A change detection method for heterogeneous remote sensing images based on self-supervised transformation

Abstract

The present invention discloses a method for detecting changes in heterogeneous remote sensing images based on self-supervised transformation, comprising the following steps: step 1): discretizing the original remote sensing image into pixels; step 2): performing superpixel segmentation and boundary integration on the discretized image in step 1), extracting and analyzing superpixel features, and obtaining labeled pixel pairs; step 3): converting the pre- and post-process heterogeneous remote sensing images into a common feature space based on the labeled pixel pairs obtained in step 2), i.e., reliable change and unchanged pixel pairs, and then obtaining two difference images through forward and backward transformation; step 4): performing evidence fusion on the difference images in step 3) and expanding training data based on a priori data; and step 5): importing the training data in step 4) into a trained classifier model to generate change detection results. The present invention has high detection accuracy, not only considering the influence of unchanged pixel pairs, but also considering the positive influence of changed pixel pairs, can well adapt to complex environments and high-noise environments, and has high robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing and pattern recognition, and in particular to a heterogeneous remote sensing image change detection method based on self-supervised conversion. Background Art

[0002] Remote sensing observation technology is capable of large-scale, long-term, and periodic detection, and change detection is a crucial technique in remote sensing observation. Remote sensing image change detection involves identifying changes in two registered remote sensing images of the same geographic area acquired at two different times. In recent years, change detection has been widely applied in various fields, such as urban planning, environmental monitoring, and disaster assessment.

[0003] Because change detection in homogeneous remote sensing images places high demands on both the satellite sensors and the ambient conditions during image acquisition, it cannot adapt to the complex environmental influences of different time periods. However, heterogeneous remote sensing image change detection methods effectively address this limitation. Existing heterogeneous remote sensing image change detection methods are mainly categorized as classification-based, deep learning-based, and similarity-based methods. Classification-based methods cannot effectively avoid the interference of noise, making it difficult to accurately classify images. Deep learning-based methods require supervision from a large amount of labeled training data, which is difficult to obtain in a timely manner in emergency scenarios or when noise levels are very high, thus failing to provide high-precision change detection results. Similarity-based methods learn only under the supervision of unchanged training samples, thus only considering unchanged pixel pairs and ignoring the positive impact of changed pixel pairs. In summary, the main shortcomings of existing heterogeneous remote sensing image change detection methods are as follows: 1) they cannot avoid the interference of noise, making it difficult to accurately classify images; 2) they have low detection accuracy and cannot adapt to complex and noisy environments; and 3) they ignore the positive impact of changed pixel pairs, considering only the influence of unchanged pixel pairs. Summary of the Invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a heterogeneous remote sensing image change detection method based on self-supervised transformation. The method has high detection accuracy, not only considers the influence of unchanged pixel pairs, but also considers the positive influence of changed pixel pairs, can adapt well to complex environments and high-noise environments, and has high robustness.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A method for heterogeneous remote sensing image change detection based on self-supervised transformation includes the following steps:

[0007] Step 1): Discretize the original remote sensing image into pixels;

[0008] Step 2): Perform superpixel segmentation and boundary integration on the discretized image in step 1, extract and analyze superpixel features, and obtain labeled pixel pairs;

[0009] Step 3): Based on the labeled pixel pairs obtained in step 2), i.e., the reliable change and unchanged pixel pairs, the before-and-after heterogeneous remote sensing images are converted into a common feature space, and then two difference images are obtained through forward and backward transformations;

[0010] Step 4): perform evidence fusion on the two difference images in step 3) and expand them based on the prior to obtain training data;

[0011] Step 5): Import the training data from step 4) into the training classifier model to generate change detection results.

[0012] The specific steps of step 1) are:

[0013] (1) The K-means algorithm is used to cluster pixels. When calculating the distance between the pixel samples within the class and the cluster center pixel, the spatial position information of the pixel is also taken into account. The weighting factor η is introduced to fuse the distance between the pixel samples within the class and the cluster center pixel based on the spectral and spatial position information. Therefore, the distance between the i-th pixel in the image and the cluster center ω is c The distance between them can be expressed as:

[0014]

[0015] in and Represents the distance between the spectrum and the spatial position information, η is as close to 1 as possible, and then the remote sensing image is divided into C classes by K-means clustering, called ω1, ω2,…, ω C ;

[0016] (2) Calculate the deviation distance between the pixel sample within the class and the cluster center pixel, and record the lth channel of pixel i as X il , calculate its difference with the cluster center ω c The absolute value deviation distance δ il As shown below:

[0017]

[0018] in,

[0019]

[0020] And |·| is the symbol of absolute value operation, M represents ω c The number of pixels in ;

[0021] (3) Calculate the correction value of the pixel spectral characteristics and calculate the absolute deviation distance δil The distance from the midpoint of the data distribution range, X il Correction value It is given by:

[0022]

[0023] in,

[0024]

[0025] and and Represents all pixels δ `l The maximum and minimum values of ;

[0026] (4) Updating discrete pixel values: The method for updating the discretized pixel value of each pixel is expressed as follows:

[0027]

[0028] Among them, X il It refers to the pixel X in the lth discretization channel i To discretize the image pixels, in order to obtain the optimized discretized image, two methods are used to stop the iteration. The first is to set the maximum number of iterations Z, and the other is to calculate the average discreteness.

[0029] The step 2) is specifically as follows:

[0030] (1) Superpixel segmentation and boundary integration: First, the discretized image in step 1) is segmented into superpixels to obtain highly uniform superpixel image blocks and highlight the boundaries of the ground objects. Secondly, a simple boundary integration strategy is used to generate superpixel pairs with consistent boundaries, that is, consistent segmentation results.

[0031] (2) Superpixel feature extraction and analysis: Using superpixels as the smallest image analysis unit, some labeled pixel pairs are automatically mined by analyzing the differences in the spectral and texture features of superpixels;

[0032] A set of spectral feature difference mappings of different operators It is derived from the following formula:

[0033]

[0034] where ‖·‖ is the Euclidean distance, is the image T t The set of all pixels in the lth channel contained in the jth superpixel, f (O) (·) represents the corresponding operation of the operator;

[0035] Texture feature difference mapping set of different operators It is derived from the following formula:

[0036]

[0037] where f(·) is the mean function, is the image T t Gray-level co-occurrence matrix (GLCM) feature map of the j-th superpixel in ; get the superpixel feature difference atlas Then a set of thresholds is used to identify The six sub-difference maps are: the texture GLCM mean and standard deviation feature difference map of the original image, the spectral mean, median, and standard deviation feature difference map of the original image, and the spectral mean feature difference map of the discretized image. A simple voting rule is then used to determine whether a superpixel has changed. The larger the feature difference, the more likely it is a changed superpixel. Conversely, the smaller the feature difference, the more likely it is an unchanged superpixel. The simple voting mechanism can be expressed as:

[0038]

[0039] Among them, σ up and σ low are the common upper and lower limits of different sub-difference maps, close to 1 and 0 respectively, σ up = 0.9 and σ low =0.1;

[0040] For each superpixel, there are s1+s2=6 votes to determine whether the superpixel belongs to the set of changed pixels. When the label corresponding to the superpixel obtains all votes, that is, 6 votes, it is considered a reliable label, which can be expressed as:

[0041]

[0042] in It is a super pixel pair S j The corresponding spatial position of C and are the sets of changed and unchanged pixel pairs respectively, u represents the set of other uncertain pixel pairs, and finally, the changed and unchanged labeled pixel pairs are obtained.

[0043] In step 3), for pixel X i ∈T1 (T1 is the prior remote sensing image), and its K nearest neighbors are defined as y1,…,y K , is the pixel point obtained from the change area in image T1, and the K nearest neighbors corresponding to the pixel points in image T2 are y′1,…,y′ K , these corresponding points provide multi-value estimates for their mapped pixels, denoted as X i,1→2, when the conversion medium is a changing area, X i The mapped pixel in another feature space is represented as:

[0044]

[0045]

[0046] Among them, λ k is the weight factor of the Kth nearest neighbor in image T1. Noise pixels are usually closer to X i Very far, so λ k It is inversely proportional to the distance to the neighbor, that is, the smaller the distance, the greater the contribution of the neighbor. Similarly, by reverse conversion to X′ in image T2 i Obtained the mapped pixel map X′ i,2→1 Obviously, when the conversion medium is a changing area, the corresponding conversion mapping pixels should be and

[0047] Then, the i-th pixel pair {X i ,X′ i The difference of} can be expressed as:

[0048]

[0049] Among them, ξ i and Represents the corresponding difference between the changed and unchanged pixels as the conversion medium. Finally, the difference images D and

[0050] The step 4) is specifically as follows:

[0051] (1) Evidence fusion expands training data.

[0052] Step 3) The pixel pair {X i ,X′ i The average distance between} and its neighboring pixel pairs obtained by converting reliable changed pixel pairs is expressed as:

[0053]

[0054] Among them, K represents the number of neighbors, y k and y′ k are X and X′ in images T1 and T2 respectively i The nearest neighbor of {X i ,X′ iThe average distance d between the neighboring pixel pairs obtained by transforming the reliable unchanged pixel pairs i , distance d i The smaller the value, the more reliable the evidence is, and the discount factor α is defined as i The smaller , the discount factor is defined as:

[0055]

[0056] m i The basic trust function of is given by:

[0057]

[0058] Further, Also through its own discount factor Discount, through the evidence fusion rule and Combined and The difference between them is denoted as μ i , making It is used to measure the reliability of the change and invariance of uncertain pixel pairs and set an upper limit and lower limit After sorting the μ of all evidences from small to large, and Take the first 1% and last 1% values respectively, if μ i Higher and Then it means pixel pair {X i ,X′i} as a changed pixel pair has a significantly greater confidence than an unchanged one, in which case it is classified as a changed pixel pair. In addition, if μ i Smaller and Pixel pair {X i ,X′ i} will be classified as an unchanged pixel pair, but if This indicates and There may not be a significant difference, and the pixel pair {X i ,X′ i} is considered an uncertain pixel pair;

[0059] (2) Expand training data based on priors.

[0060] Calculate the changed pixel pair set C and the unchanged pixel pair set respectively The average value of all pixels in is then used as the segmentation threshold and Pixel pair {X i ,X′ iAs mentioned above, the smaller the difference image value is, the greater the probability that the pixel pair at this position belongs to the same class as the transformation medium, which can be expressed as:

[0061]

[0062] Among them, ξ i and Respectively represent the pixel pairs corresponding to the changing and unchanged pixel pairs as the conversion medium {X i ,X′ i}.

[0063] Beneficial effects of the present invention:

[0064] This paper proposes a heterogeneous image change detection method based on self-supervised transformation. First, the performance of superpixel segmentation is improved by discretizing image pixels. Second, superpixels are used as the minimum image analysis unit, fully utilizing the advantage of spectral and texture features in superpixels to more reasonably describe complex geographic spatial information. Then, evidence fusion and prior-based methods are used to greatly expand the training data based on reliability. Finally, a random forest classifier model is trained based on a large amount of training data to generate high-precision binary change detection result maps.

[0065] The method proposed in the present invention has high detection accuracy and high robustness, and can still provide high-precision change detection results in complex environments and high-noise environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 Flowchart of the present invention.

[0067] Figure 2 It is a specific flow chart of the present invention.

[0068] Figure 3 Flowchart of the image pixel discretization algorithm.

[0069] Figure 4 This is a binary change detection result diagram of the present invention. DETAILED DESCRIPTION

[0070] The present invention will be described in further detail below with reference to the accompanying drawings.

[0071] Reference Figure 1 and Figure 2 , the specific steps of the present invention are described in detail.

[0072] Step 1: Discretize the original remote sensing image into pixels.

[0073] Step 1 addresses the adverse effects of noisy pixels by minimizing differences within the same class and enhancing differences between pixels of different classes. This is primarily because the spectral differences between images of different objects in reality are minimal, resulting in inherent speckle noise, which adversely affects superpixel segmentation. Pixel discretization effectively mitigates noise interference, thereby improving superpixel segmentation performance. Figure 3 This is the flow chart of the image pixel discretization algorithm, and its main steps are:

[0074] (1) K-means algorithm is used to cluster pixels. In particular, the spatial position information of the pixel is also taken into account when calculating the distance between the pixel sample within the class and the cluster center pixel. The weighting factor η is introduced to fuse the distance between the pixel sample within the class and the cluster center pixel based on the spectral and spatial position information. Therefore, the distance between the i-th pixel in the image and the cluster center ω is c The distance between them can be expressed as:

[0075]

[0076] in and Represents the distance between the spectrum and the spatial position information. Since the spectral characteristics of the image pixels play a key role in the clustering process, η is as close to 1 as possible. Then, the remote sensing image is divided into C classes by K-means clustering, called ω1, ω2, ..., ω C .

[0077] (2) Calculate the deviation distance between the pixel samples within the class and the cluster center pixel. For pixel i, the lth channel is recorded as X il , calculate its difference with the cluster center ω c The absolute value deviation distance δ il As shown below:

[0078]

[0079] in,

[0080]

[0081] And |·| is the symbol of absolute value operation, M represents ω c The number of pixels in .

[0082] (3) Calculate the correction value of the pixel spectral characteristics. The key to the image pixel discretization algorithm is how to control the correction amount of the pixel spectral characteristics in a cluster. In order to measure the degree of correction, the absolute deviation distance δ is calculated. il The distance from the midpoint of the data distribution range, so X il Correction value It is given by:

[0083]

[0084] in,

[0085]

[0086] and and Represents all pixels δ `l The maximum and minimum values of .

[0087] (4) Update the discrete pixel value. The method of updating the discretized pixel value of each pixel can be expressed as follows:

[0088]

[0089] Among them, X il It refers to the pixel X in the lth discretization channel i Discretize the image pixels. By doing so, the distribution difference of the lth channel of pixels in the same cluster is reduced, and the distribution difference of the lth channel of pixels in different clusters is enhanced. In order to obtain an optimized discretized image, two ways are used to stop the iteration. The first is to set the maximum number of iterations Z, and the other is to calculate the average discreteness

[0090] Step 2: Superpixel feature extraction and analysis.

[0091] The second step is to perform superpixel segmentation and boundary integration on the discretized image, and then perform superpixel feature extraction and analysis based on spectral and texture features. The main reason for using superpixel feature extraction and analysis is that compared with single pixels, superpixel features (such as spectral and texture features) can more reasonably describe complex geographic spatial information. The main steps of superpixel feature extraction and analysis are as follows:

[0092] (1) Superpixel segmentation and boundary integration: First, a simple linear iterative clustering method is used to segment the discretized image into superpixels to obtain highly uniform superpixel image blocks and highlight the boundaries of the ground objects; secondly, a simple boundary integration strategy is used to generate superpixel pairs with consistent boundaries, that is, consistent segmentation results.

[0093] (2) Superpixel feature extraction and analysis: Taking superpixels as the smallest image analysis unit, some labeled pixel pairs are automatically mined by analyzing the differences in the spectral and texture features of superpixels.

[0094] A set of spectral feature difference mappings of different operators It is derived from the following formula:

[0095]

[0096] where ‖·‖ is the Euclidean distance, is the image T t The set of all pixels in the lth channel contained in the jth superpixel, f (O) (·) represents the corresponding operation of the operator.

[0097] Texture feature difference mapping set of different operators It is derived from the following formula:

[0098]

[0099] where f(·) is the mean function, is the image T t Gray-level co-occurrence matrix (GLCM) feature map of the j-th superpixel in .

[0100] Through the above calculations, we can get the superpixel feature difference atlas Then a set of thresholds is used to identify A simple voting rule is used to determine whether a superpixel has changed based on the changed and unchanged areas of the six sub-difference maps. The six sub-difference maps are: the texture GLCM mean and standard deviation feature difference map of the original image, the spectral mean, median, and standard deviation feature difference map of the original image, and the spectral mean feature difference map of the pixel discretized image. Generally speaking, the larger the feature difference value, the more likely it is a changed superpixel. Conversely, the smaller the feature difference value, the more likely it is an unchanged superpixel. This simple voting mechanism can be expressed as:

[0101]

[0102] Among them, σ up and σ low are the common upper and lower limits of different sub-difference maps, which are close to 1 and 0 respectively. In practice, it is recommended that σ up = 0.9 and σ low =0.1. The closer these two values are, the more reliable the superpixel labels are, but the fewer superpixel labels there are. Obviously, these two thresholds are used to balance the reliability and quantity of superpixels. In this invention, we focus more on the reliability of superpixels, which means that these two thresholds are very close to the two extremes.

[0103] For each superpixel, there are s1+s2=6 votes to determine whether the superpixel belongs to the set of changed pixels. To further ensure the reliability of superpixel labels, the label corresponding to the superpixel is considered reliable only when it obtains all the votes, that is, 6 votes, which can be expressed as:

[0104]

[0105] in It is a super pixel pair S j The corresponding spatial position of C and are the sets of changed and unchanged pixel pairs, and u represents the set of other uncertain pixel pairs. Finally, we obtain some highly reliable changed and unchanged labeled pixel pairs.

[0106] Step 3: Multi-media bidirectional pixel conversion.

[0107] The step three converts the before-and-after heterogeneous remote sensing images into a common feature space based on the reliable labeled pixel pairs in step two, i.e., the reliable changed and unchanged pixel pairs, and then obtains two different difference images through forward and backward changes, which provides complementary information for change detection and thus improves the accuracy of change detection.

[0108] For Pixel X i ∈T1 (T1 is the prior remote sensing image), and its K nearest neighbors are defined as y1,…,y K , which are pixels obtained from the change area in image T1. These K nearest neighbors correspond to the pixels y′1,…,y′ in image T2. K , these corresponding points provide multi-value estimates for their mapped pixels, denoted as X i,1→2 When the conversion medium is a changing area, X i The mapped pixel in another feature space is represented as:

[0109]

[0110]

[0111] Among them, λ k is the weight factor of the Kth nearest neighbor in image T1. Noise pixels are usually closer to X i Very far, so λ k is inversely proportional to the distance to the nearest neighbor. That is, the smaller the distance, the greater the contribution of the nearest neighbor. Similarly, by reverse conversion to X′ in image T2 i Obtained the mapped pixel map X′ i,2→1 Obviously, when the conversion medium is a changing area, we can appropriately convert the mapped pixels accordingly. and

[0112] Then, the i-th pixel pair {X i ,X′ i The difference of} can be expressed as:

[0113]

[0114] Among them, ξ i and Denotes the corresponding difference between the changed and unchanged pixels as the conversion medium. Finally, the difference images D and Although the final change map can be generated directly from the difference image, there are still some pixels that are difficult to distinguish whether they have changed. Therefore, it is necessary to further design a more cautious strategy to perform change detection for uncertain pixel pairs.

[0115] Step 4: Evidence fusion and expansion of training data based on priors.

[0116] Step 4 first uses the evidence fusion method to perform the first data expansion on the training data, and then uses the prior-based method to perform the second data expansion on the training data, providing sufficient data preparation for training a high-performance classifier model. The main steps are:

[0117] (1) Evidence fusion expands training data.

[0118] Pixel pair {X i ,X′ i The average distance between {X} and its neighbors after multi-media bidirectional transformation is intended to evaluate the reliability of evidence. i ,X′ i The average distance between} and its neighboring pixel pairs obtained by converting reliable changed pixel pairs can be expressed as:

[0119]

[0120] Among them, K represents the number of neighbors, y k and y′ k are X and X′ in images T1 and T2 respectively i Then, the same method can be used to obtain the pixel pair {X i ,X′ i The average distance d between the neighboring pixel pairs obtained by transforming the reliable unchanged pixel pairs i . Distance d i The smaller the value, the more reliable the evidence is, and the discount factor α is defined as i The smaller , the discount factor is defined as:

[0121]

[0122] m i The basic trust function of is given by:

[0123]

[0124] Further, Also through its own discount factor Discount, through the evidence fusion rule and Combined and The difference between them is denoted as μ i , making It is used to measure the reliability of the change and invariance of uncertain pixel pairs. Set an upper limit and lower limit After sorting all the evidence from smallest to largest, and Take the first 1% and last 1% values respectively. If μ i Higher and Then it means pixel pair {X i ,X′ i} as the confidence of the pixel pair that has changed is significantly greater than the confidence of the pixel pair that has not changed, in which case it is classified as a pixel pair that has changed. In addition, if μ i Smaller and Pixel pair {X i ,X′ i} will be classified as an unchanged pixel pair. But if This indicates and There may not be a significant difference, and the pixel pair {X i ,X′ i} are considered as uncertain pixel pairs.

[0125] (2) Expand training data based on priors.

[0126] Based on prior knowledge, the training data is further expanded to calculate the set of changed pixel pairs C and the set of unchanged pixel pairs respectively. The average value of all pixels in is then used as the segmentation threshold and Pixel pair {X i ,X′ i As mentioned above, the smaller the difference image value, the greater the probability that the pixel pair at that position belongs to the same class as the transformation medium, which can be expressed as:

[0127]

[0128] Among them, ξ i and Respectively represent the pixel pairs corresponding to the changing and unchanged pixel pairs as the conversion medium {X i ,X′ i}, which is described in detail in step three.

[0129] Step 5: Train the classifier model to generate change detection results.

[0130] The step five is to train the classifier model to generate change detection results. Since the purpose of change detection is to highlight the differences between pairs of images at different times, it can essentially be converted into a binary classification task. Compared with the method of clustering or thresholding the difference image, the method based on the classification model can weaken the influence of noise in the difference image and obtain a more robust change result map. Once the number of training samples is sufficient, the self-supervised classification model will play its huge advantages. Since the random forest method uses the simplicity of the decision tree and the robustness of the integration method for classification, the random forest is selected as the self-supervised classification model, and the expanded training data is input into the model, and finally a model with high detection accuracy and high robustness is obtained. Then the pre- and post-remote sensing images are input into the random forest classifier model, and finally a binary change detection result map is output.

[0131] Example: Change detection of real-world heterogeneous remote sensing images before and after the event

[0132] The effectiveness and accuracy of the proposed method are demonstrated through a set of real-world change detection experiments on heterogeneous remote sensing images before and after the event. Figure 4 (a) and Figure 4 (b) indicates that this set of images is from Gloucester, UK, with a size of 465×827, of which the prior remote sensing images Figure 4 (a) Taken in July 2006, remote sensing image after the event Figure 4 (b) Photographed in July 2007. In addition, Figure 4 (c) is the reference image of the actual changed area.

[0133] The present invention is compared with the heterogeneous remote sensing image change detection (HPT) method based on homogeneous pixel transformation. The comparison results are as follows: Figure 4 (d) and 4(e), Figure 4 (d) is the change detection result graph generated by the HPT method. Figure 4(e) is a graph of the change detection results of the method of the present invention, in which white pixels represent changed pixels and black pixels represent unchanged pixels. Although the HPT method can detect most of the areas that have changed, a large number of false alarm pixels also appear (actually unchanged pixels but mistakenly detected as changed pixels). The graph of the change detection results obtained by the method of the present invention is very clear, and the number of missed detection and false alarm pixels in the graph is very small. This is because for heterogeneous remote sensing image change detection, the super-pixel feature extraction and analysis and multi-media bidirectional pixel conversion ideas proposed in the present invention reduce the negative impact of noise points on the detection results, making the detection results more accurate. Table 1 is an evaluation of the detection result accuracy of the two change detection methods, where OA is the overall accuracy, R m is the missed detection rate, R f The false alarm rate is the F1 score, which is an indicator to measure the accuracy of the binary classification model. The Kappa (KC) coefficient can fully reflect the accuracy of change detection. The heterogeneous remote sensing image change detection results obtained by the method of the present invention are compared with the reference image, and the missed detection rate R m and false alarm rate R f The F1 score is significantly improved, and the performance of the method proposed in the present invention in terms of Kappa coefficient is better than that of the HPT method, which is improved to 75.04%.

[0134] Table 1 Accuracy evaluation of change detection results of different methods

[0135] method OA <![CDATA[R m ]]> <![CDATA[R f ]]> <![CDATA[F1]]> KC HPT Method 0.8879 0.3672 0.6867 0.4191 0.3647 Method of the present invention 0.9725 0.2992 0.1582 0.7648 0.7504

[0136] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention's description and drawings, or directly or indirectly applied to other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for change detection in heterogeneous remote sensing images based on self-supervised transformation, characterized in that: The following steps are included: Step 1): Discretize the original remote sensing image into pixels; Step 2): Perform superpixel segmentation and boundary integration on the discretized image in step 1, extract and analyze superpixel features, and obtain labeled pixel pairs; Step 3): Based on the labeled pixel pairs obtained in step 2), i.e., the reliable change and unchanged pixel pairs, the before-and-after heterogeneous remote sensing images are converted into a common feature space, and then two difference images are obtained through forward and backward transformations; Step 4): perform evidence fusion on the two difference images in step 3) and expand them based on the prior to obtain training data; Step 5): Import the training data from step 4) into the training classifier model to generate change detection results; The specific steps of step 1) are: (1) The K-means algorithm is used to cluster pixels. When calculating the distance between the pixel samples within the class and the cluster center pixel, the spatial position information of the pixel is also taken into account. The weighting factor η is introduced to fuse the distance between the pixel samples within the class and the cluster center pixel based on the spectral and spatial position information. Therefore, the distance between the i-th pixel in the image and the cluster center ω is c The distance between them is expressed as: in and Represents the distance between the spectrum and the spatial position information, η approaches 1, and then the remote sensing image is divided into C classes by K-means clustering, called ω1, ω2,…, ω C ; (2) Calculate the deviation distance between the sample point and the cluster center pixel, and record the lth channel of pixel i as X il , calculate its difference with the cluster center ω c The absolute value deviation distance δ il As shown below: in, And |·| is the symbol of absolute value operation, M represents ω c The number of pixels in ; (3) Calculate the correction value of the pixel spectral characteristics and calculate the absolute deviation distance δ il The distance from the midpoint of the data distribution range, so X il Correction value It is given by: in, and and Represents all pixels δ `l The maximum and minimum values of ; (4) Updating discrete pixel values: The method for updating the discretized pixel value of each pixel is expressed as follows: Among them, X il It refers to the pixel X in the lth discretization channel i To discretize the image pixels, in order to obtain the optimized discretized image, two methods are used to stop the iteration. The first is to set the maximum number of iterations Z, and the other is to calculate the average discreteness. The step 2) is specifically as follows: (1) Superpixel segmentation and boundary integration: First, the discretized image in step 1) is segmented into superpixels to obtain highly uniform superpixel image blocks and highlight the boundaries of the ground objects. Secondly, a simple boundary integration strategy is used to generate superpixel pairs with consistent boundaries, that is, consistent segmentation results. (2) Superpixel feature extraction and analysis: Using superpixels as the smallest image analysis unit, some labeled pixel pairs are automatically mined by analyzing the differences in the spectral and texture features of superpixels; A set of spectral feature difference mappings of different operators It is derived from the following formula: where ‖·‖ is the Euclidean distance, is the image T t The set of all pixels in the lth channel contained in the jth superpixel, f (O) (·) represents the corresponding operation of the operator; Texture feature difference mapping set of different operators It is derived from the following formula: where f(·) is the mean function, is the image T t Gray-level co-occurrence matrix (GLCM) feature map of the j-th superpixel in ; get the superpixel feature difference atlas Then a set of thresholds is used to identify The six sub-difference maps are: the texture GLCM mean and standard deviation feature difference map of the original image, the spectral mean, median, and standard deviation feature difference map of the original image, and the spectral mean feature difference map of the image after pixel discretization. Then, a simple voting rule is used to determine whether the superpixel has changed. The larger the feature difference, the superpixel has changed. Conversely, the smaller the feature difference, the superpixel has not changed. The simple voting mechanism is expressed as: Among them, σ up and σ low are the common upper and lower limits of different sub-difference maps, close to 1 and 0 respectively, σ up = 0.9 and σ low =0.1; For each superpixel, there are s1+s2=6 votes to determine whether the superpixel belongs to the set of changed pixels. When the label corresponding to the superpixel obtains all votes, that is, 6 votes, it is considered a reliable label, which is expressed as: in It is a super pixel pair S j The corresponding spatial position of C and are the sets of changed and unchanged pixel pairs, u represents the set of other uncertain pixel pairs, and finally, the changed and unchanged labeled pixel pairs are obtained; In step 3), for pixel X i ∈T1, T1 is the prior remote sensing image, and its K nearest neighbors are defined as y1,…,y K , is the pixel point obtained from the change area in image T1, and the K nearest neighbors corresponding to the pixel points in image T2 are y′1,…,y′ K , these corresponding points provide multi-value estimates for their mapped pixels, denoted as X i,1→2 , when the conversion medium is a changing area, X i The mapped pixel in another feature space is represented as: Among them, λ k is the weight factor of the Kth nearest neighbor in image T1. Noise pixels are usually closer to X i Very far, so λ k It is inversely proportional to the distance to the neighbor, that is, the smaller the distance, the greater the contribution of the neighbor. Similarly, by reverse conversion to X′ in image T2 i Obtained the mapped pixel map X′ i,2→1 Obviously, when the conversion medium is a changing area, the corresponding conversion mapping pixels should be and Then, the i-th pixel pair {X i ,X′ i The difference of} is expressed as: Among them, ξ i and Represents the corresponding difference between the changed and unchanged pixels as the conversion medium. Finally, the difference images D and D are obtained through two different conversion media.

2. The method for heterogeneous remote sensing image change detection based on self-supervised transformation according to claim 1, characterized in that: The step 4) is specifically as follows: (1) Evidence fusion to expand training data; Step 3) The pixel pair {X i ,X′ i The average distance between} and its neighboring pixel pairs obtained by converting reliable changed pixel pairs is expressed as: Among them, K represents the number of neighbors, y k and y′ k are X in images T1 and T2 respectively. i and X′ i The nearest neighbor of {X i ,X′ i The average distance d between the neighboring pixel pairs obtained by transforming the reliable unchanged pixel pairs i , distance d i The smaller the value, the more reliable the evidence is, and the discount factor α is defined as i The smaller , the discount factor is defined as: m i The basic trust function of is given by: Further, Also through its own discount factor Discount, through the evidence fusion rule and Combined and The difference between them is denoted as μ i , making It is used to measure the reliability of the change and invariance of uncertain pixel pairs and set an upper limit and lower limit After sorting the μ of all evidences from small to large, and Take the first 1% and last 1% values respectively, if Then it means pixel pair {X i ,X′ i } As the confidence of the changed pixel pair is significantly greater than the unchanged confidence, in this case it is classified as a changed pixel pair. In addition, if Pixel pair {X i ,X′ i } will be classified as an unchanged pixel pair, but if This indicates and There is no significant difference, and the pixel pairs {X i ,X′ i } is considered an uncertain pixel pair; (2) Expanding training data based on priors; Calculate the changed pixel pair set C and the unchanged pixel pair set respectively The average value of all pixels in is then used as the segmentation threshold and Pixel pair {X i ,X′ i As mentioned above, the smaller the difference image value, the greater the probability that the corresponding pixel pair belongs to the same class as the transformation medium, which can be expressed as: Among them, ξ i and Respectively represent the pixel pairs corresponding to the changing and unchanged pixel pairs as the conversion medium {X i ,X′ i }.

Images