Complete-valued image volume integral class algorithm based on fully-polarized SAR (Synthetic Aperture Radar) image

Abstract

The invention provides a complex valued image convolution class algorithm based on a fully-polarized SAR image, and the algorithm comprises the steps: firstly carrying out the preprocessing of the fully-polarized SAR image, obtaining a polarization covariance matrix, obtaining superpixels and an average polarization covariance matrix based on the segmentation of the superpixels, calculating the similarity through WA2 distance, constructing an adjacent matrix, and carrying out the segmentation of the average polarization covariance matrix; then, a complex value graph convolutional network model supporting complex value feature input is constructed, a Leaky-zReLU complex value activation function is adopted from an input layer to a hidden layer, an AM-Softmax function and a model weight are adopted from the hidden layer to an output layer to implement complex number initialization, focusing loss Focus Loss is used as a loss function to perform gradient updating of back propagation, and a complex value graph convolutional network model supporting complex value feature input is constructed; and finally, inputting a complex element value of the polarization covariance matrix as a superpixel region feature and an adjacent matrix into a complex value graph convolutional network for classification. According to the method, the limitation that the existing GCN network architecture only supports real number operation is broken through, the complex value information contained in the fully-polarized SAR data is effectively utilized, and the precision of the GCN network in fully-polarized SAR classification is improved.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image data processing, and mainly relates to the extraction of complex attribute information of fully polarimetric SAR data and image classification based on graph convolutional neural networks. Specifically, the present invention is a complex-valued graph convolution classification algorithm based on fully polarimetric SAR images. Background Art

[0002] Synthetic Aperture Radar (SAR), an active microwave side-looking imaging system, offers stable imaging capabilities around the clock and in all weather conditions. It can acquire high-quality images even in complex weather conditions and low-light conditions, and has been widely used in fields such as topographic mapping and land use / cover classification. Polarimetric Synthetic Aperture Radar (PolSAR) further expands SAR's application boundaries. By acquiring information from multiple polarization channels, PolSAR can deeply analyze the scattering mechanisms of ground objects, providing rich information support for applications such as ship detection, target recognition, vegetation classification, and urban planning.

[0003] In the field of polarimetric SAR data processing and analysis, polarimetric SAR image classification is a key technical direction for achieving accurate interpretation of ground object information. Its core lies in classifying ground objects by quantifying the differences between targets. These differences are usually quantitatively characterized by the polarimetric covariance matrix. Existing traditional polarimetric SAR classification methods mainly rely on target decomposition techniques based on polarimetric covariance matrices or coherence matrices to extract discriminative feature parameters and combine them with traditional classification algorithms to complete ground object classification. However, with the rapid development of microwave imaging technology, traditional classification methods have gradually shown limitations such as insufficient feature expression capabilities and limited classification accuracy when processing high-resolution, large-coverage, and multi-temporal data.

[0004] In recent years, the rise of deep learning technology has brought new breakthroughs in remote sensing image processing. Among them, the Graph Convolutional Network (GCN) has demonstrated significant advantages with its unique graph structure modeling method. Unlike traditional convolutional neural networks, GCN breaks through the limitations of local square receptive fields and establishes direct or indirect connections between global image nodes by constructing a graph structure, thereby achieving effective utilization of global image information. In the task of fully polarimetric SAR image classification, GCN has demonstrated excellent performance. However, due to the inherent complex storage characteristics of polarimetric SAR data, which contains dual information dimensions of real and imaginary parts, the existing GCN network architecture only supports real number operations, making it difficult to fully explore the complex feature information contained in polarimetric SAR data. As a result, the network's performance potential in processing fully polarimetric SAR data has not been fully unleashed.

[0005] In view of the above technical status, the present invention makes innovative improvements on the traditional real-domain GCN network based on the complex-valued characteristics of fully polarimetric SAR data, and proposes a complex-valued graph convolution classification algorithm based on the complex-valued characteristics of fully polarimetric SAR data, aiming to break through the bottleneck of existing technology and achieve high-precision classification of polarimetric SAR images. Summary of the Invention

[0006] The purpose of this paper is to propose a complex-valued graph convolution classification network based on fully polarimetric SAR images. By modifying the classic GCN network, it supports the processing of complex values ​​of polarimetric covariance matrix elements, thereby making full use of the complex-valued information contained in fully polarimetric SAR data to obtain higher classification accuracy.

[0007] Therefore, a complex-valued graph convolution classification algorithm based on polarimetric SAR images is proposed. The specific steps are as follows:

[0008] Step 1: Preprocess the original fully polarimetric SAR single-look complex image, including radiometric calibration, multi-look processing, polarimetric filtering, terrain correction, and geocoding, to suppress speckle noise and eliminate geometric distortion, improve image quality, and obtain the polarimetric covariance matrix of the pixel.

[0009] Step 2: Based on the scattering characteristics represented by the polarization covariance matrix, the image is segmented into superpixels, and the arithmetic mean of the polarization covariance matrix of all pixels in each superpixel unit is calculated to obtain the regional level scattering characteristic representation;

[0010] Step 3: Combine the fully polarimetric SAR PauliRGB synthetic image with the Google Earth high-resolution optical image to determine the typical ground object categories through manual visual interpretation, and collect the polarimetric covariance matrix of each category of pixels as a training sample set;

[0011] Step 4: Combine the superpixel region features with the sample features to construct a joint feature space, calculate the ‖WA||2 distance between any two covariance matrices in the feature space, and construct the adjacency matrix in the graph convolutional neural network based on the generated distance matrix;

[0012] Step 5: Construct a complex-valued graph convolutional neural network model with a two-layer network structure, in which the Leaky-zReLU complex-valued activation function is used from the input layer to the hidden layer, and the AM-Softmax function is used from the hidden layer to the output layer. The model weights are initialized with complex numbers, and the focal loss is used as the loss function for back-propagation gradient update.

[0013] In step 6, the upper triangular elements of the polarization covariance matrix are expanded into a six-dimensional feature vector as the feature representation of the superpixel node, and input into the complex-valued graph convolutional network together with the adjacency matrix to realize the ground object classification of the polarimetric SAR image.

[0014] Furthermore, in step 2, polarimetric superpixel segmentation (Pol-ASLIC) is used to segment the fully polarimetric SAR image. The distance formula used in Pol-ASLIC segmentation is as follows:

[0015]

[0016] Where D SIRV Represents the distance of the covariance matrix between pixels, D T Represents the texture distance between pixels, D S represents the spatial distance, S represents the sampling step size, and β is used to balance the spatial distance with the other two distances.

[0017] Furthermore, in step 4, the superpixel region features and the sample features are combined to construct a joint feature space, and the ||WA||2 distance between any two covariance matrices in the feature space is calculated. The calculation formula is as follows:

[0018]

[0019] Where, d Wishhart Revised Wishart distance for symmetry, d AIRM is the affine invariant Riemannian metric (AIRM) distance, ω is the weight coefficient, and the formulas are as follows:

[0020]

[0021] Where X and Y are arbitrary polarization covariance matrices, q is a parameter related to the polarization mode, and q is 3 in full polarization imaging. i is the matrix X -1 The eigenvalues ​​of Y, W represents the Wishart distance matrix, represents the corresponding self-connected adjacency matrix, A represents the AIRM distance matrix, Represents the corresponding self-connected adjacency matrix, indicating the function Indicates 1 when X is set, otherwise 0. i,(m) represents the mth smallest element in the i-th row of matrix X, and n represents the number of rows or columns of the distance matrix;

[0022] The adjacency matrix is ​​constructed based on the ||WA||2 distance matrix, and the formula is as follows:

[0023]

[0024] Where Nei(·) represents the neighborhood operation, S i represents the i-th superpixel corresponding to the i-th polarization covariance matrix.

[0025] Furthermore, in step 5, a complex-valued graph convolutional network model with a two-layer network structure is constructed, and the mathematical architecture of the model is expressed as:

[0026]

[0027] Where, is the adjacency matrix with self-connection added, F is the feature matrix, W (0) and W (1) is the weight matrix of the two-layer neural network;

[0028] The forward model of the complex-valued graph convolutional network from the input layer to the hidden layer is Leaky - The zReLU function is used as the activation function. The Leaky-zReLU function is a variant of the Relu function in the complex domain. It takes the phase as the independent variable and introduces a leakage mechanism (Leaky mechanism) to avoid underfitting of the model caused by excessive neuron death. The formula of the Leaky-zReLU function is as follows:

[0029]

[0030] Where a is the leakage coefficient, which usually ranges from 0.001 to 0.01;

[0031] The forward model of the complex-valued graph convolutional network uses the AM-Softmax function as the activation function from the hidden layer to the output layer. The AM-Softmax function is a performance-optimized version of the Softmax function. It adds an additive cosine margin to increase the distance between classes and reduce the distance within classes, thereby enhancing the discrimination between different objects. The formula is as follows:

[0032]

[0033] Where N represents the total number of training samples, is the angle between the feature vector and the true category weight vector, s is the feature scaling factor, which is used to control the compactness of the feature distribution, and m ≥ 0 is the additional cosine interval, which is used to force the increase of the inter-class distance;

[0034] The loss function of the complex-valued graph convolutional network is defined as the focal loss FocalLoss, and the formula is as follows:

[0035]

[0036] In the formula, N represents the number of training samples, C represents the number of set ground object categories, and Y ij and P ijRepresent the true value label and predicted probability respectively, J is a constant matrix, γ is a hyperparameter, and Focal Loss controls the error contribution of easy-to-classify samples to back propagation by focusing on the parameter γ, so that the model can pay more attention to difficult-to-classify samples, thereby effectively improving the overall classification performance of the model.

[0037] The present invention provides a complex-valued graph convolution classification algorithm based on fully polarimetric SAR images, which has the following technical effects: breaking through the limitation of the existing GCN network architecture that only supports real number operations, extracting the deep complex-valued information of polarimetric SAR data, and improving the accuracy of the GCN network in polarimetric SAR classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic diagram of the process of a complex-valued graph convolution classification algorithm based on polarimetric SAR images provided by the present invention;

[0039] Figure 2 Schematic diagram of complex-valued graph convolutional network;

[0040] Figure 3 Schematic diagram of the Leaky-zReLU function;

[0041] Figure 4 Attached cosine interval diagram for AM-Softmax function. DETAILED DESCRIPTION

[0042] The following examples are only used to more clearly illustrate the technical solutions of the present invention. Figure 1 :

[0043] Step 1: Perform radiometric calibration, multi-look processing, polarimetric filtering, terrain correction, and geocoding preprocessing on a single-look fully polarimetric SAR image to suppress speckle noise and eliminate geometric distortion, improve image quality, and obtain the polarimetric covariance matrix of the pixel.

[0044] Step 2: Based on the scattering characteristics represented by the polarization covariance matrix, polarimetric superpixel segmentation (Pol-ASLIC) is used to segment the polarimetric SAR image. The distance formula used in Pol-ASLIC segmentation is as follows:

[0045]

[0046] Where D SIRV Represents the distance of the covariance matrix between pixels, D T Represents the texture distance between pixels, D S Represents the spatial distance, S represents the sampling step size, and β is used to balance the spatial distance with the other two distances. Then, the arithmetic mean of the polarization covariance matrix of all pixels in each superpixel unit is calculated to characterize the polarization scattering characteristics of the region.

[0047] In step 3, the polarimetric SAR PauliRGB synthetic image and the Google Earth high-resolution optical image are combined to determine the typical ground object categories through manual visual interpretation, and the polarimetric covariance matrix of pixels in each category is randomly collected as a training sample set.

[0048] Step 4: Combine the superpixel region features and sample features to construct a joint feature space, and calculate the ||WA||2 distance metric between any two covariance matrices in the feature space. The calculation formula is as follows:

[0049]

[0050] Where, d Wishart Revised Wishart distance for symmetry, d AIRM is the affine invariant Riemann distance, ω is the weight coefficient, and the formulas are as follows:

[0051]

[0052] Where X and Y are arbitrary polarization covariance matrices, q is a parameter related to the polarization mode, and q is 3 in full polarization imaging. i is the matrix X -1 The eigenvalues ​​of Y, W represents the Wishart distance matrix, represents the corresponding self-connected adjacency matrix, A represents the AIRM distance matrix, Represents the corresponding self-connected adjacency matrix, indicating the function Indicates 1 when X is set, otherwise 0. i，(m) represents the mth smallest element in the i-th row of matrix X, and n represents the number of rows or columns of the distance matrix;

[0053] The adjacency matrix is ​​constructed based on the ||WA||2 distance matrix, and the formula is as follows:

[0054]

[0055] Where Nei(·) represents the neighborhood operation, S i represents the i-th superpixel corresponding to the i-th polarization covariance matrix.

[0056] Step 5: Construct a complex-valued graph convolutional network model with a two-layer network structure, such as Figure 2 , the mathematical structure of the model is expressed as:

[0057]

[0058] Where, is the adjacency matrix with self-connection added, F is the feature matrix, W (0) and W(1) is the weight matrix of the two-layer neural network.

[0059] The forward model from the input layer to the hidden layer uses the Leaky-zReLU function as the activation function. The Leaky-zReLU function is a variant of the Relu function in the complex domain. It takes the phase as the independent variable and introduces a leaky mechanism to avoid underfitting of the model caused by excessive neuron death. The schematic diagram of the Leaky-zReLU function is as follows Figure 3 , the formula is as follows:

[0060]

[0061] Where a is the leakage coefficient, which usually ranges from 0.001 to 0.01.

[0062] The forward model from the hidden layer to the output layer uses the AM-softmax function as the activation function. The AM-Softmax function is a performance-optimized version of the Softmax function. It increases the inter-class distance and reduces the intra-class distance by adding an additive cosine interval, thereby enhancing the discrimination between different targets. For example, Figure 4 , the formula is as follows:

[0063]

[0064] Where N represents the total number of training samples, is the angle between the feature vector and the true category weight vector, s is the feature scaling factor, which is used to control the compactness of the feature distribution, and m ≥ 0 is the additional cosine interval, which is used to force an increase in the distance between classes.

[0065] The loss function is defined as the focal loss FocalLoss, and the formula is as follows:

[0066]

[0067] In the formula, N represents the number of training samples, C represents the number of set ground object categories, and Y ij and P ij Represent the true value label and predicted probability respectively, J is a constant matrix, γ is a hyperparameter, and Focal Loss controls the error contribution of easy-to-classify samples to back propagation by focusing on the parameter γ, so that the model can pay more attention to difficult-to-classify samples, thereby effectively improving the overall classification performance of the model.

[0068] In step 6, the upper triangular elements of the polarization covariance matrix are expanded into a six-dimensional feature vector as the feature representation of the superpixel node, and input into the complex-valued graph convolutional network together with the adjacency matrix to realize the ground object classification of the polarimetric SAR image.

Claims

1. A complex-valued graph convolution classification algorithm based on fully polarimetric SAR images, characterized by The following steps are involved: Step 1: Preprocess the original fully polarimetric SAR single-look complex image, including radiometric calibration, multi-look processing, polarimetric filtering, terrain correction, and geocoding, to suppress speckle noise and eliminate geometric distortion, improve image quality, and obtain the polarimetric covariance matrix of the pixel. Step 2: Based on the scattering characteristics represented by the polarization covariance matrix, the image is segmented into superpixels, and the arithmetic mean of the polarization covariance matrix of all pixels in each superpixel unit is calculated to obtain the regional level scattering characteristic representation; Step 3: Combine the fully polarimetric SAR PauliRGB synthetic image with the Google Earth high-resolution optical image to determine the typical ground object categories through manual visual interpretation, and collect the polarimetric covariance matrix of each category of pixels as a training sample set; Step 4: Combine the superpixel region features with the sample features to construct a joint feature space, calculate the ||WA||2 distance between any two covariance matrices in the feature space, and construct the adjacency matrix in the graph convolutional neural network based on the generated distance matrix; Step 5: Construct a complex-valued graph convolutional neural network model with a two-layer network structure, in which the Leaky-zReLU complex-valued activation function is used from the input layer to the hidden layer, and the AM-Softmax function is used from the hidden layer to the output layer. The model weights are initialized with complex numbers, and the focal loss is used as the loss function for back-propagation gradient update. In step 6, the upper triangular elements of the polarization covariance matrix are expanded into a six-dimensional feature vector as the feature representation of the superpixel node, and input into the complex-valued graph convolutional network together with the adjacency matrix to realize the ground object classification of the polarimetric SAR image.

2. The complex-valued graph convolution classification algorithm based on fully polarimetric SAR images according to claim 1 is characterized in that: In step 2, polarimetric superpixel segmentation (Pol-ASLIC) is used to segment the fully polarimetric SAR image. The distance formula used in Pol-ASLIC segmentation is as follows: Where D SIRV Represents the distance of the covariance matrix between pixels, D T Represents the texture distance between pixels, D S represents the spatial distance, S represents the sampling step size, and β is used to balance the spatial distance with the other two distances.

3. The complex-valued graph convolution classification algorithm based on fully polarimetric SAR images according to claim 1 is characterized in that: In step 4, the distance between any two polarization covariance matrices is calculated according to the following formula: Where, d Wishart Revised Wishart distance for symmetry, d AIRM is the affine invariant Riemannian metric (AIRM) distance, ω is the weight coefficient, and the formulas are as follows: Where X and Y are arbitrary polarization covariance matrices, q is a parameter related to the polarization mode, and q is 3 in full polarization imaging. i is the matrix X -1 The eigenvalues ​​of Y, W represents the Wishart distance matrix, represents the corresponding self-connected adjacency matrix, A represents the AIRM distance matrix, Represents the corresponding self-connected adjacency matrix, indicating the function Indicates 1 when X is set, otherwise 0. i，(m) represents the mth smallest element in the i-th row of matrix X, and n represents the number of rows or columns of the distance matrix; The adjacency matrix is ​​constructed based on the ||WA||2 distance matrix, and the formula is as follows: Where Nei(·) represents the neighborhood operation, S i represents the i-th superpixel corresponding to the i-th polarization covariance matrix.

4. The complex-valued graph convolution classification algorithm based on fully polarimetric SAR images according to claim 1, characterized in that: In step 5, a complex-valued graph convolutional network model with a two-layer network structure is constructed, and the mathematical architecture of the model is expressed as: Where, is the adjacency matrix with self-connection added, F is the feature matrix, W (0) and W (1) is the weight matrix of the two-layer neural network; The forward model of the complex-valued graph convolutional network from the input layer to the hidden layer uses the Leaky-zReLU function as the activation function. The Leaky-zReLU function is a variant of the ReLU function in the complex domain. It takes the phase as the independent variable and introduces a leakage mechanism (Leaky mechanism) to avoid model underfitting caused by excessive neuron death. The formula of the Leaky-zReLU function is as follows: Where a is the leakage coefficient, which usually ranges from 0.001 to 0.01; The forward model of the complex-valued graph convolutional network from the hidden layer to the output layer uses the AM-Softmax function as the activation function. The AM-Softmax function is a performance-optimized version of the Softmax function. It adds an additive cosine interval to increase the inter-class distance and reduce the intra-class distance, thereby enhancing the discrimination between different targets. The formula is as follows: Where N represents the total number of training samples, is the angle between the feature vector and the true category weight vector, s is the feature scaling factor, which is used to control the compactness of the feature distribution, and m ≥ 0 is the additional cosine interval, which is used to force the increase of the inter-class distance; The loss function of the complex-valued graph convolutional network is defined as the focal loss, and the formula is as follows: In the formula, N represents the number of training samples, C represents the number of set ground object categories, and Y ij and P ij Represent the true value label and predicted probability respectively, J is a constant matrix, γ is a hyperparameter, and Focal Loss controls the error contribution of easy-to-classify samples to back propagation by focusing on the parameter γ, so that the model can pay more attention to difficult-to-classify samples, thereby effectively improving the overall classification performance of the model.

Images