Unsupervised PolSAR classification method based on comparative learning

Abstract

The invention relates to the technical field of image processing, in particular to an unsupervised PolSAR classification method based on comparative learning, and the method comprises the steps: obtaining original PolSAR image data to form multi-view data; performing superpixel segmentation on the image to generate multi-view primitive samples in one-to-one correspondence with superpixels; obtaining potential feature representation of each view, and generating a self-representation coefficient matrix of each view through a forward self-representation module; applying cross-view relation consistency constraint by comparing loss, and optimizing a self-representation coefficient matrix of each view; fusing the self-representation coefficient matrixes of all the optimized views to generate a uniform affinity matrix, and executing spectral clustering to obtain clustering labels of element samples; according to the mapping relation between the element sample and the pixel established by the superpixel segmentation, the clustering label is mapped back to the pixel level, the image classification result is obtained, and the classification accuracy and robustness are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to an unsupervised PolSAR classification method based on contrastive learning. Background Technology

[0002] Polarimetric Synthetic Aperture Radar (PolSAR) possesses all-weather, all-time imaging capabilities, and its image classification is a crucial step in remote sensing interpretation. While existing deep learning-based supervised classification methods offer high accuracy, they rely on a large number of manually labeled samples. However, PolSAR data labeling is costly, highly specialized, and difficult to obtain, limiting its practical application. Therefore, unsupervised classification methods that do not require labeling have become an important research direction.

[0003] Deep Multi-View Subspace Clustering (DMVSC) has shown potential in unsupervised classification by fusing multi-source features such as polarization and texture in PolSAR images to learn shared subspace structures. However, its application to large-scale high-resolution PolSAR images still faces the following key technical bottlenecks: (1) Insufficient scalability: Directly constructing a global self-representation matrix at the pixel level and performing spectral clustering results in a sharp increase in computational and memory overhead as the image size grows, making it difficult to apply to million-pixel scenarios; (2) Inefficient self-representation construction: Existing methods either rely on iterative optimization (such as ADMM), which separates feature learning from self-representation solving; or parameterize the self-representation matrix into a fully connected layer, introducing redundant parameters, which is not conducive to end-to-end optimization. (3) Weak view alignment mechanism: mainstream contrastive learning methods only align sample-level feature representations, which makes it difficult to guarantee that the underlying consistent subspace structure is learned when there is strong noise or significant differences between multiple views.

[0004] Therefore, there is an urgent need for a scalable, efficient, and structure-level view-aligned unsupervised PolSAR image classification method to overcome the practical bottleneck of large-scale remote sensing image processing. Summary of the Invention

[0005] To address the key technical bottlenecks mentioned above, such as poor scalability, low efficiency in self-representation construction, and weak view alignment mechanisms, which make it difficult for existing image classification methods to achieve efficient, robust, and end-to-end optimized unsupervised classification, the present invention aims to provide an unsupervised PolSAR classification method based on contrastive learning. The specific technical solution adopted is as follows: One embodiment of the present invention provides an unsupervised PolSAR classification method based on contrastive learning, the method comprising the following steps: Acquire raw PolSAR image data and extract at least two types of pixel-level features to construct multi-view data; The image is segmented into superpixels, and the features of all pixels in each superpixel under each view are aggregated to generate multi-view meta samples that correspond one-to-one with the superpixels. The multi-view meta-samples are input into the corresponding deep feature extraction network to obtain the latent feature representation of each view. The self-representation coefficient matrix of each view is generated through the forward self-representation module, where each column corresponds to the self-representation coefficient vector of a meta-sample, representing its weights linearly reconstructed from all meta-samples. Based on the relational consistency contrastive learning mechanism, the self-representation coefficient vectors of the same meta-sample under different views are constructed as positive sample pairs. Cross-view relational consistency constraints are applied through contrastive loss to optimize the self-representation coefficient matrix of each view. The self-representation coefficient matrices of each view after fusion and optimization are used to generate a unified affinity matrix, and spectral clustering is performed to obtain the clustering labels of the meta-samples. Based on the mapping relationship between metasamples and pixels established by superpixel segmentation, clustering labels are mapped back to the pixel level to obtain image classification results.

[0006] Furthermore, the multi-view data includes polarization feature views and texture feature views.

[0007] Furthermore, the aggregation uses a truncated mean aggregation function.

[0008] Further, obtaining the self-representation coefficient matrix includes: L2 normalization is performed on the latent features of all meta-samples under each view to obtain the normalized feature matrix; Calculate the inner product between any two meta-sample features in the normalized feature matrix to generate an initial similarity matrix; An adaptive soft thresholding operator controlled by learnable parameters is applied to the initial similarity matrix, and a self-representation coefficient matrix is ​​output during forward propagation.

[0009] Furthermore, the optimization of the self-representation coefficient matrix of each view includes: The relational consistency contrastive learning mechanism adopts the InfoNCE loss function, in which the self-representation coefficient vector of the same meta-sample under different views is used as a positive sample pair, and the self-representation coefficient vector of the meta-sample and other meta-samples under any view is used as a negative sample pair. By maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs, the self-representation structures of each view are aligned in a shared subspace, thereby optimizing the self-representation coefficient matrix of each view.

[0010] Furthermore, the self-representation coefficient matrices of the fused and optimized views generate a unified affinity matrix, including: Based on the self-representation coefficient matrix of each view, the views are weighted and fused using learnable attention weights; The attention weights are dynamically generated by a lightweight neural network based on the statistical characteristics or structural information of the self-representation coefficient matrix of each view, and are used to characterize the contribution of each view to the final clustering result. The unified affinity matrix is ​​obtained by summing the weighted self-representation coefficient matrices of each view.

[0011] Furthermore, the deep feature extraction network is a multi-view autoencoder structure, and the training process of the network includes minimizing the reconstruction loss between the input and output of each view.

[0012] Furthermore, after obtaining the self-representation coefficient matrix and the latent feature representation, the method further includes: Regularization constraints are applied to the self-representation coefficient matrix and the latent feature representation, respectively. The regularization constraints include elastic net regularization and graph regularization.

[0013] The present invention has the following beneficial effects: This invention provides an unsupervised PolSAR classification method based on contrastive learning. This method introduces superpixel segmentation to aggregate the original pixel-level data into multi-view meta-samples with a significantly reduced order of magnitude. This allows subsequent self-representation matrix construction and spectral clustering operations to be performed at the meta-sample level, greatly reducing computational complexity and storage requirements. This overcomes the scalability problem of existing methods being unsuitable for high-resolution, large-scale PolSAR images. By designing a forward self-representation module, sparse self-representation coefficient matrices are directly generated during the network's forward propagation, avoiding the training fragmentation and parameter redundancy problems caused by relying on iterative optimization algorithms or introducing fully connected parameters in traditional methods, thus improving deep feature extraction. The feature learning and subspace structure discovery of the forward self-representation module are jointly updated under the same optimization objective, thereby improving the stability and convergence efficiency of deep multi-view subspace clustering network training. Through the relation consistency contrastive learning mechanism, the view alignment objective is elevated from the sample feature level to the sample relation structure level. Specifically, by constraining the self-representation coefficient vectors of the same meta-sample under different views to form positive sample pairs, the reconstructed relations of each view are aligned in the shared subspace. This mechanism effectively suppresses the structural mismatch caused by differences between views or noise interference, making the learned subspace more reflective of the essential distribution characteristics of land cover categories, thereby improving the accuracy and reliability of unsupervised classification results. Attached Figure Description

[0014] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart illustrating an unsupervised PolSAR classification method based on contrastive learning, as an embodiment of the present invention; Figure 2 This is a schematic diagram of the process for constructing multi-view meta samples in an embodiment of the present invention; Figure 3 This is a schematic diagram of the overall architecture of the deep multi-view subspace clustering network model in an embodiment of the present invention; Figure 4 These are the truth graph and PauliRGB graph in the embodiments of the present invention; Figure 5 This is a visualization comparison of the methods in the embodiments of the present invention on the Oberpfaffenhofen dataset. Detailed Implementation

[0016] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific implementation methods, structures, features, and effects of the technical solution proposed according to the present invention are described in detail below with reference to the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] This embodiment provides an unsupervised PolSAR classification method based on contrastive learning, such as... Figure 1 As shown, it includes the following steps: S1. Acquire raw PolSAR image data and extract at least two pixel-level features to form multi-view data.

[0019] First, acquire the raw PolSAR image data. Then, extract at least two types of pixel-level features from the raw PolSAR image data to construct multi-view data. The multi-view data includes a polarization feature view and a texture feature view.

[0020] As an exemplary implementation, acquiring multi-view data includes: The first step is to extract polarization features that characterize the electromagnetic scattering mechanism of ground objects, starting from the original PolSAR complex covariance matrix. These features include, but are not limited to: The real and imaginary elements of the covariance matrix, the polarization channel ratio, the depolarization ratio, the scattering energy components, and 38 polarization features obtained based on decomposition methods such as Pauli, Krogager, Freeman, HA-α, and Huynen are used. These 38 features constitute the polarization feature matrix. Where N represents the total number of pixels in the image, a polarization feature view can be obtained.

[0021] The second step involves first generating an RGB pseudo-color composite image based on the Pauli basis; then, local spatial structure features are extracted from this image using a sliding window (preferably with a window size of 11×1111×11 pixels), including: Gray-level co-occurrence matrix: Calculates statistics in four directions (0°, 45°, 90°, 135°) for contrast, correlation, energy, and homogeneity, totaling 16 dimensions; Gabor filter response: Convolution is performed using Gabor kernels with multiple scales (e.g., 3 scales) and multiple directions (e.g., 6 directions), and the response energy is taken as the feature, for a total of 18 dimensions; Optionally, auxiliary texture descriptors such as Local Binary Pattern (LBP) or Histogram of Gradient Orientation (HOG) can be added.

[0022] The above features are normalized and then concatenated to form a 48-dimensional texture feature vector, which constitutes the texture feature matrix. This allows you to obtain a view of the texture features.

[0023] It should be noted that the above dimensions (38-dimensional, 48-dimensional) are merely exemplary implementation schemes. Those skilled in the art can adjust the feature types, quantities, and combinations according to the actual PolSAR data quality, ground feature complexity, and computing resources, and all such adjustments should be considered equivalent implementations of the present invention.

[0024] S2 performs superpixel segmentation on the image, aggregates the features of all pixels in each superpixel under each view, and generates multi-view meta samples that correspond one-to-one with the superpixels.

[0025] It should be noted that, in order to address the computational bottleneck caused by large-scale data, this embodiment aggregates pixel-level data into meta-samples. The flowchart illustrating the construction of multi-view meta-samples is shown below. Figure 2 As shown.

[0026] In one embodiment of the present invention, the superpixel segmentation uses the Simple Linear Iterative Clustering (SLIC) algorithm to segment the Pauli pseudo-color composite image of the PolSAR image, generating M spatially continuous and compact superpixel regions, where M≪N (i.e., M is much smaller than N, and N is the total number of original pixels), thereby significantly reducing the data size of subsequent processing.

[0027] For each superpixel region, robust aggregation of the features of all pixels within it across different views is performed to generate the corresponding meta-sample feature vector. To suppress the influence of anomalous pixels (such as speckle noise and edge aliasing) on ​​the aggregation results, a truncated mean is preferably used as the aggregation function: that is, for each feature dimension, the highest and lowest pixel values ​​within the superpixel are removed according to a preset ratio, and the mean of the remaining pixel values ​​is calculated.

[0028] Through the above aggregation, a meta-sample level multi-view feature matrix is ​​obtained, including a polarization feature matrix and a texture feature matrix.

[0029] As an example, the formula for calculating the j-th feature vector of the k-th meta-sample under the v-th view can be: ; In the formula, This represents the j-th feature vector of the k-th meta-sample under the v-th view. This indicates the cutoff ratio (i.e., the preset ratio), such as 0.005. This represents the total number of pixels in the region corresponding to the k-th meta-sample. This indicates the removal of the highest and lowest values. The set of indices after 1 pixel This represents the j-th component of the feature of the i-th pixel in the index set under the v-th view.

[0030] Simultaneously, during the segmentation process, the original pixel index set contained in each metasample is recorded and saved, establishing a bidirectional mapping relationship between metasamples and pixels. This mapping relationship is used to accurately backfill the clustering labels of metasamples to the corresponding pixel positions after subsequent spectral clustering, thereby generating the final pixel-level classification result.

[0031] A schematic diagram of the overall architecture of the deep multi-view subspace clustering network model is shown below. Figure 3 As shown, steps S3 to S6 include the construction, training, and clustering processes of the network model.

[0032] S3 inputs the multi-view meta-samples into the corresponding deep feature extraction network to obtain the latent feature representations of each view, and generates the self-representation coefficient matrix of each view through the forward self-representation module.

[0033] Multi-view meta-samples are input into their respective deep feature extraction networks. Each deep feature extraction network can employ a convolutional neural network, graph neural network, or fully connected network structure. Its function is to map the original high-dimensional, redundant meta-sample features to a low-dimensional, highly discriminative latent feature space, thereby obtaining a more semantically expressive latent feature representation for each view. Furthermore, to avoid the computational overhead and training fragmentation problems caused by iterative optimization in traditional subspace clustering methods, this embodiment introduces a forward self-representation module. This module takes the latent feature representation of each view as input and directly generates a sparse self-representation coefficient matrix during a single forward propagation.

[0034] As an exemplary implementation, obtaining the self-representation coefficient matrix includes: The first step is to perform L2 normalization on the latent features of all meta-samples under each view to obtain the normalized feature matrix.

[0035] In this embodiment, the latent feature representations of all meta-samples under each view are normalized using the L2 norm, so that the feature vector of each meta-sample lies on the unit hypersphere. The implementation process of L2 norm normalization is existing technology and is not within the scope of this invention; therefore, it will not be described in detail here.

[0036] It should be noted that the above normalization operation can eliminate feature scale differences, making the subsequent inner product directly equivalent to cosine similarity, which is beneficial to improving the robustness of similarity measurement.

[0037] In another embodiment, an autoencoder network is constructed for each view, consisting of an encoder and a decoder, for example, a multilayer perceptron with a symmetric structure. The network input is a meta-sample feature matrix. In this embodiment, the network is trained by minimizing the reconstruction loss, enabling the encoder to learn an effective low-dimensional latent feature representation of the data. The expression for the reconstruction loss can be: In the formula, This represents the reconstruction loss, where V represents the number of views. This represents the meta-sample feature matrix under the v-th view. This represents the encoder under the v-th view. This represents the decoder under the v-th view. This represents the square of the Frobenius norm of the matrix.

[0038] The second step is to calculate the inner product between any two meta-sample features in the normalized feature matrix to generate the initial similarity matrix.

[0039] In this embodiment, the inner product between any two meta-samples is calculated based on the normalized feature matrix to generate an initial similarity matrix.

[0040] The third step is to apply an adaptive soft thresholding operator controlled by learnable parameters to the initial similarity matrix, and output a self-representation coefficient matrix during forward propagation.

[0041] In this embodiment, in order to obtain a sparse and structured self-representation coefficient matrix, an adaptive soft thresholding operator controlled by learnable parameters is applied to the initial similarity matrix, and the operator is applied element by element to the initial similarity matrix to obtain a sparse self-representation coefficient matrix.

[0042] As an example, the output self-representation coefficient matrix can be generated using the forward self-representation module, and its expression can be: In the formula, Represents the self-representation coefficient matrix. Represents the initial similarity matrix. This represents an adaptive soft thresholding operator controlled by a learnable threshold parameter, which can be defined as follows: , This represents an adaptive soft thresholding operator controlled by a learnable threshold parameter b, used to perform element-wise sparsification on the initial similarity matrix S. sgn represents the sign function, and max represents the maximum value function. denoted by , b represents the element-wise absolute value of the initial similarity matrix, and b represents the learnable threshold parameter.

[0043] The expression for the self-representation loss function can be: In the formula, Let L represent the self-representation loss, and L be the number of potential representation layers for the self-representation constraint. This indicates that the output is taken from the l-th layer of the encoder, so L can correspond to the number of encoder layers; This represents the self-representation coefficient matrix.

[0044] It should be noted that the self-representation coefficient matrix is ​​directly output during forward propagation without iterative optimization, and its sparse structure implicitly characterizes the affinity relationship between similar meta-samples, providing high-quality input for subsequent cross-view alignment and spectral clustering.

[0045] S4, based on the relational consistency contrastive learning mechanism, constructs positive sample pairs by converting the self-representation coefficient vectors of the same meta-sample under different views, and applies cross-view relational consistency constraints through contrastive loss to optimize the self-representation coefficient matrix of each view.

[0046] To achieve deep alignment between views, relational consistency contrastive learning is introduced. In this embodiment, the relational consistency contrastive learning mechanism uses the self-representation coefficient vectors describing the global relationship of the same metasample under different views as positive sample pairs, and employs a cross-view contrastive loss of the form InfoNCE for constraint, prompting different views to learn the shared subspace structure. The contrastive loss function expression is: In the formula, Indicates comparative loss, Let V represent the number of metasamples, V represent the number of views, log represents the logarithmic function, exp represents the exponential function with base to the natural constant, and sim represents the cosine similarity. Let represent the self-representation coefficient vector of the k-th meta-sample under the i-th view. Let represent the self-representation coefficient vector of the k-th meta-sample under the j-th view. For temperature coefficient, Indicates the j-th view under the first view. The self-representation coefficient vector of each sample.

[0047] S5, merge the self-representation coefficient matrices of each view after optimization to generate a unified affinity matrix, and perform spectral clustering to obtain the clustering labels of the meta-samples.

[0048] It should be noted that the self-representation coefficient matrices of each view after relational consistency comparison learning optimization already have cross-view structure alignment characteristics. In order to make full use of the complementary information of multiple views and improve the clustering robustness, they need to be merged into a unified affinity matrix.

[0049] As an exemplary implementation, a unified affinity matrix is ​​generated by fusing the optimized self-representation coefficient matrices of each view, including: To fuse multi-view information, an attention mechanism can be used to adaptively assign weights to the self-representation coefficient matrices of each view. The weighted matrices are then fused, for example, through convolution operations, resulting in a unified affinity matrix. Furthermore, to improve model performance, regularization terms can be introduced. For instance, applying elastic mesh regularization loss to the self-representation coefficient matrices combines the sparsity-inducing ability of the L1 norm with the numerical stability advantage of the L2 norm; and applying graph regularization loss to the latent feature representations preserves the inherent geometric structure of the data.

[0050] As an exemplary implementation, based on a unified affinity matrix, spectral clustering is performed to obtain cluster labels for meta-samples, including: In this embodiment, a symmetric similarity matrix is ​​calculated based on the fused unified affinity matrix, and its expression can be: In the formula, W represents the similarity matrix, and C represents the fused unified affinity matrix. This represents the transpose of the unified affinity matrix after fusion.

[0051] Applying a spectral clustering algorithm to the similarity matrix yields cluster labels for M meta-samples. The implementation process of the spectral clustering algorithm is existing technology and is not within the scope of this invention; therefore, it will not be described in detail here.

[0052] It should be noted that the process of determining cluster labels can map high-dimensional, nonlinear meta-sample relationships to a low-dimensional spectral space, achieving clustering results that are compact within clusters and separate between clusters while preserving the global structure.

[0053] S6: Based on the mapping relationship between metasamples and pixels established by superpixel segmentation, the clustering labels are mapped back to the pixel level to obtain the image classification results.

[0054] In this embodiment, based on the mapping relationship saved in step S2 above, the metasample label is assigned to all corresponding pixels, thereby obtaining the final pixel-level image classification result.

[0055] The overall objective function of the network model can be expressed as a weighted sum of the loss terms. The expression for this is: In the formula, This represents the total loss of the network model. , , All of these are hyperparameters. Indicates the reconstruction loss. Indicates self-indication of loss, This represents the regularization loss of the elastic net. This represents the graph regularization loss. This indicates a comparative loss.

[0056] Training can employ a two-stage strategy: first, pre-train the multi-view autoencoder using reconstruction loss, and then perform end-to-end optimization by combining self-representation, contrast, and regularization terms.

[0057] To verify the implementation effect of the embodiments of the present invention and to better illustrate the advantages of the present invention, the following verification and explanation are provided: First, determine the evaluation indicators.

[0058] This embodiment uses evaluation metrics widely used in clustering methods to assess clustering performance, namely Normalized Mutual Information (NMI), Adjusted Rand Index (ARI), Accuracy (ACC), Precision, Recall, and Balanced Mean (F-score). These six metrics evaluate the quality of clustering results from multiple perspectives, and higher evaluation values ​​indicate better performance.

[0059] Second, prepare the dataset.

[0060] In this embodiment, the Oberpfaffenhofen dataset was acquired by the German Aerospace Center (DLR) using its airborne synthetic aperture radar system ESAR in the Oberpfaffenhofen region of Germany. It includes a PolSAR image with a size of 1300×1200 pixels and a spatial resolution of 3 meters, labeled with three land cover classes (built-up area, woodland, and open space). The ground truth map and PauliRGB image are shown below. Figure 4 As shown.

[0061] Third, set up the experiment.

[0062] The network model is implemented using the PyTorch framework and employs the Adam optimizer. The autoencoder uses a fully connected network structure with two hidden layers and the ReLU activation function. The image is uniformly segmented into approximately 8000 superpixels (M≈8000). Hyperparameters are determined through grid search; for example, the optimal value of hyperparameter λ1 is selected from {0.01, 0.1, 1, 10}, while the optimal values ​​of λ2 and λ3 are determined from {0.01, 0.05, 0.1, 1, 5, 10}. The experimental environment includes Python 3.13, an Intel i5-12400F processor, and an NVIDIA TITAN XP graphics card.

[0063] Fourth, comparative experiments.

[0064] We employ a variety of representative deep multi-view clustering (DMVSC) or related methods as baselines, including MFLVC, D2MVSC, DMAC, DMCAG, MSESC, EDESC, DSMVC, etc., covering different fusion and consistency modeling strategies.

[0065] Table 1 shows the clustering metrics results of the comparative experimental models on the Oberpfaffenhofen dataset: Table 1 As shown in Table 1, on the Oberpfaffenhofen dataset, compared to the suboptimal method DMAC, the proposed method improves the performance of ACC, NMI, ARI, F-score, Precision, and Recall metrics by 6.9%, 2.08%, 0.83%, 1.26%, 4.94%, and 1.03%, respectively. This performance improvement over DMAC is attributed to stronger consistency constraints: by imposing structural consistency constraints at the meta-sample level through contrastive learning, this invention more effectively guides the model to learn cross-view invariant discriminative features compared to DMAC's macro-distribution alignment. Furthermore, this invention introduces an attention mechanism at the structural level for fusion, which adaptively measures and utilizes the contribution of each view, outperforming DMAC's feature-level average fusion.

[0066] The visualization comparison of each method on the Oberpfaffenhofen dataset is shown below. Figure 5 As shown. Although the suboptimal method DMAC performs reasonably well in woodland and open areas, it exhibits significant confusion and misclassification of built-up areas, resulting in a substantial loss of detail in this category. In contrast, this invention not only accurately classifies major feature categories but also more clearly preserves the outline and internal structure of built-up areas, demonstrating superior overall visual consistency and detail fidelity compared to other methods.

[0067] Fourth, ablation experiment.

[0068] To further explore the contributions of each core component in the method of this invention, ablation experiments were conducted based on the above embodiments. First, it should be noted that the superpixel metasample construction module is the foundation of the entire framework and is indispensable. Its core function is to transform large-scale high-resolution images from computationally intractable pixel-level data into efficient and feasible metasamples. Without this module, the subsequent deep clustering model would be unable to run due to enormous memory overhead. Therefore, the effectiveness of this module is not reflected in the numerical improvement of clustering metrics, but rather in solving the fundamental problem of making the model runnable. Subsequent ablation experiments were conducted based on this. Subsequently, ablation analysis was performed on the three key loss terms in the model. Table 2 shows the ablation results on the six clustering metrics after removing the contrastive learning module, the graph regularization module, and retaining only the self-representation module. Optimal performance was achieved only when all modules were used, fully demonstrating the effectiveness of each module in the method of this application.

[0069] Table 2 In summary, this invention solves the scalability problem of large-scale image processing by constructing superpixel meta-samples. By utilizing an efficient forward self-representation module and an innovative relational consistency contrastive learning mechanism, it achieves significant improvements in classification accuracy and robustness, providing an effective technical solution for the application of multi-view subspace clustering theory in unsupervised classification of PolSAR images.

[0070] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An unsupervised PolSAR classification method based on contrastive learning, characterized in that, Includes the following steps: Acquire raw PolSAR image data and extract at least two types of pixel-level features to construct multi-view data; The image is segmented into superpixels, and the features of all pixels in each superpixel under each view are aggregated to generate multi-view meta samples that correspond one-to-one with the superpixels. The multi-view meta-samples are input into the corresponding deep feature extraction network to obtain the latent feature representation of each view. The self-representation coefficient matrix of each view is generated through the forward self-representation module, where each column corresponds to the self-representation coefficient vector of a meta-sample, representing its weights linearly reconstructed from all meta-samples. Based on the relational consistency contrastive learning mechanism, the self-representation coefficient vectors of the same meta-sample under different views are constructed as positive sample pairs. Cross-view relational consistency constraints are applied through contrastive loss to optimize the self-representation coefficient matrix of each view. The self-representation coefficient matrices of each view after fusion and optimization are used to generate a unified affinity matrix, and spectral clustering is performed to obtain the clustering labels of the meta-samples. Based on the mapping relationship between metasamples and pixels established by superpixel segmentation, clustering labels are mapped back to the pixel level to obtain image classification results.

2. The unsupervised PolSAR classification method based on contrastive learning according to claim 1, characterized in that, The multi-view data includes polarization feature views and texture feature views.

3. The unsupervised PolSAR classification method based on contrastive learning according to claim 1, characterized in that, The aggregation uses a truncated mean aggregation function.

4. The unsupervised PolSAR classification method based on contrastive learning according to claim 1, characterized in that, Obtaining the self-representation coefficient matrix includes: L2 normalization is performed on the latent features of all meta-samples under each view to obtain the normalized feature matrix; Calculate the inner product between any two meta-sample features in the normalized feature matrix to generate an initial similarity matrix; An adaptive soft thresholding operator controlled by learnable parameters is applied to the initial similarity matrix, and a self-representation coefficient matrix is ​​output during forward propagation.

5. The unsupervised PolSAR classification method based on contrastive learning according to claim 1, characterized in that, The optimization of the self-representation coefficient matrix for each view includes: The relational consistency contrastive learning mechanism adopts the InfoNCE loss function, in which the self-representation coefficient vector of the same meta-sample under different views is used as a positive sample pair, and the self-representation coefficient vector of the meta-sample and other meta-samples under any view is used as a negative sample pair. By maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs, the self-representation structures of each view are aligned in a shared subspace, thereby optimizing the self-representation coefficient matrix of each view.

6. The unsupervised PolSAR classification method based on contrastive learning according to claim 1, characterized in that, The self-representation coefficient matrices of each view after fusion optimization generate a unified affinity matrix, including: Based on the self-representation coefficient matrix of each view, the views are weighted and fused using learnable attention weights; The attention weights are dynamically generated by a lightweight neural network based on the statistical characteristics or structural information of the self-representation coefficient matrix of each view, and are used to characterize the contribution of each view to the final clustering result. The unified affinity matrix is ​​obtained by summing the weighted self-representation coefficient matrices of each view.

7. The unsupervised PolSAR classification method based on contrastive learning according to claim 1, characterized in that, The deep feature extraction network is a multi-view autoencoder structure, and the training process of the network includes minimizing the reconstruction loss between the input and output of each view.

8. The unsupervised PolSAR classification method based on contrastive learning according to claim 1, characterized in that, After obtaining the self-representation coefficient matrix and the latent feature representation, the method further includes: Regularization constraints are applied to the self-representation coefficient matrix and the latent feature representation, respectively. The regularization constraints include elastic net regularization and graph regularization.

Images