Polarization SAR image classification method combining polarization rotation angle and space self-attention

Abstract

The invention discloses a polarized SAR image classification method combining a polarization rotation angle and spatial self-attention, and the method comprises the steps: obtaining a polarized coherence matrix sequence of a polarized SAR image, and extracting the neighborhood window four-dimensional data of a to-be-classified pixel from the polarized coherence matrix sequence; extracting low-level feature representation of the polarized SAR image by using the two 3D convolution blocks; establishing a polarization angle and space local feature extraction network architecture, and respectively extracting polarization angle and space local features; the method comprises the following steps: establishing a polarization angle and space global feature extraction and fusion network architecture by taking a transform coding layer as an infrastructure so as to respectively learn the global features of the polarization angle and the space and carry out global feature fusion; performing down-sampling and advanced feature extraction on the global feature fusion image by using a combination of a 3D pooling layer and a 3D convolution block to obtain a feature vector; and mapping the feature vectors to a classifier to carry out polarized SAR image category classification and outputting a classification result of the to-be-classified pixels. According to the method, the classification accuracy of the polarized SAR images is effectively improved.

Description

Technical Field

[0001] The invention belongs to the technical field of polarimetric SAR image processing, and in particular relates to a polarimetric SAR image classification method combining polarization rotation angle and spatial self-attention. Background Art

[0002] The synthetic aperture radar (SAR) system generates a two-dimensional image by transmitting and receiving electromagnetic wave pulses, recording the intensity and time difference of the radar signal return. The imaging characteristics of SAR include: all-weather imaging, certain penetration ability and other advantages. Polarimetric synthetic aperture radar (PolSAR) is an extension of SAR technology, which obtains multi-dimensional information of ground objects through different polarization modes (such as horizontal and vertical polarization). This polarization information can distinguish different types of ground objects, such as water bodies, vegetation, buildings, etc. Therefore, PolSAR images are widely used in environmental monitoring, agricultural surveys, urban planning and other fields. However, the existing polarimetric SAR image processing technology only uses a fixed polarization basis to obtain polarization information, and fails to fully tap the potential of polarization data.

[0003] With the rapid development of deep learning, PolSAR image interpretation is moving towards intelligence and automation. Convolutional neural networks (CNNs) perform well in image classification and feature extraction, and can effectively capture local features in images. Transformer is known for its self-attention mechanism, which can understand data relationships from a global perspective. The combination of these two technologies helps to improve the effect of image processing, but in polarimetric SAR image processing, their advantages have not been fully utilized due to reasons such as different information dimensions. Summary of the invention

[0004] In view of the problems of insufficient utilization of polarization information of polarization SAR images and insufficient combination with spatial information in the process of polarization SAR image processing based on deep learning, the present invention proposes a polarization SAR image classification method combining polarization rotation angle and spatial self-attention. The method enriches the polarization information of polarization SAR images by rotating the polarization basis; combines the advantages of Transformer and convolutional neural networks to solve the problem that the polarization rotation angle features and spatial features are difficult to combine due to their different dimensions, fully extracts and fuses the local and global information of the polarization rotation angle features and spatial features of polarization SAR images, and effectively improves the classification accuracy of polarization SAR image datasets.

[0005] The technical solution adopted by the present invention is:

[0006] According to a first aspect of the present invention, a polarization SAR image classification method combining polarization rotation angle and spatial self-attention is provided, the method comprising:

[0007] Obtain the polarization coherence matrix sequence of the polarization SAR image, and extract the neighborhood window four-dimensional data of the pixels to be classified;

[0008] The four-dimensional data of the neighborhood window of the pixel to be classified is input into two 3D convolution blocks, and the low-level feature representation of the polarimetric SAR image is extracted by using the two 3D convolution blocks to generate a four-dimensional local preliminary feature map z;

[0009] Establish a polarization angle local feature extraction network, input the four-dimensional local preliminary feature map z into the polarization angle local feature extraction network, and obtain the polarization angle local feature map X l_3d ; Establish a spatial local feature extraction network, merge the four-dimensional local preliminary feature map z into a three-dimensional local feature map z′ and input it into the spatial local feature extraction network to obtain the spatial local feature map X l_2d ;

[0010] Based on the transformer coding layer as the basic architecture, the polarization angle and spatial global feature extraction and fusion network architecture is established:

[0011] The polarization angle local feature map X l_3d With the spatial local feature map X l_2d The multi-head self-attention mechanism is input into the transformer encoding layer to learn the global features of polarization angle and space respectively;

[0012] The global features of the space are formatted and added in parallel with the global features of the polarization angle, and then input into the multi-layer perceptron MLP of the transformer encoding layer for global feature fusion to obtain the global feature fusion graph z″;

[0013] The global feature fusion map z″ is input into the 3D pooling layer and the 3D convolution block, and the combination of the 3D pooling layer and the 3D convolution block is used to downsample and extract high-level features of the global feature fusion map z″ to obtain a feature vector;

[0014] The feature vector is mapped to the classifier to classify the polarimetric SAR image and the classification result of the pixel to be classified is output.

[0015] In the above scheme, the polarization coherence matrix sequence of the polarization SAR image is obtained, including:

[0016] Based on Pauli decomposition, the coherence matrix of polarimetric SAR images is constructed through the polarimetric scattering matrix. The polarimetric scattering matrix contains the information of four polarization channels: HH, HV, VH, and VV. The specific formula is as follows:

[0017]

[0018] Where T3 is the coherence matrix of the polarimetric SAR image; H and V represent the horizontal and vertical polarizations, respectively; S HH , S HV , S VH , S VV are the scattering coefficients of different polarization channels respectively; * indicates complex conjugate;

[0019] By rotating the polarization basis, the polarization coherence matrix at different rotation angles is calculated to obtain the polarization coherence matrix sequence of the polarization SAR image; wherein, the formula of the polarization coherence matrix T(θ) is as follows:

[0020]

[0021] Where R3(θ) represents the polarization rotation matrix, represents the transpose; where the polarization rotation matrix R3(θ) is defined as follows:

[0022]

[0023] Wherein, θ represents the polarization direction angle POA.

[0024] In the above scheme, the angle range used is The step length is Nine different polarization direction angles POA are used to generate a polarization coherence matrix sequence, and based on the patch classification strategy, the neighborhood window four-dimensional data of each pixel to be classified is extracted from the polarization coherence matrix sequence; wherein the number of polarization coherence matrix sequences corresponds to the number of different polarization direction angles POA.

[0025] In the above scheme, both 3D convolution blocks contain a 3D convolution layer, a 3D batch normalization BN, and a sigmoid weighted linear unit SiLU activation layer; wherein SiLU is a nonlinear activation function;

[0026] The polarization coherence matrix sequence of the polarimetric SAR image is input into two 3D convolution blocks. The convolution kernel in the 3D convolution block slides in the local area of ​​the input data, calculates the weighted sum, extracts the geometric and local features in the multidimensional space of the polarimetric SAR image, and generates a four-dimensional local preliminary feature map z.

[0027] In the above scheme, the polarization angle local feature extraction network includes a 3D convolution layer; the four-dimensional local preliminary feature map z is input into the 3D convolution layer, and the local features of the polarization angle are learned by the 3D convolution layer to obtain the polarization angle local feature map X l_3d ;

[0028] The spatial local feature extraction network includes a 2D convolution layer; the dimension of the four-dimensional local preliminary feature map z is expressed as C×D×H×W, where C represents the number of channels, D represents the number of rotation domain neighborhood window data, and H and W represent the height and width of the neighborhood window; under the C and D dimensions, the four-dimensional local preliminary feature map z is merged into a three-dimensional local feature map z′, whose dimension is C′×H×W, where C′=C×D; the three-dimensional local feature map z′ is input into the 2D convolution layer, and the 2D convolution layer is used to learn the local features of the space to obtain the spatial local feature map X l_2d ∈R C′×H×W .

[0029] In the above scheme, the polarization angle local feature map X l_2d With the spatial local feature map X l_2d The multi-head self-attention mechanism is input into the transformer encoding layer to learn the global features of polarization angle and space respectively, including:

[0030] Construct a polarization angle and spatial global feature extraction network, and transform the polarization angle local feature map X l_3d Expand in two dimensions: height H and width W, and normalize the expanded feature vector to ensure the consistency of feature scale. Then, use the multi-head attention mechanism to process the normalized features. The generated global feature map of polarization angle is denoted as X m_3d ; The spatial local feature map X l_2d Divide into K small flat patches X u_2d , where the dimension of each patch is P×K×C′, P is the number of pixels in each patch, that is, P=w′×h′, where w′ and h′ are the width and height of the patch respectively, and K is the number of patches, that is, K=HW / P. Then the flat patches are normalized, and the multi-head attention mechanism is used to capture the relationship of spatial global features. The resulting spatial global feature map is X m_2d , whose dimensions are P×N×C′, where N is the number of feature dimensions.

[0031] In the above scheme, the global features of the space are formatted and added in parallel with the global features of the polarization angle, and then input into the multi-layer perceptron MLP of the transformer encoding layer for global feature fusion to obtain the global feature fusion graph z″, including:

[0032] Construct a polarization angle and spatial global feature fusion network, and transform the polarization angle global feature map X m_3d and spatial global feature map X m_2d Perform feature fusion: transform the spatial global feature map X m_2d First expand it into the global feature map X with polarization angle m_3d compatible dimensions, and then expand the X m_2d and Xm_3d Splice and get a new feature map X n_3d , feature map X n_3d A sequence of feature vectors containing polarization rotation angles and spatial self-attention information; for feature map X n_3d Normalization is performed and then input into the multi-layer perceptron MLP. The feature map processed by the multi-layer perceptron MLP is compared with the original feature map X n_3d Perform residual connection and finally obtain the global feature fusion map z″, where the dimension of z″ is a 4D tensor of C×D×H×W, the same as the original input dimension.

[0033] In the above scheme, the 3D pooling layer performs a global maximum pooling operation on the feature map to map high-dimensional features to a low-dimensional space; the pooling operation reduces the spatial dimension of the feature map to a feature vector.

[0034] In the above scheme, the classifier includes two linear layers and the middle ReLU&Dropout layer. The output of the classifier is logits, which is the raw score of the category prediction. The Softmax function then maps logits to a probability distribution, indicating the predicted probability of each category.

[0035] According to a second aspect of the present invention, a polarimetric synthetic aperture radar is provided. The polarimetric synthetic aperture radar adopts any one of the steps of the polarimetric SAR image classification method combining polarization rotation angle and spatial self-attention described above.

[0036] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0037] The present invention proposes a polarization SAR image classification method combining polarization rotation angle and spatial self-attention, establishes a polarization angle and spatial feature extraction and fusion network architecture, enriches the polarization information of the polarization SAR image by rotating the polarization basis; combines the advantages of Transformer and convolutional neural networks to solve the problem that the polarization rotation angle features and spatial features are difficult to combine due to their different dimensions, fully extracts and fuses the local and global information of the polarization rotation angle features and spatial features of the polarization SAR image, and effectively improves the classification accuracy of the polarization SAR image data set.

[0038] In addition, the present invention extracts local features of polarization angle and space respectively through the polarization angle and spatial local feature extraction network architecture, and establishes a polarization angle and spatial global feature extraction and fusion network architecture based on the transformer coding layer. By expanding and folding format conversion and other methods, the convolutional neural network (including 3D and 2D convolution) and Transformer technologies are combined to realize polarization angle and spatial local and global feature extraction and fusion, which helps to improve the image processing effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Flowchart of the polarimetric SAR image classification method combining polarization rotation angle and spatial self-attention;

[0040] Figure 2 This is the structure diagram of the polarimetric SAR image classification model;

[0041] Figure 3 This is the network structure diagram of the polarization angle and spatial feature extraction module;

[0042] Figure 4 Pauli pseudo-color images and ground truth (GT) images of the two datasets;

[0043] Figure 5 The following are the classification results of two real PolSAR datasets. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0045] The present invention discloses a polarization SAR image classification method combining polarization rotation angle and spatial self-attention, comprising the following steps: preprocessing polarization SAR image data to obtain input for model training; using 3D convolution blocks to extract low-level feature representations of polarization SAR images; establishing a polarization angle and spatial local feature extraction network architecture to extract local features of polarization angle and space respectively; using transformer coding layer as the basic architecture to establish a polarization angle and spatial global feature extraction and fusion network architecture; using a combination of pooling layer-3D convolution blocks to downsample feature maps and extract high-level features; mapping feature vectors to classifiers to classify image categories and output classification results. The polarization SAR image classification method proposed by the present invention fully combines the attention information of polarization rotation angle and space, and deeply integrates local and global information of images, effectively improving the classification accuracy of polarization SAR images.

[0046] The polarization SAR image classification method combining polarization rotation angle and spatial self-attention of the present invention is as follows: Figure 1 As shown, the following steps are included:

[0047] S1. Preprocess polarimetric SAR image data to obtain input for model training.

[0048] S11. Based on Pauli decomposition, the coherence matrix of the polarization SAR image is constructed through the polarization scattering matrix, where the polarization scattering matrix contains the information of the four polarization channels of HH, HV, VH, and VV. The formula for constructing the coherence matrix T3 of the polarization SAR image based on Pauli decomposition and through the polarization scattering matrix is ​​as follows:

[0049]

[0050] Where T3 is the coherence matrix of the polarimetric SAR image; H and V represent the horizontal and vertical polarizations, respectively; S HH , S HV , S VH , S VV are the scattering coefficients of different polarization channels; * indicates complex conjugation.

[0051] S12. By rotating the polarization basis, the polarization coherence matrix at different rotation angles is calculated to obtain the polarization coherence matrix sequence of the polarization SAR image data as the input of the model. The formula of the polarization coherence matrix T(θ) is as follows:

[0052]

[0053] The PolSAR rotation domain data is defined as the dot product between the polarization coherence matrix T3 and the rotation matrix R3(θ); where θ represents the polarization rotation angle POA, and R3(θ) is defined as follows:

[0054]

[0055] As in the above formula, use the angle range The step length is 9 different POAs are used to generate polarization coherence matrix sequences, and based on the patch classification strategy, the neighborhood window data of each pixel is extracted from the polarization coherence matrix sequence as the input for model training.

[0056] S2. Use 3D convolution blocks to extract low-level feature representations of polarimetric SAR images.

[0057] like Figure 2 As shown in the figure, the polarization coherence matrix sequence is input into two 3D convolution blocks, where the convolution kernel slides in the local area of ​​the input data, calculates the weighted sum, extracts the geometric and local features in the multidimensional space of the polarization SAR image, and generates a preliminary feature map. Each 3D convolution block contains a 3D convolution layer, a 3D batch normalization (BN), and a sigmoid weighted linear unit (SiLU) activation layer. SiLU is a nonlinear activation function with the formula:

[0058]

[0059] S3, establish the polarization angle and spatial local feature extraction network architecture, such as Figure 3 As shown, the local features of polarization angle and space are extracted respectively.

[0060] S31, construct a polarization angle local feature extraction network, and use the four-dimensional preliminary feature map output from step S2 z It means that the local features of the polarization angle are learned through the 3D convolution layer, and the local feature map c of the polarization angle is obtained.

[0061] S32. In the C and D dimensions, the four-dimensional local feature map is merged into a three-dimensional local feature map, and then the three-dimensional local feature map is input into a 2D convolution layer to learn the spatial local feature representation. The preliminary feature map z is a 4D tensor with dimensions C×D×H×W; C represents the number of channels (the number of feature maps), D represents the number of neighborhood window data in the rotation domain, and H and W represent the height and width of the neighborhood window. C and D are merged to obtain a 3D tensor z′ with dimensions C′×H×W, where C′=C×D. z′ is processed by a 2D convolution layer to extract local spatial features, and the output spatial local feature map is X l_2d ∈R C′×H×W .

[0062] S4. Based on the transformer coding layer as the basic architecture, a polarization angle and spatial global feature extraction and fusion network architecture is established.

[0063] S41. Input the polarization angle local feature map and the spatial local feature map into the multi-head self-attention mechanism in the transformer encoding layer respectively to learn the global feature representation of polarization angle and space.

[0064] like Figure 3 As shown in the figure, a polarization angle and spatial global feature extraction network is constructed, and X l_3d Expand in two dimensions H (height) and W (width), and normalize the expanded feature vector to ensure that the scale of the features is consistent. Next, apply the multi-head attention mechanism to process the normalized features, and the generated polarization angle global feature map is denoted as X m_3d Then set X l_2d Divide into K small flat patches X u_2d , where the dimension of each patch is P×K×C′, P is the number of pixels in each patch (i.e., P=w′×h′, where w′ and h′ are the width and height of the patch, respectively), and K is the number of patches (i.e., K=HW / P). These small flat patches are normalized and a multi-head attention mechanism is applied to capture the relationship of spatial global features. Finally, the resulting spatial global feature map is X m_2d , whose dimensions are P×N×C′, where N is the number of feature dimensions.

[0065] S42: The global features of the space are converted into a new format, added in parallel with the global features of the polarization angle, and input into the next MLP of the transformer coding layer for global feature fusion.

[0066] like Figure 3 As shown in the figure, a polarization angle and spatial global feature fusion network is constructed, and the polarization angle global feature map X m_3d and the spatial global feature map X in S5 m_2d To perform feature fusion, X m_2d First expand (or reshape) to the same value as X m_3d compatible dimensions, ensuring that the two are aligned in dimension, and then, the expanded X m_2d and X m_3d Splice and get a new feature map X n_3d , which contains the feature vector sequence of polarization rotation angle and spatial self-attention information. n_3d Normalize the image and then input it into the multi-layer perceptron (MLP) to enhance the feature representation capability. n_3d Perform residual connection. Finally, the output feature map z″ is obtained. At this time, the dimension of z″ will be a 4D tensor of C×D×H×W, which is the same as the original input dimension.

[0067] S5. Use the combination of pooling layer and 3D convolution block to downsample the feature map and extract high-level features.

[0068] like Figure 2 As shown in the figure, the global feature fusion map is input into the maximum pooling layer for downsampling, and the dimension reduction and feature enhancement are achieved through aggregation operation; then it is input into the 3D convolution block to further extract higher-level features. Specifically, the combination of pooling layer-3D convolution block is used to downsample the feature map and extract high-level features, and the final feature map z″ is input into the 3D pooling layer in the backbone network for processing. In this step, the pooling layer is used to perform a global maximum pooling operation on the feature map to map the high-dimensional features to the low-dimensional space. The pooling operation reduces the spatial dimensions (height and width) of the feature map to a feature vector, which contains global information. The third 3D convolution block is used to further extract higher-level features.

[0069] S6. Map the feature vector to the classifier to classify the image category and output the classification result.

[0070] like Figure 2As shown in the figure, the total feature map is input into a classifier with a linear layer, ReLU&Dropout, and a linear layer for classification. The final output of the classifier is logits, which is the raw score of the category prediction. The logits are then mapped to a probability distribution through the Softmax function, indicating the predicted probability of each category.

[0071] The present invention uses two real PolSAR datasets, namely RS-2Wuhan and RADARSAT-2San Francisco (RS-2San), to evaluate the performance of the proposed method, and the detailed information is shown in Table 4. The Pauli pseudo-color images and ground truth (GT) images of these two datasets are shown in Table 4. Figure 4 As shown in the figure, the two figures on the left are the RS-2Wuhan dataset, the two figures on the right are the RS-2San dataset, the left figure of each two figures is the Pauli pseudo-color image, and the right figure is the ground truth (GT) image. All datasets are collected by the C-band RS-2 sensor in quadrupole mode. All training pixels are randomly selected from the labeled pixels, and the remaining labeled pixels are used for performance evaluation.

[0072] The proposed method is compared with four state-of-the-art PolSAR image classification methods.

[0073] 1) MobileViT and 3D-Conv-ViT MobileViT and Rotation-3-D-ViT blocks are proposed, so they are used as comparison methods.

[0074] 2) CNN-wavelet is an open source and excellent CNN-based PolSAR image classification method.

[0075] 3) PolSARFormer is a newly published ViT-based PolSAR image classification method.

[0076] The input of MobileViT, CNN-wavelet and PolSARFormer is only the polarization coherence matrix T without PolSAR rotation domain information. The input of 3D-Conv-ViT is the same as the proposed method. Overall accuracy (OA) (%) and Kappa coefficient are used to evaluate the performance of all methods.

[0077] Table 1 Performance evaluation comparison with other methods on RS-2Wuhan dataset

[0078]

[0079]

[0080] Table 2 Performance evaluation comparison with other methods on RS-2San dataset

[0081]

[0082] As shown in Tables 1 and 2 above, our method achieves the best classification performance on RS-2Wuhan and RS-2San datasets compared with the state-of-the-art CNN-based and ViT-based PolSAR image classification methods. Figure 5 These are the classification results of two real PolSAR datasets. The experimental results prove the effectiveness of polarization rotation angle and spatial self-attention features in PolSAR image classification.

[0083] The present invention also provides a polarimetric synthetic aperture radar, which adopts any one of the steps of the polarimetric SAR image classification method combining polarization rotation angle and spatial self-attention described above to achieve pixel classification.

[0084] In summary, the present invention proposes a polarization SAR image classification method combining polarization rotation angle and spatial self-attention, and establishes a polarization angle and spatial feature extraction and fusion network architecture. The method enriches the polarization information of the polarization SAR image by rotating the polarization basis; combines the advantages of Transformer and convolutional neural networks to solve the problem that the polarization rotation angle features and spatial features are difficult to combine due to their different dimensions, fully extracts and fuses the local and global information of the polarization rotation angle features and spatial features of the polarization SAR image, and effectively improves the classification accuracy of the polarization SAR image data set.

[0085] It should be noted that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0086] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, and two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0087] It will be easily understood by those skilled in the art that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A polarimetric SAR image classification method combining polarization rotation angle and spatial self-attention, characterized in that: The method includes: Obtain the polarization coherence matrix sequence of the polarization SAR image, and extract the neighborhood window four-dimensional data of the pixels to be classified; The four-dimensional data of the neighborhood window of the pixel to be classified is input into two 3D convolution blocks, and the low-level feature representation of the polarimetric SAR image is extracted by using the two 3D convolution blocks to generate a four-dimensional local preliminary feature map z; Establish a polarization angle local feature extraction network, input the four-dimensional local preliminary feature map z into the polarization angle local feature extraction network, and obtain the polarization angle local feature map X l_3d ; Establish a spatial local feature extraction network, merge the four-dimensional local preliminary feature map z into a three-dimensional local feature map z′ and input it into the spatial local feature extraction network to obtain the spatial local feature map X l_2d ; Based on the transformer coding layer as the basic architecture, the polarization angle and spatial global feature extraction and fusion network architecture is established: The polarization angle local feature map X l_3d With the spatial local feature map X l_2d The multi-head self-attention mechanism is input into the transformer encoding layer to learn the global features of polarization angle and space respectively; The global features of the space are formatted and added in parallel with the global features of the polarization angle, and then input into the multi-layer perceptron MLP of the transformer encoding layer for global feature fusion to obtain the global feature fusion graph z″; The global feature fusion map z″ is input into the 3D pooling layer and the 3D convolution block, and the combination of the 3D pooling layer and the 3D convolution block is used to downsample and extract high-level features of the global feature fusion map z″ to obtain a feature vector; The feature vector is mapped to the classifier to classify the polarimetric SAR image and the classification result of the pixel to be classified is output.

2. The polarimetric SAR image classification method combining polarization rotation angle and spatial self-attention according to claim 1, characterized in that: Obtain polarization coherence matrix sequence of polarization SAR image, including: Based on Pauli decomposition, the coherence matrix of polarimetric SAR images is constructed through the polarimetric scattering matrix. The polarimetric scattering matrix contains the information of four polarization channels: HH, HV, VH, and VV. The specific formula is as follows: Where T3 is the coherence matrix of the polarimetric SAR image; H and V represent the horizontal and vertical polarizations, respectively; S HH , S HV , S VH , S VV are the scattering coefficients of different polarization channels respectively; * indicates complex conjugate; By rotating the polarization basis, the polarization coherence matrix at different rotation angles is calculated to obtain the polarization coherence matrix sequence of the polarization SAR image; wherein, the formula of the polarization coherence matrix T(θ) is as follows: Where R3(θ) represents the polarization rotation matrix, represents the transpose; where the polarization rotation matrix R3(θ) is defined as follows: Wherein, θ represents the polarization direction angle POA.

3. The polarimetric SAR image classification method combining polarization rotation angle and spatial self-attention according to claim 2, characterized in that: The use angle range is The step length is Nine different polarization direction angles POA are used to generate a polarization coherence matrix sequence, and based on the patch classification strategy, the neighborhood window four-dimensional data of each pixel to be classified is extracted from the polarization coherence matrix sequence; wherein the number of polarization coherence matrix sequences corresponds to the number of different polarization direction angles POA.

4. The polarimetric SAR image classification method combining polarization rotation angle and spatial self-attention according to claim 1, characterized in that: Both 3D convolution blocks contain a 3D convolution layer, a 3D batch normalization BN, and a sigmoid weighted linear unit SiLU activation layer; where SiLU is a nonlinear activation function; The polarization coherence matrix sequence of the polarimetric SAR image is input into two 3D convolution blocks. The convolution kernel in the 3D convolution block slides in the local area of ​​the input data, calculates the weighted sum, extracts the geometric and local features in the multidimensional space of the polarimetric SAR image, and generates a four-dimensional local preliminary feature map z.

5. The polarimetric SAR image classification method combining polarization rotation angle and spatial self-attention according to claim 1, characterized in that: The polarization angle local feature extraction network includes a 3D convolution layer; the four-dimensional local preliminary feature map z is input into the 3D convolution layer, and the local features of the polarization angle are learned by the 3D convolution layer to obtain the polarization angle local feature map X l_3d ; The spatial local feature extraction network includes a 2D convolution layer; the dimension of the four-dimensional local preliminary feature map z is expressed as C×D×H×W, where C represents the number of channels, D represents the number of rotation domain neighborhood window data, and H and W represent the height and width of the neighborhood window; under the C and D dimensions, the four-dimensional local preliminary feature map z is merged into a three-dimensional local feature map z′, whose dimension is C′×H×W, where C′=C×D; the three-dimensional local feature map z′ is input into the 2D convolution layer, and the 2D convolution layer is used to learn the local features of the space to obtain the spatial local feature map X l_2d ∈R C′×H×W .

6. The polarimetric SAR image classification method combining polarization rotation angle and spatial self-attention according to claim 5, characterized in that: The polarization angle local feature map X l_3d With the spatial local feature map X l_2d The multi-head self-attention mechanism is input into the transformer encoding layer to learn the global features of polarization angle and space respectively, including: Construct a polarization angle and spatial global feature extraction network, and transform the polarization angle local feature map X l_3d Expand in two dimensions: height H and width W, and normalize the expanded feature vector to ensure the consistency of feature scale. Then, use the multi-head attention mechanism to process the normalized features. The generated global feature map of polarization angle is denoted as X m_3d ; The spatial local feature map X l_2d Divide into K small flat patches X u_2d , where the dimension of each patch is P×K×C′, P is the number of pixels in each patch, that is, P=w′×h′, where w′ and h′ are the width and height of the patch respectively, and K is the number of patches, that is, K=HW / P. Then the flat patches are normalized, and the multi-head attention mechanism is used to capture the relationship of spatial global features. The resulting spatial global feature map is X m_2d , whose dimensions are P×N×C′, where N is the number of feature dimensions.

7. The polarimetric SAR image classification method combining polarization rotation angle and spatial self-attention according to claim 6, characterized in that: The global features of the space are converted into a format, added in parallel with the global features of the polarization angle, and then input into the multi-layer perceptron MLP of the transformer encoding layer for global feature fusion to obtain the global feature fusion graph z″, including: Construct a polarization angle and spatial global feature fusion network, and transform the polarization angle global feature map X m_3d and spatial global feature map X m_2d Perform feature fusion: transform the spatial global feature map X m_2d First expand it into the global feature map X with polarization angle m_3d compatible dimensions, and then expand the X m_2d and X m_3d Splice and get a new feature map X n_3d , feature map X n_3d A sequence of feature vectors containing polarization rotation angles and spatial self-attention information; for feature map X n_3d Normalization is performed and then input into the multi-layer perceptron MLP. The feature map processed by the multi-layer perceptron MLP is compared with the original feature map X n_3d Perform residual connection and finally obtain the global feature fusion map z″, where the dimension of z″ is a 4D tensor of C×D×H×W, the same as the original input dimension.

8. The polarimetric SAR image classification method combining polarization rotation angle and spatial self-attention according to claim 1, characterized in that: The 3D pooling layer performs a global maximum pooling operation on the feature map to map high-dimensional features to a low-dimensional space; the pooling operation reduces the spatial dimension of the feature map to a feature vector.

9. The polarimetric SAR image classification method combining polarization rotation angle and spatial self-attention according to claim 1, characterized in that: The classifier consists of two linear layers and a middle ReLU&Dropout layer. The output of the classifier is logits, which is the raw score of the category prediction. The Softmax function then maps logits to a probability distribution, indicating the predicted probability of each category.

10. A polarimetric synthetic aperture radar, characterized in that: The polarimetric synthetic aperture radar adopts the steps of the polarimetric SAR image classification method combining polarization rotation angle and spatial self-attention as described in any one of claims 1 to 9.

Images