PolSAR image classification method based on physical guidance and dynamic knowledge injection

Abstract

The application discloses a PolSAR image classification method based on physical guidance and dynamic knowledge injection, mainly solving the problems that the existing PolSAR classification model lacks full-process physical constraints and is easily disturbed by noise due to static fusion of physical priori. The method comprises the following steps: 1) a double-branch feature extraction network is built, global and local features are extracted in multiple scales, and cross-attention mechanism is used to realize double-branch feature interaction and fusion; 2) a center coordinate attention dynamic physical injection module is designed, nonlinear mapping and spatial re-labeling are performed on the physical priori features, and deep features are adaptively fused through a gated residual; and 3) a physical scattering constraint auxiliary branch is added, physical dominant labels are generated relying on polarization decomposition, and physical consistency constraints are applied through multi-task learning. The application can effectively improve the classification accuracy and feature physical consistency, and enhance the model generalization ability in the limited labeling scene, and can be applied to the fields of ground object classification, environment monitoring and remote sensing information interpretation.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, and further relates to Polarimetric Synthetic Aperture Radar (PolSAR) image interpretation and land cover classification technology. Specifically, it is a PolSAR image classification method based on physical guidance and dynamic knowledge injection, which can be used in the fields of land cover classification, environmental monitoring and remote sensing information interpretation. Background Technology

[0002] Polarimetric Synthetic Aperture Radar (PolSAR), as an important active microwave remote sensing imaging technology, can acquire multi-channel polarimetric scattering information of surface targets under all-weather and all-day conditions. It has significant application value in fields such as agricultural monitoring, urban area analysis, land cover classification, and disaster assessment. PolSAR image classification is a crucial component of PolSAR image interpretation. Its core function is to transform the multi-dimensional features acquired by PolSAR into usable land cover category information, providing fundamental data support for subsequent applications such as land cover identification, change detection, and macro-level decision-making.

[0003] Traditional machine learning-based PolSAR image classification typically employs a two-stage process: feature extraction and classifier design. The feature extraction stage often uses methods such as polarimetric decomposition and texture statistics to extract polarimetric scattering and image features, reducing information redundancy and generating discriminative input. The classification stage uses models such as Wishart clustering, support vector machines, and random forests to map features to land cover categories. However, these methods heavily rely on the quality of manually designed features, making it difficult to fully exploit the deep scattering information contained in PolSAR data, and their classification performance is limited by the completeness of the feature representation.

[0004] In recent years, deep learning architectures such as convolutional neural networks, recurrent neural networks, and Transformers have been widely used in polarimetric SAR (PolSAR) image classification. Relying on end-to-end autonomous feature learning, they have demonstrated feature extraction and classification performance far exceeding that of traditional methods in complex terrain scenes. Bi et al. provided a systematic review of deep learning solutions in this field, but existing mainstream methods still suffer from three core shortcomings that are difficult to overcome: First, the models rely entirely on data-driven training, failing to embed the polarimetric electromagnetic scattering mechanism as a hard constraint into the feature learning process. They rely solely on labeled samples to fit the data distribution, which easily leads to classification results that violate physical laws in areas where ground object scattering mechanisms overlap, resulting in a loss of classification accuracy and physical consistency in decision-making. For example, the application No. 201910968153.8, entitled "A Polarimetric SAR Image Classification Method Based on Deep Complex-Valued Fully Convolutional Neural Network," proposed a complex-valued convolutional classification scheme. Although it improved the utilization rate of basic features by utilizing SAR complex raw data, the overall structure is still purely data-driven, without introducing the polarimetric scattering mechanism as a constraint. In areas where ground object scattering types overlap, it is prone to generating misclassification results that do not conform to the objective laws of electromagnetic scattering. Second, existing physical information fusion schemes are generally limited to a shallow combination of input layer feature splicing and shallow network weighted fusion, failing to construct a complete physical constraint from input features, intermediate deep representations to output decisions. The binding mechanism causes deep network features to gradually deviate from the true electromagnetic scattering law. In scenarios with scarce labeled samples, this easily leads to model overfitting and a significant degradation in generalization ability. The application number 202011511666.5, entitled "Polarization SAR Image Classification Method, Medium and Device Based on Scattering Graph Convolutional Network", proposes to use polarimetric scattering features as input pre-information of graph convolutional network. It utilizes polarimetric prior knowledge to optimize graph node construction to some extent. However, it only completes physical feature splicing and fusion at the network's input end and cannot apply continuous physical correction constraints to the deep iteration process of the network. Under small sample conditions, deep features will gradually deviate from the true scattering law, and the model's generalization performance will degrade significantly. Thirdly, current physical prior injection methods mostly adopt static embedding strategies with fixed weights, lacking dynamic adjustment mechanisms to adapt to scene noise changes. They cannot effectively suppress the interference of inherent speckle noise of PolSAR on physical prior information, causing physical knowledge to continuously decay and become ineffective during the propagation of deep network. It is difficult to form a stable and accurate physical guidance for deep feature learning throughout the process, which greatly limits the synergistic optimization effect of physical prior and data-driven model.

[0005] Therefore, how to implement physical prior constraints and spatial adaptive dynamic injection in deep networks to overcome the physical consistency bottleneck of pure data-driven models is a key technical problem for improving the accuracy and generalization ability of PolSAR image classification. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a PolSAR image classification method based on physical guidance and dynamic knowledge injection. This method explicitly embeds the polarimetric electromagnetic scattering mechanism into the entire learning process of a deep neural network, achieving synergistic optimization of physical prior knowledge and data-driven feature learning. It aims to solve problems in existing deep learning methods for PolSAR classification, such as insufficient utilization of physical mechanisms, a single method for fusing physical prior knowledge and deep features, and susceptibility to noise interference. This invention extracts scattering priors with clear physical meaning through polarimetric decomposition and constructs a physical supervision signal. It extracts multi-scale features through a dual-branch cross-attention network, establishes a physical regularization mechanism through physical scattering constraint branches, and finally achieves spatial adaptive fusion of physical priors and deep semantic features through a dynamic physical injection module based on center coordinate attention. This constructs a classification framework based on physical guidance and dynamic knowledge injection, improving classification accuracy while enhancing the physical consistency of features and the model's generalization ability under limited sample conditions.

[0007] To achieve the above objectives, the technical solution of the present invention includes the following:

[0008] (1) Input polarimetric synthetic aperture radar (PolSAR) image data and corresponding labels; the PolSAR image data includes two types of input components: main input data and physical prior features; the main input data is composed of a PolSAR multi-channel coherence matrix, which is used to characterize the polarimetric scattering characteristics of ground objects; the physical prior features are generated based on the polarimetric scattering decomposition model and include surface scattering, dihedral scattering, and volume scattering feature components, which are used to introduce prior information on the electromagnetic scattering mechanism of ground objects;

[0009] (2) A dual-branch cross-attention network is used as the backbone of feature extraction to build a dual-branch cross-attention feature extraction module; the module is used to extract global context features and local fine-grained structural features respectively, and a two-dimensional spatial depth feature map is constructed.

[0010] (3) Construct a dynamic physics injection module, which includes a physics mapping layer and a gating control unit, and introduces a center coordinate attention mechanism; used to adaptively embed the physical prior features in step (1) into the deep features extracted in step (2) to improve the physical consistency of the model output and achieve adaptive fusion;

[0011] (4) Construct a physical scattering constraint module, which applies explicit physical constraints to the network output based on the polarization scattering mechanism, and corrects the physical compatibility between depth features and classification results;

[0012] (5) Input the deep features that integrate physical prior features into the classification head and output the pixel-level land cover classification results of the PolSAR image to achieve accurate identification of land cover categories.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] First, because this invention uses a dual-branch cross-attention backbone network, the large-scale branch captures the global semantic structure and the small-scale branch extracts fine-grained scattering texture. The cross-attention mechanism enables bidirectional information interaction between the two branches, allowing features of different scales to complete complementary learning while preserving global contextual association and local boundary texture information. This effectively solves the problem of insufficient feature expression in single-scale networks and significantly improves the ability to distinguish multi-scale ground objects in complex scattering scenarios.

[0015] Secondly, because the present invention adopts a dynamic physical injection module based on center coordinate attention, it completes feature space recalibration through direction-aware encoding and Gaussian center bias weighting, and combines a gated residual mechanism to adaptively adjust the injection intensity of physical information, so that physical priors can be fused with deep features, effectively suppressing the interference of speckle noise on physical priors and significantly enhancing the physical consistency of deep features.

[0016] Third, because the present invention uses a physical scattering mechanism constraint module, automatically generates physical supervision labels based on polarization decomposition, and applies global physical regularization from the loss level through a multi-task learning framework, the model training is simultaneously guided by both the ground object category label and the electromagnetic scattering law, thereby effectively reducing the risk of model overfitting in scenarios with limited labeled samples and significantly improving the robustness and generalization ability of the classification model. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention;

[0018] Figure 2 This is a schematic diagram of the overall architecture for PolSAR image classification using the method of the present invention;

[0019] Figure 3 This is a schematic diagram of the dual-branch cross-attention network structure in this invention;

[0020] Figure 4 This is a schematic diagram of the physical feature injection module based on center coordinate attention in this invention;

[0021] Figure 5 This is a schematic diagram of the physical scattering mechanism constraint in this invention;

[0022] Figure 6 This is a comparison chart of the classification performance of the present invention and existing methods on the Flevoland dataset, provided in an embodiment of the present invention.

[0023] Figure 7This is a comparison chart of the classification performance of the present invention and existing methods in the nearshore scene of the San Francisco dataset, provided in the embodiments of the present invention.

[0024] Figure 8 This is a comparison chart of the classification performance of the present invention and existing methods in a densely packed Xi'an dataset scenario, as provided in the embodiments of the present invention.

[0025] Appendix Figure 6-8 In the middle: (a)-(g) correspond to existing CNN, CV-CNN, CV-FCN, CV-3D-CNN, ViT, PGIL and CV-AW-UNet methods respectively, and (h) corresponds to the method of the present invention; Detailed Implementation

[0026] The present invention will now be further described with reference to the accompanying drawings.

[0027] Example 1: Refer to Appendix Figure 1 The present invention proposes a PolSAR image classification method based on physical guidance and dynamic knowledge injection, the specific implementation steps of which include the following:

[0028] Step 1) Input polarimetric synthetic aperture radar (PolSAR) image data and corresponding labels; the PolSAR image data includes two types of input components: main input data and physical prior features; the main input data is composed of a PolSAR multi-channel coherence matrix, which is used to characterize the polarimetric scattering characteristics of ground objects; the physical prior features are generated based on the polarimetric scattering decomposition model and include surface scattering, dihedral scattering, and volume scattering feature components, which are used to introduce prior information on the electromagnetic scattering mechanism of ground objects;

[0029] Step 2) A dual-branch cross-attention network is used as the backbone for feature extraction to build a dual-branch cross-attention feature extraction module. This module is used to extract global context features and local fine-grained structural features respectively, and a two-dimensional spatial depth feature map is constructed. The specific implementation steps are as follows:

[0030] (2.1) Perform block embedding operation on the input PolSAR image: Set up large-scale branch and small-scale branch, and configure large-size block patch1 and small-size block patch2 respectively to extract depth features; and map all blocks to token sequence through convolutional embedding operation; In this embodiment, the large-scale branch is configured with large-size block patch1 to capture the overall distribution pattern of macroscopic ground objects; the small-scale branch is configured with small-size block patch2 with a size smaller than patch1 to retain the fine scattering features corresponding to ground object edges and point targets, and complete the collaborative representation of multi-scale polarization information.

[0031] (2.2) A multi-layer Transformer encoder is set up within each branch, and a cross-attention mechanism is introduced to complete the interaction of feature information between the two branches. In this embodiment, cross-scale feature interaction is used to achieve multi-scale feature complementary learning in this step to improve the integrity of feature representation and the ability to classify and distinguish land features. Specifically, a query item is generated by projecting the feature of any branch in the two branches. The feature projection of the other branch generates key terms. AND value term Then, bidirectional semantic complementarity is achieved through attention operations; the attention calculation formula is as follows:

[0032] ;

[0033] In the formula, Let T be the feature dimension of the key item, and T be the transpose operation.

[0034] (2.3) The token sequences output from the two branches are fed into the decoder consisting of transposed convolution, batch normalization and ReLU activation functions, respectively, to restore them to feature maps with the same spatial resolution as the input image. Then, the feature maps are spliced ​​in the channel dimension to obtain a two-dimensional spatial depth feature map that integrates multi-scale information.

[0035] Step 3) Construct a dynamic physics injection module, which includes a physics mapping layer and a gating control unit, and introduces a center coordinate attention mechanism; this is used to adaptively embed the physical prior features from step 1) into the deep features extracted in step 2) to improve the physical consistency of the model output and achieve adaptive fusion; in this embodiment, the above-mentioned dynamic physics injection module is implemented according to the following steps:

[0036] (3.1) The physical mapping layer is composed of two convolutional layers, LeakyReLU activation function and batch normalization layer stacked in sequence. This layer performs nonlinear transformation on the input physical prior features, completes the nonlinear dimensionality increase and feature expression enhancement of the physical prior features, and outputs a high-dimensional physical representation.

[0037] (3.2) The high-dimensional physical representation is mapped to the large-scale and small-scale feature spaces respectively. 1×1 convolution is used to complete the channel dimension matching to obtain the physical enhancement feature aligned with the corresponding scale depth feature dimension. The physical enhancement feature is initially fused with the corresponding scale depth feature to obtain the initially fused feature, thus completing the spatial alignment of the physical prior feature.

[0038] (3.3) Introducing a center coordinate attention mechanism to perform spatial recalibration on the initially fused features: global average pooling encoding is performed on the initially fused features along both the horizontal and vertical directions to construct a two-dimensional spatial interactive feature map; a center Gaussian distribution weight is introduced to enhance the feature response of the image center region, generating an adaptive spatial attention weight map. The specific implementation steps are as follows:

[0039] (3.3.1) Global average pooling is performed on the physical prior features after nonlinear mapping along the horizontal and vertical directions to extract the spatial encoding of direction awareness; after convolution, batch normalization and activation function to aggregate the direction information, the basic coordinate attention weights of row and column dimensions are generated.

[0040] (3.3.2) A Gaussian spatial weighting function is introduced to construct a center bias weight, which focuses on the stable scattering region at the center of the image patch and suppresses edge speckle noise interference. The formula for calculating the center bias weight is:

[0041] ;

[0042] In the formula, The center coordinates of the image patch For learnable focusing scale parameters; These represent the row index and column index of the pixels within the image block, respectively.

[0043] (3.3.3) Multiply the coordinate attention weights and Gaussian center bias weights element by element to obtain a spatial weight map with centripetal focusing characteristics, thus completing the spatial recalibration.

[0044] (3.4) A gate control unit composed of convolutional layers and a sigmoid activation function is used to adjust the injection weights of physical prior features, generating an adaptive injection weight matrix that simultaneously covers the channel dimension and the spatial dimension. This matrix is ​​used to selectively adjust the magnitude of the physical prior features, and then the adjusted physical prior features are integrated into the depth features of the corresponding scale by residual superposition, realizing the hierarchical adaptive fusion of physical prior features and depth features. The expression is as follows:

[0045] ,

[0046] In the formula, The main trunk depth feature The weighted physical characteristics, This represents element-wise multiplication. This represents the adaptive injection weight matrix.

[0047] Step 4) Construct a physical scattering constraint module, which applies explicit physical constraints to the network output based on the polarization scattering mechanism, correcting the physical compatibility between deep features and classification results; the specific implementation steps include:

[0048] (4.1) The main input data in step 1) is decomposed using the Freeman-Durden polarization decomposition model to extract three types of scattering components: surface scattering, dihedral scattering, and volume scattering. Pixel-level physical labels are then constructed to characterize the dominant ground object scattering mechanism corresponding to each pixel in the image. In this embodiment, the preferred method for constructing pixel-level physical labels is: based on the surface scattering power obtained from the Freeman-Durden decomposition... Dihedral scattering power Volume scattering power According to the maximum power criterion, a corresponding dominant scattering mechanism is matched for each pixel; the calculation formula for the pixel-level physical label is expressed as follows:

[0049] ;

[0050] In the formula, For pixel sample index.

[0051] (4.2) Construct a physical-aided prediction branch, using the physical prior features from step 1) as input to complete the scattering type classification prediction. A cross-entropy loss function is used to complete the supervised learning constraints. This constraint mechanism allows the deep network to simultaneously learn the distribution characteristics of sample data and the physical structure rules of electromagnetic scattering, ensuring that the network output is consistent with the objective laws of electromagnetic scattering. In this embodiment, the cross-entropy loss function is used to complete the supervised learning constraints. Specifically, the cross-entropy loss function is used to construct physical scattering constraints, and a weighted joint loss is obtained by fusing the ground feature classification loss. :

[0052] ;

[0053] In the formula, For the cross-entropy loss of the land cover classification task, Cross-entropy loss due to physical scattering constraints. These are the optimization strength parameters used to balance the constraints of land cover classification and physical consistency;

[0054] The cross-entropy loss of the physical scattering constraint is expressed as follows:

[0055] ;

[0056] In the formula, For the first The pixel corresponds to the first Activation terms of physical scattering mechanisms; The summation dummy variable represents the traversal of all scattering mechanism categories; This is an explicit indicator function; its value is 1 when the condition within the parentheses is true, and 0 otherwise.

[0057] Step 5) Input the deep features fused with physical prior features into the classification head and output the pixel-level land cover classification results of the PolSAR image to achieve accurate identification of land cover categories.

[0058] Example 2: The overall implementation steps of the PolSAR image classification method proposed in this example are the same as in Example 1. It achieves high-precision, physically consistent PolSAR land cover classification by deeply fusing polarimetric scattering physical priors with deep feature learning. Now, referring to… Figure 1-2 Further detailed description of the specific operational steps for classification using the method of the present invention:

[0059] Step 1. Input the PolSAR image data to be classified and the corresponding land cover label information. Perform normalization preprocessing on the input data and unify the spatial size to provide standardized input for subsequent feature extraction.

[0060] Step 2. Extract the polarization coherence matrix components and Freeman-Durden scattering decomposition components to construct data-driven fundamental features and physical prior features, respectively:

[0061] (2a) PolSAR data essentially describes the amplitude and phase response information of electromagnetic waves interacting with ground objects under different polarization modes. For any pixel, its scattering characteristics can be represented by the scattering matrix as follows:

[0062] ;

[0063] According to the reciprocity theorem of electromagnetic waves (i.e.) This scattering matrix typically contains three independent complex elements. When decomposing incoherent targets, the scattering matrix is ​​usually converted into a polarimetric coherence matrix. , is represented as:

[0064] ;

[0065] The polarization coherence matrix is ​​a complex conjugate symmetric matrix with real diagonal elements. It is converted into a 9-dimensional real vector as the input feature for the data-driven branch, in the form of:

[0066] ;

[0067] in, Indicates the real part, This represents the imaginary part. This 9-dimensional vector serves as the input to the data-driven dual-branch network, providing the model with fundamental polarization statistics.

[0068] (2b) To compensate for the lack of physical constraints in the pure data-driven model, the Freeman-Durden three-component decomposition model is introduced, which decomposes the polarization coherence matrix into three scattering mechanism components with clear physical meaning:

[0069] Surface scattering describes the characteristics of micro-rough surfaces, and the corresponding scattering model can be expressed as:

[0070] ;

[0071] Dihedral scattering describes the characteristics of mutually perpendicular dihedral structures, and the corresponding scattering model can be expressed as:

[0072] ;

[0073] Volume scattering describes the scattering characteristics of dipoles composed of completely randomly distributed elements such as vegetation canopy. It can be represented as:

[0074] ;

[0075] Finally, the polarization coherence matrix This is a weighted linear combination of the three physical mechanisms mentioned above:

[0076] ;

[0077] In the formula, , , These are the surface scattering, dihedral scattering, and volume scattering coefficients, respectively. The three types of scattering power obtained through decomposition will serve as prior physical knowledge for subsequent physics injection and physics constraint modules.

[0078] Step 3. Construct a dual-branch CrossViT backbone feature extraction network to extract global contextual features and local detail features in parallel, and achieve cross-branch information interaction through cross-attention:

[0079] (3a) Input the 9-dimensional polarization feature vector obtained in step 2 into the large-scale branch and the small-scale branch respectively, and flatten the two-dimensional feature map into a token sequence through block embedding operation; the large-scale branch uses a larger block size to capture the macro-distribution pattern of ground features, and the small-scale branch uses a smaller block size to retain the fine features of edges and point targets.

[0080] (3b) Introduce a cross-branch attention mechanism in each Transformer coding layer to achieve bidirectional information interaction between two branches. The attention calculation method is as follows:

[0081] ;

[0082] Unlike traditional feature concatenation or simple weighting, the cross-attention module uses query terms ( ), key items ( ) and value terms ( The projection transformation of ) achieves feature complementarity between one branch and another branch;

[0083] (3c) Spatial reconstruction is performed on the token sequence output from the previous step to complete feature decoding and fusion. The output features of the two branches are fed into the decoder consisting of transposed convolution, batch normalization and ReLU activation function, respectively, to map the abstract token features into a two-dimensional feature map with the same spatial resolution as the input image.

[0084] Step 4. In this step, this embodiment further constructs a dynamic physics injection module based on center coordinate attention to achieve deep fusion of physical priors and deep features:

[0085] (4a) First, nonlinear mapping is performed on the physical prior features obtained by scattering decomposition to improve the physical feature representation ability;

[0086] ;

[0087] (4b) Introduce the Central Coordinate Attention Mechanism (CCAM) to spatially weight the physical prior features, enhance the feature contribution of the effective scattering region and suppress speckle noise interference.

[0088] First, the mapped physical prior features are... Perform feature aggregation in the horizontal and vertical directions to extract direction-aware spatial codes, given spatial coordinates. , No. Horizontal coding of channels With vertical encoding They are represented as follows:

[0089] ,

[0090] ;

[0091] After concatenating the encoded features in the horizontal and vertical directions, the directional information is aggregated through convolution, batch normalization, and ReLU activation. The aggregated features are then split into two independent feature tensors, row and column, and basic coordinate attention weights are generated through convolution and Sigmoid activation. and .

[0092] To focus attention on the image's central region, which has more stable scattering characteristics, and to suppress edge speckle noise interference, this paper introduces a Gaussian space weighting function. Define the center coordinates of the patch as Regarding spatial location The formula for calculating the center bias weight is as follows:

[0093] ;

[0094] in, It is a learnable focusing scale parameter that allows the network to dynamically adjust the radius of attention to the central region based on the scattering complexity of different datasets.

[0095] The generated coordinate attention weights are multiplied element-wise with the Gaussian center bias weights and fused to obtain a spatial weight map with centripetal focusing properties. This weight can accurately locate key areas of physical scattering and weaken the influence of redundant background. The formula is as follows:

[0096] ;

[0097] Obtain the spatial weight map Then, this paper utilizes weighted graphs. For physical prior features Spatial weighting is performed to enhance the feature representation of the physically effective region, resulting in weighted physical features:

[0098] ;

[0099] (4c) A gated residual fusion strategy is adopted to achieve deep fusion of physical features and deep features: Specifically, firstly, an adaptive weight matrix consistent with the dimension of physical features is generated by using a gating mechanism, and the weighted physical features are dynamically adjusted by element-wise multiplication to achieve selective injection of physical information in the spatial and channel dimensions; then, the gated physical features are superimposed on the features of the backbone network in the form of residuals.

[0100] ;

[0101] Step 5. Introduce Physical Scattering Mechanism Constraint (PSMC) to simultaneously optimize classification accuracy and physical consistency within a multi-task learning framework, thereby improving the model's generalization ability.

[0102] (5a) Constructing a physical dominant label map: based on surface scattering power extracted from Freeman decomposition Dihedral scattering power Volume scattering power Construct pixel-level physical dominant label maps according to the maximum power criterion. The formula is:

[0103] ;

[0104] (5b) Input the physical features obtained from Freeman decomposition into the physical mapping layer (PML) for nonlinear transformation and channel dimension adaptation, so that the physical feature dimension matches the input requirements of subsequent auxiliary branches, thereby further enhancing the representation ability of physical features.

[0105] (5c) Construct a physical reconstruction auxiliary branch to implement supervised constraints on the physical scattering mechanism. Using the adapted physical features as input, a physical feature extractor is used to mine the spatial correlation and intrinsic patterns of the scattering features. The output dimension is then adjusted via a convolutional layer to obtain the scattering type prediction result. After normalizing the prediction result using Softmax, the cross-entropy loss between the prediction result and the physical dominant label map is calculated and used as the physical constraint loss term.

[0106] ;

[0107] In the formula, For the first The pixel corresponds to the first Activation terms of physical scattering mechanisms; The summation dummy variable represents the traversal of all scattering mechanism categories; This is an explicit indicator function; its value is 1 when the condition within the parentheses is true, and 0 otherwise.

[0108] (5d) Constructing the main classification branch: It consists of stacked convolutional layers, batch normalization layers, ReLU activation functions, and output convolutional layers. It maps the deep features fused with physical priors to pixel-by-pixel land cover category prediction results and calculates the cross-entropy loss for land cover classification. ;

[0109] (5e) Synchronous optimization is achieved through a joint loss function, which weights and fuses the main classification loss and the physical guidance loss:

[0110] ;

[0111] in, The physical-guided loss weights are used to balance the optimization intensity of land cover classification and physical constraints, so that the model can maintain the consistency between feature representation and electromagnetic scattering laws while learning discriminative features.

[0112] Step 6. Iteratively train and optimize the model based on the training sample set. Calculate the classification loss and physical constraint loss using forward propagation, and update the network parameters through backpropagation to obtain the optimal model parameters after convergence. After training, input the test sample set into the trained model for inference and prediction.

[0113] Step 7. Input the final fused features into the classification head, and after Softmax activation, obtain the pixel-by-pixel land cover category probability. Take the category corresponding to the maximum probability as the output result, and finally complete the land cover classification of the PolSAR image.

[0114] Example 3: The overall implementation steps of the polarimetric SAR image classification method proposed in this example are the same as in Example 1 or 2. Refer to the appendix below. Figure 2-5 The implementation process of the present invention is further described in detail with specific examples:

[0115] Reference Figure 2-5 This paper presents a detailed implementation process based on publicly available experimental datasets, further refining the operational parameters and engineering implementation logic of each module. Existing pure data-driven deep networks rely solely on sample fitting of ground features, completely ignoring the objective laws of polarimetric electromagnetic scattering. This leads to class confusion, boundary fragmentation, and physical logic distortion, especially in scenarios with only 1% labeled samples. Therefore, this embodiment fully implements three core innovative modules: dual-branch cross-scale feature extraction, CCAM dynamic physical injection, and PSMC physical scattering constraints. Using three standard PolSAR experimental datasets (Flevoland, San Francisco, and Xi'an) as examples, the classification process is fully reproduced. The specific steps are as follows:

[0116] Step S1: Input the standardized PolSAR dataset and the corresponding ground truth labels; use Z-score to normalize the 9-dimensional features of the polarization coherence matrix.

[0117] Step S2: Simultaneously extract the 9-dimensional basic features of the polarization coherence matrix and the Freeman-Durden three-component physical scattering prior features;

[0118] Step S201: Transform the scattering matrix of each input pixel into a complex conjugate polarization coherence matrix, extract the real and imaginary components and concatenate them into a 9-dimensional polarization feature vector, which is used as the unified input of the dual-branch cross-attention network;

[0119] Step S202: Perform Freeman three-component polarization decomposition on the coherence matrix and solve for surface scattering power, dihedral scattering power and volume scattering power pixel by pixel. The three types of scattering power maps are used as all physical prior inputs.

[0120] Step S203: Generate a physical dominant label map pixel by pixel according to the maximum power criterion, which serves as the truth label for the subsequent PSMC auxiliary supervision branch. No manual annotation is required, and the supervision signal is automatically generated entirely based on the polarization decomposition physical rules.

[0121] Step S3, refer to Figure 3 The diagram shows the dual-branch cross-attention network structure in this invention. In this embodiment, a dual-branch cross-attention cross-scale feature extraction backbone network is built in this step to extract global context and local detail features in parallel.

[0122] Step S301: Preferably, the large-scale branch patch block size is set to 15×15 and the small branch block size is set to 7×7. Linear projection is used to generate CLS classification tokens and image patch tokens, and position codes are embedded.

[0123] Step S302: Each Transformer encoder layer has a built-in bidirectional cross-attention module. The Q query vector of the large branch retrieves the fine-grained K / V features of the small branch, and the small branch synchronously aggregates the global context of the large branch.

[0124] Step S303: The output tokens of the two Transformers are fed into the decoder, and the spatial feature map is upsampled and restored by transposed convolution, BatchNorm, and ReLU stacking.

[0125] Step S4, refer to Figure 4 The present invention provides a schematic diagram of the physical feature injection module based on center coordinate attention; a dynamic physical injection module based on center coordinate attention is constructed, and Freeman physical priors are adaptively embedded into the backbone deep features:

[0126] Step S401: The three-component scattering power map is fed into the physical mapping layer PML. Two convolutional layers + LeakyReLU + BN complete nonlinear dimensionality upscaling and channel alignment, and output physical features that adapt to the channel dimension of the network.

[0127] Step S402: Perform horizontal and vertical bidirectional global average pooling on the physical features output from the previous step to generate directional encodings. After concatenation and fusion, split the encodings to generate basic coordinate attention weights. Introduce a Gaussian center weighting function to suppress speckle noise at the patch edges. The initial value is set to 2.0 and is adaptively updated with training iterations; the coordinate weights and Gaussian weights are multiplied element-wise to obtain the centripetal spatial weight map, and the weighted physical features are used to obtain the enhanced physical features;

[0128] Step S403: Adopt a gated residual fusion mechanism to generate adaptive gated weights from the weighted physical features, perform channel-space selective injection on the weighted physical features, and superimpose the residuals to obtain the fused physical prior features.

[0129] Step S5, refer to Figure 5 The present invention provides a flowchart illustrating the physical scattering mechanism constraint, and establishes a PSMC physical scattering constraint multi-task learning framework to jointly optimize the classification loss and physical consistency loss.

[0130] Step S501: Construct the physical-aided reconstruction branch: The physical features output by PML are processed through multiple convolutions to extract the spatial correlation features of scattering. The last convolution outputs a 3-channel scattering category prediction, which is then combined with the automatically generated physical dominant label to calculate the cross-entropy physical loss.

[0131] Step S502: The classification main branch stacks convolutional layers with fused features to output the corresponding dataset category channel prediction and calculates the standard land cover classification cross-entropy loss.

[0132] Step S503: Prefer the total loss weighting parameter The value is fixed at 0.7 to balance the classification task and physical constraints, and the total loss is used as the optimization objective for backpropagation of the model.

[0133] Step S6: Input the fused features output by the model into the classification network, generate a pixel-by-pixel category response map through the convolution classifier head, obtain the probability distribution after Softmax normalization, and generate the final PolSAR land cover classification result map according to the maximum probability principle, so as to realize the spatial expression and visualization output of the classification results.

[0134] The effects of the present invention will be further explained below with reference to relevant experiments.

[0135] 1. Simulation conditions:

[0136] The simulation experiments of this invention were conducted in a hardware environment with an Intel® Core™ i7-10700K processor, 32GB of memory, and an NVIDIA GeForce RTX 3080 (10GB of video memory) graphics card, and a software environment with Python 3.9 and PyTorch 1.12.0 deep learning framework.

[0137] 2. Simulation content:

[0138] The algorithm's performance was validated using three publicly available standard PolSAR datasets: Flevoland, San Francisco, and Xi'an. The Flevoland dataset was acquired by the NASA / JPL AIRSAR sensor in the L-band (1.25 GHz), with image sizes of 750 × 1024 pixels. The San Francisco dataset was also acquired by the NASA / JPL AIRSAR sensor in the L-band, with image sizes of 1300 × 1300 pixels. The Xi'an dataset was acquired by the RADARSAT-2 satellite, with image sizes of 512 × 512 pixels. One percent of randomly labeled pixels were used as training samples, and the remaining labeled pixels were used as test samples. The algorithm was iteratively trained using the AdamW optimizer. All comparative and ablation experiments were run multiple times with a fixed random seed of 42, and the final result was the average of multiple runs to ensure fairness in the comparison.

[0139] 3. Ablation experiment content and result analysis:

[0140] This invention designs a series of ablation experiments to demonstrate and analyze the effectiveness of each module in the proposed technical solution. All experiments are deployed on three public PolSAR datasets: Flevoland, San Francisco, and Xi'an. A pure two-branch cross-attention network without any physical guidance or injection mechanism is used as the baseline model. The evaluation metrics are the overall accuracy (OA), average class accuracy (AA), and Kappa coefficient, which are common in the PolSAR classification field.

[0141] 1a) Overall effect analysis

[0142] This invention primarily designs two core components to improve the classification accuracy of PolSAR images: Physical Scattering Mechanism Constraint (PSMC) and Dynamic Physical Knowledge Injection (DPI+CCAM) based on center coordinate attention. We verified the impact of different components on the network's classification accuracy through progressive module stacking. The overall accuracy results for the three datasets are summarized in Table 1. It can be seen that each module proposed in this invention brings stable performance gains. Compared with the baseline network, the complete method improves the OA by 8.78%, 5.41%, and 2.91% on the Flevoland, San Francisco, and Xi'an datasets, respectively, achieving significant results in complex agricultural and mixed urban scenarios.

[0143] Table 1. Impact of different components on OA (%)

[0144]

[0145] 1b) Physical Scattering Mechanism Constraint Module (PSMC)

[0146] The physical scattering mechanism constraint introduces global physical regularization into the network through a multi-task learning framework, and is a core module for improving the physical consistency of the model. To verify its independent role, we added only the PSMC module to the baseline network, and the experimental results are shown in Table 1.

[0147] When only the PSMC module is added, the OA of the three datasets is improved by 6.44%, 0.90%, and 1.30%, respectively, which verifies that physical regularization constraints can effectively regulate the physical rationality of features. In particular, the regularization effect is more significant in agricultural scenarios with a wide variety of land cover types, and can directly alleviate the overfitting problem of pure data-driven approaches.

[0148] 1c) Dynamic Physics Injection Module (DPI) based on Center Coordinate Attention Mechanism (CCAM)

[0149] Adding a basic DPI module to the baseline improved OA by 6.27%, 0.49%, and 0.79%, respectively, proving that physical injection in the feature layer can directly enhance the discriminativeness of features. Further introducing CCAM improved OA by 1.13%, 3.52%, and 1.15%, respectively, verifying that center coordinate attention can accurately enhance the physical guidance of the effective scattering region, suppress noise interference, and significantly improve the accuracy of physical injection.

[0150] When PSMC and DPI+CCAM work together, the model simultaneously injects physical constraints and prior physical knowledge. The final complete model improves OA by 8.78%, 5.41%, and 2.91% compared to the baseline, respectively. This performance gain is greater than the simple sum of the gains of individual modules, demonstrating the synergistic creativity of multi-module fusion.

[0151] 4. Comparison of experimental content and results analysis

[0152] To comprehensively verify the superior classification performance of the proposed method, seven representative PolSAR image classification methods were selected for quantitative and qualitative comparisons, covering CNN, CV-CNN, CV-FCN, CV-3D-CNN, ViT, PGIL, and CV-AW-UNet. All comparative experiments were conducted on the Flevoland, San Francisco, and Xi'an standard PolSAR datasets. All methods employed identical training strategies, a 1% labeled sample ratio, and experimental configurations to ensure fairness in the comparisons. The corresponding classification visualization results are attached. Figure 6-8 .

[0153] 2a) Comparison results of Flevoland dataset

[0154] The Flevoland dataset represents a complex agricultural scenario, containing 15 categories of crops and natural features. With its diverse feature types and high similarity in scattering mechanisms, it serves as a benchmark dataset for validating the fine-grained classification capabilities of various algorithms. Table 2 shows the comparison of classification accuracy among the different methods.

[0155] Table 2. Comparison of experimental results (%) using the Flevoland dataset

[0156]

[0157] As shown in Table 2, the method of this invention achieves the best overall accuracy (OA), average category accuracy (AA), and Kappa coefficient, with an OA of 97.92%. In terms of subcategories, this invention achieves the highest classification accuracy in multiple categories such as forest land, potatoes, rapeseed, and buildings, especially in easily confused crops like rapeseed, where the accuracy approaches 100%, significantly outperforming traditional pure data-driven methods. This indicates that the introduction of physical scattering priors effectively enhances the feature discrimination between similar scattering features and improves the fine-grained classification capability in complex agricultural scenarios.

[0158] 2b) Comparison results of the San Francisco dataset

[0159] The San Francisco dataset represents a typical heterogeneous urban scene, containing five types of land features: water bodies, vegetation, low-density urban areas, high-density urban areas, and built-up areas. The spatial distribution of these features is uneven, and mixed scattering is prominent. The dataset is primarily used to verify the robustness of the algorithm in urban scene classification. Table 3 shows the comparison of classification accuracy among the various methods.

[0160] Table 3. Comparison of experimental results using the San Francisco dataset (%)

[0161]

[0162] As shown in Table 3, the method of this invention significantly outperforms all comparative methods in all three comprehensive indicators, with an OA of 99.34% and a Kappa coefficient of 99.01%. This invention achieves the best accuracy in vegetation categories and maintains extremely high classification accuracy for various urban land features.

[0163] 2c) Xi'an Dataset

[0164] The Xi'an dataset is a mixed urban-rural scene containing three types of terrain features: water, grassland, and buildings. The terrain features have intersecting edges and a high degree of scattering and mixing, making it a typical dataset for verifying the algorithm's generalization ability in mixed scenes with a small sample size. The classification accuracy comparison results of each method are shown in Table 4.

[0165] Table 4. Comparison Experiment Results of Xi'an Dataset (%)

[0166]

[0167] As shown in Table 4, the method of this invention significantly outperforms all the comparative methods in all three comprehensive indicators, with an OA of 95.14% and a Kappa coefficient of 91.97%. The classification accuracy in grassland areas reaches 95.86%, with a particularly significant improvement in the classification of mixed grassland and building boundary areas. This verifies the advantages of the dynamic physics injection mechanism in highly mixed scattering scenarios. By injecting dynamic physical prior knowledge, it effectively improves the model's generalization ability and boundary discrimination accuracy with limited labeled samples.

[0168] In summary, this invention proposes a PolSAR image classification method based on a physics-guided and dynamic knowledge injection mechanism. This method aims to address the problems of traditional data-driven PolSAR classification methods, which neglect the physical mechanism of polarization scattering and lack sufficient long-range context modeling, resulting in weak feature physical consistency and poor generalization under limited labeled samples. To this end, this invention introduces Freeman-Durden polarization decomposition to extract scattering prior features with clear physical meaning, designs a dynamic knowledge injection module based on center coordinate attention, and eliminates the distribution mismatch between physical priors and depth features through spatial adaptive weighting and gated residual fusion, achieving deep and dynamic injection of physical knowledge. Simultaneously, considering the multi-scale nature of ground objects and the complex scattering mechanism in PolSAR images, a dual-branch cross-attention backbone network is constructed to collaboratively extract global context and local detail features. Furthermore, a multi-task learning framework constrained by the physical scattering mechanism is proposed, using scattering-dominated labels to provide explicit physical regularization, enhancing the model's generalization ability under sparse labeling. Experimental results show that the proposed method achieves superior classification performance in different scenarios, exhibiting stronger robustness and physical consistency.

[0169] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0170] The above simulation analysis proves the correctness and effectiveness of the method proposed in this invention.

[0171] The parts of this invention not described in detail are common knowledge to those skilled in the art.

[0172] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Obviously, those skilled in the art, after understanding the content and principle of the present invention, may make various modifications and changes in form and detail without departing from the principle and structure of the present invention. However, these modifications and changes based on the concept of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A PolSAR image classification method based on physical guidance and dynamic knowledge injection, characterized in that, Includes the following steps: (1) Input polarimetric synthetic aperture radar (PolSAR) image data and corresponding labels; the PolSAR image data includes two types of input components: main input data and physical prior features; The main input data consists of a PolSAR multi-channel coherence matrix, used to characterize the polarization scattering features of ground objects; the physical prior features are generated based on the polarization scattering decomposition model, including surface scattering, dihedral scattering, and volume scattering feature components, used to introduce prior information on the electromagnetic scattering mechanism of ground objects. (2) A dual-branch cross-attention network is used as the backbone of feature extraction to build a dual-branch cross-attention feature extraction module; This module is used to extract global context features and local fine-grained structural features, and to construct a two-dimensional spatial depth feature map. (3) Construct a dynamic physics injection module, which includes a physics mapping layer and a gating control unit, and introduces a center coordinate attention mechanism; This is used to adaptively embed the physical prior features in step (1) into the deep features extracted in step (2) to improve the physical consistency of the model output and achieve adaptive fusion. (4) Construct a physical scattering constraint module, which applies explicit physical constraints to the network output based on the polarization scattering mechanism, and corrects the physical compatibility between depth features and classification results; (5) Input the deep features that integrate physical prior features into the classification head and output the pixel-level land cover classification results of the PolSAR image to achieve accurate identification of land cover categories.

2. The method according to claim 1, characterized in that: The two-dimensional spatial depth feature map mentioned in step (2) is obtained specifically according to the following steps: (2.1) Perform block embedding operation on the input PolSAR image: Set up large-scale branch and small-scale branch, and configure large-size block patch1 and small-size block patch2 respectively to extract depth features; And all blocks are mapped to a token sequence through convolutional embedding operations; (2.2) Set up a multi-layer Transformer encoder within each branch and introduce a cross-attention mechanism to complete the interaction of feature information between the two branches; (2.3) The token sequences output from the two branches are fed into the decoder consisting of transposed convolution, batch normalization and ReLU activation functions, respectively, to restore them to feature maps with the same spatial resolution as the input image. Then, the feature maps are spliced ​​in the channel dimension to obtain a two-dimensional spatial depth feature map that integrates multi-scale information.

3. The method according to claim 2, characterized in that: In step (2.1), the large-scale branch is configured with a large-size patch1 to capture the overall distribution pattern of macroscopic ground features; the small-scale branch is configured with a small-size patch2, which is smaller than patch1, to preserve the fine scattering features of ground feature edges and point targets, and to complete the collaborative representation of multi-scale polarization information.

4. The method according to claim 2, characterized in that: The Cross-Attention mechanism described in step (2.2) utilizes cross-scale feature interaction to achieve multi-scale feature complementarity learning, thereby improving the completeness of feature representation and the ability to classify and distinguish ground features; specifically, it generates query items by projecting the features of either branch in the two-branch process. The feature projection of the other branch generates key terms. AND value term Then, bidirectional semantic complementarity is achieved through attention operations; the attention calculation formula is as follows: ； In the formula, Let T be the feature dimension of the key item, and T be the transpose operation.

5. The method according to claim 1, characterized in that: The dynamic physics injection module described in step (3) is implemented according to the following steps: (3.1) The physical mapping layer is composed of two convolutional layers, LeakyReLU activation function and batch normalization layer stacked in sequence. This layer performs nonlinear transformation on the input physical prior features, completes the nonlinear dimensionality increase and feature expression enhancement of the physical prior features, and outputs a high-dimensional physical representation. (3.2) The high-dimensional physical representation is mapped to the large-scale and small-scale feature spaces respectively. 1×1 convolution is used to complete the channel dimension matching to obtain the physical enhancement feature aligned with the corresponding scale depth feature dimension. The physical enhancement feature is initially fused with the corresponding scale depth feature to obtain the initially fused feature, thus completing the spatial alignment of the physical prior feature. (3.3) Introducing a central coordinate attention mechanism to perform spatial recalibration on the features after initial fusion: global average pooling encoding is performed on the features after initial fusion along the horizontal and vertical directions to construct a two-dimensional spatial interactive feature map; central Gaussian distribution weights are introduced to enhance the feature response of the image center region and generate an adaptive spatial attention weight map; (3.4) A gate control unit composed of convolutional layers and sigmoid activation function is used to adjust the injection weights of physical prior features and generate an adaptive injection weight matrix that simultaneously covers the channel dimension and spatial dimension. The matrix is ​​used to selectively adjust the amplitude of physical prior features, and then the adjusted physical prior features are integrated into the depth features of the corresponding scale by residual superposition, so as to realize the hierarchical adaptive fusion of physical prior features and depth features.

6. The method according to claim 5, characterized in that: Step (3.3) introduces a center coordinate attention mechanism to perform spatial recalibration on the features after preliminary fusion. The specific implementation steps are as follows: (3.3.1) Global average pooling is performed on the physical prior features after nonlinear mapping along the horizontal and vertical directions to extract the spatial encoding of direction awareness; after convolution, batch normalization and activation function to aggregate the direction information, the basic coordinate attention weights of row and column dimensions are generated. (3.3.2) A Gaussian spatial weighting function is introduced to construct a center bias weight, which focuses on the stable scattering region at the center of the image patch and suppresses edge speckle noise interference. The formula for calculating the center bias weight is: ； In the formula, The center coordinates of the image patch For learnable focusing scale parameters; These represent the row index and column index of the pixels within the image block, respectively. (3.3.3) Multiply the coordinate attention weights and Gaussian center bias weights element by element to obtain a spatial weight map with centripetal focusing characteristics, thus completing the spatial recalibration.

7. The method according to claim 6, characterized in that: The hierarchical adaptive fusion of physical prior features and deep features described in step (3.4) is achieved according to the following formula: ， In the formula, The main trunk depth feature The weighted physical characteristics, This represents element-wise multiplication. This represents the adaptive injection weight matrix.

8. The method according to claim 1, characterized in that: The physical scattering constraint module described in step (4) is implemented according to the following steps: (4.1) The main input data in step (1) is decomposed using the Freeman-Durden polarization decomposition model. The three types of scattering components, namely surface scattering, dihedral scattering and volume scattering, are extracted and pixel-level physical labels are constructed to characterize the dominant ground object scattering mechanism corresponding to each pixel of the image. (4.2) Construct a physical auxiliary prediction branch, use the physical prior features in step (1) as input to complete the scattering type classification prediction, and use the cross-entropy loss function to complete the supervised learning constraint. This constraint mechanism enables the deep network to learn the distribution characteristics of sample data and the physical structure rules of electromagnetic scattering in sync, so that the network output results are consistent with the objective laws of electromagnetic scattering.

9. The method according to claim 8, characterized in that: The method for constructing pixel-level physical tags in step (4.1) is based on the surface scattering power obtained from Freeman-Durden decomposition. Dihedral scattering power Volume scattering power According to the maximum power criterion, a corresponding dominant scattering mechanism is matched for each pixel; the calculation formula for the pixel-level physical label is expressed as follows: ； In the formula, For pixel sample index.

10. The method according to claim 9, characterized in that: Step (4.2) describes using the cross-entropy loss function to complete the supervised learning constraints. Specifically, it uses the cross-entropy loss function to construct physical scattering constraints and integrates the ground cover classification loss to obtain a weighted joint loss. : ； In the formula, For the cross-entropy loss of the land cover classification task, Cross-entropy loss due to physical scattering constraints. These are the optimization strength parameters used to balance the constraints of land cover classification and physical consistency; The cross-entropy loss of the physical scattering constraint is expressed as follows: ； In the formula, For the first The pixel corresponds to the first Activation terms of physical scattering mechanisms; The summation dummy variable represents the traversal of all scattering mechanism categories; This is an explicit indicator function; its value is 1 when the condition within the parentheses is true, and 0 otherwise.

Images