Polarimetric SAR image classification method and system, medium and program product

Abstract

The invention discloses a polarized SAR image classification method and system, a medium and a program product, and relates to the field of polarized SAR image classification, and the method comprises the steps: obtaining a polarized data set; constructing an anchoring sample training set and a positive sample pair data set according to the polarization data set; constructing a residual neural network model; performing first-stage training on the residual neural network model by using the positive sample pair data set to obtain a middle residual neural network model which passes through the first-stage training; performing second-stage training on the intermediate residual neural network model by using the polarization data set and the anchoring sample training set to obtain a residual neural network model through the second-stage training; and clustering the to-be-classified polarized SAR images by using the residual neural network model trained in the second stage and generating a classification result graph. According to the method, the polarimetric SAR image can be effectively and accurately classified and identified under the condition that any marking cost is not needed.

Description

Technical Field

[0001] The present invention relates to the technical field of polarimetric synthetic aperture radar image interpretation, and particularly to a polarimetric SAR image classification method, system, medium and program product. Background Art

[0002] In the process of polarimetric SAR image analysis and interpretation, the classification of ground objects in polarimetric SAR images has been widely studied as a key technology. Currently, many traditional methods for polarimetric SAR image classification have been developed, such as Wishart classifier, target decomposition theory, support vector machine (SVM), and random field model-based methods. However, these traditional methods rely on manually extracted low-level features, and the suboptimal classification results generated by these features with low recognition ability can no longer meet the current high-precision ground object classification requirements. For example, although the Wishart classifier has high computational efficiency, its classification accuracy depends severely on the accuracy of the scattering model; although SVM performs well on high-dimensional data, its computational complexity is high and it is difficult to apply to large-scale PolSAR image classification tasks.

[0003] In recent years, polarimetric SAR image classification technology based on deep learning has become the mainstream and achieved high-precision classification results under the guidance of sufficient manually annotated data. However, the training of these supervised models requires a large number of labeled samples, which is time-consuming and laborious for large-scale scenes and high-resolution PolSAR images. Although a series of subsequent semi-supervised methods have been proposed to solve the over-reliance on annotation information to a certain extent, some target class labels are still required to assist in discriminative classification. In the field of image analysis, unsupervised learning methods have been widely studied because they do not require any labeled data. For PolSAR data, early unsupervised methods mainly relied on simple clustering based on scattering mechanisms. The most classic examples include Wishart clustering based on target decomposition (such as Freeman decomposition, H / A decomposition). Although these methods are fast and efficient in calculation, due to their dependence on simple scattering models, the classification results are often not ideal. With the rapid development of deep learning technology, researchers have begun to introduce end-to-end deep learning networks into unsupervised PolSAR classification, and the autoencoder-decoder network is one of the most commonly used methods. These methods compare the similarity between the decoded features and the input features through distance metrics (such as Wishart distance) to learn discriminative latent representations for classification.

[0004] Currently, the classification of polarimetric SAR images with complex scattering characteristics faces the following problems: (1) Traditional methods based on scattering mechanisms rely on simple scattering models, and the classification accuracy is difficult to meet the actual needs; (2) Due to the significant differences between PolSAR data and general image data, unsupervised methods in the field of natural image processing are difficult to directly apply, resulting in relatively few deep unsupervised learning and clustering methods for PolSAR images; (3) Currently, the mainstream unsupervised PolSAR image classification methods based on autoencoder-decoder networks have serious semantic discontinuity problems during the classification process, and additional classification constraints usually need to be introduced to ensure performance. Summary of the Invention

[0005] To solve the technical problems in the background art, the present invention proposes a polarimetric SAR image classification method, system, medium, and program product.

[0006] In a first aspect, a polarimetric SAR image classification method proposed by the present invention includes:

[0007] Obtain a polarimetric dataset; wherein, the polarimetric dataset includes polarimetric SAR images and corresponding ground truth maps; construct an anchor sample training set and a positive sample pair dataset according to the polarimetric dataset;

[0008] Construct a residual neural network model;

[0009] Use the positive sample pair dataset to perform the first-stage training on the residual neural network model to obtain an intermediate residual neural network model that has passed the first-stage training; use the polarimetric dataset and the anchor sample training set to perform the second-stage training on the intermediate residual neural network model to obtain a residual neural network model that has passed the second-stage training;

[0010] Use the residual neural network model that has passed the second-stage training to cluster the polarimetric SAR image to be classified to obtain a polarimetric SAR image clustering result; use the Hungarian matching algorithm to map the polarimetric SAR image clustering result to the true label to obtain a mapping result; generate a classification result map according to the mapping result.

[0011] Preferably, constructing an anchor sample training set and a positive sample pair dataset according to the polarimetric data specifically includes: constructing an anchor sample training set according to the polarimetric dataset; constructing a positive sample pair dataset according to the anchor sample training set.

[0012] Preferably, an anchored sample training set is constructed according to the polarization data set, specifically including: obtaining a complex-valued feature matrix of the polarimetric SAR image according to the polarimetric SAR image; wherein, the complex-valued feature matrix includes the complex-valued feature vectors of each pixel; generating labeled data according to the ground truth map; constructing an initial ground truth matrix, and performing pixel-level label classification and annotation on the initial ground truth matrix according to the labeled data to generate a ground truth matrix; obtaining the complex-valued feature vectors corresponding to the coordinate values in the complex-valued feature matrix as the first complex-valued feature vectors according to the coordinate values of the labeled pixels in the ground truth matrix; taking the first complex-valued feature vectors as the center points, and respectively selecting a plurality of complex-valued feature vectors around as the second complex-valued feature vectors; combining the first complex-valued feature vectors and the second complex-valued feature vectors to form a complex-valued feature matrix block; separating the real and imaginary parts of the upper triangular elements in the complex-valued feature vectors in the complex-valued feature matrix block to form real-valued feature vectors; generating a high-dimensional real-valued feature matrix based on the real-valued feature vectors; and correspondingly combining the pixel-level label classification and annotation information in the ground truth matrix and the high-dimensional real-valued feature matrix block to form an anchored sample training set.

[0013] Preferably, a positive sample pair data set is constructed according to the anchored sample training set, specifically including: performing random data augmentation on the anchored sample training set to generate positive samples with the same semantics as the anchored samples in the anchored sample training set, and combining the anchored samples with their corresponding positive samples to form positive sample pairs; and combining all the positive sample pairs to form a positive sample pair data set.

[0014] Preferably, the residual neural network model includes an input layer, a residual layer, a global average pooling layer, and a clustering head; wherein, the input layer includes a first convolutional layer and a first batch normalization layer, the number of input feature channels of the first convolutional layer is 9, which is the same as the real-valued feature matrix of the sample, the number of feature maps of the output feature data is 64, the size of the convolutional filter in the first convolutional layer is 3×3, the padding is 1, and the stride is 1; wherein, the residual layer includes 4 residual structure levels, each level includes a number of residual blocks, and each residual block includes two convolutional layers and a shortcut connection layer; wherein, the global average pooling layer is used to compress the spatial dimension of the feature map to 1×1 using adaptive average pooling and the output feature dimension is 512; wherein, the clustering head is used to map the 512-dimensional feature to an n-dimensional space using a linear layer to obtain the probability distribution of each sample belonging to n clustering categories.

[0015] Preferably, the four residual structure levels are layer1, layer2, layer3, and layer4 in sequence; among them, layer1 includes a first residual block and a second residual block, and the number of input and output channels of both residual blocks is 64. The first residual block includes a second convolutional layer, a third convolutional layer, and a second batch normalization layer, and the second residual block includes a fourth convolutional layer, a fifth convolutional layer, and a third batch normalization layer. The sizes of the convolutional filters in the second, third, fourth, and fifth convolutional layers are all 3×3, the padding is 1, the stride is 1, and there is no downsampling operation;

[0016] layer2 includes a third residual block and a fourth residual block. The third residual block is used to expand the number of channels from 64 to 128. The third residual block includes a sixth convolutional layer, a seventh convolutional layer, and a fourth batch normalization layer. The sizes of the convolutional filters in the sixth and seventh convolutional layers are 3×3, the padding is 1, the stride is 1, and there is no downsampling operation; at the same time, the output of layer1 and the output of layer2 are channel-matched and summed through a first shortcut connection layer as the input of layer3; the first shortcut connection layer includes an eighth convolutional layer and a fifth batch normalization layer. The number of input feature channels of the eighth convolutional layer is 64, and the number of feature maps of the output feature data is 128. The size of the convolutional filter in the eighth convolutional layer is 1×1, there is no padding, and the stride is 1; the fourth residual block maintains the number of channels at 128. The fourth residual block includes a ninth convolutional layer, a tenth convolutional layer, and a sixth batch normalization layer. The sizes of the convolutional filters in the ninth and tenth convolutional layers are 3×3, the padding is 1, the stride is 1, and there is no downsampling operation;

[0017] layer3 includes a fifth residual block and a sixth residual block; the fifth residual block expands the number of channels from 128 to 256. The fifth residual block includes an eleventh convolutional layer, a twelfth convolutional layer, and a seventh batch normalization layer. The size of the convolutional filter in the eleventh convolutional layer is 3×3, the padding is 1, the stride is 2 for downsampling operation, and the size of the convolutional filter in the twelfth convolutional layer is 3×3, the padding is 1, the stride is 1; at the same time, the output of layer2 and the output of layer3 are channel-matched and summed through a second shortcut connection layer as the input of layer4; the second shortcut connection layer includes a thirteenth convolutional layer and an eighth batch normalization layer. The number of input feature channels of the thirteenth convolutional layer is 128, the number of feature maps of the output feature data is 256, and the size of the convolutional filter is 1×1, there is no padding, and the stride is 1; the sixth residual block maintains the number of channels at 256. The sixth residual block includes a fourteenth convolutional layer, a fifteenth convolutional layer, and a ninth batch normalization layer. The sizes of the convolutional filters in the fourteenth and fifteenth convolutional layers are both 3×3, the padding is 1, the stride is 1, and there is no downsampling operation;

[0018] Layer 4 includes the seventh residual block and the eighth residual block. The seventh residual block expands the number of channels from 128 to 256. The seventh residual block contains the sixteenth convolutional layer, the seventeenth convolutional layer, and the tenth batch normalization layer. The sizes of the convolutional layer filters in the sixteenth convolutional layer and the seventeenth convolutional layer are both 3×3, the padding is 1, and the stride is 1. At the same time, the output of layer 3 and the output of layer 4 are channel-matched and summed through the third shortcut connection layer as the final output. The third shortcut connection layer contains the eighteenth convolutional layer and the eleventh batch normalization layer. The number of input feature channels of the eighteenth convolutional layer is 128, and the number of feature maps of the output feature data is 256. The size of the convolutional filter is 1×1, there is no padding, and the stride is 1. The eighth residual block maintains the number of channels at 256. The eighth residual block includes the nineteenth convolutional layer, the twentieth convolutional layer, and the eleventh normalization layer. The sizes of the convolutional filters in the nineteenth convolutional layer and the twentieth convolutional layer are both 3×3, the padding is 1, the stride is 1, and there is no downsampling operation.

[0019] Preferably, the residual neural network model is trained in the first stage using the positive sample pair dataset to obtain an intermediate residual neural network model trained in the first stage, specifically including: inputting the positive sample pair dataset into the residual neural network model to obtain the anchor sample features and their positive sample features; calculating the feature similarity between the anchor sample and the positive sample based on the anchor sample features and their positive sample features, and calculating the contrast loss between the anchor sample and the positive sample based on the feature similarity between the anchor sample and the positive sample; repeating the optimization of the parameters of the residual neural network model according to the contrast loss between the anchor sample and the positive sample through the backpropagation algorithm until the residual neural network model converges or reaches a preset number of training epochs; using the residual neural network model that converges or reaches the preset number of training epochs as the intermediate residual neural network model trained in the first stage.

[0020] Preferably, the intermediate residual neural network model is trained in the second stage using the polarized data and the anchor sample training set to obtain a residual neural network model trained in the second stage, specifically including:

[0021] Based on the polarization dataset, a Pauli image set is obtained; among them, the superpixel block set contains multiple superpixel blocks, and each superpixel block includes multiple pixel point coordinates and corresponding superpixel block index information; according to the superpixel block index information and the coordinates of the central pixel point of each anchor sample in the anchor sample dataset, the superpixel block to which each anchor sample belongs is determined; taking all non-edge region pixels of the superpixel block to which each anchor sample belongs as the center, a sample block set of the same size as the anchor sample is taken as the sample candidate set of the nearest neighbor sample cluster of the anchor sample; the samples in the candidate set corresponding to each anchor sample are input into the intermediate residual neural network model to obtain the candidate feature library of the nearest neighbor sample cluster of each anchor sample; each anchor sample is input into the intermediate residual neural network model to obtain the anchor feature; the Euclidean distance between the anchor feature of each anchor sample and all feature tensors in the candidate feature library is calculated for each anchor sample; according to the Euclidean distance, the K nearest neighbor samples with the closest distance are selected for each anchor sample; the K nearest neighbor samples selected for each anchor sample are combined to form the nearest neighbor sample cluster of each anchor sample;

[0022] According to the nearest neighbor sample cluster of each anchor sample, the cluster feature contrast loss and the soft assignment contrast loss are calculated; a total loss function is constructed according to the cluster feature contrast loss and the soft assignment contrast loss; according to the cluster feature contrast loss and the soft assignment contrast loss, the parameters of the intermediate residual neural network model are optimized by the backpropagation algorithm until the intermediate residual neural network model converges or reaches the preset number of optimization times; the intermediate residual neural network model that converges or reaches the preset number of optimization times is used as the residual neural network model trained in the second stage.

[0023] Preferably, according to the nearest neighbor sample cluster of each anchor sample, the cluster feature contrast loss and the soft assignment contrast loss are calculated, specifically including: according to the similar semantic relationship between the nearest neighbor sample clusters of each anchor sample, the contrast loss is used to calculate the cluster feature contrast loss of the nearest neighbor sample cluster; the soft assignment probability of the intermediate residual neural network model for the nearest neighbor sample cluster is obtained through the Softmax classifier, and the cross-entropy loss function is used to calculate the soft assignment contrast loss.

[0024] Preferably, In the formula, represents the total loss, represents the soft assignment loss, represents the cluster feature contrast loss, and λ represents the weight coefficient;

[0025]

[0026] In the formula, represents the i-th anchor sample x i and its nearest neighbor sample cluster Positive sample pairs composed of any nearest neighbor samples, Group represents the nearest neighbor sample cluster, x k represents a sample that does not belong to the i-th anchor sample x i Any sample in the nearest neighbor sample cluster, that is, x i is the negative sample of represents a positive sample pair After the feature representation by the residual neural network model, (z i , z k ) represents the feature representation of the negative sample pair (x i , x k ) after passing through the residual neural network model, τ represents the temperature coefficient;

[0027] In the formula, N represents the total number of anchor samples, N x represents the number of nearest neighbor samples in each nearest neighbor sample cluster, p i represents the feature representation z i of the anchor sample x i The predicted vector output by passing through the softmax layer, p j represents the nearest neighbor sample cluster of the anchor sample x i Any nearest neighbor sample x j The predicted vector output by the feature representation z j of passing through the softmax layer, · represents the dot product operator, represents the probability that the sample x i is assigned to the cluster c, c represents a certain cluster category, C represents the set of dataset categories, and β represents the weight coefficient of the entropy term

[0028] In a second aspect, the present invention also provides a polarimetric SAR image classification system, including a memory, a processor, and a computer program stored on the memory. The processor executes the computer program to implement the steps of the polarimetric SAR image classification method described in any one of the first aspects.

[0029] In a third aspect, the present invention also proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the polarimetric SAR image classification method described in any one of the first aspects.

[0030] In a fourth aspect, the present invention also proposes a computer program product, including a computer program. When the computer program is executed by a processor, it implements the steps of the polarimetric SAR image classification method described in any one of the first aspects.

[0031] ​​In the present invention, a proposed polarization SAR image classification method, system, medium, and program product construct an anchor sample training set and a positive sample pair dataset according to a polarization dataset, and use the anchor sample training set and the positive sample pair dataset to perform two-stage training on a constructed residual neural network model to obtain a residual neural network model that passes the two-stage training; then use the residual neural network model that passes the second-stage training to cluster the polarization SAR image to be classified to obtain a polarization SAR image clustering result; use the Hungarian matching algorithm to map the polarization SAR image clustering result to a true label to obtain a mapping result; and generate a classification result map according to the mapping result. The present invention avoids the problems of a large dependence on labeled data in existing polarization SAR classification technologies and the accuracy problems of traditional unsupervised algorithms, and can effectively and accurately classify and identify polarization SAR images without any labeling cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic flowchart of a polarization SAR image classification method in an embodiment proposed by the present invention.

[0033] Figure 2 It is a polarization SAR image in an embodiment proposed by the present invention.

[0034] Figure 3 is Figure 2 corresponding true ground object marking map.

[0035] Figure 4 is Figure 2 superpixel segmentation schematic diagram of.

[0036] Figure 5 It is a classification result map obtained by using the Wishart clustering method based on target decomposition in the prior art for classification.

[0037] Figure 6 It is a classification result map obtained by using the polarization SAR classification method based on an autoencoder-decoder network in the prior art for classification.

[0038] Figure 7 It is a classification result map obtained by using the polarization SAR image classification method proposed by the present invention for classification. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments may be combined with each other. The present invention will be described in detail below with reference to the drawings and in combination with the embodiments.

[0040] First aspect, referring to Figure 1 , a polarization SAR image classification method proposed by the present invention includes:

[0041] Obtain a polarization dataset; wherein, the polarization dataset includes polarization SAR images and corresponding true ground object labeling maps; construct an anchor sample training set and a positive sample pair dataset according to the polarization dataset;

[0042] Construct a Residual Neural Network (ResNet) model; perform the first-stage training on the ResNet model using the positive sample pair dataset to obtain an intermediate ResNet model that has passed the first-stage training; perform the second-stage training on the intermediate ResNet model using the polarization dataset and the anchor sample training set to obtain a ResNet model that has passed the second-stage training;

[0043] Use the ResNet model that has passed the second-stage training to cluster the polarization SAR image to be classified to obtain a polarization SAR image clustering result; use the Hungarian matching algorithm to map the polarization SAR image clustering result to the true label to obtain a mapping result; generate a classification result map according to the mapping result.

[0044] In the present invention, an anchor sample training set and a positive sample pair dataset are constructed according to the polarization dataset, and the constructed ResNet model is trained in two stages using the anchor sample training set and the positive sample pair dataset to obtain a ResNet model that has passed the two-stage training; then use the ResNet model that has passed the second-stage training to cluster the polarization SAR image to be classified to obtain a polarization SAR image clustering result; use the Hungarian matching algorithm to map the polarization SAR image clustering result to the true label to obtain a mapping result; generate a classification result map according to the mapping result. The present invention avoids the problems of heavy dependence on labeled data in existing polarization SAR classification technologies and the accuracy problems of traditional unsupervised algorithms, and can effectively and accurately classify and identify polarization SAR images without any labeling cost.

[0045] In this embodiment, constructing an anchor sample training set and a positive sample pair dataset according to the polarization data specifically includes: constructing an anchor sample training set according to the polarization dataset; constructing a positive sample pair dataset according to the anchor sample training set.

[0046] In a further embodiment, an anchored sample training set is constructed based on the polarization dataset, specifically including: obtaining a complex-valued feature matrix of the polarimetric SAR image from the polarimetric SAR image; wherein, the complex-valued feature matrix includes complex-valued feature vectors of each pixel; generating labeled data according to the ground truth map; constructing an initial ground truth matrix, and performing pixel-level label classification and annotation on the initial ground truth matrix according to the labeled data to generate a ground truth matrix; obtaining the complex-valued feature vector corresponding to the coordinate value in the complex-valued feature matrix as the first complex-valued feature vector according to the coordinate value of the labeled pixel in the ground truth matrix; taking the first complex-valued feature vector as the center point, and respectively selecting a number of complex-valued feature vectors around it as the second complex-valued feature vectors; combining the first complex-valued feature vector and the second complex-valued feature vectors into a complex-valued feature matrix block; separating the real part and the imaginary part of the upper triangular elements in the complex-valued feature vectors in the complex-valued feature matrix block to form real-valued feature vectors; generating a high-dimensional real-valued feature matrix based on the real-valued feature vectors; performing corresponding combination on the pixel-level label classification and annotation information in the ground truth matrix and the high-dimensional real-valued feature matrix block to form an anchored sample training set.

[0047] Wherein, obtaining a complex-valued feature matrix of the polarimetric SAR image from the polarimetric SAR image specifically includes: calculating the polarimetric complex-valued coherence matrix of each pixel point in the preprocessed polarimetric SAR image according to the polarimetric SAR image; calculating the mean and variance of each element in the polarimetric complex-valued coherence matrix of each pixel point, and calculating the normalized value of each element in the polarimetric complex-valued coherence matrix of each pixel point according to the calculation results of the mean and variance, and integrating the normalized values of each element to generate the normalized polarimetric complex-valued coherence matrix of each pixel point; extracting the normalized values of several upper triangular elements in the normalized polarimetric complex-valued coherence matrix of each pixel point as complex-valued feature vectors; integrating the complex-valued feature vectors to generate a complex-valued feature matrix of the polarimetric SAR image.

[0048] Wherein, In the formula, T (ij) represents the polarimetric complex-valued coherence matrix corresponding to the pixel point in the i-th row and j-th column in the polarimetric SAR image, i = 1, 2,..., a, j = 1, 2,..., b; wherein, a and b respectively represent the length and width of the polarimetric SAR image, represents the element in the u-th row and v-th column in the polarimetric coherence matrix, u, v ∈ {1, 2, 3} and except for the diagonal elements, the rest of the elements in the matrix are complex-valued.

[0049] Wherein, In the formula, · represents the multiplication operation, represents the average value of each element in the polarimetric complex-valued coherence matrix of each pixel point in the polarimetric SAR image.

[0050] Among them, In the formula, represents the conjugate operation, represents the variance value obtained by each element in the polarization complex coherence matrix of each pixel of the polarimetric SAR image with respect to the entire polarization complex coherence matrix.

[0051] Among them, In the formula, represents the normalization value corresponding to each element in the polarization complex coherence matrix T corresponding to each pixel of the polarimetric SAR image (ij) in.

[0052] Among them, In the formula, T (ij) represents the normalized polarization complex coherence matrix corresponding to the pixel at the i-th row and j-th column of the polarimetric SAR image.

[0053] Since the elements other than the diagonal elements in the normalized polarization coherence matrix corresponding to each pixel in the polarimetric SAR image are all complex values and have the conjugate transpose identity property, in order to facilitate network processing and reduce the network feature operation complexity, in this example, during the process of generating the high-dimensional real-valued feature matrix, the 6 elements in the upper triangle of the complex-valued feature matrix block are selected for real and imaginary part separation to form the 9-dimensional real-valued feature vector of this pixel; among them, In the formula, I (ij) represents the 9-dimensional real-valued feature vector of each pixel of the polarimetric SAR image, i = 1, 2,..., a, j = 1, 2,..., b; then, according to the 9-dimensional real-valued feature vector of each pixel of the polarimetric SAR image, a high-dimensional real-valued feature matrix of the polarimetric SAR image is constructed; among them, F = (I (ij) ) a×b High-dimensional real-valued feature matrix of polarimetric SAR image.

[0054] In order to ensure the randomness and diversity of sample pairs during the training process, in a further embodiment, a positive sample pair data set is constructed according to the anchor sample training set, specifically including: performing random data augmentation on the anchor sample training set to generate positive samples with the same semantics as the anchor samples in the anchor sample training set, and combining the anchor samples with their corresponding positive samples to form positive sample pairs; combining all the positive sample pairs to form a positive sample pair data set.

[0055] Among them, performing random data augmentation on the anchor sample training set specifically includes: performing random data augmentation on the anchor samples in the anchor sample training set using a random data augmentation strategy; among them, the random data augmentation strategy includes random rotation, flipping, and random noise.

[0056] Specifically, the anchor sample data set is defined as D anchor ={x1,...,x|D|}, where x i represents the high-dimensional real-valued feature matrix of each pixel block sample, with a dimension of 21 * 21 * 9. The random data augmentation strategy is selected from random rotation, flipping, and random noise, and can be defined as T = random{Rotation(θ), Flipping(α), Noise(μ, σ)}, where θ is a randomly generated angle parameter with a range of -180° to 180°; α is a random boolean value controlling horizontal or vertical flipping; μ, σ are the mean and variance parameters for adding random Gaussian noise. For each anchor sample, a positive sample T[x i can be generated through any random data augmentation operation in T to form a positive sample pair dataset.

[0057] Of course, in different embodiments, the intensity of data augmentation can be adjusted according to actual task requirements. For example, the range of the rotation angle (such as -90° to 90°) or the variance of the noise (such as σ = 0.02) can be restricted to further improve the quality of the anchor sample dataset.

[0058] To adapt the existing ResNet network structure to the polarimetric SAR pixel block input, in this embodiment, the residual neural network model includes an input layer, a residual layer, a global average pooling layer, and a clustering head; among them, the input layer includes a first convolutional layer and a first batch normalization layer. The number of input feature channels of the first convolutional layer is 9, the same as the sample real-valued feature matrix, the number of feature map numbers of the output feature data is 64, the size of the convolutional filter in the first convolutional layer is 3×3, the padding is 1, and the stride is 1; among them, the residual layer includes 4 residual structure levels, each level includes several residual blocks (BasicBlock), and each residual block includes two convolutional layers and a shortcut connection (Shortcut Connection); among them, the global average pooling layer uses adaptive average pooling (AdaptiveAvgPool2d) to compress the spatial dimension of the feature map to 1×1, and the output feature dimension is 512; among them, the clustering head (Cluster Head) uses a linear layer (Linear) to map the 512-dimensional feature to an n-dimensional space (the number of clustering categories is n, which is 5 in this embodiment), representing the probability distribution of each sample belonging to n clustering categories.

[0059] In a further embodiment, the four residual structure levels are layer1, layer2, layer3, and layer4 in sequence; among them, layer1 includes a first residual block and a second residual block. The number of input and output channels of both residual blocks is 64. The first residual block includes a second convolutional layer, a third convolutional layer, and a second batch normalization layer. The second residual block includes a fourth convolutional layer, a fifth convolutional layer, and a third batch normalization layer. The sizes of the convolutional filters in the second, third, fourth, and fifth convolutional layers are all 3×3, the padding is all 1, the stride is all 1, and there is no downsampling operation;

[0060] layer2 includes a third residual block and a fourth residual block. The third residual block is used to expand the number of channels from 64 to 128. The third residual block includes a sixth convolutional layer, a seventh convolutional layer, and a fourth batch normalization layer. The sizes of the convolutional filters in the sixth and seventh convolutional layers are 3×3, the padding is 1, the stride is 1, and there is no downsampling operation; at the same time, the output of layer1 and the output of layer2 are channel-matched and summed through a first shortcut connection layer as the input of layer3; the first shortcut connection layer includes an eighth convolutional layer and a fifth batch normalization layer. The number of input feature channels of the eighth convolutional layer is 64, and the number of feature maps of the output feature data is 128. The size of the convolutional filter in the eighth convolutional layer is 1×1, there is no padding, and the stride is 1; the fourth residual block maintains the number of channels at 128. The fourth residual block includes a ninth convolutional layer, a tenth convolutional layer, and a sixth batch normalization layer. The sizes of the convolutional filters in the ninth and tenth convolutional layers are 3×3, the padding is 1, the stride is 1, and there is no downsampling operation;

[0061] layer3 includes a fifth residual block and a sixth residual block; the fifth residual block expands the number of channels from 128 to 256. The fifth residual block includes an eleventh convolutional layer, a twelfth convolutional layer, and a seventh batch normalization layer. The size of the convolutional filter in the eleventh convolutional layer is 3×3, the padding is 1, the stride is 2 for downsampling operation, and the size of the convolutional filter in the twelfth convolutional layer is 3×3, the padding is 1, the stride is 1; at the same time, the output of layer2 and the output of layer3 are channel-matched and summed through a second shortcut connection layer as the input of layer4; the second shortcut connection layer includes a thirteenth convolutional layer and an eighth batch normalization layer. The number of input feature channels of the thirteenth convolutional layer is 128, the number of feature maps of the output feature data is 256, and the size of the convolutional filter is 1×1, there is no padding, and the stride is 1; the sixth residual block maintains the number of channels at 256. The sixth residual block includes a fourteenth convolutional layer, a fifteenth convolutional layer, and a ninth batch normalization layer. The sizes of the convolutional filters in the fourteenth and fifteenth convolutional layers are both 3×3, the padding is 1, the stride is 1, and there is no downsampling operation;

[0062] Layer 4 includes the seventh residual block and the eighth residual block. The seventh residual block expands the number of channels from 128 to 256. The seventh residual block contains the sixteenth convolutional layer, the seventeenth convolutional layer, and the tenth batch normalization layer. The sizes of the convolutional layer filters in the sixteenth convolutional layer and the seventeenth convolutional layer are both 3×3, the padding is 1, and the stride is 1. At the same time, the output of layer 3 and the output of layer 4 are channel-matched and summed through the third shortcut connection layer as the final output. The third shortcut connection layer contains the eighteenth convolutional layer and the eleventh batch normalization layer. The number of input feature channels of the eighteenth convolutional layer is 128, and the number of feature map of the output feature data is 256. The size of the convolutional filter is 1×1, there is no padding, and the stride is 1. The eighth residual block keeps the number of channels at 256. The eighth residual block includes the nineteenth convolutional layer, the twentieth convolutional layer, and the eleventh normalization layer. The sizes of the convolutional filters in the nineteenth convolutional layer and the twentieth convolutional layer are both 3×3, the padding is 1, the stride is 1, and there is no downsampling operation.

[0063] In this embodiment, the residual neural network model is trained in the first stage using the positive sample pair dataset to obtain an intermediate residual neural network model trained in the first stage, which specifically includes:

[0064] Input the positive sample pair dataset into the residual neural network model to obtain the anchor sample features and their positive sample features; calculate the feature similarity between the anchor sample and the positive sample according to the anchor sample features and their positive sample features, and calculate the contrast loss between the anchor sample and the positive sample according to the feature similarity between the anchor sample and the positive sample; according to the contrast loss between the anchor sample and the positive sample, and repeat the optimization of the parameters of the residual neural network model through the backpropagation algorithm until the residual neural network model converges or reaches the preset number of training rounds; use the residual neural network model that converges or reaches the preset number of training rounds as the intermediate residual neural network model trained in the first stage.

[0065] Among them, In the formula, L slr represents the feature similarity between the anchor sample and the positive sample, sim(u, v) represents the cosine similarity calculation function, τ represents the temperature parameter, N is the total number of samples, z i represents the feature representation of the anchor sample x i , represents the positive sample i after the anchor sample x has undergone random data augmentation, represents the feature representation of any other sample except the anchor sample x i and its positive sample , and τ represents the temperature coefficient.

[0066] In this embodiment, the parameters of the residual neural network model are repeatedly optimized through the backpropagation algorithm, enabling the residual neural network model to learn the feature representations of semantically similar samples. The learning objective can be expressed as min θ d(Φ θ (x i ), Φ θ (T[x i )), where Φ θ is the embedding function learned by the model in a self-supervised manner. During the repeated optimization process, the AdamW optimizer is used to optimize the ResNet network model. The learning rate is set to 0.01, and the weight decay coefficient is set to 0.0001 until the model converges or reaches the preset number of training epochs.

[0067] In this embodiment, the intermediate residual neural network model is trained in the second stage using the polarized data and the anchor sample training set to obtain the residual neural network model trained in the second stage, which specifically includes:

[0068] According to the polarized data set, a Pauli image set is obtained; the superpixel segmentation algorithm is used to segment each Pauli image in the Pauli image set to obtain a superpixel block set; among them, the superpixel block set contains multiple superpixel blocks, and each superpixel block includes multiple pixel point coordinates and corresponding superpixel block index information; according to the superpixel block index information and the coordinates of the central pixel point of each anchor sample in the anchor sample data set, the superpixel block to which each anchor sample belongs is determined; taking all non-edge region pixels of the superpixel block to which each anchor sample belongs as the center, a sample block set of the same size as the anchor sample is taken as the sample candidate set of the nearest neighbor sample cluster of the anchor sample; the samples in the candidate set corresponding to each anchor sample are input into the intermediate residual neural network model to obtain the candidate feature library of the nearest neighbor sample cluster of each anchor sample; each anchor sample is input into the intermediate residual neural network model to obtain the anchor feature; the Euclidean distance between the anchor feature of each anchor sample and all feature tensors in the candidate feature library is calculated; according to the Euclidean distance, the K nearest neighbor samples with the closest distance are selected for each anchor sample; the K nearest neighbor samples selected for each anchor sample are combined to form the nearest neighbor sample cluster of each anchor sample;

[0069] According to the nearest neighbor sample cluster of each anchor sample, the cluster feature contrast loss and the soft assignment contrast loss are calculated; the total loss function is constructed based on the cluster feature contrast loss and the soft assignment contrast loss; according to the cluster feature contrast loss and the soft assignment contrast loss, the parameters of the intermediate residual neural network model are optimized through the backpropagation algorithm until the intermediate residual neural network model converges or reaches the preset number of optimization times; the intermediate residual neural network model that converges or reaches the preset number of optimization times is used as the residual neural network model trained in the second stage.

[0070] In this embodiment, the PolSAR image is divided into multiple local regions with similar scattering characteristics through superpixel segmentation technology, and a nearest neighbor sample cluster under superpixel constraints is constructed in combination with a contrastive learning framework, so as to learn a discriminative feature representation and achieve high-precision unsupervised classification.

[0071] In a further embodiment, according to the polarization dataset, a Pauli image set is obtained, specifically including: extracting the intensity information of the HH, HV, and VV channels from the polarimetric SAR image to generate a Pauli image; performing normalization processing on the generated Pauli image to scale the pixel values to the range of [0, 255].

[0072] The specific formula is: R = |HH - W|, G = 2|HV|, B = |HH + VV|.

[0073] In a further embodiment, each Pauli image in the Pauli image set is segmented using a superpixel segmentation algorithm to obtain a set of superpixel blocks, specifically including: setting the number K of superpixels and calculating the initial grid size of each superpixel according to the size of the Pauli image; uniformly initializing K clustering centers on the Pauli image, and the position of each clustering center is the center point of the grid; for each pixel, calculating the distance between it and the clustering centers within the surrounding 2S×2S area; according to the distance between each pixel and the clustering centers within the surrounding 2S×2S area, assigning each pixel to the nearest clustering center and updating the position of the clustering center to the average position of the affiliated pixels; repeating the above process until the change in the clustering centers is less than the threshold or the maximum number of iterations is reached, and taking each grid when the change in the clustering centers is less than the threshold or the maximum number of iterations is reached as a superpixel block, and combining all the superpixel blocks to form a set of superpixel blocks.

[0074] Among them, In the formula, S represents the size of the grid, and N is the total number of pixels in the image. In this embodiment, K = 300. Among them, In the formula, D represents the distance between each pixel and the clustering centers within the surrounding 2S×2S area, d s and d c are the color distance (based on the Pauli RGB value) and the spatial distance respectively, and m is the weight parameter.

[0075] The superpixel segmentation algorithm in this embodiment is the simple linear iterative clustering (SLIC) method, which divides the image into multiple superpixel regions, and the pixels within each region have similar scattering characteristics.

[0076] To prevent the interference of blurred samples at the edges of the superpixel blocks, in this embodiment, according to the Euclidean distance, the K nearest neighbor samples closest to each anchor sample are selected from its corresponding superpixel block, specifically including:

[0077] Form a similarity set \(S\) for each anchor sample by combining the anchor samples within the preset distance \(d\) of the Euclidean distance for each anchor sample. i ; Based on the L2 norm between feature representations, select the top \(K\) nearest neighbors of the anchor sample \(x\) i from the similarity set \(S\) of each anchor sample i to facilitate the construction of the nearest neighbor sample cluster for each anchor sample.

[0078] Specifically, let \(z\) i represent the feature output by the residual neural network model for the anchor sample \(x\) i . Then the similarity between \(x\) i and \(x\) k is calculated as follows: \(S\) i =\(\{x\) k \in g\) i \mid\|\left(x\right.\) k , g\) c )\|_2\leq d\}\); where \(\|\cdot\|_2\) represents the L2 norm, \(g\) i represents the superpixel block, and \(g\) c represents the superpixel center of the superpixel block. The nearest neighbor sample cluster is denoted as

[0079]

[0080] In a further embodiment, according to the nearest neighbor sample cluster of each anchor sample, the cluster feature contrast loss and the soft assignment contrast loss are calculated, specifically including:

[0081] Let the group represent the nearest neighbor sample cluster of the anchor sample \(x\) i , where is the nearest neighbor of \(x\) i ; Consider any pair of samples in the same nearest neighbor cluster as a positive sample pair, and the anchor samples that do not belong to the same nearest neighbor cluster as \(x\) i are regarded as negative samples of \(x\) i ; According to the similar semantic relationship between the nearest neighbor sample clusters of each anchor sample, use the contrast loss to calculate the cluster feature contrast loss of the nearest neighbor sample cluster; Obtain the soft assignment probability of the intermediate residual neural network model for the sample cluster through the Softmax classifier, and use the cross-entropy loss function to calculate the soft assignment contrast loss.

[0082] Among them, In the formula, represents the total loss, represents the soft assignment loss, represents the cluster feature contrast loss, and \(\lambda\) represents the weight coefficient;

[0083]

[0084] In the formula, represents the i-th anchor sample x i and any nearest neighbor sample in its nearest neighbor sample cluster constitute a positive sample pair. Group represents the nearest neighbor sample cluster, and x k represents any sample that does not belong to the nearest neighbor sample cluster of the i-th anchor sample x i , that is, the negative sample of x i . represents the feature representation of the positive sample pair after passing through the residual neural network model. (z i , z k ) represents the feature representation of the negative sample pair (x i , x k ) after passing through the residual neural network model. τ represents the temperature coefficient, which can affect the gradient of the loss function. The larger τ is, the relatively stronger the influence of the negative sample is;

[0085]

[0086] In the formula, N represents the total number of anchor samples, and N x represents the number of nearest neighbor samples in each nearest neighbor sample cluster. p i represents the predicted vector output by the softmax layer after the feature representation z i of the anchor sample x i . p j represents the predicted vector output by the softmax layer after the feature representation i of any nearest neighbor sample x in the nearest neighbor sample cluster j of the anchor sample x j . · represents the dot product operator. represents the probability that the sample x i is assigned to the cluster c. c represents a certain cluster category, C represents the set of cluster categories, and β represents the weight coefficient of the entropy term, which distributes the prediction evenly among each cluster.

[0087] It should be noted that in the process of generating the classification result map, different colors are assigned according to the category to which each pixel belongs.

[0088] In a second aspect, the present invention also proposes a polarimetric SAR image classification system, including a memory, a processor, and a computer program stored on the memory. The processor executes the computer program to implement the steps of the polarimetric SAR image classification method described in any one of the first aspects.

[0089] In a third aspect, the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the polarization SAR image classification method described in any one of the first aspects are implemented.

[0090] In a fourth aspect, the present invention further provides a computer program product, including a computer program. When the computer program is executed by a processor, the steps of the polarization SAR image classification method described in any one of the first aspects are implemented.

[0091] The effects of the present invention can be further illustrated by the following simulations.

[0092] In this embodiment, the polarization SAR image selected for the simulation experiment is a polarization SAR image with an actual ground object marking map, as Figure 2 shown. This polarization SAR image is a composite image of polarization SAR data of the San Francisco Bay Area obtained by the Canadian spaceborne C-band RADARSAT-2 system. The image contains 1800×1380 pixels. The image mainly includes five ground object types: forest, water area, high-density urban area, low-density urban area, and developing urban area.

[0093] Among them, Figure 2 the actual ground object marking map of Figure 3 is as shown in

[0094] Simulation experiment 1: The Wishart clustering based on target decomposition (Freeman decomposition) of the prior art is used to classify the Figure 2 target polarization SAR image, and the classification result is as shown in Figure 5 shown. Among them, the ground object classification accuracy and the overall classification accuracy are shown in Table 1.

[0095] Simulation experiment 2: The polarization SAR classification method based on the autoencoder-decoder network of the prior art is used to classify the Figure 2 target polarization SAR image, and the classification result is as shown in Figure 6 shown. Among them, the ground object classification accuracy and the overall classification accuracy are shown in Table 1.

[0096] Simulation experiment 3: The method of the present invention is used to classify the Figure 2 target polarization SAR image. The superpixel segmentation result is as shown in Figure 4 shown, and the classification result map is as shown in Figure 7 shown. Among them, the ground object classification accuracy and the overall classification accuracy are shown in Table 1.

[0097] Table 1

[0098]

[0099] As can be seen from Table 1, compared with the existing Wishart polarimetric SAR clustering method based on target decomposition (Freeman decomposition), the present invention not only significantly improves the overall accuracy OA, but also greatly improves the classification accuracy of forest and urban ground cover types, verifying that the deep unsupervised learning and clustering method proposed by the present invention is significantly superior to the method based on scattering mechanism; compared with the existing polarimetric SAR classification method based on autoencoder-decoder network, the overall accuracy OA is increased by 1.77%, and the classification accuracy of forest and low-density urban ground cover types is greatly improved, verifying that the method proposed by the present invention can not only effectively avoid the problem of semantic discontinuity in the classification process of the existing deep unsupervised method, but also achieve a higher overall classification accuracy. It provides an advanced and effective solution for the existing unsupervised PolSAR image classification.

[0100] From Figure 5 、 Figure 6 and Figure 7 it can be seen that Figure 7 is closer to Figure 5 the true ground truth map of the polarimetric SAR image in the target domain, and Figure 7 has better regional consistency in the classification results. This is mainly because a polarimetric SAR image classification method adopted by the present invention can learn a wider and more discriminative feature representation, thereby effectively improving the accuracy of polarimetric SAR image classification and recognition.

[0101] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A method for classifying polarimetric SAR images, characterized in that, Including: Obtain a polarization dataset; wherein, the polarization dataset includes a polarimetric SAR image and a corresponding ground truth map; construct an anchor sample training set and a positive sample pair dataset according to the polarization dataset; Construct a residual neural network model; Perform the first-stage training on the residual neural network model using the positive sample pair dataset to obtain an intermediate residual neural network model trained in the first stage; perform the second-stage training on the intermediate residual neural network model using the polarization dataset and the anchor sample training set to obtain a residual neural network model trained in the second stage; Use the residual neural network model trained in the second stage to cluster the polarimetric SAR image to be classified to obtain a clustering result of the polarimetric SAR image; use the Hungarian matching algorithm to map the clustering result of the polarimetric SAR image to the true label to obtain a mapping result; generate a classification result map according to the mapping result.

2. The polarization SAR image classification method according to claim 1, characterized in that, Construct an anchor sample training set and a positive sample pair dataset according to the polarization data, specifically including: construct an anchor sample training set according to the polarization dataset; construct a positive sample pair dataset according to the anchor sample training set; Among them, constructing an anchor sample training set according to the polarization dataset specifically includes: obtaining a complex-valued feature matrix of the polarimetric SAR image according to the polarimetric SAR image; wherein, the complex-valued feature matrix includes a complex-valued feature vector of each pixel; generating labeled data according to the ground truth map; constructing an initial ground truth matrix, and performing pixel-level label classification and annotation on the initial ground truth matrix according to the labeled data to generate a ground truth matrix; obtaining the complex-valued feature vector corresponding to the coordinate value in the complex-valued feature matrix as the first complex-valued feature vector according to the coordinate value of the labeled pixel in the ground truth matrix; using the first complex-valued feature vector as the center point, and respectively selecting several complex-valued feature vectors around as the second complex-valued feature vectors; combining the first complex-valued feature vector and the second complex-valued feature vectors to form a complex-valued feature matrix block; separating the real part and the imaginary part of the upper triangular elements in the complex-valued feature vectors in the complex-valued feature matrix block to form real-valued feature vectors; generating a high-dimensional real-valued feature matrix based on the real-valued feature vectors; performing corresponding combination on the pixel-level label classification and annotation information in the ground truth matrix and the high-dimensional real-valued feature matrix block to form an anchor sample training set; Among them, constructing a positive sample pair dataset according to the anchor sample training set specifically includes: performing random data augmentation on the anchor sample training set to generate positive samples with the same semantics as the anchor samples in the anchor sample training set, and combining the anchor samples with their corresponding positive samples to form positive sample pairs; combining all the positive sample pairs to form a positive sample pair dataset.

3. The polarimetric SAR image classification method according to claim 1, wherein The residual neural network model includes an input layer, a residual layer, a global average pooling layer, and a clustering head; wherein, the input layer includes a first convolutional layer and a first batch normalization layer, the number of input feature channels of the first convolutional layer is 9, which is the same as the sample real-valued feature matrix, the number of feature map of the output feature data is 64, the size of the convolutional filter in the first convolutional layer is 3×3, the padding is 1, and the moving stride is 1; Among them, the residual layer contains 4 levels of residual structures. Each level includes a number of residual blocks, and each residual block contains two convolutional layers and a shortcut connection layer. Among them, the global average pooling layer is used to compress the spatial dimension of the feature map to 1×1 using adaptive average pooling and output a feature dimension of 512. Among them, the clustering head is used to map the 512-dimensional feature to an n-dimensional space using a linear layer to obtain the probability distribution of each sample belonging to n clustering categories.

4. The polarimetric SAR image classification method according to claim 3, wherein The 4 levels of residual structures are layer1, layer2, layer3, and layer4 in sequence. Among them, layer1 contains the first residual block and the second residual block. The number of input and output channels of the two residual blocks is 64. The first residual block includes the second convolutional layer, the third convolutional layer, and the second batch normalization layer. The second residual block includes the fourth convolutional layer, the fifth convolutional layer, and the third batch normalization layer. The size of the convolutional filters in the second, third, fourth, and fifth convolutional layers is 3×3, the padding is 1, the stride is 1, and there is no downsampling operation. layer2 contains the third residual block and the fourth residual block. The third residual block is used to expand the number of channels from 64 to 128. The third residual block includes the sixth convolutional layer, the seventh convolutional layer, and the fourth batch normalization layer. The size of the convolutional filters in the sixth and seventh convolutional layers is 3×3, the padding is 1, the stride is 1, and there is no downsampling operation. At the same time, the output of layar1 and the output of layer2 are matched in terms of the number of channels through the first shortcut connection layer and summed as the input of layer3. The first shortcut connection layer includes the eighth convolutional layer and the fifth batch normalization layer. The number of input feature channels of the eighth convolutional layer is 64, and the number of feature map of the output feature data is 128. The size of the convolutional filter in the eighth convolutional layer is 1×1, there is no padding, and the stride is 1. The fourth residual block maintains the number of channels at 128. The fourth residual block includes the ninth convolutional layer, the tenth convolutional layer, and the sixth batch normalization layer. The size of the convolutional filters in the ninth and tenth convolutional layers is 3×3, the padding is 1, the stride is 1, and there is no downsampling operation. Layer 3 includes a fifth residual block and a sixth residual block; the fifth residual block expands the number of channels from 128 to 256. The fifth residual block includes an eleventh convolutional layer, a twelfth convolutional layer, and a seventh batch normalization layer. The convolutional filter in the eleventh convolutional layer has a size of 3×3, a padding of 1, and a stride of 2 for downsampling operations. The convolutional filter in the twelfth convolutional layer has a size of 3×3, a padding of 1, and a stride of 1. At the same time, the output of layer 2 is channel-matched with the output of layer 3 through a second shortcut connection layer and summed as the input of layer 4. The second shortcut connection layer includes a thirteenth convolutional layer and an eighth batch normalization layer. The input feature channels of the thirteenth convolutional layer are 128, the number of feature maps of the output feature data is 256, the convolutional filter has a size of 1×1, no padding, and a stride of 1. The sixth residual block maintains the number of channels at 256. The sixth residual block includes a fourteenth convolutional layer, a fifteenth convolutional layer, and a ninth batch normalization layer. The convolutional filters of the fourteenth convolutional layer and the fifteenth convolutional layer both have a size of 3×3, a padding of 1, and a stride of 1, without downsampling operations. Layer 4 includes a seventh residual block and an eighth residual block. The seventh residual block expands the number of channels from 128 to 256. The seventh residual block includes a sixteenth convolutional layer, a seventeenth convolutional layer, and a tenth batch normalization layer. The convolutional filters in the sixteenth convolutional layer and the seventeenth convolutional layer both have a size of 3×3, a padding of 1, and a stride of 1. At the same time, the output of layer 3 is channel-matched with the output of layer 4 through a third shortcut connection layer and summed as the final output. The third shortcut connection layer includes an eighteenth convolutional layer and an eleventh batch normalization layer. The input feature channels of the eighteenth convolutional layer are 128, the number of feature maps of the output feature data is 256, the convolutional filter has a size of 1×1, no padding, and a stride of 1. The eighth residual block maintains the number of channels at 256. The eighth residual block includes a nineteenth convolutional layer, a twentieth convolutional layer, and an eleventh normalization layer. The convolutional filters in the nineteenth convolutional layer and the twentieth convolutional layer both have a size of 3×3, a padding of 1, and a stride of 1, without downsampling operations.

5. The polarization SAR image classification method according to claim 1, characterized in that Using the positive sample pair dataset to perform the first-stage training on the residual neural network model, obtaining an intermediate residual neural network model trained in the first stage, specifically including: Inputting the positive sample pair dataset into the residual neural network model to obtain the anchor sample features and their positive sample features; calculating the feature similarity between the anchor sample and the positive sample according to the anchor sample features and their positive sample features, and calculating the contrast loss between the anchor sample and the positive sample according to the feature similarity between the anchor sample and the positive sample. According to the contrast loss between the anchor sample and the positive sample, and repeatedly optimizing the parameters of the residual neural network model through the backpropagation algorithm until the residual neural network model converges or reaches a preset number of training rounds; taking the residual neural network model that converges or reaches the preset number of training rounds as the intermediate residual neural network model trained in the first stage.

6. The polarization SAR image classification method according to claim 1, wherein The intermediate residual neural network model is trained in the second stage using the polarization data and the anchored sample training set to obtain a residual neural network model trained in the second stage, specifically including: Based on the polarization data set, a Pauli image set is obtained; the superpixel segmentation algorithm is used to segment each Pauli image in the Pauli image set to obtain a superpixel block set; among them, the superpixel block set contains multiple superpixel blocks, and each superpixel block includes multiple pixel point coordinates and corresponding superpixel block index information; according to the superpixel block index information and the coordinates of the central pixel point of each anchored sample in the anchored sample data set, the superpixel block to which each anchored sample belongs is determined; taking all non-edge region pixels of the superpixel block to which each anchored sample belongs as the center, a sample block set of the same size as the anchored sample is taken as the sample candidate set of the nearest neighbor sample cluster of the anchored sample; the samples in the candidate set corresponding to each anchored sample are input into the intermediate residual neural network model to obtain the candidate feature library of the nearest neighbor sample cluster of each anchored sample; each anchored sample is input into the intermediate residual neural network model to obtain the anchored feature; the Euclidean distance between the anchored feature of each anchored sample and all feature tensors in the candidate feature library is calculated for each anchored sample; according to the Euclidean distance, the K nearest neighbor samples with the closest distance are selected for each anchored sample; the K nearest neighbor samples selected for each anchored sample are combined to form the nearest neighbor sample cluster of each anchored sample; According to the nearest neighbor sample cluster of each anchored sample, the cluster feature contrast loss and the soft assignment contrast loss are calculated; a total loss function is constructed according to the cluster feature contrast loss and the soft assignment contrast loss; according to the cluster feature contrast loss and the soft assignment contrast loss, the parameters of the intermediate residual neural network model are optimized by the backpropagation algorithm until the intermediate residual neural network model converges or reaches the preset number of optimization times; the intermediate residual neural network model that converges or reaches the preset number of optimization times is used as the residual neural network model trained in the second stage.

7. The polarization SAR image classification method according to claim 6, wherein The total loss of the second-stage training is wherein, represents the total loss, represents the soft assignment loss, represents the cluster feature contrast loss, and λ represents the weight coefficient; In the formula, represents the i-th anchored sample x i and any nearest neighbor sample in its nearest neighbor sample cluster to form a positive sample pair. Group represents the nearest neighbor sample cluster, and x k represents any sample that does not belong to the nearest neighbor sample cluster of the i-th anchored sample x i , that is, x i is the negative sample of represents the feature representation of the positive sample pair after passing through the residual neural network model. (z i , z k ) represents the feature representation of the negative sample pair (x i , x k ) after passing through the residual neural network model, and τ represents the temperature coefficient; Wherein, N represents the total number of anchor samples, N x represents the number of nearest neighbor samples in each nearest neighbor sample cluster, p i represents the feature representation z i of the anchor sample x i is the predicted vector output after passing through the softmax layer, p j represents the nearest neighbor sample cluster i of the anchor sample x and any nearest neighbor sample x j in it, the feature representation z j is the predicted vector output after passing through the softmax layer, · represents the dot product operator, represents the probability that the sample x i is assigned to the clustering category c, c represents a certain clustering category, C represents the set of clustering categories, and β represents the entropy term is the weight coefficient.

8. A polarimetric SAR image classification system, comprising a memory, a processor, and a computer program stored on the memory, characterized in that, The processor executes the computer program to implement the steps of the polarization SAR image classification method according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the polarization SAR image classification method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the polarization SAR image classification method according to any one of claims 1-7.

Images