An unsupervised change detection method for SAR and optical images based on pseudo-sample learning

Abstract

The present invention discloses an unsupervised SAR and optical image change detection method based on pseudo-sample learning. The method directly learns the change knowledge by constructing pseudo-samples, continuously improves the quality of pseudo-sample sets, and consolidates and generalizes the learned knowledge to improve the reliability of the network. First, a pseudo-sample set and a real sample training set are constructed, and then a dual-branch U-Net network is constructed. Then, the pseudo-sample learning branch and the real sample learning branch of the network are trained respectively. After several trainings, the dual-time images are respectively input into the two encoders of the network, and the decoder outputs a change probability map. Finally, the three probability maps are fused using a weighted fusion method, and then threshold segmentation is performed to obtain the final detection binary map. The present invention constructs a labeled data set for training the network by constructing a change region, avoiding the complex feature design of the traditional unsupervised change detection method and simplifying the training process.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing images, and in particular relates to an unsupervised SAR and optical image change detection method. Background Art

[0002] Remote sensing image change detection is a technology that identifies changes in the surface or objects by comparing remote sensing images from different periods. Understanding the changes in the surface is of great significance to many fields. Specifically, it can be used to monitor urban expansion, land use changes and building growth for urban planning; monitor environmental changes such as deforestation, wetland degradation, and land degradation to protect ecosystems and assess environmental risks; monitor damage caused by natural disasters (such as floods, earthquakes, and forest fires) and man-made disasters (such as industrial accidents and oil spills) to help emergency response agencies respond quickly and rescue, and assess the extent of losses; monitor and assess changes in infrastructure, such as roads, railways, bridges, and power lines, which are crucial for infrastructure planning, maintenance, and renewal. With the rapid development of remote sensing technology, we can obtain high-resolution, multi-spectral, multi-temporal, multi-sensor, and multi-platform remote sensing image data, which provides a large amount of data foundation for remote sensing image change detection. At the same time, diverse remote sensing data also promotes the development of multi-source image processing tasks.

[0003] In the change detection task, compared with homogeneous remote sensing images, heterogeneous remote sensing images are more flexible in acquisition. Images of a certain place obtained by any sensor or platform can be selected from the image library to detect changes. Therefore, heterogeneous change detection technology is often used in emergency response. In addition, heterogeneous remote sensing images have different imaging mechanisms, which can make up for the shortcomings of single sensor images in some occasions. For example, in foggy weather or at night, visible light sensors cannot capture clear images of objects, while SAR sensors can work all day long. The comprehensive use of the two sensors can fully grasp the situation of objects. Although heterogeneous image change detection has many advantages, due to the large differences in imaging between different sensors and platforms, the same object has huge differences in different images, which brings great difficulties to the change detection task. The traditional homogeneous change detection method is not suitable for heterogeneous remote sensing images. In addition, due to the high cost of annotation, the labeled data that can be used for heterogeneous change detection is scarce, which increases the difficulty of heterogeneous change detection tasks.

[0004] Existing methods can be roughly divided into four categories: classification-based methods, that is, first classify the image, and then compare whether the dual-time image categories are consistent; image conversion-based methods, that is, convert the image to another phase image domain or convert each other, so that they can be directly compared; based on feature space alignment, that is, project heterogeneous images into the same feature space for comparison; based on similarity measurement, this type of method needs to find a robust image structure representation method to compare whether the structural information of two images is similar to determine whether there is a change. The existing four types of methods require the design of complex proxy tasks or feature representation methods. Summary of the invention

[0005] In order to overcome the deficiencies of the prior art, the present invention provides an unsupervised SAR and optical image change detection method based on pseudo sample learning, which directly learns change knowledge by constructing pseudo samples, continuously improves the quality of pseudo sample sets, and consolidates and generalizes the learned knowledge to improve the reliability of the network. First, a pseudo sample set and a real sample training set are constructed, and then a dual-branch U-Net network is constructed. Then, the pseudo sample learning branch and the real sample learning branch of the network are trained respectively. After several trainings, the dual-time images are input into the two encoders of the network respectively, and the decoder outputs a change probability map. Finally, the weighted fusion method is used to fuse the three probability maps, and then threshold segmentation is performed to obtain the final detection binary map. The present invention constructs a labeled data set for training the network by constructing a change region, avoids the complex feature design of the traditional unsupervised change detection method, and simplifies the training process.

[0006] The technical solution adopted by the present invention to solve the technical problem includes the following steps:

[0007] Step 1: Construct a pseudo sample set;

[0008] Step 1-1: Build the paste library;

[0009] Perform unsupervised classification on the selected phase and divide the pixels of the entire image into C categories. Use median filtering with a filter kernel size of K to filter the classification values, and then search the entire image in a sliding window manner. Each position includes three window sizes: large, medium, and small. As long as a category in the window is greater than the threshold T bank is added to the corresponding library, and finally a paste library with C categories is obtained, each category including multiple paste pieces of different sizes;

[0010] Step 1-2: Screening samples;

[0011] The dataset is divided into patches as basic training images, and the number of changed pixels in the prior change map is calculated to see whether it is less than the threshold T. unch Screen the patches used to construct the pseudo sample set;

[0012] Step 1-3: Select the pasting area and pasting piece;

[0013] Randomly select the position, size, and shape of the paste to generate the paste area, and select the category consistency within the paste area that exceeds the threshold T pasted area;

[0014] Calculate the pixel center in the pasting area and select any pasting patch in the c categories farthest from it;

[0015] After pasting the patch, the corresponding position of the prior change map will also be set to 1 as the change class, thus obtaining the label of the pseudo sample;

[0016] Step 2: Construct a real sample training set;

[0017] Randomly select patches from all patches that are equal to the number of pseudo samples, and use the strong data augmentation method of Cut-Mix to paste the random regions in the corresponding pseudo samples into the real samples, and paste the two phases and labels at the same time;

[0018] Step 3: Build the network structure;

[0019] A dual-branch U-Net network is used, in which the decoder is a dual-branch structure without shared parameters, which is used to extract image features of the two phases. The two encoders have the same structure, and the decoder only performs downsampling once; Dropout is added to the convolution module; the features of each level of the two encoders are subtracted and the absolute value is taken, and then input into the decoder, where the first set of feature maps is input into the decoder through a jump connection, and this process remains consistent with U-Net; the last set of feature maps is pooled with the label as a mask to obtain two feature vectors: representing changes and invariance respectively; when masking pooling, bilinear interpolation is used to reduce the label to the size of the feature map;

[0020] Step 4: Pseudo sample learning branch;

[0021] In the pseudo sample learning branch, a weight decay pseudo sample learning branch loss is used, and the weight is set to increase linearly. Its loss L p for:

[0022]

[0023] Where CE(·) represents the cross entropy loss, c prior represents the area where changes occurred in the prior map, y p represents the label of the pseudo sample set, represents the predicted value; α represents the coefficient; i and j represent the pixels in the corresponding area;

[0024] Step 5: Real sample learning branch;

[0025] The real sample learning branch and the pseudo sample learning branch are trained in parallel and share network parameters. The real sample learning branch uses the prior change map of the whole image reasoning as a label and uses the cross entropy loss for supervision. This branch is trained with any patch and its loss L r for:

[0026]

[0027] where y r is the prior label;

[0028] Step 6: Contrastive feature learning;

[0029] Use contrast loss to constrain the distance between different branches and different categories of features;

[0030] The encoding results are masked and pooled according to the labels, and each training branch obtains two feature vectors, one changed and one unchanged. The cosine similarity is used as the distance measurement function for the features, and the four feature vectors can obtain the following four distance relationships.

[0031]

[0032] Where D(·,·) represents the cosine similarity, represents the intra-class similarity of unchanged regions, represents the intra-class similarity of the change region, represents the inter-class difference of the pseudo sample learning branch, Indicates the inter-class differences of the real sample learning branch;

[0033] A prototype vector is maintained for each changing class and the same class respectively. The new feature vector obtained in each iteration will update the prototype vector by averaging.

[0034] Step 7: The total training loss L is as follows;

[0035]

[0036]

[0037] Among them, β is the weight;

[0038] Step 8: Reasoning process;

[0039] After several trainings, the dual-time images are input into the two encoders of the network respectively, and finally the decoder outputs the change probability map p1; p1 is binarized with a threshold of 0.5, and the new change map will update the prior change map, and training will continue until the specified number of iterations is reached.

[0040] After training, the cosine similarity is calculated and normalized using the changed and unchanged prototype vectors and the features encoded by the encoder in the inference process to obtain p2 and p3. Finally, the weighted fusion method is used to fuse the three probability maps:

[0041] p=λ1p1+λ2p2+λ3(1-p3) (7)

[0042] Among them, λ1, λ2, and λ3 are weight coefficients respectively;

[0043] The threshold value of p is also 0.5 to obtain the final detection binary image.

[0044] Preferably, the e is the current round, e max is the maximum number of training rounds;

[0045] Preferably, C is 5, K is 7, T unch is 0.01, and c is 3.

[0046] Preferably, the λ1, λ2, and λ3 are 0.7, 0.2, and 0.1, respectively.

[0047] The beneficial effects of the present invention are as follows:

[0048] 1. The present invention constructs a labeled data set for training the network by constructing the change region, avoiding the complex feature design of the traditional unsupervised change detection method and simplifying the training process.

[0049] 2. The present invention designs a dual-branch training framework and combines it with feature contrast learning to effectively improve the detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is an overall schematic diagram of the present invention.

[0051] Figure 2 This is a schematic diagram of network data flow.

[0052] Figure 3 Schematic diagram of the network encoder and decoder structure, (a) encoder, (b) decoder, (c) basic convolutional block.

[0053] Figure 4 The present invention is a flowchart for the specific implementation of the method. DETAILED DESCRIPTION

[0054] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0055] The purpose of the present invention is to provide a method capable of more accurately analyzing the change areas of SAR and optical remote sensing images without labels.

[0056] The present invention adopts the following technical solution: using the Cut-Paste method to perform pseudo-sample learning on a certain phase structural change area, and strengthening and generalizing the learned knowledge on real samples.

[0057] The present invention uses a dual-branch U-Net structure, and the two branches input dual-time images respectively, and finally the decoder is used to predict the change probability map. The training process also includes two branches: a pseudo-sample learning branch and a real sample learning branch. The former uses the Cut-Paste method to construct pseudo-samples and labels for training, and the labels of real samples are the prior labels predicted by the network. The latter uses real samples and corresponding labels of partial area mixed pseudo-samples for training. The other two branches are trained with contrast loss, and the two branches are trained and optimized at the same time. The overall method is as shown in the attached figure. Figure 1 After the training reaches the condition, the dual-time image is input into the network to obtain the change probability map, and the change map is obtained after thresholding. The specific invention details are as follows.

[0058] (1) Construct a pseudo sample set.

[0059] The present invention divides the image into overlapping patches as samples. To construct a pseudo sample set, it is necessary to first construct a pasting library, and then select appropriate pasting patches from the library according to certain rules and paste them to reasonable positions.

[0060] 1) Build the paste library.

[0061] If you want to simulate the changed area by the Cut-Paste method, you must ensure that the semantic categories of the pasted piece and the pasted area are different. In order to ensure the unity of the pasted piece and the entire data set, and the changed area has a prototype in the image, the present invention extracts uniform areas from the original data set to form a paste library. Since the changed area appears in the post-event phase, the phase after the event is generally selected as the paste phase.

[0062] First, the selected phase is unsupervisedly classified. The present invention uses a classic clustering method to divide the pixels of the entire image into C categories. Considering the spatial local consistency of the categories, the present invention uses a median filter with a filter kernel size of K to filter the classification values, while greatly reducing the impact of SAR noise on the classification results. Then the entire image is searched in a sliding window manner, and each position includes three window sizes of large, medium and small. As long as a category in the window is greater than the threshold T bank Finally, a sticker library with C categories is obtained, each category including multiple stickers of different sizes.

[0063] 2) Screen samples.

[0064] Since there are changing areas and unchanged areas in the entire data set, ideally, only by pasting on the unchanged areas can accurate labels be obtained. Therefore, the present invention selects patches that may not have changed through the prior change map to construct a pseudo sample set. Specifically, the present invention calculates whether the number of changed pixels in the prior change map is less than a certain threshold T unch Filter patches.

[0065] 3) Select the paste area and paste patch.

[0066] For each training sample, a suitable position needs to be selected for pasting. In order to generate diverse samples, the present invention randomly selects the position, size and shape of the pasting to generate the pasting area. In order to reduce the semantic inaccuracy caused by inconsistent pasting area categories, the present invention selects the area with category consistency exceeding the threshold T. pasted area.

[0067] The pasting piece should be selected to have a different category from the pasted area. The present invention selects an arbitrary pasting piece in the c categories farthest from the center of the pixel in the area by calculating the center of the pixel in the area.

[0068] After selecting the training patch, its prior change map is obtained at the same time. After pasting the patch, the corresponding position of the prior change map will also be set to 1 as the change class. In this way, the label of the pseudo sample is obtained.

[0069] (2) Pseudo sample learning branch.

[0070] In (1), the pseudo sample set and its label have been obtained. The label combines the prior change information and the pasted change information. Since the prior change information is not necessarily accurate, the present invention uses a weight to attenuate its loss. As the number of training rounds increases, the prior information will become more and more reliable, so the present invention sets this weight to increase linearly. Its loss L p for:

[0071]

[0072] Where CE(·) represents the cross entropy loss, c prior represents the area where changes occurred in the prior map, y p represents the label of the pseudo sample set, Represents the predicted value.

[0073] (3) Real sample learning branch.

[0074] This branch is trained in parallel with the pseudo sample learning branch. They share network parameters and are used for training on real datasets. This branch is mainly used to consolidate the knowledge learned by the pseudo sample branch. At the same time, pseudo samples are used to randomly mix some areas to enhance the generalization of training. This branch uses the prior change map of full-image reasoning as a label and uses cross entropy loss for supervision. This branch uses arbitrary patches for training. Its loss L r for:

[0075]

[0076] in is the predicted value of the branch, y r is the prior label.

[0077] (4) Contrastive feature learning.

[0078] As attached Figure 2 The network model of the present invention is a dual-branch U-Net, which is downsampled only once and has a jump connection, so the spatial information can be greatly retained. In order for the network to better distinguish between the change class and the unchanged class, the present invention uses a contrast loss to constrain the distance between different branches and different categories of features. The feature constraint is only applied before the decoder, so that the contrast loss can assist the decoder in classification.

[0079] Specifically, the encoding results are masked and pooled according to the labels, and each training branch can obtain two feature vectors, one changed and one unchanged. The distance metric function for the features uses cosine similarity, and the four feature vectors can obtain the following four distance relationships.

[0080]

[0081] Where D(·,·) represents the cosine similarity, represents the intra-class similarity of unchanged regions, represents the intra-class similarity of the change region, represents the inter-class difference of the pseudo sample learning branch, Represents the inter-class diversity of the real sample learning branch.

[0082] At the same time, the present invention maintains a prototype vector for the change class and the unchanged class respectively, and the new feature vector obtained in each iteration will update the prototype vector in an average manner.

[0083] (5) Total training loss.

[0084] The total training loss L is as follows,

[0085]

[0086]

[0087] Since the knowledge learned by the pseudo-samples is not accurate at the beginning, in order to prevent misleading the learning of real samples, its weight β is also dynamically increased.

[0088] (6) Reasoning process

[0089] After several trainings, the dual-time images are input into the two encoders of the network respectively, and finally the decoder outputs the change probability map p1. p1 is binarized with a threshold of 0.5. The new change map will update the prior change map, and training will continue until the specified number of iterations is reached.

[0090] After training, the cosine similarity is calculated and normalized using the changed and unchanged prototype vectors and the features encoded by the encoder in the inference process to obtain p2 and p3. Finally, the three probability maps are fused using the weighted fusion method:

[0091] p= λ1 p1+λ2 p2+λ3(1- p3) (7)

[0092] p also uses 0.5 for threshold segmentation to obtain the final detection binary image.

[0093] Example:

[0094] The specific implementation process is as follows Figure 4 The detailed description is as follows:

[0095] (1) Build a network structure.

[0096] The present invention uses a dual-branch U-Net network, and its specific network data flow is as shown in the attached Figure 2 As shown in the figure, the decoder is a dual-branch structure that does not share parameters. It is used to extract image features in both time phases. The two encoders have the same structure. The decoder only performs downsampling once. In order to improve the generalization ability of the network, Dropout is added to the convolution module. The features of each level of the two encoders are subtracted and the absolute value is taken, and then input into the decoder. The first set of feature maps is input into the decoder through a jump connection. This process remains consistent with the original U-Net. The structure of the encoder and decoder is shown in the figure. Figure 3 In addition, the last set of feature maps will be pooled with the label as the mask to obtain two feature vectors: one representing the change and the other representing the invariant. Since the last set of feature maps has undergone a downsampling compared to the label size, the label is first reduced to the size of the feature map using bilinear interpolation during mask pooling.

[0097] (2) Generate a clustering graph and build a paste library.

[0098] After selecting the phase, the C-Means clustering algorithm is used to classify the image. In this embodiment, category C is 5. Since the C-Means clustering algorithm does not consider the spatial information of the image, especially the serious noise interference in the SAR image, the classification results are very scattered. Considering the local consistency of the category, the present invention uses median filtering to filter the classification map. In this embodiment, the filter window size K is 7.

[0099] Next, a sliding window search is performed on the classification map. In this embodiment, the window size is set to be smaller than patch_size (the size of the patch segmentation): patch_size / 8, patch_size / 4, and patch_size / 2. For each window, if the proportion of a category in the classification map exceeds the threshold T bank , then add it to the corresponding category in the paste library. In this embodiment, T bank After searching the complete graph, there will be C categories in the paste library, and each category has multiple paste sheets of multiple sizes.

[0100] (3) Initialize the prior graph and divide the training samples.

[0101] a priori Figure 1 The first function is to filter unchanged patches to improve the quality of pseudo samples. Since the unchanged areas account for the majority, the prior map is initialized to all 0s, indicating that nothing has changed. In subsequent training, the prior map will be updated. The prior map and the dual-time image are simultaneously segmented according to the given patch size and step size to obtain all patches. Different data allows different patch sizes and step sizes.

[0102] (4) Generate pseudo sample training set.

[0103] The basis for filtering unchanged patches is the prior map of the patch. When the number of changed pixels is less than T unch It is considered that the patch has not changed. In this embodiment, T unch The unchanged patch will be pasted to generate the pseudo sample.

[0104] The present invention uses a method of looping a fixed number of times to find the target area on the unchanged patch. Each loop will randomly generate the shape, position and size of the target. If the category map of a certain area meets the requirement that a certain category accounts for more than T pasted It is considered to be a region to be pasted that meets the requirements. In this embodiment, the threshold is a linear growth value in the range of 0.8-0.85, that is, it increases with the rounds of training. In this embodiment, the maximum number of cycles is set to 150.

[0105] For the area to be pasted, its average pixel value is calculated through the classification map, and then compared with the C category centers, and the c categories farthest away are selected. In this embodiment, c is 3, and then one is randomly selected from the c categories in the paste library for pasting, and the prior change map of the patch is copied and the pasted area is set to 1 to obtain the label. If there is no area to be pasted that meets the requirements at the end of the cycle, no pasting is performed.

[0106] At this point, the pasted phase, another phase and the constructed label constitute a pseudo sample, and the same operation is performed on each unchanged patch to obtain a pseudo sample set.

[0107] (5) Construct a real sample training set.

[0108] Since the real samples and pseudo samples are trained together, and the real samples come from all patches, it is necessary to randomly select patches from all patches that are equal to the number of pseudo sample sets. In order for the real samples to learn additional generalization knowledge, the present invention uses the strong data enhancement method of Cut-Mix to paste the random areas in the corresponding pseudo samples into the real samples. This process pastes two phases and labels at the same time.

[0109] (6) Dual-branch training network.

[0110] When training the network, both pseudo samples and real samples are input. To prevent the network from overfitting to pseudo samples and improve the generalization performance of the network, all training samples are randomly subjected to conventional data augmentation, including random flipping, random center cropping, Gaussian blurring, or no augmentation. Next, the two learning branches construct the loss function according to formula (2) and formula (3), respectively. In formula (2), e is the current round, e max is the maximum number of training rounds. During the training process, there may be a problem that the feature difference between the changed area and the unchanged area of ​​the pseudo sample is not large enough, and the learning difficulty is high. In order to assist the above two supervision tasks, the present invention constructs a contrast loss through four intrinsic relationships between two changed vectors and two unchanged vectors, namely: intra-class similarity of unchanged areas, intra-class similarity of unchanged areas, inter-class difference of pseudo sample learning branches, and inter-class difference of real sample learning branches. Finally, the overall loss calculation and back propagation are performed according to formula (6). Similarly, in formula (6)

[0111] (7) Update the prior graph and reconstruct the sample.

[0112] After several rounds of training, a binary change map is obtained according to the process of the invention content (6) and used to update the prior map. In this embodiment, the prior map is updated after 5 rounds of training. Since the selection of unchanged patches, the labels of pseudo samples and real samples in the above process all depend on the prior map, the samples need to be regenerated after the prior map is updated.

[0113] (8) Get the final change graph.

[0114] The training is terminated after the specified rounds of training. In this embodiment, 100 rounds of training are performed. Then, according to the method described in the invention content (6), two feature distance graphs p2 and p3 are obtained by measuring the maintained prototype vector, and the predicted p1 is added. The final change probability graph is obtained by simple weighted fusion of formula (7), and the final change graph is obtained by thresholding with 0.5. In this embodiment, the weights λ1, λ2, and λ3 of the three are 0.7, 0.2, and 0.1 respectively.

Claims

1. An unsupervised SAR and optical image change detection method based on pseudo sample learning, characterized in that: The steps include: Step 1: Construct a pseudo sample set; Step 1-1: Build the paste library; Perform unsupervised classification on the selected phase and divide the pixels of the entire image into C categories; The classification value is filtered using a median filter with a filter kernel size of K, and then the entire image is searched in a sliding window manner. Each position includes three window sizes: large, medium, and small. As long as a category in the window is greater than the threshold T bank It is added to the corresponding library, and finally a paste library with C categories is obtained, each category includes multiple paste pieces of different sizes; Step 1-2: Screening samples; The dataset is divided into patches as basic training images, and the number of changed pixels in the prior change map is calculated to see whether it is less than the threshold T. unch Screen the patches used to construct the pseudo sample set; Step 1-3: Select the pasting area and pasting piece; Randomly select the position, size, and shape of the paste to generate the paste area, and select the category consistency within the paste area that exceeds the threshold T pasted area; Calculate the pixel center in the pasting area and select any pasting patch in the c categories farthest from it; After pasting the patch, the corresponding position of the prior change map will also be set to 1 as the change class, thus obtaining the label of the pseudo sample; Step 2: Construct a real sample training set; Randomly select patches from all patches that are equal to the number of pseudo samples, and use the data augmentation method of Cut-Mix to paste the random regions in the corresponding pseudo samples into the real samples, and paste the two phases and labels at the same time; Step 3: Build the network structure; A dual-branch U-Net network is used, in which the decoder is a dual-branch structure without shared parameters, which is used to extract image features of the two phases. The two encoders have the same structure, and the decoder only performs downsampling once; Dropout is added to the convolution module; the features of each level of the two encoders are subtracted and the absolute value is taken, and then input into the decoder, where the first set of feature maps is input into the decoder through a jump connection, and this process remains consistent with U-Net; the last set of feature maps is pooled with the label as a mask to obtain two feature vectors: representing changes and invariance respectively; when masking pooling, bilinear interpolation is used to reduce the label to the size of the feature map; Step 4: Pseudo sample learning branch; In the pseudo sample learning branch, a weight decay pseudo sample learning branch loss is used, and the weight is set to increase linearly. Its loss L p for: Where CE(·) represents the cross entropy loss, c prior represents the area where changes occurred in the prior map, y p represents the label of the pseudo sample set, represents the predicted value; α represents the coefficient; i and j represent pixels in the corresponding area; Step 5: Real sample learning branch; The real sample learning branch and the pseudo sample learning branch are trained in parallel and share network parameters; The real sample learning branch uses the prior change map of the whole image reasoning as a label and uses the cross entropy loss for supervision; this branch is trained using any patch, and its loss L r for: where y r is the prior label; Step 6: Contrastive feature learning; Use contrast loss to constrain the distance between different branches and different categories of features; The encoding results are masked and pooled according to the labels, and each training branch obtains two feature vectors, one changed and one unchanged. The cosine similarity is used as the distance metric function for the features, and the four feature vectors can obtain the following four distance relationships. Where D(·,·) represents the cosine similarity, represents the intra-class similarity of unchanged regions, represents the intra-class similarity of the change region, represents the inter-class difference of the pseudo sample learning branch, Indicates the inter-class differences of the real sample learning branch; A prototype vector is maintained for each changing class and the same class respectively. The new feature vector obtained in each iteration will update the prototype vector by averaging. Step 7: The total training loss L is as follows; Among them, β is the weight; Step 8: Reasoning process; After several trainings, the dual-time images are input into the two encoders of the network respectively, and finally the decoder outputs the change probability map p1; p1 is binarized with a threshold of 0.5, and the new change map will update the prior change map, and continue training until the specified number of iterations is reached; After training, the cosine similarity is calculated and normalized using the changed and unchanged prototype vectors and the features encoded by the encoder in the inference process to obtain p2 and p3. Finally, the three probability maps are fused using the weighted fusion method: p= λ1 p1+λ2 p2+λ3(1- p3) (7) Among them, λ1, λ2, and λ3 are weight coefficients respectively; The threshold value of p is also 0.5 to obtain the final detection binary image.

2. The unsupervised SAR and optical image change detection method based on pseudo sample learning according to claim 1, characterized in that: Said e is the current round, e max is the maximum number of training rounds; 3. The unsupervised SAR and optical image change detection method based on pseudo sample learning according to claim 1, characterized in that: The C is 5, K is 7, and c is 3.

4. The unsupervised SAR and optical image change detection method based on pseudo sample learning according to claim 1, characterized in that: The λ1, λ2, and λ3 are 0.7, 0.2, and 0.1, respectively.

Images