Unsupervised sar change detection method based on cooperative change consistency network model

Abstract

The present application belongs to the technical field of change detection, and in particular to an unsupervised SAR change detection method based on a collaborative change consistency network model. First, the present application obtains pseudo labels by pre-classifying dual-time SAR images, and then randomly samples supervised sample sets and self-supervised sample sets according to the pseudo labels. Then, a collaborative change consistency network model is constructed: an encoder extracts deep feature vectors from samples through cascaded residual blocks, the feature vectors corresponding to the supervised sample sets are input into a supervised learning branch for supervised learning in combination with the pseudo labels, and the feature vectors corresponding to the self-supervised sample sets are input into a self-supervised learning branch for self-supervised learning by minimizing feature distance. Then, a change consistency loss function is established between the two branches to realize collaborative training of the two branches and enhance the model's representation ability for change features. Finally, the SAR images to be detected are input into the trained model to realize change detection that is not easily affected by noise.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of change detection, and specifically relates to an unsupervised SAR change detection method based on a collaborative change consistency network model. BACKGROUND

[0002] Change detection aims to identify changes in consistent spatial locations in different time instances, which can be regarded as a binary classification task, classifying each location as changed 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] Synthetic Aperture Radar (SAR) change detection is an important technology 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 all day. 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 will produce significant speckle noise in SAR images, which may cause high false positive rates in the change detection task, thus posing a challenge to change detection. Therefore, it is crucial to develop robust change detection techniques that 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 labeling or prior knowledge, making them the mainstream method of recent research. Traditional unsupervised SAR change detection methods mainly involve generating difference images (DIs) using various difference operators, and then classifying DIs through thresholding or clustering algorithms.

[0005] Although traditional methods are computationally efficient and widely applicable, they are limited by feature representation capabilities and are easily affected by the inherent speckle noise in SAR. Due to the powerful feature representation capability of deep learning, some researchers have successfully applied deep learning to unsupervised change detection, mainly involving two steps: 1) First, use clustering methods to generate pseudo labels; 2) The generated pseudo labels are treated as ground truth values, and a certain proportion of samples are sampled as a training set, thus realizing the training process. However, current mainstream methods often fail to fully exploit the feature information of samples and are still susceptible to noise, making it difficult to accurately obtain change information.

[0006] Through the above analysis, the problems and defects of the prior art are that the mainstream unsupervised SAR change detection method is prone to be affected by noise and the change detection accuracy is low. SUMMARY

[0007] In view of the problems of the prior art, the present application 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: obtaining pseudo-labels by pre-classifying the dual-time SAR images.

[0010] Step 2: obtaining a supervised sample set by sampling the pseudo-labels obtained in step 1, and obtaining a self-supervised sample set by sampling the supervised sample set.

[0011] Step 3: constructing a collaborative change consistency network, including the construction of an encoder, the construction of a supervised branch, and the construction of a self-supervised branch. The specific construction process is as follows:

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

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

[0014] Step 4: training the collaborative change consistency network using the two sample sets, and performing gradient backpropagation through the collaborative loss function L of the collaborative change consistency network to complete the training of the collaborative change consistency network.

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

[0016] The supervised loss function is defined as a binary cross-entropy loss, which is expressed by the formula as follows:

[0017]

[0018] where 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, and N sl represents the total number of samples in the supervised sample set.

[0019] Using the self-supervised sample set, the features are extracted by the encoder, and the self-supervised learning is carried out without labels, and the parameters are optimized by the self-supervised loss function;

[0020] The self-supervised loss function is constructed by cosine similarity, and is expressed by formula as follows:

[0021]

[0022] Where, stopgrad represents 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 feature between the supervised learning branch and the self-supervised learning branch, the similarity between the probability distributions of the change feature is measured by KL divergence, and the change consistency loss function is constructed. Expressed by formula as follows:

[0024]

[0025] F i,d is the change feature, D i,1 and D i,2 are a pair of mutually symmetric feature vector distances; combining all the loss functions, the proportion of each loss is limited by weight parameters λ ssl and λ cons , and the collaborative loss function L of the whole collaborative change consistency network is obtained as:

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

[0027] The parameter optimization of the collaborative change consistency network is realized by gradient back propagation of L.

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

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

[0030] Further, the step 2 is specifically: sampling the pseudo label obtained in step 1 to obtain a supervised sample set, and then sampling the supervised sample set to obtain a self-supervised sample set.

[0031] The SAR image is expanded before sampling, and T1, T2∈R H×W Mirror filling three rows / columns T1', T2'∈R in the up, down, left and right directions. (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, the neighborhood of the two SAR images at different time is taken as a patch with a 7*7 sliding window and a pixel as the center. Then, 15% of the patches are randomly sampled to form a supervised sample set Ω sl .

[0032] Then, according to the pseudo label, it is judged whether the corresponding supervised sample is a change sample, and all supervised samples are divided into two categories {Ω0, Ω1}, wherein Ω1 is composed of all samples with pseudo labels as changes, and Ω0 is composed of all samples with pseudo labels as no changes. Ω0 is used as a self-supervised sample set.

[0033] Further, 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, and the input patch is convolved to obtain a preliminary shallow feature representation.

[0035] The residual layer is stacked by four 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 down-sampled 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 a Relu activation, and the calculation process is as follows:

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

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

[0039] Further, the supervised learning branch includes change feature extraction and change feature classification, and the dual-phase feature vector F i,1 and F i,2 are input into a multi-layer perception after vector splicing to obtain fused change features F i,d . F i,d is input into a classifier for dimension reduction to obtain a vector with a length of 2, and finally a prediction result is obtained through Softmax function activation.

[0040] Further, the change feature extraction can be specifically expressed by a formula as follows:

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

[0042] Further, the self-supervised learning branch includes two multi-layer perceptions with shared weights, which are respectively referred to as a projection head and a prediction head. The feature vector output by the projection head is Z i,t , and the feature vector output by the prediction head is P i,t , wherein t=0, 1 represents different phases, and i represents the index of a sample. The self-supervised learning branch is trained by maximizing the cosine similarity between the two.

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

[0044] To sum up, the present application firstly obtains pseudo labels through pre-classification of dual-phase SAR images, and then randomly samples a supervised sample set and a self-supervised sample set according to the pseudo labels; then, a collaborative change consistency network model is constructed: an encoder extracts a deep feature vector from a sample through cascaded residual blocks, a feature vector corresponding to the supervised sample set is input into a supervised learning branch for supervised learning in combination with pseudo labels, and a feature vector corresponding to the self-supervised sample set is input into a self-supervised learning branch for self-supervised learning by minimizing feature distance; change consistency loss functions are established between the two branches to realize collaborative training of the two branches and enhance the representation ability of the model to change features; finally, a SAR image to be detected is input into the trained model to realize change detection which is not easily affected by noise. The present application effectively solves the problem that the existing unsupervised SAR change detection method is easily affected by noise and has low change detection precision. BRIEF DESCRIPTION OF DRAWINGS

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

[0046] Figure 2 is a specific implementation flowchart of the embodiment;

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

[0048] Figure 4 is a structural diagram of an encoder in the embodiment. DETAILED DESCRIPTION

[0049] The application will be further described in detail below in combination with the embodiments and the accompanying drawings.

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

[0051] Step 1, applying a mean ratio difference operator and fuzzy c-means clustering to the dual-time SAR image to obtain pseudo labels.

[0052] 1.1) Obtaining a difference image by a mean ratio operator (LR) on the dual-time image in the data set.

[0053] 1.2) Obtaining a clustering result as a pseudo label by applying a fuzzy c-means clustering algorithm (FCM) on the basis of the DI.

[0054] Step 2, randomly sampling to obtain a sample set according to the pseudo label obtained in step 1, dividing the sample with the pseudo label as unchanged into a self-supervised sample set, and dividing the sample with the pseudo label as changed into a supervised sample set.

[0055] 2.1) Filling 3 rows / columns of data in four directions of rows and columns of the dual-time SAR image respectively, then selecting a 7x7 neighborhood window around each pixel position in the SAR image of the two time phases as a sample, and taking the pseudo label of the pixel position as the label of the sample.

[0056] 2.2) Randomly selecting 15% from all samples as a training sample set. According to the pseudo label as changed or unchanged, all changed samples and unchanged samples in the sample set are taken as a supervised sample set, and all unchanged samples in the sample set are taken as a self-supervised sample set.

[0057] Step 3, constructing a collaborative change consistency network structure, combining a supervised branch and a self-supervised branch to construct a parallel structure, and applying consistency constraints between the two to obtain a change detection model based on a collaborative change consistency network. Including: construction of the supervised branch, construction of the self-supervised branch, construction of change consistency, and construction of the collaborative change consistency network.

[0058] Referring toFigure 3 Figure 3 The structural diagram of the network model based on the collaborative change consistency in the embodiment is shown in FIG. 2. The network is composed of two weight-shared encoders and corresponding two branches, i.e., a supervised learning branch and a self-supervised learning branch. The components of each part are as follows:

[0059] The encoder is constructed as follows: Figure 4 Figure 4 The structural diagram of the encoder in the embodiment is shown in FIG. 3;

[0060] The network structure of the encoder is as follows: input layer→first layer of convolutional layer (3×3, step size 2, number of convolutional kernels 64→first layer of batch normalization layer→ReLU activation function→residual block 1 (containing two 3×3 convolutional layers, number of convolutional kernels 64 in each layer, step sizes 2 and 1, with batch normalization and ReLU activation function→residual block 2 (containing two 3×3 convolutional layers, number of convolutional kernels 64 in each layer, step sizes 2 and 1, with batch normalization and ReLU activation function)→residual block 3 (containing two 3×3 convolutional layers, number of convolutional kernels 128 in each layer, step sizes 2 and 1, with batch normalization and ReLU activation function)→residual block 4 (containing two 3×3 convolutional layers, number of convolutional kernels 128 in each layer, step sizes 2 and 1, with batch normalization and ReLU activation function)→adaptive average pooling layer (output size 1×1)→flattening layer. The momentum of all batch normalization layers is the default value, 1×1 convolution is used for down-sampling 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 with a length of 128.

[0061] The supervised learning branch is constructed as follows:

[0062] The supervised learning branch is composed of change feature extraction and classification, and its role is to enable the network to obtain preliminary discrimination ability through the supervision of pseudo labels. Specifically, the supervised learning branch is composed of two weight-shared multilayer perceptrons and a classifier. The output dimensions of each layer of the multilayer perceptron are 256-512-512, respectively, and the output dimension of the classifier is 512-128-2, i.e., a vector with a length of 2, and the prediction probability is obtained through the Softmax activation function.

[0063] The self-supervised learning branch is constructed as follows:

[0064] ​​The self-supervised learning branch includes the mapping and similarity of the feature vectors, and the role is to help the network to mine more feature information as a supplement to the supervised branch under the condition of no label. The self-supervised branch is composed of two projection heads sharing weights and two prediction heads sharing weights. Among them, 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, the supervised training set is used to train the supervised branch, and the self-supervised training set is used to train the self-supervised branch, and the parameters are optimized through gradient back propagation, and finally the training of the collaborative change consistency network is completed. Supervised learning branch training:

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

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

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

[0069] The change feature is input to the classifier and classified through the softmax activation function, and the prediction result is output.

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

[0071] y i ′=softmax(MLP(F i,d ))

[0072] Combined with the prediction result and the pseudo label, the supervised learning branch loss is calculated through the cross entropy loss function, and the parameter optimization is carried out through the gradient back propagation algorithm.

[0073] The calculation process of the 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 through the prediction head to obtain the final feature representation.

[0077] The calculation process of the feature vector mapping Z i,t is:

[0078]

[0079] wherein, denotes the feature vector of the t-th time phase of the i-th self-supervised sample, Z i,t denotes the feature high-dimensional representation of the t-th time phase of the i-th self-supervised sample;

[0080] The calculation process of the high-dimensional feature vector prediction P i,t is as follows:

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

[0082] wherein, P i,t denotes the prediction result of the high-dimensional feature vector of the t-th time phase of the i-th sample;

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

[0084] The calculation process of the cosine similarity is as follows:

[0085]

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

[0087]

[0088] wherein, stopgrad denotes stopping gradient propagation, N ssl denotes 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, then the similarity between the high-dimensional feature distance and the change feature extracted by the supervised branch is measured through KL divergence, and finally the change consistency loss is obtained by maximizing the KL divergence;

[0091] The calculation process of the high-dimensional feature distance is as follows:

[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] where D i,1 and D i,2 represent a pair of mutually symmetric eigenvector distances;

[0095] The calculation process of KL divergence is as follows:

[0096]

[0097] where P and Q represent the probability distribution of random variable ξ;

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

[0099]

[0100] The collaborative loss function is:

[0101] The collaborative loss of the network includes 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 as follows:

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

[0104] where λ ssl and λ cons are self-defined weight parameters used to adjust the proportion of each loss function in the collaborative loss. Among them, λ ssl = 0.7, λ cons = 0.5.

[0105] Finally, the adaptive moment estimation (Adam) algorithm is used to minimize the collaborative loss function through iteration and training, and the iteration number is 50 times.

[0106] Step 5, input the SAR images to be detected at different time into the trained collaborative change consistency network to obtain the final detection result.

[0107] 5.1) The 7x7 size neighborhood window of each pixel position of the original SAR image is taken as the output of the network. After feature extraction by the encoder, it is input into the supervised learning branch to obtain the prediction probability. The index of the position with higher probability is taken as the prediction result corresponding to each pixel.

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

[0109] As can be seen from the above embodiments, the present application 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 a deep feature vector from the samples through cascaded residual blocks, the feature vector corresponding to the supervised sample set is input into a supervised learning branch for supervised learning combined with a pseudo label, and the feature vector corresponding to the self-supervised sample set is input into a 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, so as to realize double-branch collaborative training, enhance the model's representation ability for change characteristics, and more effectively suppress the interference of noise. The problem that the existing unsupervised SAR change detection method is easily affected by noise and the change detection precision is low is effectively solved.

Claims

1. An unsupervised SAR change detection method based on a synergistic change consistency network model, characterized in that, The method comprises the following steps: Step 1: obtaining pseudo labels by pre-classifying dual-temporal SAR images; Step 2: obtaining a supervised sample set by sampling the pseudo labels obtained in step 1, and obtaining a self-supervised sample set by sampling the supervised sample set; Step 3: constructing a collaborative change consistency network, including construction of an encoder, construction of a supervised branch, and construction of a self-supervised branch; the specific process is as follows: First, an encoder with two shared weights is constructed to extract features of samples in two time phases, and feature vectors corresponding to the two time phases are obtained; The dual-temporal samples in the supervised sample set and the self-supervised sample set are input into two encoders for feature extraction, respectively; then a supervised learning branch and a self-supervised learning branch are constructed, and the feature vectors extracted by the two encoders are input into the corresponding learning branches, respectively; Step 4: training the collaborative change consistency network using the two sample sets, and performing gradient back propagation on the collaborative loss function L of the collaborative change consistency network as a whole to complete the training of the collaborative change consistency network; After the features are extracted by the encoder, the supervised learning branch is supervised by the pseudo labels to perform supervised learning, and the parameters are optimized by the supervised loss function; The supervised loss function is defined as a binary cross-entropy loss, which is expressed by a formula as follows: wherein 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; After the features are extracted by the encoder, the self-supervised learning branch is supervised without labels to perform self-supervised learning, and the parameters are optimized by the self-supervised loss function; The self-supervised loss function is constructed by cosine similarity, which is expressed by a formula as follows: where stopgrad denotes stop gradient propagation, N ssl denotes 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, the consistency loss function of the change feature is constructed by limiting the consistency between the supervised learning branch and the self-supervised learning branch, and the similarity between the probability distributions of the change features is measured by KL divergence; the consistency loss function is expressed by a formula as follows: F i,d For the changing feature, D i,1 and D i,2 represent a pair of mutually symmetrical feature vector distances; in combination with all loss functions, the proportion of each loss is limited by a weight parameter, and the collaborative loss function L of the collaborative change consistency network as a whole is obtained as: L = L sl + λ ssl L ssl + λ cons L cons The parameter optimization of the collaborative change consistency network is realized by gradient back propagation of L; Step 5: inputting 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 synergic change consistency network model according to claim 1, wherein, The step 1 is specifically as follows: The dual-temporal SAR images are defined as T1 and T2, respectively, and a difference image DI is generated by a logarithmic ratio operator, and the calculation process is as follows: DI = abs(log((T1+1) / (T2+1))); Then, the 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.

3. The unsupervised SAR change detection method based on the synergic change consistency network model according to claim 2, wherein, The step 2 is specifically as follows: The pseudo labels obtained in step 1 are sampled to obtain a supervised sample set, and the supervised sample set is sampled to obtain a self-supervised sample set; H×W (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; finally, a 7*7 sliding window is used to take a neighborhood as a patch with a pixel as a center for the SAR images of two different time phases; then, 15% of the patches are randomly sampled to form a supervised sample set Ω sl ;​​ Then, according to the pseudo label, it is judged whether the corresponding supervised sample is a change sample, and all the supervised samples are divided into two categories {Ω0, Ω1}, wherein Ω1 is composed of all samples with pseudo labels as changes, Ω0 is composed of all samples with pseudo labels as no changes, and Ω0 is used as a self-supervised sample set.

4. The unsupervised SAR change detection method based on the synergic change consistency network model according to claim 1, characterized in that: The encoder comprises an input layer, a residual layer, and an output layer; The input layer comprises a normal convolution layer with a convolution kernel size of 3x3 and a Relu activation function, and the input patch is convolved to obtain a preliminary shallow feature representation; The residual layer is stacked by four residual blocks to represent deep features; Each residual block first calculates the input feature map F through a common convolution layer with a kernel size of 3*3, a Relu activation function and a BatchNorm layer to obtain F1, and then calculates F2 through a common convolution layer with a kernel size of 3*3 and a BatchNorm layer; then, the input feature is down-sampled through a common convolution layer with a 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 a Relu activation, and the calculation process is as follows: F o = Relu(F2 + Conv(F)) After sequentially passing through the four residual blocks, the deep feature F' is obtained, and finally converted into a feature vector through the output layer; wherein the output layer sequentially includes a global average pooling operation and a flattening operation, so as to reduce the dimension of the extracted deep feature and convert it into a vector.

5. The unsupervised SAR change detection method based on the collaborative change consistency network model according to claim 2, wherein: 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 splicing, 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 activation through the Softmax function.

6. The unsupervised SAR change detection method based on the synergic change consistency network model according to claim 5, wherein, The change feature extraction is expressed by a formula as F i,d = MLP(concat(F i,1 , F i,2 )).

7. The unsupervised SAR change detection method based on the synergic change consistency network model according to claim 2, wherein: The self-supervised learning branch includes two multi-layer perceptrons sharing weights, namely a projection head and a prediction head; the feature vector output by the projection head is Z i,t , and the feature vector output by the prediction head is P i,t , wherein t=0, 1 represents different time phases, and i represents the index of a sample; the self-supervised learning branch is trained by maximizing the cosine similarity between Z and P.

8. The unsupervised SAR change detection method based on the collaborative change consistency network model according to claim 1, wherein: The step 5 is specifically: inputting the patch corresponding to all pixel positions of the SAR image into the encoder, performing prediction through the supervised learning branch to obtain the prediction result of each patch, and assigning all patch prediction results to the corresponding pixel positions to obtain a complete change map.

Images