deep features to obtain the adjacency matrix, degree matrix, feature matrix and weight matrix;

[0010] S3: Input the adjacency matrix, degree matrix, feature matrix and weight matrix into a two-channel graph convolution to obtain new feature information;

[0011] S4: Upsample the new feature information to obtain the upsampled feature information;

[0012] S5: The upsampled feature information is mapped through dimensionality reduction calculation and normalization processing to output the final remote sensing image landslide area detection result.

[0013] Furthermore, in the above detection method, in step S1, a fully convolutional network is used to extract deep-level features from the remote sensing image of the landslide area, specifically as follows:

[0014] Construct a fully convolutional neural network; the fully convolutional neural network includes several superimposed downsampling modules; the downsampling module includes a 1×1 convolution operation unit and three parallel 1×1, 3×3, and 5×5 convolution operation units;

[0015] The fully convolutional neural network is used to compress the feature information of the input landslide remote sensing image to obtain high-dimensional deep features R. H×W×C .

[0016] Furthermore, in the above detection method, in step S2, the triplet information of the graph structure constructed from the deep features, nodes, edges, and distances between nodes, are obtained; specifically, the adjacency matrix, degree matrix, feature matrix, and weight matrix are obtained:

[0017] S1. Using a graph construction method, construct a feature graph G = (V, E) for the deep features; where... Represents the association of nodes. Let E represent the set of edges of the graph, where e∈E is the M edge of the undirected graph, and r, s∈E represent the index values ​​of the vertices at the receiving end and the input end, respectively.

[0018] S2. Treat each point on the feature map as a graph node, and randomly initialize the weights of the graph nodes. Constructing the weight matrix

[0019] S3, Calculate v for every two graph nodes. i and v j Euclidean distance;

[0020] S4. Construct the adjacency matrix for each node based on the Euclidean distance;

[0021] S5. For the l-th layer, construct the degree matrix. elementa ij For the l-th layer node v i and v j The weights between them are given by the formula:

[0022]

[0023] Where d(*) represents node v i and v j The Euclidean distance function between them; exp represents the exponential function with base e; ∑ represents the summation function; v i For the i-th graph node, v j Let j be the j-th graph node;

[0024] S6. Construct a symmetric normalized adjacency matrix using the adjacency matrix and degree matrix;

[0025] S7. Construct the feature matrix based on the symmetric normalized adjacency matrix and the weight matrix, specifically as follows:

[0026]

[0027] Where σ(*) is a nonlinear activation function; The value is randomly initialized. The symmetric normalized adjacency matrix representing the l-th layer;

[0028] S8. For the weight matrix W (l) The matrix is ​​updated to obtain a new weight matrix.

[0029] S9. Update the feature matrix based on the symmetric normalized adjoint matrix and the new weight matrix, specifically as follows:

[0030]

[0031] in, Let be the feature matrix.

[0032] Furthermore, in the above detection method, the adjacency matrix of each node is specifically:

[0033]

[0034]

[0035] Where l represents the number of layers; This represents the value of the adjacency matrix in the i-th row and j-th column at level l; v i v j ∈E, representing node v i v j There are edges between them; R represents an N×N matrix.

[0036] Furthermore, in the above detection method, the symmetric normalized adjacency matrix is ​​defined by the following formula:

[0037]

[0038]

[0039]

[0040]

[0041] in, A represents the symmetric normalized adjacency matrix of the l-th layer; (l) D represents the adjacency matrix of the l-th layer; (l) The degree matrix representing the l-th layer; a ij For the l-th layer node v i and v j The weights between nodes; d(*) represents node v i and v j The Euclidean distance function between them; exp represents the exponential function with base e; ∑ represents the summation function; v i For the i-th graph node, v j Let j be the j-th graph node.

[0042] Furthermore, in the above detection method, in step S3, the adjacency matrix, degree matrix, feature matrix, and weight matrix are input into a two-channel graph convolution to obtain new feature information, specifically as follows:

[0043]

[0044]

[0045]

[0046] Among them, X OUT Representing new feature layer information, X C，DRP-GCN For the feature results of the convolutional branch of the depth residual simplified graph, concat represents the channel fusion operation; The feature result of the convolution branch of the parallel channel map; the v(*) function is used to reconstruct the size information; The feature information of the l-th layer graph convolution in the parallel channel graph convolution branch; The feature information of the l-th layer graph convolution in the simplified graph convolution branch of the depth residual is given.

[0047] Furthermore, in the above detection method, the feature information of the l-th layer graph convolution in the parallel channel graph convolution branch and Specifically:

[0048]

[0049]

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056] in, a dd represents the addition operation of the feature matrices of nodes in layer l of the parallel channel graph convolution branch; The feature matrix representing node i in the l-th layer of the parallel channel graph convolutional branch; The normalized adjacency matrix representing the l-th layer of the parallel channel graph convolutional branch; This represents the new weight matrix for the l-th layer of the parallel channel graph convolution branch; The degree matrix representing the l-th layer of the parallel channel graph convolution branch; This represents the adjacency matrix of the l-th layer in the parallel channel graph convolution branch; This represents the value of the adjacency matrix in the i-th row and j-th column of the l-th layer in the parallel channel graph convolution branch; element For the convolution branch of the parallel channel graph, the node of the l-th layer and The weights between them; Graph nodes representing parallel channel graph convolution branches and The Euclidean distance function between them.

[0057] Furthermore, in the above detection method, the graph nodes of the parallel channel graph convolution branch... and Specifically, the deep features are convolved using a filter of size K, with the internal channels set to C / 2, thus transforming the deep features into... Then, graph construction methods are used to construct feature graphs for deeper features, resulting in nodes of the graph structure.

[0058] Furthermore, in the above detection method, the feature information of the l-th layer graph convolution in the depth residual simplified graph convolution branch and Specifically:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067] Where, add represents the addition operation of the feature matrices of the nodes in the l-th layer of the deep residual simplified graph convolution branch; The feature matrix representing the l-th layer node i of the convolutional branch of the depth residual simplified graph; The normalized adjacency matrix representing the l-th layer of the convolutional branch of the depth residual simplified graph; The new weight matrix represents the l-th layer of the convolutional branch of the depth residual simplified graph; The degree matrix representing the l-th layer of the convolutional branch of the depth residual simplified graph; The adjacency matrix of the l-th layer of the convolutional branch of the depth residual simplified graph; In the convolutional branch representing the depth residual simplified graph, the value of the adjacency matrix in the i-th row and j-th column of the l-th layer; element To simplify the graph convolution branch for depth residuals, the node v in the l-th layer... Ci and v Cj The weights between them; Representative depth residual simplified graph convolution branch node v Ci and vCj The Euclidean distance function between them; v Ci v Cj The graph nodes representing the convolutional branches of the deep residual simplified graph are obtained from deep features using graph construction methods.

[0068] Furthermore, in the above detection method, in step S4, the new feature information is upsampled to obtain upsampled feature information, specifically as follows:

[0069] The new pair of feature information is upsampled using several cascaded upsampling modules to obtain upsampled feature information;

[0070] The upsampling module includes a 2×2 convolutional unit and a parallel unit. The output of the 2×2 convolutional unit serves as the input of the parallel unit. The parallel unit includes three parallel 1×1 convolutional units, a 3×3 convolutional unit, and a 5×5 convolutional unit with a stride of 1.

[0071] Furthermore, in the above detection method, in step S5, the upsampled feature information is mapped by dimensionality reduction calculation and normalization processing to output the final remote sensing image landslide area detection result. Specifically, the upsampled feature information is subjected to dimensionality reduction calculation using a 1x1 convolutional layer to obtain the dimensionality-reduced feature information; the dimensionality-reduced feature information is subjected to a softmax function operation to map the feature information of each pixel onto a probability distribution to obtain the final remote sensing image landslide area detection result.

[0072] The advantages of this invention compared to the prior art are:

[0073] (1) This invention proposes an architecture that combines CNN and GCN, where CNN is used to extract high-order features and GCN converts the features into a graph structure for learning and propagation to improve detection performance.

[0074] (2) This invention designs an improved residual convolutional block for the backbone network, extracts rich spatial background information through feature compression operation, and extracts high-order features for semantic initialization of graph nodes.

[0075] (3) This invention proposes a mechanism to interactively supplement information between two different channels, using five layers of optimized graph convolution to learn and propagate graph features, and using graph channel convolution to refine and extract channel features, thereby improving the final detection results of the proposed model. Attached image description:

[0076] Figure 1 This is a diagram illustrating the overall architecture of the method of the present invention;

[0077] Figure 2 This is a diagram of the downsampling module of the present invention;

[0078] Figure 3 This is a diagram of the upsampling module of the present invention;

[0079] Figure 4 This is a simplified convolutional branch graph of the depth residual connection in this invention;

[0080] Figure 5 This is the lightweight bibranch graph convolution graph of the present invention;

[0081] Figure 6 (a) is the original satellite image; (b) is the visualization result of manually annotated landslide labels; (c) is the visualization result of landslide areas in remote sensing images based on a lightweight bi-branch graph convolutional network model that maximizes node connections globally, where yellow represents landslide areas. Detailed Implementation

[0082] The working principle and process of the present invention will be further explained and described below with reference to the accompanying drawings.

[0083] This embodiment provides a landslide area detection method based on dual-channel graph convolution, including the following steps:

[0084] S1: Use a fully convolutional network to extract deep features from remote sensing images of landslide areas;

[0085] S2: Use graph construction methods to construct the nodes, edges and distances between nodes of the graph structure for deep features, and obtain the adjacency matrix, degree matrix, feature matrix and weight matrix;

[0086] S3: Input the adjacency matrix, degree matrix, feature matrix and weight matrix into the two-channel graph convolution to obtain new feature information;

[0087] S4: Upsample the new feature information to obtain the upsampled feature information;

[0088] S5: The upsampled feature information is mapped through dimensionality reduction calculation and normalization processing to output the final remote sensing image landslide area detection results.

[0089] Preferably, in step S1, a fully convolutional network is used to extract deep-level features from the remote sensing image of the landslide area, specifically as follows:

[0090] Construct a fully convolutional neural network; the fully convolutional neural network includes several superimposed downsampling modules; the downsampling module includes a 1×1 convolution operation unit and three parallel 1×1, 3×3, and 5×5 convolution operation units;

[0091] A fully convolutional neural network is used to compress the feature information of the input landslide remote sensing image to obtain high-dimensional deep features R. H×W×C.

[0092] Preferably, in step S2, the triplet information of the graph structure is constructed from the deep features, including nodes, edges, and distances between nodes; the adjacency matrix, degree matrix, feature matrix, and weight matrix are obtained, specifically as follows:

[0093] S1. Use graph construction methods to construct a feature map G = (V, E) for deep-level features; where... Represents the association of nodes. Let E represent the set of edges of the graph, where e∈E is the M edge of the undirected graph, and r, s∈E represent the index values ​​of the vertices at the receiving end and the input end, respectively.

[0094] S2. Treat each point on the feature map as a graph node and randomly initialize the weights of the graph nodes. Constructing the weight matrix

[0095] S3, Calculate v for every two graph nodes. i and v j Euclidean distance;

[0096] S4. Construct the adjacency matrix for each node based on the Euclidean distance;

[0097] S5. For the l-th layer, construct the degree matrix. elementa ij For the l-th layer node v i and v j The weights between them are given by the formula:

[0098]

[0099] Where d(*) represents node v i and v j The Euclidean distance function between them; exp represents the exponential function with base e; ∑ represents the summation function; v i For the i-th graph node, v j Let j be the j-th graph node;

[0100] S6. Construct a symmetric normalized adjacency matrix using the adjacency matrix and degree matrix;

[0101] S7. Construct the feature matrix based on the symmetric normalized adjacency matrix and the weight matrix, specifically as follows:

[0102]

[0103] Where σ(*) is a nonlinear activation function; The value is randomly initialized. The symmetric normalized adjacency matrix representing the l-th layer;

[0104] S8. Regarding the weight matrix W (l) The matrix is ​​updated to obtain a new weight matrix.

[0105] S9. Update the feature matrix based on the symmetric normalized adjoint matrix and the new weight matrix, specifically as follows:

[0106]

[0107] in, This is the characteristic matrix.

[0108] Preferably, the adjacency matrix of each node is as follows:

[0109]

[0110]

[0111] Where l represents the number of layers; This represents the value of the adjacency matrix in the i-th row and j-th column at level l; v i vj∈E, representing node v i v j There are edges between them; R represents an N×N matrix.

[0112] Preferably, the symmetric normalized adjacency matrix is ​​formulated as follows:

[0113]

[0114]

[0115]

[0116]

[0117] in, A represents the symmetric normalized adjacency matrix of the l-th layer; (l) D represents the adjacency matrix of the l-th layer; (l) The degree matrix representing the l-th layer; a ij For the l-th layer node v i and v j The weights between nodes; d(*) represents node v i and v j The Euclidean distance function between them; exp represents the exponential function with base e; ∑ represents the summation function; v i For the i-th graph node, v j Let j be the j-th graph node.

[0118] Preferably, in step S3, the adjacency matrix, degree matrix, feature matrix, and weight matrix are input into a two-channel graph convolution to obtain new feature information, specifically as follows:

[0119]

[0120]

[0121]

[0122] Among them, X OUT Representing new feature layer information, X C，DRP-GCN For the feature results of the convolutional branch of the depth residual simplified graph, concat represents the channel fusion operation; The feature result of the convolution branch of the parallel channel map; the v(*) function is used to reconstruct the size information; The feature information of the l-th layer graph convolution in the parallel channel graph convolution branch; The feature information of the l-th layer graph convolution in the simplified graph convolution branch of the depth residual is given.

[0123] Preferably, the feature information of the l-th layer graph convolution in the parallel channel graph convolution branch and Specifically:

[0124]

[0125]

[0126]

[0127]

[0128]

[0129]

[0130]

[0131]

[0132] Where, add represents the operation of adding the feature matrices of nodes in the l-th layer of the parallel channel graph convolution branch; The feature matrix representing node i in the l-th layer of the parallel channel graph convolutional branch; The normalized adjacency matrix representing the l-th layer of the parallel channel graph convolutional branch; This represents the new weight matrix for the l-th layer of the parallel channel graph convolution branch; The degree matrix representing the l-th layer of the parallel channel graph convolution branch; This represents the adjacency matrix of the l-th layer in the parallel channel graph convolution branch; This represents the value of the adjacency matrix in the i-th row and j-th column of the l-th layer in the parallel channel graph convolution branch; element For the convolution branch of the parallel channel graph, the node of the l-th layer and The weights between them; Graph nodes representing parallel channel graph convolution branches and The Euclidean distance function between them.

[0133] Preferably, the graph nodes of the parallel channel graph convolution branch and Specifically, the deep features are convolved using a filter of size K, with the internal channels set to C / 2, thus transforming the deep features into... Then, graph construction methods are used to construct feature graphs for deeper features, resulting in nodes of the graph structure.

[0134] Preferably, the feature information of the l-th layer graph convolution in the depth residual simplified graph convolution branch and Specifically:

[0135]

[0136]

[0137]

[0138]

[0139]

[0140]

[0141]

[0142]

[0143] Where, add represents the addition operation of the feature matrices of the nodes in the l-th layer of the deep residual simplified graph convolution branch; The feature matrix representing the l-th layer node i of the convolutional branch of the depth residual simplified graph; The normalized adjacency matrix representing the l-th layer of the convolutional branch of the depth residual simplified graph; The new weight matrix represents the l-th layer of the convolutional branch of the depth residual simplified graph; The degree matrix representing the l-th layer of the convolutional branch of the depth residual simplified graph; The adjacency matrix of the l-th layer of the convolutional branch of the depth residual simplified graph; In the convolutional branch representing the depth residual simplified graph, the value of the adjacency matrix in the i-th row and j-th column of the l-th layer; element To simplify the graph convolution branch for depth residuals, the node v in the l-th layer... Ci and v Cj The weights between them; Representative depth residual simplified graph convolution branch node v Ci and v Cj The Euclidean distance function between them; v Ci v Cj The graph nodes representing the convolutional branches of the deep residual simplified graph are obtained from deep features using graph construction methods.

[0144] Preferably, in step S4, the new feature information is upsampled to obtain upsampled feature information, specifically as follows:

[0145] Several cascaded upsampling modules are used to upsample the new pair of feature information to obtain the upsampled feature information;

[0146] The upsampling module includes 2×2 convolutional units and parallel units. The output of the 2×2 convolutional units serves as the input of the parallel units. The parallel units include three parallel 1×1 convolutional units, 3×3 convolutional units, and 5×5 convolutional units with a stride of 1.

[0147] Preferably, in step S5, the upsampled feature information is mapped through dimensionality reduction and softmax function operations to output the final remote sensing image landslide area detection result, specifically:

[0148] The upsampled feature information is reduced in dimensionality using a 1x1 convolutional layer to obtain the dimensionality-reduced feature information.

[0149] The reduced feature information is normalized and input into the softmax function to map the feature information of each pixel onto the probability distribution, thus obtaining the final landslide area detection result from the remote sensing image.

[0150] Example

[0151] To better understand the solution of this invention, the dataset used in this specific embodiment is the Bijie City landslide dataset and the Sichuan Province landslide dataset, totaling 829 images. The batch size is set to 2, and the input image size is 256*256*3. Figure 1 As shown in (a). The deep learning framework used in this embodiment is PyTorch 1.2.0, and the graphics card used is NVIDIA GeForce GTX 1080 Ti.

[0152] Step 1: Using, for example Figure 1 (b) shows a fully convolutional network that extracts deep features from remote sensing images of landslide areas;

[0153] Step 2: Construct triple information of the graph structure from deep features, including nodes, edges, and distances between nodes; obtain the adjacency matrix, degree matrix, feature matrix, and weight matrix;

[0154] Step 3: Input the adjacency matrix, degree matrix, feature matrix, and weight matrix as follows: Figure 1 (c) shows the convolution of the two-channel image to obtain new feature information;

[0155] Step 4: Process the new feature information as follows Figure 1 The upsampling operation shown in (d) yields the upsampled feature information;

[0156] Step 5: Map the upsampled feature information using dimensionality reduction and softmax functions, such as... Figure 1 As shown in (e), the final remote sensing image of the landslide area detection results is output.

[0157] Based on step 1, the deep features of the landslide area image in the remote sensing image are extracted using a fully convolutional network. The specific steps are as follows:

[0158] (1) First, downsampling modules are constructed as components of the fully convolutional neural network. Each downsampling module is as follows: Figure 2 As shown, the input information is first batch normalized to reduce internal covariate bias; then a 1×1 convolution operation is performed; subsequently, three parallel 1×1, 3×3, and 5×5 convolution operations with a stride of 1 are used to enrich the semantic features of the image; then another 1×1 convolution operation is performed; finally, three parallel 1×1, 3×3, and 5×5 convolution operations with a stride of 2 are used to compress the feature map to obtain high-dimensional features.

[0159] (2) The entire fully convolutional network consists of five superimposed downsampling modules, which are used to compress the feature information of the input landslide remote sensing image to obtain deep high-dimensional features.

[0160] Based on the triplet information of the graph structure constructed from deep features in step 2, including the nodes, edges, and distances between nodes, the adjacency matrix, degree matrix, feature matrix, and weight matrix are obtained. The specific steps are as follows:

[0161] (1) Using a graph construction method, a graph structure G = (V, E) is established for the deep features extracted by the fully convolutional network; where, Represents the association of nodes. Let E represent the set of edges of the graph, where e∈E is the M edge of the undirected graph, and r, s∈E represent the index values ​​of the vertices at the receiving end and the input end, respectively.

[0162] (2) Treat each point on the feature map as a graph node and randomly initialize the weights of the graph nodes. Constructing the weight matrix For every two nodes v i and v j Calculate the Euclidean distance between them; for each vertex, select the three nearest neighbor nodes to construct an adjacency matrix. Furthermore, for the l-th layer, the degree matrix It is a diagonal matrix. elementa ij Defined as the node v of the l-th layer i and v j The weights between them are calculated using the following formula:

[0163]

[0164] Where d(*) represents node v i and v j The Euclidean distance function between them; exp represents the exponential function with base e; ∑ represents the summation function; v i v j Represents a graph node.

[0165] (3) Construct a symmetric normalized adjacency matrix using the adjacency matrix and degree matrix. The calculation formula is as follows:

[0166]

[0167] (4) The feature matrix of a graph node consists of the symmetric normalization of the adjacency matrix, the feature transformation of the weight matrix, and the nonlinear activation function. The calculation formula is as follows:

[0168]

[0169] Where σ(*) is a nonlinear activation function, value for random initialization

[0170] According to step 3, the adjacency matrix, degree matrix, feature matrix, and weight matrix are input as follows: Figure 5 The dual-channel image convolution shown below yields new feature information. The specific steps are as follows:

[0171] (1) First, a global maximization node connection algorithm was designed to improve graph convolution by using positive and negative connections in order to simplify the parameters.

[0172] (2) Specifically, for a node from layer 1 to layer l, its three neighboring neighborhoods H (l) In, if a connection exists This relation is called a positive connection of nodes, where ij∈H (l) Represents the neighborhood H of layer l (l)Any two nodes in the symbol This means that the values ​​passed from the current layer to the next layer are equal. The equation above shows that for a given node, when the connections of surrounding edges can be passed layer by layer, a positive connection can effectively activate the incoming data to a node in that layer; conversely, for a node from layer l to layer 1, if a connection exists... This relationship is called a negative connection of a node, which can effectively activate the outgoing flow of a certain node in the layer. If both the positive and negative connections of the i-th node in the l-th layer exist, it means that the incoming and outgoing flows of this node can be effectively activated, which means that it has a spatial global connection.

[0173] (3) Spatial global connectivity means that the topological space of a graph can remain unchanged after its shape changes. Nodes and their connections without this property have no meaning and can be deleted without affecting the entire graph.

[0174] (4) Update the new weight set of the graph convolutional layer by globally maximizing node connections, and then apply the weight matrix W. (l) Updated to The feature update process based on simplified graph convolution with globally maximized node connectivity is calculated using the following formula:

[0175]

[0176] in It is a symmetric normalized adjacency matrix. This is the updated weight matrix. It is the characteristic matrix.

[0177] (5) The feature information of the convolution in this layer is fused using the add(*) function, and the calculation formula is as follows:

[0178]

[0179] (6) Applying residual connections to simplified graph convolutions based on globally maximizing node connections alleviates the problem of over-smoothing and ensures that the network has more stable gradients. The calculation formula is as follows:

[0180]

[0181] (7) Design a... Figure 4 The deep residual connection shown simplifies the graph convolutional branch, increases the number of layers for graph structure feature learning, and improves feature aggregation capability. The calculation formula is as follows:

[0182]

[0183] The feature result of the convolution branch of the depth residual simplified graph is X. C,DRP-GCNconcat represents the channel merging operation.

[0184] (8) Convolve the deep features extracted in step 1 using a fully convolutional network with a filter of size K, setting the internal channels to C / 2, and input features. Channel features after convolution And construct the adjacency matrix, degree matrix, feature matrix and weight matrix according to step 2; then calculate the channel features according to steps 3(1)-(7), the calculation formula is as follows:

[0185]

[0186]

[0187] Among them, subscript c The variables and matrices represented have the same meanings as those in the convolution branch of the depth residual simplified graph, with different channels distinguished only by subscripts.

[0188] (9) Design a parallel channel graph convolution branch, which is connected in parallel with the depth residual connection simplified graph convolution branch. This branch is used to weight the original channel, thereby concentrating the relevant channel information and weakening the interference of irrelevant information. The calculation formula is as follows:

[0189]

[0190] The v(*) function is used to reconstruct the size information. The feature results represent the convolutional branches of the parallel channel graph.

[0191] (10) The results of the parallel channel map convolution branch and the depth residual connection simplified map convolution branch are fused to obtain new feature layer information. The calculation formula is as follows:

[0192]

[0193] Where X OUT This represents new feature layer information.

[0194] As described in step 4, the new feature information is upsampled to obtain the upsampled feature information. The specific steps are as follows:

[0195] (1) Construct an upsampling module to upsample the new feature information and restore the feature map size. Each upsampling module is as follows: Figure 3 As shown, the feature information is first upsampled by 2×2 to restore the size of the feature map; then, three parallel 1×1, 3×3, and 5×5 convolutions with a stride of 1 are used to enrich the semantic features.

[0196] (2) The entire upsampling process involves using five consecutive upsampling modules to generate upsampled feature information.

[0197] As described in step 5, the upsampled feature information is mapped through dimensionality reduction and normalization to output the final remote sensing image landslide area detection result. The specific steps are as follows:

[0198] (1) Perform dimensionality reduction on the upsampled feature information by using a 1x1 convolutional layer to obtain the dimensionality-reduced feature information.

[0199] (2) The reduced feature information is normalized by using the normalization exponential function softmax to map the feature information of each pixel onto the probability distribution, thus obtaining the final remote sensing image landslide area detection result. The normalization exponential function softmax maps the output value range of each output node to [0,1] and constrains the sum of the output values ​​of each output node to 1.

[0200] Results analysis:

[0201] Figure 6 The Sentinel-2B Level-1C product was obtained at 3:35:49 AM on October 3, 2022. The ground resolution is 10m, acquired via Tile48RWT during relative orbit 061, and processed using PDGS baseline 04.00. By extracting color images of the Sentinel-2 satellite over the region of 29°38ˊN, 105°56ˊ18″E, landslide detection was performed using a lightweight bi-branch graph convolutional network based on globally maximized node connections. The vast majority of landslide areas were identified, demonstrating that this invention can meet the actual on-orbit satellite requirements for landslide detection in remote sensing imagery.

[0202] The above embodiments are merely explanations of the present invention and should not be construed as limiting the present invention. Therefore, any implementation methods similar to the present invention or implementation methods used in other similar structures but with similar concepts to the present invention are within the protection scope of the present invention.

[0203] The above description is only the best specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the protection scope of the present invention.

[0204] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A landslide area detection method based on dual-channel image convolution, characterized in that, include: Deep features of landslide areas in remote sensing images were extracted using a fully convolutional network. The deep features are used to construct a graph structure of nodes, edges, and distances between nodes using graph construction methods, resulting in an adjacency matrix, degree matrix, feature matrix, and weight matrix. The adjacency matrix, degree matrix, feature matrix, and weight matrix are input into a two-channel graph convolution to obtain new feature information; The new feature information is upsampled to obtain the upsampled feature information; The upsampled feature information is mapped through dimensionality reduction calculation and normalization processing to output the final remote sensing image landslide area detection result; The adjacency matrix, degree matrix, feature matrix, and weight matrix are input into a two-channel graph convolution to obtain new feature information, specifically: Among them, X OUT For new feature layer information, X C，DRP-GCN The convolutional branch of the depth residual simplified graph is used to represent the feature result, and concat is the channel fusion operation. The feature result of the convolution branch of the parallel channel map; the v(*) function is used to reconstruct the size information; The feature information of the l-th layer graph convolution in the parallel channel graph convolution branch; The feature information of the l-th layer graph convolution in the simplified graph convolution branch of the depth residual is given.

2. The landslide area detection method based on dual-channel graph convolution according to claim 1, characterized in that: Deep features of landslide areas in remote sensing images are extracted using a fully convolutional network, specifically: Construct a fully convolutional neural network; the fully convolutional neural network includes several superimposed downsampling modules; the downsampling module includes a 1×1 convolution operation unit and three parallel 1×1, 3×3, and 5×5 convolution operation units; The fully convolutional neural network is used to compress the feature information of the input landslide remote sensing image to obtain high-dimensional deep features R. H×W×C .

3. The landslide area detection method based on dual-channel graph convolution according to claim 1, characterized in that: For the nodes, edges, and distances between nodes in the deep feature graph structure, the adjacency matrix, degree matrix, feature matrix, and weight matrix are obtained, specifically: S1. Using a graph construction method, construct a feature graph G = (V, E) for the deep features; where... For the combination of nodes, Let E be the set of edges of the graph, e∈E be the edge of the feature graph, and r, s∈E be the index values ​​of the vertex at the receiving end and the input end, respectively. S2. Treat each point on the feature map as a graph node, and randomly initialize the weights of the graph nodes. Constructing the weight matrix S3, Calculate v for every two graph nodes. i and v j Euclidean distance; S4. Construct the adjacency matrix for each node based on the Euclidean distance; S5. Construct a symmetric normalized adjacency matrix using the adjacency matrix; S7. Construct the feature matrix based on the symmetric normalized adjacency matrix and the weight matrix, specifically as follows: Where σ(*) is a nonlinear activation function; The value is randomly initialized. Let be the symmetric normalized adjacency matrix of the l-th layer; S8. For the weight matrix W (l) The matrix is ​​updated to obtain a new weight matrix. S9. Update the feature matrix based on the symmetric normalized adjacency matrix and the new weight matrix, specifically as follows: in, The feature matrix; Let be the symmetric normalized adjacency matrix of the l-th layer.

4. The landslide area detection method based on dual-channel image convolution according to claim 3, characterized in that: The adjacency matrix of each node is as follows: Where l represents the number of layers; v is the value of the adjacency matrix in the i-th row and j-th column of the l-th layer; i v j ∈E, for node v i v j There are edges between them; R is an N×N matrix.

5. The landslide area detection method based on dual-channel graph convolution according to claim 3, characterized in that: The formula for the symmetric normalized adjacency matrix is: in, A is the symmetric normalized adjacency matrix of the l-th layer; (l) D is the adjacency matrix of the l-th layer; (l) Let l be the degree matrix of the l-th layer; For the l-th layer node v i and v j The weights between nodes; d(*) represents the weights of node v. i and v j The Euclidean distance function between them; exp is an exponential function with the natural constant e as the base; ∑ is the summation function; v i For the i-th graph node, v j Let j be the j-th graph node.

6. The landslide area detection method based on dual-channel graph convolution according to claim 1, characterized in that: The feature information of the l-th layer graph convolution in the parallel channel graph convolution branch and Specifically: Where add(*) is the operation of adding the feature matrices of the nodes in the l-th layer of the parallel channel graph convolution branch; The feature matrix of node i in the l-th layer of the parallel channel graph convolution branch; The normalized adjacency matrix of the l-th layer of the parallel channel graph convolutional branch; This is the new weight matrix for the l-th layer of the parallel channel graph convolution branch; The degree matrix of the l-th layer of the parallel channel graph convolution branch; The adjacency matrix of the l-th layer of the parallel channel graph convolution branch; In the parallel channel graph convolution branch, the adjacency matrix is ​​the value of the i-th row and j-th column in the l-th layer; element For the convolution branch of the parallel channel graph, the node of the l-th layer and The weights between them; Graph nodes for parallel channel graph convolution branches and The Euclidean distance between them.

7. The landslide area detection method based on dual-channel image convolution according to claim 6, characterized in that: The graph nodes of the parallel channel graph convolution branch and Specifically, the deep features are convolved using a filter of size K, with the internal channels set to C / 2, where C is the number of channels. The deep features then become... Then, graph construction methods are used to construct feature graphs for the deeper features, resulting in nodes of the graph structure. and 8. The landslide area detection method based on dual-channel graph convolution according to claim 1, characterized in that: The new feature information is upsampled to obtain the upsampled feature information, specifically as follows: The new feature information is upsampled using several cascaded upsampling modules to obtain upsampled feature information. The upsampling module includes a 2×2 convolutional unit and a parallel unit. The output of the 2×2 convolutional unit is used as the input of the parallel unit. The parallel unit includes three parallel 1×1 convolutional units, a 3×3 convolutional unit and a 5×5 convolutional unit with a stride of 1.

9. The landslide area detection method based on dual-channel image convolution according to claim 1, characterized in that: The upsampled feature information is mapped through dimensionality reduction and normalization to output the final remote sensing image landslide area detection result, specifically: The upsampled feature information is reduced in dimensionality using a 1x1 convolutional layer to obtain the dimensionality-reduced feature information. The reduced feature information is normalized, and the feature information of each pixel is mapped onto a probability distribution to obtain the final remote sensing image landslide area detection result.

Images