Remote sensing image change detection method based on unsupervised learning and storage medium

Abstract

The application discloses a remote sensing image change detection method based on unsupervised learning and a storage medium, and comprises the following steps: obtaining overlapping images of two time-series remote sensing images; unifying the sizes and affine transformation matrices of the two overlapping images; calculating the difference value images of the two unified overlapping images; performing twice sliding window filtering on the difference value images, and generating a first feature vector, a feature vector mean value and a second feature vector; performing principal component analysis on the first feature vector to obtain a feature vector space; mapping the second feature vector to the feature vector space, and subtracting the first feature vector mean value to obtain a mapped feature vector; clustering the mapped feature vector, and setting labels for each feature according to the corresponding cluster; projecting the mapped feature vector into the corresponding geographical space of the difference value image according to the unified affine transformation matrix, and generating a binary mask image according to the labels. The application has stronger generalization ability, automation degree and integration degree.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a remote sensing image change detection method based on unsupervised learning and a storage medium. Background Art

[0002] At present, there are three main methods for the change detection problem of remote sensing time series images:

[0003] 1. Extract and analyze the changed areas through visual interpretation and manual delineation of patches.

[0004] 2. Based on supervised learning, this type of method requires the production of samples and the establishment of a sample library of relevant land feature change maps. It then uses supervised learning methods such as neural networks, random forests, and support vector machines to train the model. After model optimization and hyperparameter adjustment, the final version that can be used for prediction is obtained.

[0005] 3. Based on unsupervised learning, this type of method does not require manual pattern drawing or sample production, and can complete the detection task through image processing methods such as window filtering, matrix operations and feature mapping.

[0006] However, the labor and time costs required for visual interpretation methods are too high. Since remote sensing images usually cover a wide area, if the change detection within the geographical scope of time-series images relies entirely on visual recognition, the workload is large and effective results cannot be obtained in a short time.

[0007] Supervised learning methods have similar shortcomings to visual interpretation methods. Although they can automatically implement change detection algorithms, they require professionals to spend time in the early stages to screen features, label samples, and other tasks. This process is relatively cumbersome and cannot meet the requirements of some rapid detection scenarios.

[0008] Traditional unsupervised learning methods have relatively stringent requirements for image input. The geographic scope, height and width of the preceding and subsequent images must be identical, with pixel-to-pixel correspondence and the two images also requiring the same resolution. Furthermore, window filter parameters must be manually calculated and entered into the program in advance, resulting in a certain amount of clutter in the resulting images and an unintuitive display. This suggests that traditional unsupervised learning methods suffer from weak generalization, low integration, and limited automation. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a remote sensing image change detection method and storage medium based on unsupervised learning, which can solve the problems of traditional unsupervised learning methods such as weak generalization ability, low integration, and limited automation.

[0010] In order to solve the above technical problems, the technical solution adopted by the present invention is: a remote sensing image change detection method based on unsupervised learning, comprising:

[0011] Determining the overlapping area of ​​the two time-series remote sensing images according to the affine transformation matrix of the two time-series remote sensing images to be measured, and extracting the overlapping areas of the two time-series remote sensing images respectively to obtain two overlapping images;

[0012] Unifying the sizes and affine transformation matrices of the two overlapping images and determining the side length of the filtering window;

[0013] Calculate the difference map based on the two unified overlapping images;

[0014] According to the side length of the filtering window and a first sliding step length, where the first sliding step length is the same as the side length of the filtering window, the difference image is subjected to sliding window filtering to obtain a first filtering result, and a first eigenvector and a eigenvector mean thereof are generated according to the first filtering result;

[0015] Performing principal component analysis on the first eigenvector to obtain a eigenvector space;

[0016] Performing sliding window filtering on the difference image according to the side length of the filtering window and a preset second sliding step size to obtain a second filtering result, and generating a second feature vector according to the second filtering result;

[0017] Mapping the second eigenvector to the eigenvector space, and performing difference calculation with the mean of the first eigenvector to obtain a mapped eigenvector;

[0018] Clustering the mapped feature vector to obtain a preset number of clusters, and setting a label for each feature in the mapped feature vector according to the cluster to which it belongs;

[0019] According to the affine transformation matrix of the unified overlapping image, the mapped feature vectors with labels are projected into the geographic space corresponding to the difference map in order from left to right and from top to bottom, and a binary mask map is generated according to the labels.

[0020] The present invention further provides a computer-readable storage medium having a computer program stored thereon, which implements the above-mentioned method when the program is executed by a processor.

[0021] The beneficial effects of the present invention are: the present invention can process two time-series remote sensing images with different resolutions and image sizes, and has stronger generalization ability; it can automatically determine the filtering parameters and automatically adjust the sizes of the two overlapping images, and has a higher degree of automation; it can detect change results without external intervention, thereby improving the operating efficiency of the overall change detection process and having a higher degree of integration.

[0022] Based on multispectral remote sensing image data, this invention utilizes unsupervised machine learning algorithms, combined with satellite data processing algorithms, to automatically extract overlapping regions between two sequential images and automatically detect regions of change. This invention can effectively identify changes in objects in high-resolution, medium-resolution, and low-resolution remote sensing time-series images that have undergone atmospheric and geometric corrections. Because it eliminates the need for manual annotation and the establishment of a sample database, the invention can effectively complete change detection tasks while significantly reducing task costs. In practice, it can provide users with change detection services more conveniently and efficiently. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of a remote sensing image change detection method based on unsupervised learning according to the present invention;

[0024] Figure 2 This is a flow chart of a method according to embodiment 1 of the present invention;

[0025] Figure 3 Schematic diagram of two time-series remote sensing images according to the first embodiment of the present invention;

[0026] Figure 4 This is a schematic diagram of the overlap of two time-series remote sensing images according to the first embodiment of the present invention;

[0027] Figure 5 is a schematic diagram of two unified overlapping images according to the first embodiment of the present invention;

[0028] Figure 6 Schematic diagram of a difference map according to the first embodiment of the present invention;

[0029] Figure 7 Schematic diagram of a binary mask image according to the first embodiment of the present invention;

[0030] Figure 8 Schematic diagram of a filtered binary mask image according to the first embodiment of the present invention;

[0031] Figure 9 This is a schematic diagram of change detection results of two time-series remote sensing images according to the first embodiment of the present invention;

[0032] Figure 10 This is a schematic diagram of overlapping change detection results of two time-series remote sensing images according to the first embodiment of the present invention. DETAILED DESCRIPTION

[0033] To illustrate the technical content, achieved objectives and effects of the present invention in detail, the following is a detailed description in conjunction with the embodiments and accompanying drawings.

[0034] See also Figure 1 , a remote sensing image change detection method based on unsupervised learning, comprising:

[0035] Determining the overlapping area of ​​the two time-series remote sensing images according to the affine transformation matrix of the two time-series remote sensing images to be measured, and extracting the overlapping areas of the two time-series remote sensing images respectively to obtain two overlapping images;

[0036] Unifying the sizes and affine transformation matrices of the two overlapping images and determining the side length of the filtering window;

[0037] Calculate the difference map based on the two unified overlapping images;

[0038] According to the side length of the filtering window and a first sliding step length, where the first sliding step length is the same as the side length of the filtering window, the difference image is subjected to sliding window filtering to obtain a first filtering result, and a first eigenvector and a eigenvector mean thereof are generated according to the first filtering result;

[0039] Performing principal component analysis on the first eigenvector to obtain a eigenvector space;

[0040] Performing sliding window filtering on the difference image according to the side length of the filtering window and a preset second sliding step size to obtain a second filtering result, and generating a second feature vector according to the second filtering result;

[0041] Mapping the second eigenvector to the eigenvector space, and performing difference calculation with the mean of the first eigenvector to obtain a mapped eigenvector;

[0042] Clustering the mapped feature vector to obtain a preset number of clusters, and setting a label for each feature in the mapped feature vector according to the cluster to which it belongs;

[0043] According to the affine transformation matrix of the unified overlapping image, the mapped feature vectors with labels are projected into the geographic space corresponding to the difference map in order from left to right and from top to bottom, and a binary mask map is generated according to the labels.

[0044] From the above description, it can be seen that the beneficial effects of the present invention are: it has stronger generalization ability, higher degree of automation and higher integration, and can effectively complete the change detection task while significantly reducing the task cost.

[0045] Furthermore, after generating the binary mask image according to the label, the method further includes:

[0046] The binary mask image is filtered by using an erosion filter operator.

[0047] From the above description, it can be seen that by using the corrosion filter operator for filtering, the inherent problems of speckle, noise and "false change" in change detection can be solved, and their adverse effects on the change detection results can be reduced to a certain extent.

[0048] Furthermore, after filtering the binary mask image by the corrosion filter operator, the method further includes:

[0049] The filtered binary mask image is format-converted to obtain a vector result image, and the vector result image is superimposed on the two time-series remote sensing images in the form of a highlighted boundary and a hollow interior.

[0050] From the above description, it can be seen that users can more clearly and intuitively understand the location and range of the image change area and the area of ​​the land corresponding to the specific change patch.

[0051] Furthermore, the overlapping area of ​​the two time-series remote sensing images to be measured is determined according to the affine transformation matrix of the two time-series remote sensing images, specifically:

[0052] Acquire two time-series remote sensing images to be measured, and perform atmospheric correction and geometric correction on the two time-series remote sensing images;

[0053] Extracting the affine transformation matrices of the two time-series remote sensing images respectively, wherein the parameters in the affine transformation matrix include longitude resolution, a first rotation coefficient, the longitude of the upper left corner, a second rotation coefficient, latitude resolution and the latitude of the upper left corner;

[0054] Determine the geographic coordinates of corner points of the two time-series remote sensing images according to the affine transformation matrix of the two time-series remote sensing images, wherein the corner points include an upper left corner point and a lower right corner point;

[0055] An overlapping area of ​​the two time-series remote sensing images is determined according to the geographic coordinates of the corner points of the two time-series remote sensing images.

[0056] Furthermore, the two time-series remote sensing images include a first image and a second image;

[0057] The step of determining the overlapping area of ​​the two time-series remote sensing images according to the geographic coordinates of the corner points of the two time-series remote sensing images is specifically as follows:

[0058] If the longitude of the upper left corner of the first image is less than the longitude of the lower right corner of the second image, the longitude of the lower right corner of the first image is greater than the longitude of the upper left corner of the second image, the latitude of the upper left corner of the first image is greater than the latitude of the lower right corner of the second image, and the latitude of the lower right corner of the first image is greater than the longitude and latitude of the upper left corner of the second image, then it is determined that the two time series remote sensing images overlap;

[0059] If the longitude of the upper left corner of the first image is greater than the longitude of the upper left corner of the second image, then the longitude of the upper left corner of the overlapping area is equal to the longitude of the upper left corner of the first image; otherwise, the longitude of the upper left corner of the overlapping area is equal to the longitude of the upper left corner of the second image;

[0060] If the latitude of the upper left corner of the first image is less than the latitude of the upper left corner of the second image, then the latitude of the upper left corner of the overlapping area is equal to the latitude of the upper left corner of the first image; otherwise, the latitude of the upper left corner of the overlapping area is equal to the latitude of the upper left corner of the second image;

[0061] If the longitude of the lower right corner of the first image is smaller than the longitude of the lower right corner of the second image, then the longitude of the lower right corner of the overlapping area is equal to the longitude of the lower right corner of the first image; otherwise, the longitude of the lower right corner of the overlapping area is equal to the longitude of the lower right corner of the second image;

[0062] If the latitude of the lower right corner point of the first image is greater than the latitude of the lower right corner point of the second image, then the latitude of the lower right corner point of the overlapping area is equal to the latitude of the lower right corner point of the first image; otherwise, the latitude of the lower right corner point of the overlapping area is equal to the latitude of the lower right corner point of the second image.

[0063] From the above description, it can be seen that the overlapping area of ​​the two images is extracted based on the affine transformation information of the two images and combined with the topological knowledge.

[0064] Furthermore, the two overlapping images are unified in size and affine transformation matrix, and the side length of the filtering window is determined, specifically:

[0065] Using the overlapping image with lower resolution among the two overlapping images as the first overlapping image, and using the other overlapping image as the second overlapping image;

[0066] Determining whether there is a common divisor within a preset value range based on the width and height of the first overlapping image;

[0067] If so, obtaining the greatest common divisor within a preset value range, using the greatest common divisor as the side length of the filtering window, and using the width and height of the first overlapping image as the new width and new height respectively;

[0068] If not, performing an upper neighborhood search on the width and height of the first overlapping image until a neighborhood value of the width and a neighborhood value of the height have a common divisor within a preset value range, and taking the maximum value of the common divisor as the side length of the filtering window, and taking a neighborhood value of the width and a neighborhood value of the height as the new width and new height respectively;

[0069] According to the width and height of the first overlapping image and the new width and height, the two overlapping images are resampled and registered by a bilinear interpolation method to obtain two unified overlapping images.

[0070] From the above description, it can be seen that by unifying the two overlapping images, the input requirements of the unsupervised learning-based method for images can be reduced, and the processing of two time-series images with different resolutions and image sizes can be achieved, thereby improving the generalization ability of change detection; by automatically determining the filtering parameters, the degree of automation of change detection is improved.

[0071] Furthermore, generating a first eigenvector and its eigenvector mean according to the first filtering result is specifically:

[0072] Arrange the first filtering results in a sliding generation order to obtain a first feature vector, where the number of rows of the first feature vector is (H×W) / (L×L) and the number of columns is L×L;

[0073] Calculate the average value of the first filtering results in the same column of the first feature vector, and obtain a feature vector mean according to the average values ​​corresponding to each column of the first feature vector, where the number of rows of the feature vector mean is 1 and the number of columns is L×L;

[0074] Wherein, H and W are the height and width of the difference image respectively, and L is the side length of the filtering window.

[0075] From the above description, it can be seen that the accuracy of change detection is improved by optimizing the calculation method of the feature vector mean.

[0076] Furthermore, the width and height of the feature vector space are both L×L; the second sliding step is 1, the number of rows of the second feature vector is H-L+1, and the number of columns is W-L+1; the number of rows of the mapped feature vector is H-L+1, and the number of columns is W-L+1.

[0077] Furthermore, the number of cluster centers is 3;

[0078] The mapped feature vector is clustered to obtain a preset number of clusters, and labels are set for each feature in the mapped feature vector according to the cluster to which it belongs, specifically:

[0079] Clustering the mapped feature vectors using the K-Means method according to the minimum Euclidean distance to obtain three clusters;

[0080] The three clusters are ranked according to the number of features they contain;

[0081] The labels of the features in the cluster with the least number of features are set as changed labels, and the labels of the features in the other two clusters are set as unchanged labels.

[0082] From the above description, we can see that by dividing the features into three categories, better classification results can be achieved. Since the changed pixels in the image account for a small number in most cases of change detection scenarios, the features in the cluster with the least number of features are assigned a changed label, and the features in the other two clusters are assigned an unchanged label.

[0083] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which implements the above-mentioned method when the program is executed by a processor.

[0084] Example 1

[0085] Please refer to Figure 2-10 , embodiment 1 of the present invention is: a remote sensing image change detection method based on unsupervised learning, which can be applied to remote sensing image processing and analysis scenarios.

[0086] This method consists of three main parts: preprocessing, change extraction, and postprocessing. The preprocessing part includes extraction of overlapping regions of the previous and next phases, image registration of overlapping regions, fine-tuning of image resolution, and automatic setting of filter windows. The change extraction part combines window filtering operators, K-Means clustering, and PCA principal component analysis to perform correlation processing on feature vectors and feature vector space and obtain preliminary detection results. Postprocessing uses methods such as erosion operators and image masks to obtain raster processing results, which are then converted into vector format through functions.

[0087] Specifically, if Figure 2 As shown, this method includes the following steps:

[0088] S1: determining the overlapping area of ​​the two time-series remote sensing images according to the affine transformation matrix of the two time-series remote sensing images to be measured, and extracting the overlapping areas of the two time-series remote sensing images respectively to obtain two overlapping images.

[0089] Specifically, first, two time-series remote sensing images to be measured are obtained, namely, a front time-series image and a back time-series image (named as the first image a1 and the second image a2 in this embodiment), and atmospheric correction and geometric correction are performed on these two images.

[0090] Then, the affine transformation matrices are respectively extracted from the metadata of the two temporal remote sensing images. The affine transformation matrix includes six parameters, and the specific format is: (longitude resolution, first rotation coefficient, longitude of the upper left corner, second rotation coefficient, latitude resolution, latitude of the upper left corner).

[0091] According to the parameters in the affine transformation matrix, the relevant corner geographical coordinates of the two temporal remote sensing images can be calculated. In this embodiment, the corners include the upper left corner and the lower right corner. Among them, the longitude and latitude of the upper left corner of the first image a1 are (a1_left, a1_top), and the longitude and latitude of the lower right corner are (a1_right, a1_bottom). The longitude and latitude of the upper left corner of the second image a2 are (a2_left, a2_top), and the longitude and latitude of the lower right corner are (a2_right, a2_bottom). And, the longitude and latitude of the upper left corner of the overlapping area are defined as (new_left, new_top), and the longitude and latitude of the lower right corner are (new_right, new_bottom).

[0092] If a1_left < a2_right and a1_right > a2_left and a1_top > a2_bottom and a1_bottom < a2_top, it is considered that there is an overlap between the two temporal remote sensing images; if this condition is not met, it is prompted that there is no overlap between the two temporal remote sensing images, and the input temporal remote sensing images need to be adjusted.

[0093] After confirming the existence of an overlap, determine the relevant corner geographical coordinates of the overlapping area.

[0094] Specifically, if a1_left > a2_left, then new_left = a1_left; if a1_left ≤ a2_left, then new_left = a2_left;

[0095] if a1_top < a2_top, then new_top = a1_top; if a1_top ≥ a2_top, then new_top = a2_top;

[0096] if a1_right < a2_right, then new_right = a1_right; if a1_right ≥ a2_right, then new_right = a2_right;

[0097] if a1_bottom > a2_bottom, then new_bottom = a1_bottom; if a1_bottom ≤ a2_bottom, then new_bottom = a2_bottom.

[0098] According to the geographic coordinates of the relevant corner points in the overlapping area, the row and column number intervals corresponding to the overlapping area in the two time-series remote sensing images can be calculated, and combined with two new affine transformation matrices (the geographic coordinates of the upper left corner points of the original two affine transformation matrices are adjusted, and the resolution and rotation coefficient are not adjusted), the two time-series remote sensing images are cropped to obtain two overlapping images.

[0099] For example, suppose two time series remote sensing images are Figure 3 As shown in the figure, the framed part is the overlapping area; the relative positions of the two time series remote sensing images are as follows Figure 4 As shown, the middle framed part is the overlapping area.

[0100] S2: unifying the sizes and affine transformation matrices of the two overlapping images, and determining the side length of the filtering window.

[0101] Specifically, first, one of the two overlapping images with a lower resolution is used as the first overlapping image, and the other overlapping image is used as the second overlapping image.

[0102] Based on the width and height of the first overlapping image, it is determined whether there is a common divisor within a preset value range. In this embodiment, the preset value range is within 10, that is, it is determined whether the width and height of the overlapping image with lower resolution have a common divisor within 10.

[0103] If so, the greatest common divisor within 10 is used as the side length L of the filtering window, and the original width and original height of the first overlapping image are used as the new width and new height respectively.

[0104] If there is no common divisor within 10, an upper neighborhood search is performed on the original width and height of the first overlapping image, that is, a search is performed in the neighborhood range larger than the original width and height in turn until a neighborhood value of the original width and a neighborhood value of the original height have a common divisor within 10. The maximum value of these common divisors is used as the side length L of the filtering window, and the two neighborhood values ​​obtained by the search are used as the new width and new height respectively.

[0105] Then, according to the resolution, new width and new height of the first overlapping image, the two overlapping images are resampled and registered by bilinear interpolation method to obtain two unified overlapping images. That is, the width, height and affine transformation matrix of the two overlapping images are aligned and unified, as shown in FIG. Figure 5 As shown, it can be used to generate the result graph of the change detection output.

[0106] S3: Calculate the difference map based on the two unified overlapping images. That is, the pixel values ​​in the two unified overlapping images are subtracted one by one to obtain the difference map, such as Figure 6 shown.

[0107] S4: Based on the side length of the filtering window and a first sliding step length, where the first sliding step length is the same as the side length of the filtering window, the difference image is subjected to sliding window filtering to obtain a first filtering result, and based on the first filtering result, a first eigenvector and its eigenvector mean are generated.

[0108] Specifically, a sliding window filter is performed on the difference image using the determined filter window side length L. The sliding window is a square, the first sliding step is equal to the filter window side length L, the window after the step does not overlap with the previous window, the sliding direction is from left to right and from top to bottom, and the window does not change the original difference image value.

[0109] A first eigenvector and a mean eigenvector are formed based on the difference map filtering results. The eigenvector is obtained by tiling the filtering results in a sliding generation order. That is, the first filtering results are arranged in the sliding generation order to obtain the first eigenvector. In this embodiment, assuming the width of the difference map is W and the height is H, the number of rows of the first eigenvector is (H×W) / (L×L), and the number of columns is L×L.

[0110] The eigenvector mean is obtained by averaging the elements in different rows but the same column of the eigenvector. That is, the first filtering results in the same column of the first eigenvector are averaged, and the eigenvector mean is obtained based on the corresponding average values ​​of each column in the first eigenvector. In this embodiment, the eigenvector mean has 1 row and L×L columns.

[0111] S5: Performing principal component analysis (PCA) on the first eigenvector to obtain a eigenvector space.

[0112] In this embodiment, the width and height of the feature vector space are both L×L.

[0113] S6: Perform sliding window filtering on the difference image according to the side length of the filtering window and a preset second sliding step size to obtain a second filtering result, and generate a second feature vector according to the second filtering result.

[0114] Specifically, a sliding window filter is performed on the difference image again. The side length of the filter window of this sliding window filter is still L, the sliding step is 1, and the step is different from the window height and width. The window overlaps with the previous window after the step, and the sliding direction is from left to right and from top to bottom. The window does not change the original difference image value.

[0115] Based on the new filtering results, a new feature vector, i.e., the second feature vector, is generated. In this embodiment, the second feature vector has H - L + 1 rows and W - L + 1 columns. The size of the new feature vector is slightly different from that of the difference map, but this does not affect its ability to reflect the distribution of the difference map's range.

[0116] S7: Map the second eigenvector to the eigenvector space, and perform difference calculation with the mean of the first eigenvector to obtain a mapped eigenvector.

[0117] The mapping in this step involves calculating the dot product of the second eigenvector with the eigenvector space, then subtracting it from the first eigenvector to obtain the mapped eigenvector, which serves as input to the subsequent clustering algorithm. In this embodiment, the size of the mapped eigenvector is the same as the size of the new eigenvector before mapping, with H - L + 1 rows and W - L + 1 columns.

[0118] S8: Clustering the mapped feature vector to obtain a preset number of clusters, and setting a label for each feature in the mapped feature vector according to the cluster to which it belongs.

[0119] Specifically, the mapped feature vectors are clustered according to the minimum Euclidean distance through the K-Means method. In this embodiment, the preset number of clusters is 3 (although it is a binary classification problem, experiments have shown that a clustering number of 3 has a better classification effect), that is, 3 clusters are obtained.

[0120] Based on the clustering results, the three clusters are then sorted by the number of features they contain. Since in most cases of change detection, the changed pixels in the image are a minority, the features in the cluster with the least number of features among the three clusters are assigned a changed label, and the features in the other two clusters are assigned unchanged labels.

[0121] S9: According to the affine transformation matrix of the unified overlapping image, the mapped feature vectors with labels are projected into the geographic space corresponding to the difference map in order from left to right and from top to bottom, and a binary mask map is generated according to the labels.

[0122] Specifically, the mapped feature vectors with labels are combined with the unified affine transformation matrix obtained in step S2, and projected back into the geographic space of the difference map in the order from left to right and from top to bottom, and a binary (0, 255) black and white mask map is produced based on the labels (changed labels, unchanged labels), as shown in the following example: Figure 7 shown.

[0123] The binary black and white mask image obtained at this time is the preliminary detection result of change detection.

[0124] S10: filtering the binary mask image by using an erosion filter operator.

[0125] Specifically, the binary black and white mask image is filtered using an erosion filter operator. In this embodiment, the erosion filter operator is composed of 0 and 1, and the window size is close to the side length L of the above-mentioned filter window and is also an integer within 10. In this embodiment, the value is 5. The binary mask image after filtering is as follows Figure 8 shown.

[0126] Experimental verification shows that the operator can effectively reduce the speckles and noise in the binary mask image, thereby eliminating false changes to a certain extent and improving the accuracy of change detection.

[0127] S11: performing format conversion on the filtered binary mask image to obtain a vector result image, and superimposing the vector result image on the two time series remote sensing images in the form of highlighted boundaries and hollow interiors.

[0128] Specifically, for the convenience of display, the binary mask image after filtering by the erosion filter operator is converted into a format, and the raster data is converted into vector data. The final vector result image can be superimposed on the two input time series remote sensing images in a way that the boundaries are highlighted and the interior is hollow, so that users can more clearly and intuitively understand the location and range of the image change area and the area of ​​the land corresponding to the specific change patch, such as Figure 9-10 shown.

[0129] This embodiment, based on multispectral remote sensing image data, utilizes unsupervised machine learning methods, combined with satellite data processing algorithms, to automatically extract overlapping areas of two sequential images and automatically detect changed areas. By combining the processing of geographic location information with a simple image processing algorithm, the parameters of the affine transformation matrix can be updated and adjusted accordingly while processing the image. At the same time, the underlying principles of the filter parameter setting process are accurately grasped, thereby integrating and automating this process with the resizing of the overlapping images. Compared to detection methods based on supervised learning, the unsupervised method of this embodiment does not require manual labeling or the establishment of a sample database, and can effectively complete the change detection task while significantly reducing the task cost.

[0130] Compared with traditional change detection methods based on unsupervised learning, this method has the following advantages:

[0131] 1. This method has stronger generalization capabilities and can process two time-series images with different resolutions and sizes. In existing technologies, remote sensing images from different time series often exhibit resolution and size differences after atmospheric and geometric correction. Sometimes these differences are subtle, but the input change detection module will cause errors and prevent the program from running properly. This method can solve this problem and enhance the overall robustness of the algorithm.

[0132] 2. This method has a higher degree of automation. The traditional method requires manual calculation of filtering parameters and adjustment of the sizes of the two images according to the filtering parameters. This method automates this process, facilitates operation, and allows users with limited knowledge of remote sensing expertise to use it without obstacles.

[0133] 3. This method optimizes some algorithms in the change detection process, including modifying the calculation method of the feature vector mean and adding a corrosion filter operator. These optimizations improve the detection accuracy.

[0134] 4. This method is more integrated. From image registration and filter parameter calculation in the preprocessing stage to the final mask generation and format conversion, this method can calculate the result image in a "one-stop" manner without external intervention, improving the operating efficiency of the overall change detection process and saving time for users.

[0135] Example 2

[0136] This embodiment is a computer-readable storage medium corresponding to the above embodiment, on which a computer program is stored. When the program is executed by the processor, it implements the various steps of the remote sensing image change detection method based on unsupervised learning in the above embodiment, and can achieve the same technical effect, which will not be repeated here.

[0137] In summary, the present invention provides a remote sensing image change detection method and storage medium based on unsupervised learning, which can process two time-series images with different resolutions and image sizes, has stronger generalization capabilities, and enhances the overall robustness of the algorithm; can automatically determine the filter parameters and automatically unify the sizes of overlapping images, and has a higher degree of automation; from the image registration and filter parameter calculation in the preprocessing link to the final mask generation and format conversion, the result map can be calculated without external intervention, which improves the operating efficiency of the overall change detection process and has a higher degree of integration; by optimizing the calculation method of the feature vector mean and filtering the mask map through the corrosion filter operator, the detection accuracy of the change detection can be improved. Compared with the detection method based on supervised learning, the unsupervised method of the present invention does not require manual labeling, does not need to establish a sample database, and can effectively complete the change detection task while significantly reducing the task cost.

[0138] 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 transformations made using the contents of the present invention's description and drawings, or directly or indirectly applied in related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A remote sensing image change detection method based on unsupervised learning, characterized in that: include: Determining the overlapping area of ​​the two time-series remote sensing images according to the affine transformation matrix of the two time-series remote sensing images to be measured, and extracting the overlapping areas of the two time-series remote sensing images respectively to obtain two overlapping images; Unifying the sizes and affine transformation matrices of the two overlapping images and determining the side length of the filtering window; Calculate the difference map based on the two unified overlapping images; According to the side length of the filtering window and a first sliding step length, where the first sliding step length is the same as the side length of the filtering window, the difference image is subjected to sliding window filtering to obtain a first filtering result, and a first eigenvector and a eigenvector mean thereof are generated according to the first filtering result; Performing principal component analysis on the first eigenvector to obtain a eigenvector space; Performing sliding window filtering on the difference image according to the side length of the filtering window and a preset second sliding step size to obtain a second filtering result, and generating a second feature vector according to the second filtering result; Mapping the second eigenvector to the eigenvector space, and performing difference calculation with the mean of the first eigenvector to obtain a mapped eigenvector; Clustering the mapped feature vector to obtain a preset number of clusters, and setting a label for each feature in the mapped feature vector according to the cluster to which it belongs; According to the affine transformation matrix of the unified overlapping images, the mapped feature vectors with labels are projected into the geographic space corresponding to the difference image in a left-to-right and top-to-bottom order, and a binary mask image is generated according to the labels; The method of determining the overlapping area of ​​the two time-series remote sensing images according to the affine transformation matrix of the two time-series remote sensing images to be measured is specifically as follows: Acquire two time-series remote sensing images to be measured, and perform atmospheric correction and geometric correction on the two time-series remote sensing images; Extracting the affine transformation matrices of the two time-series remote sensing images respectively, wherein the parameters in the affine transformation matrix include longitude resolution, a first rotation coefficient, the longitude of the upper left corner, a second rotation coefficient, latitude resolution and the latitude of the upper left corner; Determine the geographic coordinates of corner points of the two time-series remote sensing images according to the affine transformation matrix of the two time-series remote sensing images, wherein the corner points include an upper left corner point and a lower right corner point; Determining an overlapping area of ​​the two time-series remote sensing images according to the geographic coordinates of the corner points of the two time-series remote sensing images; The step of generating a first eigenvector and a eigenvector mean thereof according to the first filtering result is specifically as follows: Arrange the first filtering results in a sliding generation order to obtain a first feature vector, where the number of rows of the first feature vector is (H×W) / (L×L) and the number of columns is L×L; Calculate the average value of the first filtering results in the same column of the first feature vector, and obtain a feature vector mean according to the average values ​​corresponding to each column of the first feature vector, where the number of rows of the feature vector mean is 1 and the number of columns is L×L; Wherein, H and W are the height and width of the difference image respectively, and L is the side length of the filtering window.

2. The remote sensing image change detection method based on unsupervised learning according to claim 1, characterized in that: After generating a binary mask image according to the label, the method further includes: The binary mask image is filtered by using an erosion filter operator.

3. The remote sensing image change detection method based on unsupervised learning according to claim 2, characterized in that: After filtering the binary mask image by the corrosion filter operator, the method further includes: The filtered binary mask image is format-converted to obtain a vector result image, and the vector result image is superimposed on the two time-series remote sensing images in the form of a highlighted boundary and a hollow interior.

4. The remote sensing image change detection method based on unsupervised learning according to claim 1, characterized in that: The two time-series remote sensing images include a first image and a second image; The step of determining the overlapping area of ​​the two time-series remote sensing images according to the geographic coordinates of the corner points of the two time-series remote sensing images is specifically as follows: If the longitude of the upper left corner of the first image is less than the longitude of the lower right corner of the second image, the longitude of the lower right corner of the first image is greater than the longitude of the upper left corner of the second image, the latitude of the upper left corner of the first image is greater than the latitude of the lower right corner of the second image, and the latitude of the lower right corner of the first image is greater than the longitude and latitude of the upper left corner of the second image, then it is determined that the two time series remote sensing images overlap; If the longitude of the upper left corner of the first image is greater than the longitude of the upper left corner of the second image, then the longitude of the upper left corner of the overlapping area is equal to the longitude of the upper left corner of the first image; otherwise, the longitude of the upper left corner of the overlapping area is equal to the longitude of the upper left corner of the second image; If the latitude of the upper left corner of the first image is less than the latitude of the upper left corner of the second image, then the latitude of the upper left corner of the overlapping area is equal to the latitude of the upper left corner of the first image; otherwise, the latitude of the upper left corner of the overlapping area is equal to the latitude of the upper left corner of the second image; If the longitude of the lower right corner of the first image is smaller than the longitude of the lower right corner of the second image, then the longitude of the lower right corner of the overlapping area is equal to the longitude of the lower right corner of the first image; otherwise, the longitude of the lower right corner of the overlapping area is equal to the longitude of the lower right corner of the second image; If the latitude of the lower right corner point of the first image is greater than the latitude of the lower right corner point of the second image, then the latitude of the lower right corner point of the overlapping area is equal to the latitude of the lower right corner point of the first image; otherwise, the latitude of the lower right corner point of the overlapping area is equal to the latitude of the lower right corner point of the second image.

5. The remote sensing image change detection method based on unsupervised learning according to claim 1, characterized in that: The unification of the sizes and affine transformation matrices of the two overlapping images and determination of the side length of the filtering window is specifically as follows: Using the overlapping image with lower resolution among the two overlapping images as the first overlapping image, and using the other overlapping image as the second overlapping image; Determining whether there is a common divisor within a preset value range based on the width and height of the first overlapping image; If so, obtaining the greatest common divisor within a preset value range, using the greatest common divisor as the side length of the filtering window, and using the width and height of the first overlapping image as the new width and new height respectively; If not, performing an upper neighborhood search on the width and height of the first overlapping image until a neighborhood value of the width and a neighborhood value of the height have a common divisor within a preset value range, and taking the maximum value of the common divisor as the side length of the filtering window, and taking a neighborhood value of the width and a neighborhood value of the height as the new width and new height respectively; According to the width and height of the first overlapping image and the new width and height, the two overlapping images are resampled and registered by a bilinear interpolation method to obtain two unified overlapping images.

6. The remote sensing image change detection method based on unsupervised learning according to claim 1, characterized in that: The width and height of the feature vector space are both L×L; The second sliding step size is 1, the number of rows of the second eigenvector is H-L+1, and the number of columns is W-L+1; The number of rows of the mapped feature vector is H-L+1, and the number of columns is W-L+1.

7. The remote sensing image change detection method based on unsupervised learning according to claim 1, characterized in that: The number of cluster centers is 3; The mapped feature vector is clustered to obtain a preset number of clusters, and labels are set for each feature in the mapped feature vector according to the cluster to which it belongs, specifically: Clustering the mapped feature vectors using the K-Means method according to the minimum Euclidean distance to obtain three clusters; The three clusters are ranked according to the number of features they contain; The labels of the features in the cluster with the least number of features are set as changed labels, and the labels of the features in the other two clusters are set as unchanged labels.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images