A polarimetric SAR image change detection method and system against noise interference

Abstract

The application discloses a kind of anti-noise interference polarimetric SAR image change detection method and system, the change detection method includes the following steps: obtaining polarimetric SAR image;The polarimetric SAR image obtained is input into the denoising and change detection model trained, and the denoising result and change detection prediction result are output;The training of the denoising and change detection model includes: using the covariance matrix component of polarimetric SAR image and polarimetric decomposition component constructs data set;Change detection network is constructed for denoising network, and the change detection network is transformer network with embedded spatial attention and channel attention;The data set is used to train the denoising network, and the denoising result of polarimetric SAR image at the same place at different times is obtained, the denoising result is input into change detection network for training, while the deep feature extracted by denoising network is converted into attention weight and input into change detection network to recalibrate recognition ability;The application has the advantages of accurate detection and good denoising effect.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, specifically relating to a method and system for detecting changes in polarimetric SAR images under strong noise interference. Background Technology

[0002] Change detection is the process of identifying differences in the state of objects or phenomena observed at different times. It is one of the major problems in Earth observation and has been extensively studied in recent decades. Synthetic Aperture Radar (SAR) is widely used in various fields because its imaging process is unaffected by sunlight, cloud cover, and atmospheric conditions. Compared to single-polarization SAR imagery, multi-polarization SAR imagery can provide richer scattering information to identify land use and land cover differences in a specific area over a period of time. This is crucial in various applications such as urban planning, environmental monitoring, agricultural surveys, disaster assessment, and map revision.

[0003] Existing SAR image change detection methods mainly fall into three categories. The first category is traditional methods based on clustering / classification. These methods typically use conventional clustering or classification algorithms to divide the data into changed and unchanged regions, but they suffer from high computational complexity and poor accuracy. The second category is deep learning methods based on Convolutional Neural Networks (CNNs). These methods use CNNs to automatically learn feature representations from the data, thereby capturing complex patterns and spatial dependencies between pixels. Compared to traditional methods, this approach has proven to have better performance, but it is limited by the receptive field. The third category is deep learning methods based on Transformers. These methods utilize self-attention mechanisms to weight different regions of the input image based on the relevance of the input image to the task, or to capture long-range dependencies between pixels and generate feature representations. Some researchers have attempted to combine CNN and Transformer structures to fully leverage the advantages of both. These methods have demonstrated superior performance compared to previous approaches, representing the latest advancements in remote sensing image change detection.

[0004] However, these existing methods still face some problems. First, SAR images inherently contain speckle noise due to their imaging mechanism, which greatly interferes with the accuracy of change detection. Second, current SAR image change detection methods do not adequately utilize the rich scattering information in SAR image data, leading to a decrease in detection accuracy. Therefore, it is essential to study how to fully utilize the scattering information in SAR images and develop change detection methods that are resistant to noise interference. Summary of the Invention

[0005] This invention provides a noise-resistant polarimetric SAR image change detection method and system. Addressing the problem of insufficient utilization of polarimetric scattering information in existing SAR image change detection methods, this invention replaces the common intensity image dataset with a dataset composed of the covariance matrix components and polarimetric decomposition components of polarimetric SAR images. Simultaneously, a transformer change denoising and detection model combining spatial attention, channel attention, and cross-attention is constructed and trained. To address the negative impact of inherent speckle noise in SAR images on change detection accuracy, a denoising network is used to perform preliminary denoising of the noisy image, resulting in more accurate change detection results. This approach offers advantages such as high detection accuracy and effective denoising.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a method for detecting changes in polarimetric SAR images with resistance to noise interference, comprising the following steps:

[0007] Acquire polarimetric SAR images;

[0008] The acquired polarimetric SAR imagery is input into the trained denoising and change detection model, which outputs denoising results and change detection prediction results; wherein, the training of the denoising and change detection model includes:

[0009] A dataset was constructed using the covariance matrix components and polarization decomposition components of polarimetric SAR images;

[0010] Construct a denoising network and a change detection network. The denoising network is a residual convolutional network with embedded channel attention and cross attention, and the change detection network is a transformer network with embedded spatial attention and channel attention.

[0011] The denoising network is trained using the dataset to obtain denoising results of polarimetric SAR images of the same location at different times. The denoising results are then input into a change detection network for training. Simultaneously, the deep features extracted by the denoising network are converted into attention weights and input into the change detection network to recalibrate its recognition capabilities.

[0012] In the above scheme, during the testing phase, the trained denoising and change detection models perform denoising and change region identification on polarimetric SAR images under strong noise conditions, obtaining corresponding denoising results and change detection prediction results. During model training, the model is trained using a dataset constructed from covariance matrix components and polarimetric decomposition components, which fully utilizes the polarimetric scattering information of the image. A constructed residual convolutional network is used for self-supervised denoising of polarimetric SAR, and the denoising results are used to train the change detection network. The denoising results and the deep features extracted by the denoising network are input into the change detection network, with denoising results from different times serving as training data for the change detection network. By combining this with the deep features extracted by the denoising network, image information is effectively extracted and utilized, improving the model's recognition accuracy.

[0013] Furthermore, the steps for constructing the dataset include:

[0014] For two polarimetric SAR images at different times in the same scene, calculate their covariance matrix components and polarimetric decomposition components respectively;

[0015] The dataset is constructed by concatenating the covariance matrix components and polarization decomposition components of each time-varying SAR image along the channel dimension.

[0016] Furthermore, the steps for constructing the denoising network include:

[0017] The residual convolutional network is constructed by extracting features hierarchically from multiple consecutive dense residual modules and connecting them together through residual connections.

[0018] Construct a channel attention module 1 to reweight the feature layers of the denoising network in the channel direction;

[0019] A cross-attention module is constructed to fuse feature maps of polarimetric SAR images before and after denoising.

[0020] By using a series of dense residual modules, the number of network layers is increased, and deeper features can be extracted. The channel attention module enhances the ability to extract polarization information, and the cross attention module utilizes the advantages of the features before and after denoising to complement each other, achieving a balance between smooth information and texture information.

[0021] Furthermore, the step of constructing the denoising network includes constructing the following denoising loss function and calculating the preliminary denoising results:

[0022] Denoising loss function:

[0023] (1)

[0024] in Represents the mean square error. represent Denoising results of time-varying polarimetric SAR data. The mask represents polarimetric SAR data for the same region at another time; it is a mask calculated based on the change detection results.

[0025] Mask-guided denoising networks place greater emphasis on areas with less variation.

[0026] Furthermore, the steps for constructing the change detection network include:

[0027] A feature extraction module is constructed to extract features from polarimetric SAR images of the same location at different times based on the ResNet network;

[0028] A spatial attention module and a channel attention module 2 are constructed to reweight the features of each extracted temporal polarimetric SAR image in terms of channel and spatial dimensions.

[0029] Furthermore, the step of constructing the change detection network also includes:

[0030] The change detection network transforms the polarimetric SAR image feature map at each time step after recalibration into a semantic token.

[0031] Global information enhancement is performed on semantic tokens at different times using a Transformer encoder structure;

[0032] The enhanced semantic token is reconstructed using a twin Transformer decoder structure to obtain the reconstructed feature map for the corresponding time.

[0033] Furthermore, the step of constructing the change detection network also includes:

[0034] Construct the change detection loss function as shown below, calculate the difference features, and output the change detection prediction results:

[0035] (2)

[0036] in Represents cross-entropy loss, This is the result predicted by the model at position (h, w). It is the label at position (h, w), where h and w are the position coordinates, and H and W are the upper limits of the position coordinates.

[0037] Furthermore, the network training is constrained by the following bi-branch joint loss function:

[0038] (3)

[0039] in and These are the loss functions for the change detection network and the denoising network, respectively. and These are the weights of the change detection loss and the denoising loss, respectively.

[0040] A noise-resistant polarimetric SAR image change detection system, comprising:

[0041] The data acquisition module is used to acquire polarimetric SAR images;

[0042] The change detection module is used to input the acquired polarimetric SAR imagery into the trained denoising and change detection model, and output the denoising result and the change detection prediction result; wherein, the training of the denoising and change detection model includes:

[0043] A dataset was constructed using the covariance matrix components and polarization decomposition components of polarimetric SAR images;

[0044] Construct a denoising network and a change detection network. The denoising network is a residual convolutional network with embedded channel attention and cross attention, and the change detection network is a transformer network with embedded spatial attention and channel attention.

[0045] The denoising network is trained using the dataset to obtain denoising results of polarimetric SAR images of the same location at different times. The denoising results are then input into a change detection network for training. Simultaneously, the deep features extracted by the denoising network are converted into attention weights and input into the change detection network to recalibrate its recognition capabilities.

[0046] A noise-resistant polarimetric SAR image change detection device includes a memory and a processor. The memory stores program instructions that are executed by the processor, and the processor calls the program instructions to perform the steps of the method described above.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] (1) By using the covariance matrix and polarization decomposition components as training data, we can make full use of the rich polarization scattering information in polarimetric SAR images, thereby more accurately identifying changing areas.

[0049] (2) By embedding spatial attention and channel attention in the transformer structure, a change detection network that can effectively extract and utilize polarization scattering information was constructed;

[0050] (3) To address the problem that high-intensity speckle noise in polarimetric SAR may prevent the change detection network from accurately distinguishing change areas, a denoising network system combining channel attention, cross attention, and residual convolutional networks was designed to denoise noisy images at different times. At the same time, the extracted deep denoising features were converted into attention weights to guide the change detection network to effectively identify the contours and boundaries of ground features. By combining the denoising network with auxiliary training, change detection with noise interference resistance was achieved. In addition, the dual-branch joint loss function was used for constrained training to achieve synergistic optimization between the change detection and denoising networks. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of a polarimetric SAR image change detection method with noise interference resistance provided in an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of the network framework of a polarimetric SAR image change detection method with noise interference resistance provided in an embodiment of the present invention;

[0053] Figure 3 This is a flowchart of the training and testing process for a polarimetric SAR image change detection method with noise interference resistance provided in an embodiment of the present invention.

[0054] Figure 4 This is the structure of the transformer encoder and decoder provided in the embodiments of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0056] like Figure 1 , Figure 3 As shown, a method for detecting changes in polarimetric SAR images with resistance to noise interference includes the following steps:

[0057] Acquire polarimetric SAR images;

[0058] The acquired polarimetric SAR imagery is input into the trained denoising and change detection model, which outputs denoising results and change detection prediction results; wherein, the training of the denoising and change detection model includes:

[0059] A dataset was constructed using the covariance matrix components and polarization decomposition components of polarimetric SAR images;

[0060] Construct a denoising network and a change detection network. The denoising network is a residual convolutional network with embedded channel attention and cross attention, and the change detection network is a transformer network with embedded spatial attention and channel attention.

[0061] The denoising network is trained using the dataset to obtain denoising results of polarimetric SAR images of the same location at different times. The denoising results are then input into a change detection network for training. Simultaneously, the deep features extracted by the denoising network are converted into attention weights and input into the change detection network to recalibrate its recognition capabilities.

[0062] In the above scheme, during the testing phase, the trained denoising and change detection models perform denoising and change region identification on polarimetric SAR images under strong noise conditions, obtaining corresponding denoising results and change detection prediction results. During model training, the model is trained using a dataset constructed from covariance matrix components and polarimetric decomposition components, which fully utilizes the polarimetric scattering information of the image. In the step of constructing the model denoising and change detection network framework, a residual convolutional network is constructed for self-supervised denoising of polarimetric SAR. The denoising results are then used to train the change detection network. The denoising results and the deep features extracted by the denoising network are input into the change detection network. Denoising results from different times serve as training data for the change detection network. By combining this with the deep features extracted by the denoising network, image information is effectively extracted and utilized, improving the model's recognition accuracy.

[0063] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0064] The images undergo preprocessing including radiometric correction, topographic correction, and multi-look processing to obtain corrected images. A dataset is constructed by calculating the covariance matrix components and polarization decomposition components of two polarimetric SAR images taken at different times within the same scene. Currently, multipolar SAR includes two modes: dual-polarization and full-polarization. Full-polarization SAR data can ultimately be represented by nine covariance matrix components and three polarization decomposition components, while dual-polarization SAR consists of four covariance matrix components and three polarization decomposition components. Depending on the polarization mode, the corresponding SAR image is selected to calculate the two parameters.

[0065] For two polarimetric SAR images of the same scene at different times ( , The scattering matrix is ​​typically used to characterize the information of each element in a SAR image.

[0066] A. Assuming it is a fully polarimetric SAR image, then the scattering matrix It can be represented as:

[0067] (4)

[0068] In the formula, express Launch and The received scattering element. Indicates a horizontally polarized wave. This indicates a vertically polarized wave.

[0069] In practical applications, a two-dimensional scattering matrix cannot effectively characterize all the properties of a target. Therefore, a second-order statistic of the scattering matrix is ​​introduced to analyze its electromagnetic scattering characteristics. Assuming the scattering matrix satisfies the reciprocity theorem, when... by When represented by a base, that is When, its covariance matrix It can be represented as:

[0070] (5)

[0071] in The symbol represents the set averaging operator, and * represents the complex conjugate operation. The modulus representing a complex signal.

[0072] By extracting the real and imaginary parts of the upper triangular matrix of the covariance matrix, the SAR covariance value matrix is ​​obtained. The components of the input data. The formula for the covariance matrix is ​​as follows:

[0073] (6)

[0074] The subscript indicates its position in the covariance matrix. Represents the real part value, Represents the imaginary part of the value.

[0075] To further analyze the data, Yamaguchi three-component (YAM3) polarization decomposition was used to extract scattering components representing various object features, which were then used as an additional component of the input data. YAM3 polarization decomposition has three components: , ,and . Corresponding to single scattering, representing surfaces, water bodies, and flat terrain. This corresponds to double scattering, and is commonly seen at the dihedral corners of buildings. Corresponding volume scattering primarily represents vegetation in the scene.

[0076] The constructed fully polarimetric SAR dataset contains two time periods, with each time period consisting of 9 covariance matrix components and 3 polarimetric decomposition components.

[0077] B. Assuming it is a dual-polarization SAR image, the scattering matrix should be adjusted accordingly based on its polarization. For example, for VV and VH polarization, the scattering matrix... Adjusted to:

[0078] (7)

[0079] Its covariance matrix The corresponding adjustments are as follows:

[0080] (8)

[0081] The covariance matrix is:

[0082] (9)

[0083] To further analyze the data, the Raney polarization decomposition method was used to extract scattering components representing various object characteristics, which were then used as an additional component of the input data. Raney polarization decomposition has three components: , ,and . Corresponding to single scattering, representing surfaces, water bodies, and flat terrain. This corresponds to double scattering, and is commonly seen at the dihedral corners of buildings. Corresponding volume scattering primarily represents vegetation in the scene.

[0084] The constructed dual-polarization SAR dataset contains two time periods, with each time period consisting of four covariance matrix components and three polarization decomposition components.

[0085] The dataset is constructed by concatenating the covariance matrix components and polarization decomposition components of the polarimetric SAR images at each time point along the channel dimension.

[0086] A denoising network and a change detection network are constructed, along with a residual convolutional network with embedded channel attention and cross-attention for self-supervised denoising of polarimetric SAR. For example... Figure 2 As shown in the denoising network, a self-supervised denoising process is performed using one time-based polarimetric SAR image of the same location as the noisy image and another time-based polarimetric SAR image as the reference image.

[0087] A residual convolutional network framework is constructed. The residual convolutional network first performs initial feature extraction using two consecutive 3×3 convolutional layers. Next, three consecutive dense residual modules extract residual features hierarchically, and these hierarchical features are connected together via residual connections to ensure deeper feature extraction. Then, a 1×1 convolution adjusts the number of channels to match the initial feature extraction. A 3×3 convolution is then used to obtain deeper residual features. These deeper features are added to the initial features to obtain enhanced features. Finally, a 3×3 convolutional layer outputs the final result.

[0088] Construct Channel Attention Module 1. Channel attention is applied after each dense residual module in the residual convolutional network framework to recalibrate the extracted polarimetric SAR features along the polarimetric channel dimension, thereby effectively preserving polarimetric information. This module can be defined as:

[0089] (10)

[0090] in, This indicates a feature layer that has undergone one weighting step through the channel attention module. This indicates that the channel attention input feature layer is used. This indicates the average pooling operation. and These represent the ReLU activation function and the Sigmoid activation function, respectively. , and and The corresponding terms represent the weight and bias terms. Element-wise multiplication. Channel attention is used. Calculate the channel attention feature map and multiply it with the input feature to recalibrate it along the channel dimension.

[0091] A cross-attention module is constructed. This module is used to fuse feature maps of polarimetric SAR images before and after denoising. Typically, the features before denoising contain more spatial information, while the features after denoising are smoother. The cross-attention module uses the rich spatial information from the pre-denoising image to guide the generation of the post-denoising features, and uses the polarimetric information from the post-denoising image to guide the generation of the pre-denoising features, thus achieving a balance between smooth features and textured features. This module can be defined as follows:

[0092] (11)

[0093] in and These represent the features before and after denoising, respectively. and The feature maps before and after denoising were obtained respectively, and then... , The weights of the features before and after denoising are obtained respectively.

[0094] The cross-attention module uses the weights of the denoised features to recalibrate the feature map before denoising, thereby enhancing the polarization information of the features before denoising. By recalibrating the feature maps after denoising using the feature weights from before denoising, the spatial information of the denoised features is enhanced. The two enhanced features were obtained by fusion. .

[0095] Construct the denoising loss function as shown below and calculate the preliminary denoising results.

[0096] Denoising loss function:

[0097] (12)

[0098] in Represents the mean square error. represent Denoising results of time-varying polarimetric SAR data. The polarimetric SAR data represents the same region at another time. The mask is a mask calculated based on the change detection results and is used to guide the denoising network to pay more attention to areas with less change.

[0099] The denoising network is trained using the dataset to obtain denoising results for polarimetric SAR images of the same location at different times. A transformer network with embedded spatial attention and channel attention is then constructed for change detection. Figure 2 The change detection network shown in the example is characterized by using two denoised images of the same location at different times as input data, with the change detection label between the two images as a reference, to train the change detection network. This change detection network consists of a feature extraction module, channel and spatial attention modules, a transformer encoder, and two transformer decoders.

[0100] A feature extraction module is constructed. The feature extraction module is an improved ResNet18 network. The ResNet18 network has 5 stages. The stride of the last 3 stages is replaced with 1, and the feature dimension is reduced by pointwise convolution (32 output channels). Features are extracted from polarimetric SAR images of the same location at different times.

[0101] Construct two spatial and channel attention modules. The spatial attention module is used to recalibrate the extracted polarimetric SAR features along the polarimetric channel dimension, thereby effectively preserving polarimetric information. This module can be defined as:

[0102] (13)

[0103] in, This indicates the feature layer that has been reweighted by the spatial attention module. This represents the spatial attention input feature layer.

[0104] Spatial attention through Calculate the spatial attention weight map and multiply it with the input features to perform spatial dimension recalibration.

[0105] The channel attention module 2 in the change detection network has the same structure and function as the channel attention module 1 in the denoising network.

[0106] The deep features extracted by the denoising network are transformed into attention weights and input into the change detection network through 1×1 convolution and sigmoid activation function. This guides the recalibration of features after spatial and channel attention, thereby enhancing the change detection network's ability to resist noise.

[0107] The feature maps at each time step after multiple calibrations are transformed into corresponding semantic tokens and concatenated using 3×3 convolution and softmax functions.

[0108] The concatenated semantic tokens are augmented with global information using a Transformer encoder structure. The encoder structure is as follows: Figure 4 As shown, it consists of a multi-head self-attention system and a multilayer perceptron. The multi-head self-attention system comprises multiple attention heads, each of which computes a query (Q), key (K), and value (V) based on an intermediate token. Q and K are then used to calculate an attention score, which is subsequently used to weight V. The multi-head self-attention system concatenates the outputs of all attention heads and performs a linear transformation. This process can be represented by the following formula:

[0109]

[0110]

[0111] (14)

[0112] in Represents the bulls' self-attention. This represents the concatenation of token sets from the previous layer's input. It is a transpose operation. It is a calculation structure The process of focusing attention; , , and It is a linear projection matrix; Indicates the number of attention heads. The parameters are The channel dimension.

[0113] Then, a multilayer perceptron is used to process the input sequence to generate a new token. The MLP mainly performs nonlinear transformations on the input sequence to enhance its expressive power.

[0114] A twin Transformer decoder structure is constructed to reconstruct the enhanced semantic token. The decoder structure is as follows: Figure 4 As shown, it consists of multi-head cross-attention and a multilayer perceptron. Multi-head cross-attention also consists of multiple attention heads, each of which combines the new token with the input features at the corresponding time to compute the Q, K, and V matrices. Multi-head cross-attention also concatenates the outputs of all attention heads and performs a linear transformation.

[0115]

[0116] (15)

[0117] in It is multi-headed cross-attention. and , These are the features after decoding and reconstruction at two different times, and the enhanced token.

[0118] Then, a multilayer perceptron is used to reconstruct new data containing global context information for the corresponding time period.

[0119] Construct the change detection loss function as shown below, calculate the difference features, and output the change detection prediction results.

[0120] Change detection loss function:

[0121] (16)

[0122] in Represents cross-entropy loss, This is the result predicted by the model at position (h,w). It is the label at position (h,w), where H and W are constants, representing the upper limit of the position coordinates respectively.

[0123] like Figure 3 As shown. During the training phase, the constructed denoising network is first trained using the constructed training dataset to obtain denoising results of polarimetric SAR images of the same location at different times. Then, the denoising results of polarimetric SAR images of the same location at different times are input into the constructed change detection network for training. At the same time, the deep features extracted by the denoising network are converted into attention weights to guide the change detection network to effectively identify the contours and boundaries of ground features.

[0124] During the training phase, the constructed denoising network is first trained using the constructed training dataset. Here, we will use... , The time-varying polarimetric SAR imagery is noisy. , Denoising is performed on the temporal polarimetric SAR image as a reference image to obtain... , Denoising results of time-polarized SAR images , Then , The input is used to train the constructed change detection network, while the deep features extracted by the denoising network are transformed into attention weights to guide the change detection network in effectively recognizing ground feature outlines and boundaries. The labels used for change detection at this point are... , polarization covariance matrix Input the change detection results generated by the PolSARpro software developed by the European Space Agency.

[0125] When training the denoising network, we will use the following methods respectively: , The time-varying polarimetric SAR imagery is noisy. , Denoising is performed on the temporal polarimetric SAR image as a reference image to obtain... , Denoising results of time-polarized SAR images , In the first epoch of training, the mask for calculating the denoising loss is 1. Starting from the second epoch, the change detection result of the previous epoch is used as the mask for calculating the denoising loss.

[0126] Then , The input is used to train the change detection network, while the deep features extracted by the denoising network are transformed into attention weights. The labels used for change detection at this point are... , polarization covariance matrix Input change detection results generated by the PolSARpro software developed by the European Space Agency. The number of layers in the Transformer encoder is set to 1, and the number of layers in the Transformer decoder is set to 8. The number of heads h for MSA and MA is set to 8, and the channel dimension of each head is set to 8.

[0127] The network training is constrained by the following bi-branch joint loss function.

[0128] Total loss function:

[0129] (17)

[0130] in and These are the loss functions for the change detection network and the denoising network, respectively. and These are the weights of the change detection loss function and the denoising loss function, respectively.

[0131] , .

[0132] During the testing phase, the trained denoising and change detection network was used to denoise and identify change regions in the polarimetric SAR image, and the corresponding denoising results and change detection prediction results were obtained.

[0133] This application also provides a polarimetric SAR image change detection system resistant to noise interference, including:

[0134] The first main module is used to construct a dataset using the covariance matrix components and polarization decomposition components of polarimetric SAR images;

[0135] The second main module is used to train the dataset to obtain the denoising results of polarimetric SAR images at different times at the same location, and to extract the deep features of the polarimetric SAR images.

[0136] The third main module is used for training based on the denoising results and calibrating the recognition capability based on the attention weights transformed from the deep features.

[0137] This application also provides a polarimetric SAR image change detection device with noise interference resistance, including a memory and a processor. The memory stores program instructions that are executed by the processor, and the processor calls the program instructions to perform the steps of the above-described method.

[0138] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting changes in polarimetric SAR images with resistance to noise interference, characterized in that, Includes the following steps: Acquire polarimetric SAR images; The acquired polarimetric SAR imagery is input into the trained denoising and change detection model, which outputs denoising results and change detection prediction results; wherein, the training of the denoising and change detection model includes: A dataset is constructed using the covariance matrix components and polarization decomposition components of polarimetric SAR images. The steps for constructing the dataset include: calculating the covariance matrix components and polarization decomposition components of two polarimetric SAR images at different times in the same scene; and stitching the covariance matrix components and polarization decomposition components of the SAR images at each time along the channel dimension to construct the dataset. Construct a denoising network and a change detection network. The denoising network is a residual convolutional network with embedded channel attention and cross attention, and the change detection network is a transformer network with embedded spatial attention and channel attention. The steps for constructing the denoising network include: constructing the residual convolutional network, wherein the residual convolutional network is composed of multiple consecutive dense residual modules that extract features hierarchically and are connected together through residual connections; constructing a channel attention module 1, which is used to reweight the feature layers of the denoising network in the channel direction; and constructing a cross attention module, which is used to fuse the feature maps of polarimetric SAR images before and after denoising. The steps for constructing the change detection network include: constructing a feature extraction module to extract features from polarimetric SAR images of the same location at different times based on a ResNet network; constructing a spatial attention module and a channel attention module II to reweight the features of each extracted polarimetric SAR image at each time in terms of channel and spatial dimensions; converting the recalibrated polarimetric SAR image feature map at each time into a semantic token through the change detection network; performing global information enhancement on the semantic tokens at different times through a Transformer encoder structure; and reconstructing the enhanced semantic tokens through a Siamese Transformer decoder structure to obtain the reconstructed feature map for the corresponding time. The denoising network is trained using the dataset to obtain denoising results of polarimetric SAR images of the same location at different times. The denoising results are then input into a change detection network for training. Simultaneously, the deep features extracted by the denoising network are converted into attention weights and input into the change detection network to recalibrate its recognition capabilities.

2. The method for detecting changes in polarimetric SAR images with noise resistance according to claim 1, characterized in that: The steps for constructing the denoising network include constructing the following denoising loss function and calculating the preliminary denoising results: in Represents the mean square error. represent Denoising results of time-varying polarimetric SAR data. The mask represents polarimetric SAR data for the same region at another time; it is a mask calculated based on the change detection results.

3. The method for detecting changes in polarimetric SAR images with resistance to noise interference according to claim 1, characterized in that: The steps of constructing the change detection network further include: Construct the change detection loss function as shown below, calculate the difference features, and output the change detection prediction results: in Represents cross-entropy loss, This is the result predicted by the model at position (h, w). It is the label at position (h, w), where h and w represent the position of the pixel in the feature map, and H and W represent the size of the feature map.

4. The method for detecting changes in polarimetric SAR images with resistance to noise interference according to claim 1, characterized in that... The network training is constrained by the following bi-branch joint loss function: in and These are the loss functions for the change detection network and the denoising network, respectively. and These are the weights of the change detection loss and the denoising loss, respectively.

5. A polarimetric SAR image change detection system resistant to noise interference, characterized in that, include: The data acquisition module is used to acquire polarimetric SAR images; The change detection module is used to input the acquired polarimetric SAR imagery into the trained denoising and change detection model, and output the denoising result and the change detection prediction result; wherein, the training of the denoising and change detection model includes: A dataset is constructed using the covariance matrix components and polarization decomposition components of polarimetric SAR images. The steps for constructing the dataset include: calculating the covariance matrix components and polarization decomposition components of two polarimetric SAR images at different times in the same scene; and stitching the covariance matrix components and polarization decomposition components of the SAR images at each time along the channel dimension to construct the dataset. Construct a denoising network and a change detection network. The denoising network is a residual convolutional network with embedded channel attention and cross attention, and the change detection network is a transformer network with embedded spatial attention and channel attention. The steps for constructing the denoising network include: constructing the residual convolutional network, wherein the residual convolutional network is composed of multiple consecutive dense residual modules that extract features hierarchically and are connected together through residual connections; constructing a channel attention module 1, which is used to reweight the feature layers of the denoising network in the channel direction; and constructing a cross attention module, which is used to fuse the feature maps of polarimetric SAR images before and after denoising. The steps for constructing the change detection network include: constructing a feature extraction module to extract features from polarimetric SAR images of the same location at different times based on a ResNet network; constructing a spatial attention module and a channel attention module II to reweight the features of each extracted polarimetric SAR image at each time in terms of channel and spatial dimensions; converting the recalibrated polarimetric SAR image feature map at each time into a semantic token through the change detection network; performing global information enhancement on the semantic tokens at different times through a Transformer encoder structure; and reconstructing the enhanced semantic tokens through a Siamese Transformer decoder structure to obtain the reconstructed feature map for the corresponding time. The denoising network is trained using the dataset to obtain denoising results of polarimetric SAR images of the same location at different times. The denoising results are then input into a change detection network for training. Simultaneously, the deep features extracted by the denoising network are converted into attention weights and input into the change detection network to recalibrate its recognition capabilities.

6. A polarimetric SAR image change detection device resistant to noise interference, characterized in that, The method includes a memory and a processor, the memory storing program instructions that are executed by the processor, the processor invoking the program instructions to perform the steps of the method according to any one of claims 1 to 4.

Images