Polarimetric sar relative similarity measurement and classification algorithm based on graph convolution network

Abstract

The application provides a polarimetric SAR relative similarity measurement and classification algorithm based on a graph convolution network, performs polarimetric superpixel segmentation (Pol-ASLIC) on a preprocessed polarimetric SAR image, then uses a proposed ||WA||2 distance which fuses a symmetric revised Wishart distance and an AIRM distance to calculate the similarity degree between polarimetric covariance matrices and construct an adjacency matrix, and finally inputs the polarimetric covariance matrix as a superpixel region feature into a GCN graph convolution network for semi-supervised classification. The measurement effect of the polarimetric SAR data in the graph convolution network is better than that of the currently popular symmetric revised Wishart distance and AIRM distance, and better polarimetric SAR classification effect and precision can be achieved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image data processing, and mainly involves the similarity measurement calculation of polarimetric SAR data and polarimetric SAR image classification. Specifically, it is a polarimetric SAR relative similarity measurement and classification algorithm based on graph convolutional networks. Background Technology

[0002] Synthetic Aperture Radar (SAR) possesses a unique active microwave side-looking imaging method, unaffected by lighting, cloud cover, or weather conditions, enabling near-all-day, all-weather imaging. This compensates for the limitations of optical imaging, which is unable to image at night or under adverse weather conditions, or suffers from poor image quality. Furthermore, polarimetric SAR can additionally provide radar wave polarization information, allowing for the interpretation of rich physical scattering mechanisms of targets. Therefore, it has broad research value in areas such as ship detection, target identification, forest extraction, and even urban planning.

[0003] Polarimetric SAR image classification is an important research direction in the field of polarimetric SAR interpretation. Extracting the differences between targets is a prerequisite for ground feature classification, and these differences can be measured by calculating the similarity between targets. In polarimetric SAR images, the similarity between targets can be quantitatively analyzed using the polarimetric covariance matrix. In traditional polarimetric SAR classification methods, most methods rely on target decomposition based on the polarimetric covariance matrix or coherence matrix to obtain discriminative features, and then use classification algorithms to achieve ground feature classification. With the continuous development of microwave imaging technology, traditional classification algorithms are gradually showing limitations when processing high-resolution, large-area, and multi-temporal polarimetric SAR images.

[0004] Deep learning has already achieved large-scale applications in remote sensing image processing, and graph convolutional networks (GCNs) based on graph structures are gradually revealing their enormous potential. The core of GCN networks lies in the construction of graph structures. Wishart distance, derived from the statistical properties of polarimetric SAR data, is commonly used to construct polarimetric SAR graph structures due to its excellent metric performance. However, due to the computational characteristics of GCN networks, Wishart distance does not participate in the convolution process; it only provides a relative distance relationship for the construction of the adjacency matrix, thus weakening its excellent metric capability in GCN networks. Further research shows that the polarimetric covariance matrix space is defined on a Riemannian manifold, so its geodesic distance on the manifold can be calculated to determine geometric similarity, thereby obtaining the affine invariant Riemannian metric (AIRM distance) derived from the geometric properties of polarimetric SAR. AIRM distance has been proven to have good metric capabilities. When constructing adjacency matrices using AIRM distance in GCN networks, the classification accuracy is comparable to that of adjacency matrices constructed based on Wishart distance. However, there are significant differences between the adjacency matrices constructed using both methods.

[0005] Based on the graph structure characteristics of GCN networks, this invention proposes a novel local relative similarity measurement algorithm that combines the statistical and geometric characteristics of polarimetric SAR data to determine the degree of similarity of polarimetric covariance matrices under graph structure. Then, based on the proposed distance metric, an adjacency matrix is ​​constructed to achieve semi-supervised classification of polarimetric SAR images. Summary of the Invention

[0006] The purpose of this invention is to propose a polarimetric SAR relative similarity measurement and classification algorithm based on graph convolutional networks. By integrating the statistical and geometric properties of polarimetric SAR data, a new similarity measurement method is constructed to more accurately measure the degree of similarity of the polarimetric covariance matrix in the graph structure. The adjacency matrix constructed using this distance metric achieves higher classification accuracy in the GCN deep learning network.

[0007] Therefore, a polarimetric SAR relative similarity measurement and classification algorithm based on graph convolutional networks is presented, with the following specific steps:

[0008] Step 1: Acquire a single-look complex image of a fully polarimetric SAR image, and perform orbit correction, radiometric calibration, multi-look processing, polarimetric filtering, terrain correction, and geocoding to reduce the effects of speckle noise and geometric distortion. Obtain the polarimetric covariance matrix, as shown in the following formula:

[0009]

[0010] In the formula, <·> represents the average of a spatial or temporal set, and k 3L The scattering vector after distinctness correction. S represents the conjugate transpose of a vector. HH S HV S VH and S VV For complex data with four polarization channels, here S HV =S VH * represents complex conjugation;

[0011] Step 2: Polarimetric superpixel segmentation (Pol-ASLIC) is performed using the polarization covariance matrix as the pixel attribute value. The distance formula used in this segmentation algorithm integrates polarization, texture, and spatial information in the polarimetric SAR image, and is defined as:

[0012]

[0013] In the formula, D SIRV D represents the similarity of the covariance matrices between corresponding pixels. T D represents the texture distance between pixels. Sβ represents spatial similarity, S represents the sampling step size, and β is a parameter used to balance the spatial similarity measure with the other two similarity measures.

[0014] Step 3: Calculate the average polarization covariance matrix of the superpixel. The averaging formula is as follows:

[0015]

[0016] In the formula, N k C represents the number of pixels contained in the k-th object. i Let represent the polarization covariance matrix of the i-th pixel;

[0017] Step 4: Determine the number of land cover categories through visual interpretation using Google Earth, select training sample pixels, and extract the polarization covariance matrix of the sample pixels;

[0018] Step 5: Merge the average polarization covariance matrix with the polarization covariance matrix of the training samples into a matrix set. Calculate the symmetric revised Wishart distance and AIRM distance between any two polarization covariance matrices in the matrix set, using the following formulas:

[0019]

[0020]

[0021] In the formula, X and Y are any two polarization covariance matrices, and q is a parameter related to the polarization mode. In fully polarized imaging, q is taken as 3.

[0022] Step 6: Linearly normalize the symmetric revised Wishart distance and AIRM distance, and construct the Wishart-AIRM vector. The linear normalization formula is as follows:

[0023]

[0024] In the formula, N(·) represents the linear normalization function, X is the original data, and X0 is the standard deviation of the standard deviation. min With X max These are the minimum and maximum values ​​in the dataset, respectively.

[0025] Step 7: Calculate the L2 norm distance of the Wishart-AIRM vectors, i.e., the ||WA||2 distance, and construct the adjacency matrix using the ||WA||2 distance. Use the average polarization covariance matrix as the target feature input to the GCN network for semi-supervised classification. The formula for the ||WA||2 distance is as follows:

[0026]

[0027] The formula for constructing the adjacency matrix is ​​as follows:

[0028]

[0029] In the formula, Nei(·) represents the neighborhood operation, and S i This represents the i-th superpixel corresponding to the i-th polarization covariance matrix;

[0030] The average polarization covariance matrix is ​​used as the target feature input to the GCN network for semi-supervised classification. The GCN network consists of two neural networks: the forward model from the input layer (I) to the hidden layer (H) uses ReLU as the activation function, and the forward model from the hidden layer (H) to the output layer (O) uses softmax as the activation function. The cross-entropy error function is used to correct the difference between the predicted results and the labels of the training samples, as shown in the following formula:

[0031]

[0032] In the formula, N represents the number of training samples, C represents the number of land cover categories, and Y represents the number of land cover categories. ij and P ij These represent the truth label and the predicted probability, respectively.

[0033] This invention provides a polarimetric SAR relative similarity measurement and classification algorithm based on graph convolutional networks. The technical effects are: it integrates the statistical and geometric properties of polarimetric SAR data, and has superior performance in measuring the similarity of polarimetric covariance matrices in graph structures, which can improve the accuracy of GCN semi-supervised classification models in polarimetric SAR classification tasks. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating a polarimetric SAR relative similarity measurement and classification algorithm based on graph convolutional networks provided by the present invention.

[0035] Figure 2 This is a schematic diagram of the polarimetric SAR image structure conversion provided by the present invention;

[0036] Figure 3 A schematic diagram illustrating the principle of constructing the ||WA||2 distance provided by this invention;

[0037] Figure 4 This is a schematic diagram of a GCN network. Detailed Implementation

[0038] The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and the flowchart is as follows. Figure 1 :

[0039] Step 1: Acquire a single-look complex image of a fully polarimetric SAR image, and perform orbit correction, radiometric calibration, multi-look processing, polarimetric filtering, terrain correction, and geocoding to reduce the effects of speckle noise and geometric distortion, obtaining the preprocessed polarimetric covariance matrix. The polarimetric covariance matrix is ​​as follows:

[0040]

[0041] In the formula, <·> represents the average of a spatial or temporal set, and k 3L The scattering vector after distinctness correction. S represents the conjugate transpose of a vector. HH S HV S VH and S VV For complex data with four polarization channels, here S HV =S VH * represents complex conjugation;

[0042] Step 2: Polarimetric superpixel segmentation is performed using the polarization covariance matrix as the pixel attribute value. The distance formula used for segmentation is as follows:

[0043]

[0044] In the formula, D SIRV D represents the similarity of the covariance matrices between corresponding pixels. T D represents the texture distance between pixels. S β represents spatial similarity, S represents the sampling step size, and β is a parameter used to balance the spatial similarity measure with the other two similarity measures.

[0045] Step 3: Calculate the average polarization covariance matrix of the segmented superpixels. The calculation formula is as follows:

[0046]

[0047] In the formula, N k C represents the number of pixels contained in the k-th object. i Let represent the polarization covariance matrix of the i-th pixel;

[0048] Step 4: Determine the number of land cover categories through visual interpretation using Google Earth, select training sample pixels, with 50 samples selected for each category, and extract the polarization covariance matrix of the sample pixels;

[0049] Step 5: Merge the average polarization covariance matrix of the superpixel with the polarization covariance matrix of the training sample pixels into a matrix set. Calculate the symmetric revised Wishart distance and AIRM distance between any two polarization covariance matrices in the matrix set. The formulas for calculating the symmetric revised Wishart distance and AIRM distance are as follows:

[0050]

[0051] Where X and Y are any two polarization covariance matrices, and q is a parameter related to the polarization mode, which is taken as 3 in fully polarimetric imaging;

[0052] Step 6: Linearly normalize the symmetric revised Wishart distance and AIRM distance, and construct the Wishart-AIRM vector. The linear normalization formula is as follows;

[0053]

[0054] In the formula, N(·) represents the linear normalization function, X is the original data, and X0 is the standard deviation of the standard deviation. min With X max These are the minimum and maximum values ​​in the dataset, respectively.

[0055] Step 7, calculate the L2 norm distance of the Wishart-AIRM vectors, i.e., the ||WA||2 distance, as follows: Figure 2 Then, the adjacency matrix is ​​constructed using the ||WA||2 distance. The formula for the ||WA||2 distance is as follows:

[0056]

[0057] The formula for constructing the adjacency matrix is ​​as follows:

[0058]

[0059] In the formula, Nei(·) represents the neighborhood operation, and S i This represents the i-th superpixel corresponding to the i-th polarization covariance matrix, such as... Figure 3 .

[0060] The average polarization covariance matrix is ​​used as the target feature input to the GCN network for semi-supervised classification. Here, the GCN network consists of two neural network layers, such as... Figure 4 The forward model from the input layer (I) to the hidden layer (H) uses the ReLU function as the activation function:

[0061]

[0062] In the formula, Let F be the normalized adjacency matrix, and W be the average polarization covariance eigenma matrix. (0)This is the trainable filter matrix from the input layer to the hidden layer.

[0063] The forward model from the hidden layer (H) to the output layer (O) uses the softmax function as the activation function:

[0064]

[0065] In the formula, W (1) This is the trainable filter matrix from the hidden layer to the output layer.

[0066] The cross-entropy error function is used to correct the discrepancy between the predictions and the labels of the training samples:

[0067]

[0068] In the formula, N represents the number of training samples, C represents the number of land cover categories, and Y represents the number of land cover categories. ij and P ij These represent the truth label and the predicted probability, respectively.

Claims

1. A method for measuring and classifying the relative similarity of polarimetric SAR based on graph convolutional networks, characterized in that, Includes the following steps: Step 1: Perform orbit correction, radiometric calibration, multi-view, polarimetric filtering, terrain correction, and geocoding on the acquired single-view polarimetric SAR complex image to reduce speckle noise and geometric distortion, improve the visual interpretation effect of the image, and obtain the preprocessed polarimetric covariance matrix. Step 2: Perform polarization superpixel segmentation using the polarization covariance matrix as the pixel attribute value; Step 3: Calculate the average polarization covariance matrix of the superpixel; Step 4: Determine the number of land cover categories through visual interpretation using Google Earth, select training sample pixels, and extract the polarization covariance matrix of the sample pixels; Step 5: Merge the average polarization covariance matrix of the superpixel with the polarization covariance matrix of the sample pixel into a matrix set, and calculate the symmetric revised Wishart distance and AIRM distance between any two polarization covariance matrices in the matrix set. Step 6: Linearly normalize the symmetric revised Wishart distance and AIRM distance, and construct the Wishart-AIRM vector; Step 7, calculate the L2 norm distance of the Wishart-AIRM vectors, i.e., the ||WA||2 distance, using the following formula: X and Y are any two polarization covariance matrices, and an adjacency matrix is ​​constructed using the ‖WA‖2 distance. The average polarization covariance matrix is ​​used as the target feature input to the GCN deep neural network for semi-supervised classification.

2. The polarimetric SAR relative similarity measurement and classification method based on graph convolutional networks according to claim 1, characterized in that, In step 1, the polarimetric SAR single-look complex image is preprocessed to obtain the polarimetric covariance matrix characterizing the polarimetric wave scattering properties of the target. The polarimetric covariance matrix is ​​defined by the following formula: In the formula, <·> represents the average of a spatial or temporal set, and k 3L The scattering vector after distinctness correction. S represents the conjugate transpose of a vector. HH S HV S VH and S VV For complex data with four polarization channels, here S HV =S VH , * represents complex conjugation.

3. The polarimetric SAR relative similarity measurement and classification method based on graph convolutional networks according to claim 1, characterized in that, In step 2, polarimetric superpixel segmentation (Pol-ASLIC) is used to segment the polarimetric SAR image. Pol-ASLIC fully utilizes the polarization, texture, and spatial information of the polarimetric SAR image, and the generated superpixels can accurately represent a single type of land cover and have good homogeneity. The distance formula used for Pol-ASLIC segmentation is as follows: In the formula, D SIRV D represents the similarity of the covariance matrices between corresponding pixels. T D represents the texture distance between pixels. S β represents spatial similarity, S represents the sampling step size, and β is a parameter used to balance the spatial similarity measure with the other two similarity measures.

4. The polarimetric SAR relative similarity measurement and classification method based on graph convolutional networks according to claim 1, characterized in that, In step 3, the average polarization covariance matrix of the superpixel is calculated to characterize the radar wave scattering features of the object region. The averaging formula is as follows: In the formula, N k C represents the number of pixels contained in the k-th object. i Let represent the polarization covariance matrix of the i-th pixel.

5. The polarimetric SAR relative similarity measurement and classification method based on graph convolutional networks according to claim 1, characterized in that, In step 5, the symmetric revised Wishart distance and AIRM distance between any two polarization covariance matrices are calculated using the following formula: In the formula, q is a parameter related to the polarization mode, and q is taken as 3 in fully polarized imaging.

6. The polarimetric SAR relative similarity measurement and classification method based on graph convolutional networks according to claim 1, characterized in that, In step 6, the symmetric revised Wishart distance and AIRM distance are linearly normalized to eliminate the difference in their metric scales. The linear normalization formula is as follows: In the formula, N(·) represents the linear normalization function, X is the original data, and X0 is the standard deviation of the standard deviation. min With X max These are the minimum and maximum values ​​in the dataset, respectively.

7. The polarimetric SAR relative similarity measurement and classification method based on graph convolutional networks according to claim 1, characterized in that, In step 7, the adjacency matrix is ​​constructed using the ‖WA‖2 distance and GCN semi-supervised classification is performed. The ‖WA‖2 distance integrates the statistical properties contained in the symmetric revised Wishart distance and the geometric properties contained in the AIRM distance, characterizing the similarity of the polarization covariance matrix in the graph structure. The formula for the constructed self-connected adjacency matrix is ​​as follows: In the formula, Nei(·) represents the neighborhood operation, and S i This represents the i-th superpixel corresponding to the i-th polarization covariance matrix; The average polarization covariance matrix is ​​used as the target feature input to a GCN deep learning network for semi-supervised classification. The GCN network consists of two layers: the forward model from the input layer to the hidden layer uses ReLU as the activation function, and the forward model from the hidden layer to the output layer uses softmax as the activation function. The cross-entropy error function is used to correct the difference between the prediction results and the labels of the training samples, as shown in the following formula: In the formula, N represents the number of training samples, C represents the number of land cover categories, and Y represents the number of land cover categories. ij and P ij These represent the truth label and the predicted probability, respectively.

Images