End-to-end polarimetric SAR image classification method based on superpixels and graph convolution

Abstract

The present invention discloses an end-to-end polarimetric SAR image classification method based on superpixels and graph convolution, comprising the following steps: step 1, inputting polarimetric SAR images to be classified and cropping them into a uniform size; step 2, dividing the cropped images into a training set and a test set in proportion; step 3, decomposing the complex scattering matrix of each pixel point of each image of the test set training set, generating a polarimetric coherence matrix and converting it into a row vector as the polarimetric feature of the pixel point; step 4, splicing the polarimetric feature with the horizontal and vertical coordinates of the pixel point, and splicing the row vector as the pixel point feature; step 5, building an end-to-end network based on a full convolutional network, a graph convolutional network and a convolutional neural network; step 6, sending the training set into the end-to-end network for joint training, and sending the test set into the trained end-to-end network to obtain the result. The present invention can further improve the classification accuracy of polarimetric SAR images.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing and remote sensing, and relates to an end-to-end polarimetric SAR image classification method based on superpixels and graph convolution. Background Art

[0002] Polarimetric synthetic aperture radar (PolSAR) image classification is one of the hottest research directions in the field of PolSAR, which can provide basic support for many fields such as land use survey, geographical situation monitoring, and urban and rural planning. The so-called polarimetric SAR image classification problem is to determine the category of each pixel in the image through an algorithm to determine the corresponding ground object.

[0003] In the traditional method of classifying polarimetric SAR images based on superpixel segmentation, superpixel segmentation is used as a separate task to preprocess the image, and then the image results after superpixel segmentation preprocessing are input into the polarimetric SAR classification network to obtain the polarimetric SAR image classification results.

[0004] Since PolSAR images themselves contain a lot of random coherent speckle noise and have a low resolution, and different types of land objects tend to show similar characteristics, the traditional superpixel segmentation method will be greatly disturbed, and the superpixel segmentation results will directly affect the output results of the downstream polarimetric SAR classification network. Summary of the invention

[0005] The purpose of the present invention is to provide an end-to-end polarimetric SAR image classification method based on superpixels and graph convolution, which can further improve the classification accuracy of polarimetric SAR images.

[0006] The technical solution adopted by the present invention is:

[0007] The end-to-end polarimetric SAR image classification method based on superpixels and graph convolution includes the following steps:

[0008] Step 1, input the polarimetric SAR image to be classified and crop it into a uniform size;

[0009] Step 2: Divide the cropped images into training sets and test sets according to the proportion;

[0010] Step 3, decomposing the complex scattering matrix of each pixel point of each image in the test set training set, generating a polarization coherence matrix and converting it into a row vector as the polarization feature of the pixel point;

[0011] Step 4, splice the polarization feature with the horizontal and vertical coordinates of the pixel point, and splice the row vector as the pixel point feature;

[0012] Step 5: Build an end-to-end network based on full convolutional network, graph convolutional network and convolutional neural network;

[0013] Step 6: Send the training set into the end-to-end network for joint training, and send the test set into the trained end-to-end network to obtain the results.

[0014] The specific steps of step 3 are:

[0015] Decompose the complex scattering matrix of each pixel of each image cropped in step 2, generate a polarization coherence matrix and convert it into a row vector of size 1×9 as the polarization feature of the pixel. The expression of the generated polarization coherence matrix T is as follows:

[0016]

[0017] The polarization coherence matrix T is denoted as Then the polarization coherence matrix is ​​converted into a row vector to obtain the characteristic matrix: T′=[T 11 , T 12 , T 13 , T 21 , T 22 , T 23 , T 31 , T 32 , T 33 ], and the polarization coherence matrix is ​​a complex conjugate matrix, and the complex conjugate matrix is ​​preprocessed to obtain the polarization eigenvector F of the pixel point:

[0018]

[0019] Where Re represents the real part of the complex number, Im represents the imaginary part of the complex number, and T ij Represents the data in the i-th row and j-th column of the polarization coherence matrix.

[0020] Step 4 specifically includes: concatenating the polarization feature vector of each pixel point generated in step 3 with the horizontal and vertical coordinates of the pixel point to obtain the pixel point feature.

[0021] The structure of the full convolutional network in step 5 is a once-connected input layer, a first downsampling layer, a second downsampling layer, a third downsampling layer, a fourth downsampling layer, a fifth downsampling layer, a first upsampling layer, a second upsampling layer, a third upsampling layer, a fourth upsampling layer, and a softMax output layer. The loss function of the full convolutional network is:

[0022]

[0023] Where Φ represents the pixel feature before updating, Represents the features of the updated pixel, Represents the cross entropy loss function between the two;

[0024] The structure of the graph convolutional neural network is the input layer, the first graph convolutional layer, the second graph convolutional layer, the third graph convolutional layer and the softmax output layer connected in sequence. The activation function of each graph convolutional layer is the tanh function.

[0025] The convolutional neural network structure is an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, and a softMax output layer connected in sequence. The activation function of each convolutional layer is the LeakyRelu function.

[0026] The specific steps of step 6 are:

[0027] Step 6.1, initialize the training set and test set into superpixel blocks.

[0028] Step 6.2, the pixel features of the training set obtained in steps 3 and 4 are sent to the full convolutional network to obtain the output result. The output matrix Q is the soft correlation matrix of superpixels and pixels;

[0029] Step 6.3, obtaining the adjacency matrix A, feature matrix B, and transformation matrix C between superpixels and pixels of the superpixel block through the superpixel and pixel soft association matrix;

[0030] According to the superpixels and pixel soft association matrix Q obtained in step 6.2, each pixel point is assigned to the surrounding superpixel blocks with the highest probability to obtain the superpixel segmentation result of the entire image, and the adjacency matrix A and feature matrix B of each image are obtained according to the superpixel segmentation result;

[0031] Where A in the adjacency matrix A i,j Represents the element in the i-th row and j-th column of the adjacency matrix A:

[0032]

[0033] Feature matrix B i Represents the features of the i-th superpixel block; represents the feature of the jth pixel in superpixel block i, and n represents the number of pixels in the i-th superpixel block.

[0034]

[0035] The transformation matrix between superpixels and pixels is the transformation matrix C

[0036] Step 6.4, input the adjacency matrix and feature matrix of each image into the graph convolutional network. The input is the adjacency matrix and feature matrix of the image, and the output is a two-dimensional tensor H which is the superpixel feature of the image.

[0037] Step 6.5, convert the superpixel features of the image into pixel features:

[0038] H gcn =C·H (14)

[0039] Where C represents the transformation matrix and H represents the output of the graph convolutional network. gcn The pixel features are obtained by converting the superpixel features output by the graph convolutional network through the superpixel and pixel conversion matrix.

[0040] Step 6.6, the convolutional neural network obtains the pixel-level features of the image;

[0041] The polarization features of the training set obtained in step 3 are sent as input to the convolutional neural network to obtain pixel-level features;

[0042] Step 6.7, fusing the image pixel-level features obtained in step 6.6 and the pixel-level features obtained in step 6.5 through the transformation matrix of superpixels and pixels for classification;

[0043] Step 6.8, calculate the total loss function and back propagate iteratively to update the network until convergence. The total loss function of the end-to-end network is:

[0044] loss = loss1 + loss2 (15)

[0045] Among them, loss1 is the full convolutional network loss function of superpixel segmentation; loss2 is the classification loss function. Finally, the network is iteratively updated until the network converges, and the end-to-end network model training is completed;

[0046] In step 6.9, the test set is fed into the trained end-to-end network model to obtain the classification result.

[0047] The beneficial effects of the present invention are:

[0048] The present invention uses a full convolutional network to perform superpixel segmentation on an image and combines it with a downstream classification network. The superpixel segmentation network and the downstream classification network are jointly trained end-to-end. The downstream classification network is composed of two networks: a graph convolutional network and a convolutional neural network. The result of the superpixel segmentation network can affect the output result of the downstream classification network, and the result of the downstream classification network can in turn affect the superpixel segmentation result. The image classification work is completed by iteratively training the network model and a better classification result is achieved.

[0049] On the other hand, the downstream classification network uses graph convolutional networks and convolutional neural networks to extract image features simultaneously, and fuses the features extracted by the two. In this way, graph convolution can learn the global information of polarimetric SAR images, and convolutional neural networks can learn the pixel-level features of polarimetric SAR images. Fusion of the features extracted by the two can further improve the classification accuracy of polarimetric SAR images. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0051] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] The present invention provides an end-to-end polarimetric SAR image classification method based on superpixels and graph convolution. Figure 1 , including the following steps:

[0053] Step 1: Input the polarimetric SAR image to be classified, and crop the polarimetric SAR image into 500 images with a width and height of 512 pixels each;

[0054] Step 2: Divide the cropped images into 20% as the training set and 80% as the test set.

[0055] Step 3, decompose the complex scattering matrix of each pixel point of each image in the training set and the test set, generate a polarization coherence matrix and convert it into a row vector of size 1×9 as the polarization feature of the pixel point;

[0056] Step 4: concatenate the polarization feature with the horizontal and vertical coordinates of the pixel point, and use the concatenated row vector of size 1×11 as the pixel point feature.

[0057] Step 5: Build an end-to-end network based on full convolutional network, graph convolutional network and convolutional neural network

[0058] Step 6: Send the training set to the end-to-end network for joint training, and send the test set to the trained end-to-end network to obtain the classification result.

[0059] The specific steps of step 3 are:

[0060] The complex scattering matrix of each pixel point of each image in the training set and the test set is decomposed to generate a polarization coherence matrix and convert it into a row vector of size 1×9 as the polarization feature of the pixel point. The expression of the generated polarization coherence matrix T is as follows:

[0061]

[0062] The polarization coherence matrix T is denoted as Then the polarization coherence matrix is ​​converted into a row vector to obtain the characteristic matrix: T′=[T 11 , T 12 , T 13 , T 21 , T 22 , T 23 , T 31 , T 32 , T 33 ], and the polarization coherence matrix is ​​a complex conjugate matrix, and the complex conjugate matrix is ​​preprocessed: the polarization eigenvector F of the pixel point is obtained:

[0063]

[0064] Where Re represents the real part of the complex number, Im represents the imaginary part of the complex number, and T ij Represents the data in the i-th row and j-th column of the polarization coherence matrix.

[0065] The specific steps of step 4 are:

[0066] The 9-dimensional polarization feature vector of each pixel point generated in step 3 is concatenated with the 2-dimensional position feature vector of the pixel point, i.e., the horizontal and vertical coordinates of the pixel point, and the concatenated 11-dimensional vector S is used as the pixel point feature:

[0067]

[0068] Where x and y represent the horizontal and vertical coordinates of the pixel.

[0069] The structures of the full convolutional network, graph convolutional network and convolutional neural network in the specific steps of step 5 are:

[0070] Build a fully convolutional network:

[0071] a) Build a fully convolutional network, whose structure is: input layer (1,11,512,512) → first downsampling layer (1,16,512,512) → second downsampling layer (1,32,512,512) → third downsampling layer (1,64,512,512) → fourth downsampling layer (1,128,512,512) → fifth downsampling layer (1,256,512,512) → first upsampling layer (1,128,512,512) → second upsampling layer (1,64,512,512) → third upsampling layer (1,32,512,512) → fourth upsampling layer (1,16,512,512) → softMax output layer.

[0072] The input data size of the input layer is: (1, 11, 512, 512), where 1 represents the batch size, 11 represents the 1×11 row vector of each pixel, and 512 is the width and height of the image.

[0073] The output size of the softMax output layer is: (1,9,512,512). We record the output data as Q, which is the soft association matrix of superpixels and pixels. 1 is the batchsize, and 9 represents the probability value of the pixel belonging to the superpixel block where the pixel is located and the superpixel block where the pixel is located and its eight adjacent superpixel blocks (sit, top, top right, left, right, bottom left, bottom right, bottom right). The sum of these 9 probability values ​​is 1.

[0074] b) Construct the loss function of the full convolutional network:

[0075] Get the polarization features of the superpixel block:

[0076]

[0077] in, is the polarization feature of superpixel m, f i,j is the polarization feature of pixel (i, j), where q is the soft correlation matrix between superpixel and pixel, Represents the probability value of pixel point (i, j) belonging to superpixel block m.

[0078] Get the coordinate features of the superpixel block:

[0079]

[0080] According to formula (4) and formula (5), the superpixel update process can be written as:

[0081]

[0082] in is the polarization feature of superpixel m; is the position feature of superpixel m, p i，j is the position feature of pixel point (i, j), Φ represents the pixel feature before updating, Represents the probability value that pixel point (i, j) belongs to superpixel block m.

[0083] Combining the superpixel block features and the soft correlation matrix of superpixels and pixels, the polarization features of each pixel are obtained:

[0084]

[0085] in is the updated polarization feature of pixel point (i, j), is the polarization feature of superpixel m, Represents the probability value that pixel point (i, j) belongs to superpixel block m.

[0086] Combining the superpixel block features and the soft correlation matrix between superpixels and pixels, the position features of each pixel are obtained:

[0087]

[0088] According to formula (7) and formula (8), the pixel feature update process can be written as:

[0089]

[0090] in is the updated position feature of pixel point (i, j), is the position feature of superpixel m, Represents the probability value that pixel point (i, j) belongs to superpixel block m. Represents the updated pixel features.

[0091] The full convolutional network loss function for superpixel segmentation is:

[0092]

[0093] Where Φ represents the pixel feature before updating, Represents the features of the updated pixel, Represents the cross entropy loss function between the two;

[0094] in Represents the cross entropy loss function between the two.

[0095] Build a graph convolutional neural network: Its structure is: input layer → first graph convolution layer → second graph convolution layer → third graph convolution layer → softmax output layer, and the activation function of each graph convolution layer is the tanh function.

[0096] Constructing a convolutional neural network: Its structure is: input layer → first convolution layer → first pooling layer → second convolution layer → second pooling layer → softMax output layer. The activation function of each convolution layer is the LeakyRelu function.

[0097] The specific steps of step 6 are:

[0098] The training set is sent to the end-to-end network for training, and the test set is sent to the trained end-to-end network to obtain the classification accuracy of the model.

[0099] Step 6.1, initialize superpixel;

[0100] Initialize the superpixel. The superpixel initialization size is 16 pixels in height and width. Then the image with a length and width of 512 pixels is divided into 1024 superpixel blocks.

[0101] Step 6.2, obtaining the soft correlation matrix Q between superpixels and pixels;

[0102] The pixel features obtained from the training set in step 3 and step 4 are sent to the fully convolutional network to obtain the output result. The output result size is: (1, 9, 512, 512). We record the output matrix as Q, which is the soft association matrix of superpixels and pixels. Among them, 1 is the batchsize size, and 9 represents the probability value of the pixel belonging to the superpixel block where the pixel is located and the superpixel block where the pixel is located and its eight adjacent superpixel blocks (sitting on, directly above, upper right, left, right, lower left, directly below, and lower right). The sum of these 9 probability values ​​is 1.

[0103] Step 6.3, obtaining the adjacency matrix A, feature matrix B, and transformation matrix C between superpixels and pixels of the superpixel block through the superpixel and pixel soft association matrix;

[0104] According to the superpixel and pixel soft association matrix Q obtained in step 6.2, each pixel point is assigned to the surrounding superpixel block with the highest probability to obtain the superpixel segmentation result of the entire image. According to the superpixel segmentation result, the adjacency matrix A and feature matrix B of each image are obtained.

[0105] The size of the adjacency matrix A is (1024,1024), A i,j Represents the element in the i-th row and j-th column of the adjacency matrix A.

[0106]

[0107] The feature matrix B has a size of (1024,9), B i Represents the features of the i-th superpixel block; represents the feature of the jth pixel in superpixel block i, and n represents the number of pixels in the i-th superpixel block.

[0108]

[0109] The size of the transformation matrix C between superpixels and pixels is (512×512,1024), C i,j Represents the element in the i-th row and j-th column of the superpixel and pixel transformation matrix C.

[0110]

[0111] Where i represents the pixel index of the image, and j represents the superpixel block index of the image. i∈[1,512×512]; j∈[1,1024].

[0112] Step 6.4, graph convolution obtains the superpixel features of the image;

[0113] The adjacency matrix and feature matrix of each image are input into the graph convolutional network. The input is the adjacency matrix and feature matrix of the image. The output is a two-dimensional tensor H, which is the superpixel feature of the image; the tensor size is (1024, 64), where the first dimension 1024 represents that the image has 1024 superpixel blocks, and the second dimension 64 represents the features of each superpixel block.

[0114] Step 6.5, converting the superpixel features of the image into pixel features;

[0115] H gcn =C·H (14)

[0116] Where C represents the transformation matrix between superpixels and pixels obtained in step 6.3, and H represents the output of the graph convolutional network, with a size of (1024, 64). gcn The superpixel features output by the graph convolutional network are converted into pixel features through the transformation matrix of superpixels and pixels, and the size is: (512×512,64).

[0117] Step 6.6, the convolutional neural network obtains the pixel-level features of the image;

[0118] The polarization features of the training set obtained in step 3 are sent as input to the convolutional neural network to obtain pixel-level features;

[0119] The input of the convolutional neural network is a four-dimensional tensor of size (1,512,512,9). The first dimension 1 of the input data represents the batch size; the second and third dimensions represent the width and height of the image; and the fourth dimension 9 represents the polarization feature of each pixel in the image.

[0120] The output of the convolutional neural network is a four-dimensional tensor of size: (1, 256, 256, 64). The fourth dimension 64 represents the features extracted from each pixel.

[0121] Step 6.7, fusing the image pixel-level features obtained in step 6.6 and the pixel-level features obtained in step 6.5 through the transformation matrix of superpixels and pixels for classification;

[0122] The superpixel feature size obtained in step 6.5 is (512×512,64); the pixel feature size obtained in step 6.6 is (512×512,64). The pixel-level features and superpixel-level features are fused to obtain the final feature size of the image: (512×512,128), and finally softmax classification is performed. The classification result is obtained and the classification loss loss2 is calculated. The classification loss here uses the cross entropy loss function.

[0123] In step 6.8, calculate the total loss function and iteratively update the network in reverse until convergence.

[0124] The total loss function of the end-to-end network is:

[0125] loss = loss1 + loss2 (15)

[0126] Among them, loss1 is the full convolutional network loss function of superpixel segmentation; loss2 is the classification loss function. Finally, the network is iteratively updated until the network converges.

[0127] In step 6.9, the test set is fed into the trained end-to-end network model to obtain the classification result.

Claims

1. An end-to-end polarimetric SAR image classification method based on superpixels and graph convolution, characterized in that: The following steps are involved: Step 1, input the polarimetric SAR image to be classified and crop it into a uniform size; Step 2: Divide the cropped images into training sets and test sets according to the proportion; Step 3, decomposing the complex scattering matrix of each pixel point of each image in the test set and the training set, generating a polarization coherence matrix and converting it into a row vector as the polarization feature of the pixel point; Step 4, splice the polarization feature with the horizontal and vertical coordinates of the pixel point, and splice the row vector as the pixel point feature; Step 5: Build an end-to-end network based on full convolutional network, graph convolutional network and convolutional neural network; Step 6: Send the training set to the end-to-end network for joint training, and send the test set to the trained end-to-end network to obtain the result; The specific steps of sending the training set into the end-to-end network for joint training are as follows: Step 6.1, initialize superpixel; Initialize superpixels. The initial size of superpixels is 16 pixels in height and width. Then, an image with a length and width of 512 pixels is divided into 1024 superpixel blocks. Step 6.2, obtaining the soft correlation matrix Q between superpixels and pixels; Send the pixel features of the training set obtained in steps 3 and 4 to the fully convolutional network to obtain the output result; Step 6.3, obtaining the adjacency matrix A, feature matrix B, and transformation matrix C between superpixels and pixels of the superpixel block through the superpixel and pixel soft association matrix; According to the superpixel and pixel soft correlation matrix Q obtained in step 6.2, each pixel point is assigned to the surrounding superpixel block with the highest probability to obtain the superpixel segmentation result of the whole image; Obtain the adjacency matrix A and feature matrix B of each image based on the superpixel segmentation results; The size of the adjacency matrix A is (1024, 1024), A i，j Represents the element in the i-th row and j-th column of the adjacency matrix A; The feature matrix B has a size of (1024, 9), x Represents the features of the x-th superpixel block; represents the feature of the yth pixel in the superpixel block x, and n represents the number of pixels in the xth superpixel block; The size of the transformation matrix C between superpixels and pixels is (512×512, 1024), C p，q represents the element in the p-th row and q-th column of the superpixel and pixel transformation matrix C; in, p represents the pixel subscript of the image, q represents the superpixel subscript of the image; p∈[1,512×512]; q∈[1,1024]; Step 6.4, graph convolution obtains the superpixel features of the image; The adjacency matrix and feature matrix of each image are input into the graph convolutional network. The input is the adjacency matrix and feature matrix of the image; the output is a two-dimensional tensor H, which is the superpixel feature of the image; Step 6.5, converting the superpixel features of the image into pixel features; H gcn =C·H; Where C represents the transformation matrix between superpixels and pixels obtained in step 6.3, and H represents the output of the graph convolutional network; Step 6.6, the convolutional neural network obtains the pixel-level features of the image; The polarization features of the training set obtained in step 3 are sent as input to the convolutional neural network to obtain pixel-level features; Step 6.7, fusing the image pixel-level features obtained in step 6.6 and the pixel-level features obtained in step 6.5 through the transformation matrix of superpixels and pixels for classification; In step 6.8, calculate the total loss function and iteratively update the network in reverse until convergence.

2. The end-to-end polarimetric SAR image classification method based on superpixels and graph convolution as claimed in claim 1, characterized in that: The specific steps of step 3 are: The complex scattering matrix of each pixel point of each image cropped in step 2 is decomposed to generate a polarization coherence matrix and convert it into a row vector of size 1×9 as the polarization feature of the pixel point; the expression of the generated polarization coherence matrix T is as follows: The polarization coherence matrix T is denoted as Then the polarization coherence matrix is ​​converted into a row vector to obtain the characteristic matrix: T′=[T 11 , T 12 , T 13 , T 21 , T 22 , T 23 , T 31 , T 32 , T 33 ], and the polarization coherence matrix is ​​a complex conjugate matrix, and the complex conjugate matrix is ​​preprocessed to obtain the polarization eigenvector F of the pixel point: F=(T 11 ,T 22 ,T 33 ,Re[T 12 ],Re[T 13 ],Re[T 23 ],Im[T 21 ],Im「T 31 ],Im[T 32 ]) (2) Where Re represents the real part of the complex number, Im represents the imaginary part of the complex number, and T ij Represents the data in the i-th row and j-th column of the polarization coherence matrix.

3. The end-to-end polarimetric SAR image classification method based on superpixels and graph convolution as claimed in claim 1, characterized in that: The step 4 specifically includes: splicing the polarization feature vector of each pixel point generated in step 3 with the horizontal and vertical coordinates of the pixel point to obtain the pixel point feature.

4. The end-to-end polarimetric SAR image classification method based on superpixels and graph convolution as claimed in claim 1, characterized in that: The structure of the fully convolutional network in step 5 is a once-connected input layer, a first downsampling layer, a second downsampling layer, a third downsampling layer, a fourth downsampling layer, a fifth downsampling layer, a first upsampling layer, a second upsampling layer, a third upsampling layer, a fourth upsampling layer and a softMax output layer. The loss function of the fully convolutional network is: Where Φ represents the pixel feature before updating, Represents the features of the updated pixel, Represents the cross entropy loss function between the two; The structure of the graph convolutional neural network is an input layer, a first graph convolutional layer, a second graph convolutional layer, a third graph convolutional layer and a softmax output layer connected in sequence, and the activation function of each graph convolutional layer is a tanh function; The convolutional neural network structure is an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer and a softMax output layer connected in sequence, and the activation function of each convolutional layer is a LeakyRelu function.

5. The end-to-end polarimetric SAR image classification method based on superpixels and graph convolution as claimed in claim 1, characterized in that: The specific steps of step 6 are: Step 6.1, initialize the training set and test set into superpixel blocks; Step 6.2, the pixel features of the training set obtained in steps 3 and 4 are sent to the full convolutional network to obtain the output result. The output matrix Q is the soft correlation matrix of superpixels and pixels; Step 6.3, obtaining the adjacency matrix A, feature matrix B, and transformation matrix C between superpixels and pixels of the superpixel block through the superpixel and pixel soft association matrix; According to the superpixels and pixel soft association matrix Q obtained in step 6.2, each pixel point is assigned to the surrounding superpixel blocks with the highest probability to obtain the superpixel segmentation result of the entire image, and the adjacency matrix A and feature matrix B of each image are obtained according to the superpixel segmentation result; Where A in the adjacency matrix A i，j Represents the element in the i-th row and j-th column of the adjacency matrix A: Feature matrix B x Represents the features of the x-th superpixel block; represents the feature of the yth pixel in the superpixel block x, and n represents the number of pixels in the xth superpixel block; The transformation matrix between superpixels and pixels is the transformation matrix C; Step 6.4, input the adjacency matrix and feature matrix of each image into the graph convolutional network. The input is the adjacency matrix and feature matrix of the image, and the output is a two-dimensional tensor H which is the superpixel feature of the image. Step 6.5, convert the superpixel features of the image into pixel features: H gcn =C·H (14); Where C represents the transformation matrix, H represents the output of the graph convolutional network; H gcn The pixel features are obtained by converting the superpixel features output by the graph convolutional network through the transformation matrix between superpixels and pixels; Step 6.6, the convolutional neural network obtains the pixel-level features of the image; The polarization features of the training set obtained in step 3 are sent as input to the convolutional neural network to obtain pixel-level features; Step 6.7, fusing the image pixel-level features obtained in step 6.6 and the pixel-level features obtained in step 6.5 through the transformation matrix of superpixels and pixels for classification; Step 6.8, calculate the total loss function and iterate and update the network backward until convergence; The total loss function of the end-to-end network is: loss = loss1 + loss2 (15); Among them, loss1 is the full convolutional network loss function of superpixel segmentation; loss2 is the classification loss function; finally, the network is iteratively updated until the network converges, and the end-to-end network model training is completed; In step 6.9, the test set is fed into the trained end-to-end network model to obtain the classification result.

Images