Unsupervised SAR change detection method based on collaborative change consistency network model

Abstract

The invention belongs to the technical field of change detection, and particularly relates to an unsupervised SAR change detection method based on a collaborative change consistency network model. The method comprises the following steps: firstly, pre-classifying a dual-time-phase SAR image to obtain a pseudo label, and randomly sampling according to the pseudo label to obtain a supervision sample set and a self-supervision sample set; then, constructing a collaborative change consistency network model: extracting a deep feature vector from a sample through a cascaded residual block by an encoder, inputting a feature vector corresponding to a supervised sample set into a supervised learning branch, and performing supervised learning in combination with a pseudo tag; inputting a feature vector corresponding to the self-supervised sample set into a self-supervised learning branch, and carrying out self-supervised learning by minimizing a feature distance; establishing a change consistency loss function between the two branches to realize double-branch cooperative training and enhance the characterization capability of the model to change features; and finally, inputting a to-be-detected SAR image into the trained model so as to realize change detection which is not easily influenced by noise.

Description

Technical Field

[0001] The present invention belongs to the technical field of change detection, and in particular relates to an unsupervised SAR change detection method based on a collaborative change consistency network model. Background Art

[0002] Change detection aims to identify changes in consistent spatial locations at different time instances, which can be regarded as a binary classification task, classifying each location as changing or unchanged. SAR imaging is not affected by atmospheric conditions and can operate in all weather conditions, making SAR change detection widely used in applications such as environmental monitoring and land use detection.

[0003] Change detection in Synthetic Aperture Radar (SAR) images is an important technique for identifying changes in the same area over time. Compared with optical images, SAR imaging is not affected by sunlight, clouds, or atmospheric conditions and can work around the clock. Therefore, SAR has been widely used in change detection research for environmental monitoring, disaster assessment, and land use detection. However, the unique imaging principle of SAR sensors can produce significant speckle noise in SAR images, which may cause high false alarm rates in change detection tasks, thus posing challenges to change detection. Therefore, it is crucial to develop robust change detection techniques that can effectively suppress speckle noise.

[0004] SAR change detection techniques can be roughly divided into two categories: supervised methods and unsupervised methods. Supervised methods require manual labeling of ground truth as supervision information. However, manual labeling is costly and cannot meet the needs of many practical applications, such as landslides, floods, and earthquakes. In contrast, unsupervised methods can perform change detection without the need for manual labels or prior knowledge, making it the mainstream method in recent research. Traditional unsupervised SAR change detection methods mainly involve generating difference images (DI) using various difference operators, and then performing DI classification through thresholding or clustering algorithms.

[0005] Although traditional methods are computationally efficient and widely used, they are limited by their feature representation capabilities and are easily affected by the inherent speckle noise in SAR. Due to the powerful feature representation capabilities of deep learning, some researchers have successfully applied deep learning to unsupervised change detection, which mainly involves two steps: 1) first generate pseudo labels using clustering methods; 2) the generated pseudo labels are regarded as ground truth values, and a specific proportion of samples are sampled as training sets to achieve the training process. However, the current mainstream methods often cannot fully mine the feature information of the samples, are still easily affected by noisy images, and have difficulty in accurately obtaining change information.

[0006] Through the above analysis, the problems and defects of the existing technology are as follows: the current mainstream unsupervised SAR change detection method is easily affected by noise and the change detection accuracy is affected. Summary of the invention

[0007] In view of the problems existing in the prior art, the present invention provides an unsupervised SAR change detection method based on a collaborative change consistency network model.

[0008] An unsupervised SAR change detection method based on a collaborative change consistency network model comprises the following steps:

[0009] Step 1: Pre-classify the dual-phase SAR images to obtain pseudo labels.

[0010] Step 2: Sample the pseudo labels obtained in step 1 to obtain a supervised sample set, and then sample from the supervised sample set to obtain a self-supervised sample set.

[0011] Step 3: Construct the co-variation consistency network: This includes the construction of the encoder, the construction of the supervision branch, and the construction of the self-supervision branch. The specific construction process is as follows:

[0012] First, an encoder with two shared weights is constructed to extract features from samples of two time phases respectively, and obtain feature vectors corresponding to the two time phases.

[0013] The dual-phase samples in the supervised sample set and the self-supervised sample set are respectively input into two encoders for feature extraction. Then, the supervised learning branch and the self-supervised learning branch are respectively constructed, and the feature vectors extracted by the two encoders are respectively input into their corresponding learning branches.

[0014] Step 4: Use the two sample sets to train the co-variation consistency network respectively, and perform gradient back propagation through the overall co-variation consistency function L of the co-variation consistency network to complete the training of the co-variation consistency network;

[0015] Using the supervised sample set, after the encoder extracts features, the supervised learning branch is supervised by pseudo labels, and the parameters are optimized by the supervised loss function;

[0016] The supervised loss function is defined as the binary cross entropy loss, which is expressed as:

[0017]

[0018] Among them, y i represents the pseudo label corresponding to the i-th supervised sample, y i ′ represents the prediction result corresponding to the i-th supervised sample, N sl Represents the total number of samples in the supervised sample set.

[0019] Using the self-supervised sample set, after the encoder extracts features, self-supervised learning is performed without labels, and the parameters are optimized through the self-supervised loss function;

[0020] The self-supervised loss function is constructed through cosine similarity and is expressed as:

[0021]

[0022] Among them, stopgrad means stopping gradient propagation, N ssl Represents the total number of all samples in the self-supervised sample set, P i,t is the high-dimensional feature vector prediction, Z i,t is the feature vector mapping, t represents the phase and takes 1 or 2.

[0023] In addition, by limiting the consistency of the change features between the supervised learning branch and the self-supervised learning branch, the similarity between the probability distributions of the change features is measured by KL divergence, and a change consistency loss function is constructed. It can be expressed as:

[0024]

[0025] F i,d is the change characteristic, D i,1 and D i,2 is a pair of mutually symmetric feature vector distances; combining all loss functions, through the weight parameter λ ssl and λ cons By limiting the proportion of each loss, the overall collaborative loss function L of the collaborative change consistency network is obtained as follows:

[0026] L=L sl +λ ssl L ssl +λ cons L cons

[0027] Parameter optimization of the covariance consistency network is achieved through gradient backpropagation through L.

[0028] Step 5: The two SAR images to be detected are predicted by using the collaborative change consistency network trained in step 4 to obtain the final change detection result.

[0029] Furthermore, the step 1 is specifically as follows: the dual-phase SAR images are defined as T1 and T2 respectively, and a difference map DI is first generated by a logarithmic ratio operator, and the calculation process is as follows: DI=abs(log((T1+1) / (T2+1))); then DI is classified by a fuzzy C-means clustering algorithm to obtain a clustering result, and the clustering result is used as a pseudo label.

[0030] Furthermore, the step 2 specifically includes: sampling the pseudo labels obtained in step 1 to obtain a supervised sample set, and then sampling from the supervised sample set to obtain a self-supervised sample set.

[0031] Before sampling, the SAR image is expanded. For T1, T2∈R H×W Mirror it in four directions: up, down, left, and right to fill three rows / columns with values ​​T1', T2'∈R (H+6)×(W+6) , H represents the number of rows of the SAR image, and W represents the number of columns of the SAR image; then a 7×7 sliding window is used to take the neighborhood of the two SAR images of different phases with the pixel as the center as the patch. Then 15% of the patches are randomly sampled to form the supervised sample set Ω sl .

[0032] Then, according to the pseudo-label, determine whether the corresponding supervised sample is a changed sample, and divide all supervised samples into two categories {Ω0,Ω1}, where Ω1 consists of all samples whose pseudo-labels are changed, and Ω0 consists of all samples whose pseudo-labels are unchanged, and Ω0 is used as the self-supervised sample set.

[0033] Furthermore, the encoder includes an input layer, a residual layer and an output layer.

[0034] The input layer includes a normal convolution layer with a convolution kernel size of 3×3 and a Relu activation function, which convolves the input patch to obtain a preliminary shallow feature representation.

[0035] The residual layer is composed of four stacked residual blocks to represent deep features;

[0036] Each residual block first calculates the input feature map F through a normal convolution layer with a convolution kernel size of 3×3, a Relu activation function, and a BatchNorm layer to obtain F1, and then calculates F2 through a normal convolution layer with a convolution kernel size of 3×3 and a BatchNorm layer. Then, the input feature is downsampled through a normal convolution layer with a convolution kernel size of 1×1, and then summed with the input feature map F through a jump connection; finally, the output feature map is obtained through Relu activation. The calculation process is as follows:

[0037] F o =Relu(F2+Conv(F))

[0038] After passing through four residual blocks in sequence, the deep feature F' is obtained, and finally converted into a feature vector through the output layer. The output layer includes global average pooling operation (Avgpool) and flattening operation in sequence, so as to reduce the dimension of the extracted deep features and convert them into vectors.

[0039] Furthermore, the supervised learning branch includes change feature extraction and change feature classification, and the bi-phase feature vector F extracted by the encoder is i,1 and F i,2 , after vector concatenation, it is input into the multi-layer perceptron to obtain the fused change feature F i,d . i,d The input is sent to the classifier for dimensionality reduction to obtain a vector of length 2, and finally the prediction result is obtained by activating the Softmax function.

[0040] Furthermore, the specific formula for extracting the change feature can be expressed as:

[0041] F i,d =MLP(concat(F i,1 ,F i,2 ))

[0042] Furthermore, the self-supervised learning branch includes two multi-layer perceptrons with shared weights, respectively called the projection head and the prediction head. The feature vector output by the projection head is Z i,t , the feature vector output by the prediction head is P i,t , where t = 0, 1 represents different time phases and i represents the index of the sample. The self-supervised learning branch is trained by maximizing the cosine similarity between the two.

[0043] Furthermore, the step 5 is specifically as follows: the patches corresponding to all pixel positions of the SAR image are input into the encoder, and the prediction is performed through the supervised learning branch to obtain the prediction result of each patch; all the patch prediction results are assigned to the corresponding pixel positions, so as to obtain a complete change map.

[0044] In summary, the present invention first obtains pseudo labels from dual-phase SAR images through pre-classification, and obtains supervised sample sets and self-supervised sample sets by random sampling based on the pseudo labels; then, a collaborative change consistency network model is constructed: the encoder extracts deep feature vectors from the samples through cascaded residual blocks, the feature vectors corresponding to the supervised sample set are input into the supervised learning branch for supervised learning in combination with the pseudo labels, and the feature vectors corresponding to the self-supervised sample set are input into the self-supervised learning branch for self-supervised learning by minimizing the feature distance; then, a change consistency loss function is established between the two branches to achieve dual-branch collaborative training and enhance the model's ability to characterize change features; finally, the SAR image to be detected is input into the trained model to achieve change detection that is not easily affected by noise. The present invention effectively solves the problem that the accuracy of change detection in the existing unsupervised SAR change detection method is easily affected by noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flowchart of the present invention;

[0046] Figure 2 is a specific implementation flow chart of an embodiment;

[0047] Figure 3 is a schematic diagram of the structure of the collaborative change consistency network model in the embodiment;

[0048] Figure 4 Schematic diagram of the structure of the encoder in the embodiment. DETAILED DESCRIPTION

[0049] The present invention is further described in detail below in conjunction with embodiments and drawings.

[0050] An unsupervised SAR change detection method based on a collaborative change consistency network model, see Figure 2 , including the following steps:

[0051] Step 1: Apply the mean ratio difference operator and fuzzy c-means clustering to the dual-phase SAR images to obtain pseudo labels.

[0052] 1.1) The difference image is obtained by using the mean ratio operator (LR) for the dual-phase images in the dataset.

[0053] 1.2) Based on DI, the fuzzy c-means clustering algorithm (FCM) is applied to obtain the clustering results as pseudo labels.

[0054] Step 2: Randomly sample the pseudo labels obtained in step 1 to obtain a sample set, divide the samples whose pseudo labels are unchanged into the self-supervised sample set, and divide the samples whose pseudo labels are changed into the supervised sample set.

[0055] 2.1) Fill the dual-phase SAR image with 3 rows / columns of data in four directions respectively, and then select a 7×7 neighborhood window around each pixel position in the two-phase SAR image as the center as the sample, and use the pseudo label of the pixel position as the label of the sample.

[0056] 2.2) Randomly select 15% of all samples as the training sample set. According to whether the pseudo label is changed or unchanged, all the changed samples and unchanged samples in the sample set are used as the supervised sample set, and all the unchanged samples in the sample set are used as the self-supervised sample set.

[0057] Step 3: Construct a collaborative change consistency network structure, combine the supervision branch and the self-supervision branch to build a parallel structure, and impose consistency constraints between the two to obtain a change detection model based on the collaborative change consistency network. This includes: construction of the supervision branch, construction of the self-supervision branch, construction of the change consistency, and construction of the collaborative change consistency network.

[0058] See also Figure 3 , Figure 3 Schematic diagram of the structure of the collaborative change consistency network model in the embodiment. The network consists of two weight-sharing encoders and two corresponding branches, the two branches are respectively a supervised learning branch and a self-supervised learning branch, and the components of each part are as follows:

[0059] Building an encoder: See Figure 4 , Figure 4 It is a structural schematic diagram of the encoder in the embodiment;

[0060] The network structure of the encoder is as follows: input layer → first convolution layer (3×3, stride 2, number of convolution kernels is 64 → first batch normalization layer → ReLU activation function → residual block 1 (contains two 3×3 convolution layers, the number of convolution kernels in each layer is 64, the stride is 2 and 1 respectively, with batch normalization and ReLU activation function → residual block 2 (contains two 3×3 convolution layers, the number of convolution kernels in each layer is 64, the stride is 2 and 1 respectively, with batch normalization and ReLU activation function) → residual block 3 (contains two 3×3 convolution layers, the number of convolution kernels in each layer is 128, with step sizes of 2 and 1, with batch normalization and ReLU activation function) → residual block 4 (containing two 3×3 convolution layers, with 128 convolution kernels in each layer, with step sizes of 2 and 1, with batch normalization and ReLU activation function) → adaptive average pooling layer (output size is 1×1) → flattening layer. Among them, the momentum of all batch normalization layers is the default value, 1×1 convolution is used for downsampling in the residual block, and the activation function of each layer is the ReLU function. The features extracted by the encoder are finally flattened to obtain a feature vector of length 128.

[0061] Constructing the supervised learning branch:

[0062] The supervised learning branch consists of change feature extraction and classification, and its role is to enable the network to obtain preliminary discrimination capabilities through pseudo-label supervision. Specifically, the supervised learning branch consists of two weight-sharing multilayer perceptrons and a classifier. The output dimensions of each layer of the multilayer perceptron are 256-512-512, and the output dimensions of the classifier are 512-128-2. The output length is 2 vectors, and the prediction probability is obtained through the Softmax activation function.

[0063] Building a self-supervised learning branch:

[0064] The self-supervised learning branch includes feature vector mapping and similarity pulling, which helps the network to mine more feature information without labels as a supplement to the supervision branch. The self-supervised branch consists of two shared-weight projection heads and two shared-weight prediction heads. The projection head is a multi-layer perceptron with an output dimension of 128-512-512, and the prediction head is a multi-layer perceptron with an output dimension of 512-128-512.

[0065] Step 4: Use the supervised training set to train the supervised branch, and use the self-supervised training set to train the self-supervised branch. The two optimize parameters through gradient back propagation, and finally complete the training of the co-variation consistency network. Training of supervised learning branch:

[0066] The bi-phase feature vector F corresponding to the i-th supervised sample extracted by the encoder i,1 ,F i,2 , channel splicing is performed separately, and then the change features are extracted through a multi-layer perceptron.

[0067] Change feature F i,d The calculation process of extraction is:

[0068] F i,d =MLP(concat(F i,1 ,F i,2 ))

[0069] The change features are input into the classifier for classification through the softmax activation function, and the prediction results are output;

[0070] The calculation process of change feature prediction is:

[0071] y i ′=softmax(MLP(F i,d ))

[0072] Combining the prediction results and pseudo labels, the supervised learning branch loss is calculated through the cross entropy loss function, and the parameters are optimized through the gradient back propagation algorithm.

[0073] The calculation process of supervised learning branch loss is:

[0074]

[0075] Self-supervised learning branch training:

[0076] The feature vector extracted by the encoder is mapped to a higher dimension through the projection head, and then the high-dimensional feature vector is further predicted by the prediction head to obtain the final feature representation.

[0077] Eigenvector Map Z i,t The calculation process is:

[0078]

[0079] in, represents the feature vector of the tth phase of the i-th self-supervised sample, Z i,t Represents the high-dimensional representation of the features of the t-th phase of the i-th self-supervised sample;

[0080] High-dimensional feature vector prediction P i,t The calculation process is:

[0081] P i,t =MLP(Z i,t )

[0082] Among them, P i,t Represents the prediction result of the high-dimensional feature vector of the t-th phase of the i-th sample;

[0083] The self-supervised learning branch calculates the loss through the cosine similarity function and optimizes the network parameters by maximizing the cosine similarity between feature vectors;

[0084] The calculation process of cosine similarity is:

[0085]

[0086] The self-supervised loss function is calculated by i,t and P i,t The negative cosine similarity between is obtained, and the calculation process is:

[0087]

[0088] Among them, stopgrad means stopping gradient propagation, N ssl Represents the total number of all samples in the self-supervised sample set.

[0089] Change consistency loss:

[0090] First, the distance between the high-dimensional features of the self-supervised branch is calculated, and then the KL divergence is used to measure the similarity between the high-dimensional feature distance and the change features extracted by the supervised branch. Finally, the change consistency loss is obtained by maximizing the KL divergence.

[0091] The high-dimensional feature distance calculation process is:

[0092] D i,1 =||p i,1 -stopgrad(z i,2 )||2

[0093] D i,2 =||p i,2 -stopgrad(z i,1)||2

[0094] Among them, D i,1 and D i,2 Represents the distance between a pair of mutually symmetric eigenvectors;

[0095] The calculation process of KL divergence is:

[0096]

[0097] Where P and Q represent the probability release for the random variable ξ;

[0098] The calculation process of change consistency loss is:

[0099]

[0100] Collaborative loss function:

[0101] The collaborative loss of the network consists of the cross entropy loss of the supervised learning branch, the cosine similarity loss of the self-supervised learning branch, and the change consistency loss between the two branches;

[0102] The calculation process of the collaborative loss function is:

[0103] L=L sl +λ ssl L ssl +λ cons L cons

[0104] Among them, λ ssl and λ cons is a custom weight parameter used to adjust the proportion of each loss function in the collaborative loss. ssl =0.7,λ cons =0.5.

[0105] Finally, the adaptive moment estimation (Adam) algorithm is used to iteratively minimize the collaborative loss function and perform iterative training for 50 generations.

[0106] Step 5: Input the SAR images to be detected at different phases into the trained collaborative variation consistency network to obtain the final detection results.

[0107] 5.1) The 7×7 neighborhood window of each pixel position in the original SAR image is taken as the output of the network. After the encoder extracts the features, it is input into the supervised learning branch to obtain the prediction probability, and the index of the position with a higher probability is taken as the prediction result corresponding to each pixel.

[0108] 5.2) Combine all prediction results according to the corresponding pixel positions to obtain the final change map.

[0109] It can be seen from the above embodiments that the present invention builds a collaborative change consistency network model based on a supervised sample set and a self-supervised sample set, and uses an encoder to extract deep feature vectors from samples through cascaded residual blocks. The feature vectors corresponding to the supervised sample set are input into the supervised learning branch for supervised learning in combination with pseudo labels, and the feature vectors corresponding to the self-supervised sample set are input into the self-supervised learning branch for self-supervised learning by minimizing the feature distance; then a change consistency loss function is established between the two branches, thereby realizing dual-branch collaborative training, enhancing the model's ability to characterize change features, and better suppressing noise interference. It effectively solves the problem that the existing unsupervised SAR change detection method is easily affected by noise in terms of change detection accuracy.

Claims

1. An unsupervised SAR change detection method based on a collaborative change consistency network model, characterized in that: The following steps are involved: Step 1: Pre-classify the dual-phase SAR images to obtain pseudo labels; Step 2: Sample the pseudo labels obtained in step 1 to obtain a supervised sample set, and then sample from the supervised sample set to obtain a self-supervised sample set; Step 3: Construct a collaborative variation consistency network: This includes the construction of the encoder, the construction of the supervision branch, and the construction of the self-supervision branch. The specific process is as follows: First, an encoder with two shared weights is constructed to extract features from samples of two time phases respectively, and obtain feature vectors corresponding to the two time phases; The dual-phase samples in the supervised sample set and the self-supervised sample set are respectively input into two encoders for feature extraction; then the supervised learning branch and the self-supervised learning branch are respectively constructed, and the feature vectors extracted by the two encoders are respectively input into their corresponding learning branches; Step 4: Use the two sample sets to train the co-variation consistency network respectively, and perform gradient back propagation through the overall co-variation consistency function L of the co-variation consistency network to complete the training of the co-variation consistency network; Using the supervised sample set, after the encoder extracts features, the supervised learning branch is supervised by pseudo labels, and the parameters are optimized by the supervised loss function; The supervised loss function is defined as the binary cross entropy loss, which is expressed as: Among them, y i represents the pseudo label corresponding to the i-th supervised sample, y i ′ represents the prediction result corresponding to the i-th supervised sample, N sl Represents the total number of samples in the supervised sample set; Using the self-supervised sample set, after the encoder extracts features, self-supervised learning is performed without labels, and the parameters are optimized through the self-supervised loss function; The self-supervised loss function is constructed through cosine similarity and is expressed as: Among them, stopgrad means stopping gradient propagation, N ssl Represents the total number of all samples in the self-supervised sample set, P i,t is the high-dimensional feature vector prediction, Z i,t is the feature vector mapping, t represents the time phase and takes 1 or 2; In addition, by restricting the consistency of the change features between the supervised learning branch and the self-supervised learning branch, the similarity between the probability distributions of the change features is measured by KL divergence, and a change consistency loss function is constructed; it can be expressed as: F i,d is the change characteristic, D i,1 and D i,2 Represents a pair of mutually symmetrical feature vector distances; combining all loss functions, limiting the proportion of each loss through weight parameters, the overall collaborative loss function L of the collaborative change consistency network is obtained as follows: L=L sl +λ ssl L ssl +λ cons L cons The parameter optimization of the co-variation consistency network is achieved by gradient back propagation through L; Step 5: Input the two SAR images to be detected into the collaborative change consistency network trained in step 4 to obtain the final change detection result.

2. The unsupervised SAR change detection method based on the collaborative change consistency network model as claimed in claim 1, characterized in that: The step 1 is specifically as follows: The dual-phase SAR images are defined as T1 and T2 respectively, and the difference map DI is first generated by the logarithmic ratio operator. The calculation process is as follows: DI = abs(log((T1+1) / (T2+1))); Then, DI is classified by the fuzzy C-means clustering algorithm to obtain the clustering results, which are used as pseudo labels.

3. The unsupervised SAR change detection method based on the collaborative change consistency network model as claimed in claim 2, characterized in that: The specific steps of step 2 are as follows: The pseudo labels obtained in step 1 are sampled to obtain a supervised sample set, and then a self-supervised sample set is sampled from the supervised sample set; Before sampling, the SAR image is expanded. For T1, T2∈R H×W Mirror it in four directions: up, down, left, and right to fill three rows / columns with values ​​T1', T2'∈R (H+6)×(W+6) , H represents the number of rows of SAR images, and W represents the number of columns of SAR images; finally, a 7×7 sliding window is used to take the neighborhood of two SAR images of different phases with the pixel as the center as the patch; then 15% of the patches are randomly sampled from them to form the supervision sample set Ω sl ; Then, according to the pseudo-label, determine whether the corresponding supervised sample is a changed sample, and divide all supervised samples into two categories {Ω0,Ω1}, where Ω1 consists of all samples whose pseudo-labels are changed, and Ω0 consists of all samples whose pseudo-labels are unchanged, and Ω0 is used as the self-supervised sample set.

4. The unsupervised SAR change detection method based on the collaborative change consistency network model as claimed in claim 1, characterized in that: The encoder comprises an input layer, a residual layer and an output layer; The input layer includes a common convolution layer with a convolution kernel size of 3×3 and a Relu activation function, which convolves the input patch to obtain a preliminary shallow feature representation; The residual layer is composed of four stacked residual blocks to represent deep features; Each residual block first calculates the input feature map F through a normal convolution layer with a convolution kernel size of 3×3, a Relu activation function, and a BatchNorm layer to obtain F1, and then calculates F2 through a normal convolution layer with a convolution kernel size of 3×3 and a BatchNorm layer; then, the input feature is downsampled through a normal convolution layer with a convolution kernel size of 1×1, and then summed with the input feature map F through a jump connection; finally, the output feature map is obtained through Relu activation. The calculation process is as follows: <h2 style=";text-align:left;direction:ltr">F<h2 style=";text-align:left;direction:ltr"> o <h2 style=";text-align:left;direction:ltr"> =Relu(F2+Conv(F)) After passing through four residual blocks in sequence, the deep feature F' is obtained, and finally converted into a feature vector through the output layer; the output layer includes global average pooling operation and flattening operation in sequence, so as to reduce the dimension of the extracted deep features and convert them into vectors.

5. The unsupervised SAR change detection method based on the collaborative change consistency network model as claimed in claim 2, characterized in that: The supervised learning branch includes change feature extraction and change feature classification, and the bi-phase feature vector F extracted by the encoder is i,1 and F i,2 , after vector concatenation, it is input into the multi-layer perceptron to obtain the fused change feature F i,d ; F i,d The input is sent to the classifier for dimensionality reduction to obtain a vector of length 2, and finally the prediction result is obtained by activating the Softmax function.

6. The unsupervised SAR change detection method based on the collaborative change consistency network model as claimed in claim 5, characterized in that: The change feature extraction is expressed as: F i,d =MLP(concat(F i,1 ,F i,2 )).

7. The unsupervised SAR change detection method based on the collaborative change consistency network model as claimed in claim 2, characterized in that: The self-supervised learning branch includes two multi-layer perceptrons with shared weights, namely the projection head and the prediction head; the feature vector output by the projection head is Z i,t , the feature vector output by the prediction head is P i,t , where t = 0, 1 represents different time phases, i represents the index of the sample, and the self-supervised learning branch is trained by maximizing the cosine similarity between the two.

8. The unsupervised SAR change detection method based on the collaborative change consistency network model as claimed in claim 1, characterized in that: The step 5 is specifically as follows: the patches corresponding to all pixel positions of the SAR image are input into the encoder, and prediction is performed through the supervised learning branch to obtain the prediction result of each patch; all the patch prediction results are assigned to the corresponding pixel positions, so as to obtain a complete change map.

Images