Polarization SAR relative similarity measurement and classification algorithm based on graph convolution network

Abstract

The invention provides a polarized SAR (Synthetic Aperture Radar) relative similarity measurement and classification algorithm based on a graph convolutional network, which comprises the following steps: carrying out polarized superpixel segmentation (Pool-ASLIC) on a preprocessed polarized SAR image, then calculating the similarity degree between polarized covariance matrixes by using a proposed WA2 distance fusing a symmetric revised Wishart distance and an AIRM distance, and constructing an adjacent matrix; and finally, the polarization covariance matrix is used as a superpixel region feature to be input into a GCN graph convolutional network for semi-supervised classification. According to the method, the measurement effect of the polarimetric SAR data in the graph convolution network is better than that of the current popular symmetric revised Wishart distance and AIRM distance, and better polarimetric SAR classification effect and precision can be realized.

Description

Technical Field

[0001] The invention belongs to the field of remote sensing image data processing, and mainly relates to similarity measurement calculation of polarimetric SAR data and polarimetric SAR image classification, and specifically is a polarimetric SAR relative similarity measurement and classification algorithm based on graph convolutional network. Background Art

[0002] Synthetic aperture radar (SAR) has a unique active microwave side-view imaging method, which is not restricted by conditions such as light, clouds, and weather, and can perform imaging almost all day and all weather, making up for the defects of optical imaging that cannot be imaged or the imaging quality is low at night and in bad weather conditions. On this basis, polarimetric SAR can also provide additional polarization information of radar waves, from which the rich physical scattering mechanism of the target can be interpreted, so it has extensive research value in ship detection, target recognition, forest extraction and even urban construction planning.

[0003] Polarimetric SAR image classification is an important research direction in the field of polarimetric SAR interpretation. Extracting the differences between targets is the prerequisite for achieving ground object classification, and this difference can be measured by calculating the similarity between targets. In polarimetric SAR images, the similarity between targets can be quantitatively analyzed by the polarimetric covariance matrix. In traditional polarimetric SAR classification methods, most of them are based on polarimetric covariance matrix or coherence matrix to decompose the target to obtain discriminant features, and then use classification algorithms to achieve ground object classification. With the continuous development of microwave imaging technology, traditional classification algorithms have gradually revealed their limitations when processing high-resolution, large-scale and multi-phase polarimetric SAR images.

[0004] Deep learning has been widely used in the field of remote sensing image processing, and graph convolutional networks (GCNs) based on graph structures have gradually revealed their great potential. The core of the GCN network lies in the construction of graph structures. The Wishart distance derived from the statistical properties of polarimetric SAR data is often used to construct polarimetric SAR graph structures due to its extremely excellent metric performance. However, due to the operational characteristics of the GCN network, the Wishart distance does not participate in the convolution process, but only provides a relative distance relationship for the construction of the adjacency matrix, resulting in the excellent metric ability of the Wishart distance being weakened in the GCN network. 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 the geometric similarity, and then the affine invariant Riemannian metric (AIRM distance) derived from the geometric characteristics of polarimetric SAR is obtained. The AIRM distance has been proven to have good metric ability. When the adjacency matrix is ​​constructed with the AIRM distance in the GCN network, its classification accuracy is comparable to that of the adjacency matrix constructed based on the Wishart distance. However, there are significant differences between the adjacency matrices constructed based on the two.

[0005] According to the graph structure characteristics of the GCN network, this paper proposes a new local relative similarity measurement algorithm that combines the statistical and geometric characteristics of polarimetric SAR data to determine the similarity degree of polarimetric covariance matrices in the form of a graph structure, and then constructs an adjacency matrix based on the proposed distance metric to achieve semi-supervised classification of polarimetric SAR images. Summary of the invention

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

[0007] Therefore, a polarimetric SAR relative similarity measurement and classification algorithm based on graph convolutional network is proposed. The specific steps are as follows:

[0008] Step 1: Obtain a fully polarimetric SAR single-view complex image, perform orbit correction, radiation calibration, multi-view, polarization filtering, terrain correction, geocoding, reduce the influence of coherent speckle noise and geometric distortion, and obtain the polarization covariance matrix. The formula is as follows:

[0009]

[0010] In the formula, <·> represents the spatial or temporal ensemble average, k 3L is the scattering vector after reciprocity correction, Represents the conjugate transpose of a vector, S HH , S HV , S VH and S VV is the complex data of the four polarization channels, where S HV =S VH , * represents complex conjugation;

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

[0012]

[0013] Where D SIRV represents the similarity of the covariance matrix between corresponding pixels, D T Represents the texture distance between pixels, D Srepresents 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 superpixels. The averaging formula is as follows:

[0015]

[0016] Where N k Indicates the number of pixels contained in the kth object, C i represents the polarization covariance matrix of the i-th pixel;

[0017] Step 4, determine the number of ground object categories through Google Earth visual interpretation, select training sample pixels, and extract the polarization covariance matrix of the sample pixels;

[0018] Step 5: Combine the average polarization covariance matrix and the polarization covariance matrix of the training sample into a matrix set, and calculate the symmetric revised Wishart distance and AIRM distance between any two polarization covariance matrices in the matrix set. The calculation formula is as follows:

[0019]

[0020]

[0021] Where X and Y are any two polarization covariance matrices, q is a parameter related to the polarization mode, and q is 3 in full polarization imaging;

[0022] Step 6: Linearly normalize the symmetric revised Wishart distance and AIRM distance and construct a 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 X min With X max are the minimum and maximum values ​​in the data set respectively;

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

[0026]

[0027] The adjacency matrix formula is constructed as follows:

[0028]

[0029] In the formula, Nei(·) represents the neighborhood operation, S i 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 into the GCN network for semi-supervised classification. The GCN network here contains two layers of neural networks. The forward model from the input layer (I) to the hidden layer (H) uses the ReLU function as the activation function, and the forward model from the hidden layer (H) to the output layer (O) uses the softmax function 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. The formula is as follows:

[0031]

[0032] In the formula, N represents the number of training samples, C represents the number of set ground object categories, and Y ij and P ij Represent the true value label and predicted probability respectively.

[0033] The present invention provides a polarimetric SAR relative similarity measurement and classification algorithm based on graph convolutional network, which has the technical effects of integrating the statistical and geometric properties of polarimetric SAR data, having more excellent performance in measuring the similarity of polarimetric covariance matrices in graph structures, and being able to improve the accuracy of GCN semi-supervised classification models in polarimetric SAR classification tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A flowchart of a polarimetric SAR relative similarity measurement and classification algorithm based on a graph convolutional network provided by the present invention;

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

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

[0037] Figure 4 Schematic diagram of the GCN network. DETAILED DESCRIPTION

[0038] The following examples are only used to more clearly illustrate the technical solution of the present invention. Figure 1 :

[0039] Step 1: Obtain a fully polarimetric SAR single-look complex image, perform orbit correction, radiation calibration, multi-look, polarization filtering, terrain correction, and geocoding to reduce the influence of coherent speckle noise and geometric distortion, and obtain the preprocessed polarization covariance matrix. The polarization covariance matrix is:

[0040]

[0041] In the formula, <·> represents the spatial or temporal ensemble average, k 3L is the scattering vector after reciprocity correction, Represents the conjugate transpose of a vector, S HH , S HV , S VH and S VV is the complex data of the four polarization channels, where S HV =S VH , * represents complex conjugation;

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

[0043]

[0044] Where D SIRV represents the similarity of the covariance matrix between corresponding pixels, D T Represents the texture distance between pixels, D S 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] Where N k represents the number of pixels contained in the kth object, C i represents the polarization covariance matrix of the i-th pixel;

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

[0049] Step 5: Combine the average polarization covariance matrix of the superpixel and the polarization covariance matrix of the training 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. The calculation formulas for the symmetric revised Wishart distance and AIRM distance are as follows:

[0050]

[0051] Where X and Y are any two polarization covariance matrices, q is a parameter related to the polarization mode, and is taken as 3 in full polarization 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 X min With X max are the minimum and maximum values ​​in the data set respectively;

[0055] Step 7, calculate the L2 norm distance of the Wishart-AIRM vector, that is, ||WA||2 distance, such as Figure 2 . And use ||WA||2 distance to construct the adjacency matrix. The formula of ||WA||2 distance is as follows:

[0056]

[0057] The adjacency matrix formula is constructed as follows:

[0058]

[0059] In the formula, Nei(·) represents the neighborhood operation, S i 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 into the GCN network for semi-supervised classification. The GCN network here contains two layers of neural networks, 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, is the normalized adjacency matrix, F is the average polarization covariance feature matrix, W (0)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] Where W (1) is the trainable filter matrix from the hidden layer to the output layer.

[0066] Use the cross entropy error function to correct the difference between the prediction results and labels of the training samples:

[0067]

[0068] In the formula, N represents the number of training samples, C represents the number of set ground object categories, and Y ij and P ij Represent the true value label and predicted probability respectively.

Claims

1. A polarimetric SAR relative similarity measurement and classification algorithm based on graph convolutional network, characterized in that: The following steps are involved: Step 1: perform orbit correction, radiation calibration, multi-view, polarization filtering, terrain correction, and geocoding on the acquired polarimetric SAR single-view complex image to reduce coherent speckle noise and geometric distortion, improve the visual interpretation effect of the image, and obtain the pre-processed polarimetric covariance matrix; Step 2, using the polarization covariance matrix as the pixel attribute value to perform polarization superpixel segmentation; Step 3, calculating the average polarization covariance matrix of superpixels; Step 4, determine the number of ground object categories through Google Earth visual interpretation, select training sample pixels, and extract the polarization covariance matrix of the sample pixels; Step 5, merging the average polarization covariance matrix of the superpixel and the polarization covariance matrix of the sample pixel into a matrix set, and calculating 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 the AIRM distance and construct a Wishart-AIRM vector; Step 7: Calculate the L2 norm distance of the Wishart-AIRM vector, i.e., the ||WA||2 distance, and use the ||WA||2 distance to construct the adjacency matrix. Use the average polarization covariance matrix as the target feature and input it into the GCN deep neural network for semi-supervised classification.

2. According to claim 1, a polarimetric SAR relative similarity measurement and classification algorithm based on graph convolutional network is characterized in that: In step 1, the polarimetric SAR single-view complex image is preprocessed to obtain a polarimetric covariance matrix that can characterize the target polarimetric wave scattering characteristics. The polarimetric covariance matrix is ​​defined by the following formula: In the formula, <·> represents the spatial or temporal ensemble average, k 3L is the scattering vector after reciprocity correction, S represents the conjugate transpose of the vector, HH , S HV , S VH and S VV is the complex data of the four polarization channels, where S HV =S VH , * represents complex conjugation.

3. The relative similarity measurement and classification algorithm for polarimetric SAR based on graph convolutional network according to claim 1, characterized in that: In step 2, polarization superpixel segmentation (Pol-ASLIC) is used to segment the polarimetric SAR image. This algorithm can make full use of the polarization, texture and spatial information of the polarimetric SAR image. The generated superpixels can more accurately express a single type of ground objects and have good homogeneity. The distance formula used by Pol-ASLIC segmentation is as follows: Where D SIRV represents the similarity of the covariance matrix between corresponding pixels, D T Represents the texture distance between pixels, D S represents the 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 relative similarity measurement and classification algorithm for polarimetric SAR based on graph convolutional network 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 characteristics of the object area. The averaging formula is as follows: Where N k Indicates the number of pixels contained in the kth object, C i Represents the polarization covariance matrix of the i-th pixel.

5. The relative similarity measurement and classification algorithm for polarimetric SAR based on graph convolutional network according to claim 1, characterized in that: In step 5, the symmetric modified Wishart distance and AIRM distance between any two polarization covariance matrices are calculated, and the calculation formula is as follows: Where X and Y are any two polarization covariance matrices, q is a parameter related to the polarization mode, and q is 3 in full polarization imaging.

6. The relative similarity measurement and classification algorithm for polarimetric SAR based on graph convolutional network according to claim 1, characterized in that: In step 6, the symmetric revised Wishart distance and the AIRM distance are linearly normalized to eliminate the difference in the measurement scale between the two. The linear normalization formula is as follows: In the formula, N(·) represents the linear normalization function, X is the original data, and X min With X max are the minimum and maximum values ​​in the data set, respectively.

7. The relative similarity measurement and classification algorithm for polarimetric SAR based on graph convolutional network according to claim 1, characterized in that: In step 7, the ||WA||2 distance is used to construct an adjacency matrix and perform GCN semi-supervised classification. The ||WA||2 distance combines the statistical characteristics contained in the symmetric revised Wishart distance and the geometric characteristics contained in the AIRM distance, and can more accurately characterize the similarity of the polarization covariance matrix in the graph structure. The formula of the ||WA||2 distance is as follows: The formula for constructing the self-connected adjacency matrix is ​​as follows: In the formula, Nei(·) represents the neighborhood operation, S i 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 into the GCN deep learning network for semi-supervised classification. The GCN network used contains two layers of neural networks. The forward model from the input layer (I) to the hidden layer (H) uses the ReLU function as the activation function, and the forward model from the hidden layer (H) to the output layer (O) uses the softmax function 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. The formula is as follows: In the formula, N represents the number of training samples, C represents the number of set ground object categories, and Y ij and P ij Represent the true value label and predicted probability respectively.

Images