Satellite image change detection post-processing method based on invariance analysis

Abstract

The application relates to a satellite image change detection post-processing method based on invariance analysis, and relates to the technical field of high-resolution satellite image processing. The method comprises the following steps: for each change detection result spot, corresponding spot images are cut from the front and rear images; feature point extraction and matching are performed on each pair of spot images; the position deviation between the feature point pairs is calculated to determine the geometric position difference between the front and rear images; local position correction is performed on the rear spot image; dense matching is performed on the corrected front and rear spot images; and the invariance of the current spot is analyzed according to the feature point matching result and the dense matching result to determine whether the spot is still changed. The application eliminates the false changes caused by geometric displacement in the detection result by constructing the deviation calculation, position correction and invariance measurement modules, and effectively improves the accuracy of the change detection result.

Description

Technical Field

[0001] This invention relates to the field of high-resolution satellite image processing technology, and in particular to a post-processing method for detecting changes in satellite images based on invariance analysis. Background Technology

[0002] With the continuous development of aerospace technology, the application of satellite technology in land use change information has been significantly enhanced. Satellite technology can acquire large-scale data on the land surface, landforms, and vegetation cover through satellite remote sensing, enabling the monitoring and analysis of land use changes. This data can be used not only in urban planning, agricultural development, and resource management, but also in environmental protection and natural disaster monitoring.

[0003] Change detection involves analyzing dual-temporal satellite imagery of the same region to detect and quantify land use cover changes. To obtain accurate change detection results, it is essential to ensure good geometric registration accuracy among the input satellite imagery. Related research indicates that change detection methods can achieve an accuracy of 90% only when the registration accuracy is controlled within 0.25 pixels. Furthermore, high spatial and spectral variability can lead to higher detection errors in high-resolution satellite data due to the same level of registration error. However, due to differences in acquisition platforms, processes, and imaging conditions, as well as limitations in the accuracy of current image registration methods, geometric displacement often occurs between the input imagery, introducing spurious changes in the change detection results due to positional deviations in the comparison areas.

[0004] To address the spurious changes caused by geometric registration bias, traditional methods primarily focus on two aspects: satellite image preprocessing and neural network architecture design. Satellite image preprocessing reduces positional bias by performing geometrical fine correction on the input dual-temporal images; however, this process is complex and time-consuming, and polynomial geometric correction functions struggle to fit all control points in large-area images. Neural network architecture design incorporates spatial attention mechanisms and pyramid strategies to reduce registration interference by introducing contextual spatial information during model training; however, it remains inadequate for mountainous regions with significant positional shifts. Summary of the Invention

[0005] The present invention aims to solve the technical problems in the prior art by providing a satellite image change detection post-processing method based on invariance analysis.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] A post-processing method for satellite image change detection based on invariance analysis includes the following steps:

[0008] Step 1: For each change detection result patch, crop the corresponding patch image from the previous and later images;

[0009] Step 2: Extract and match feature points for each pair of image patches to obtain corresponding feature point pairs in the previous and later image patches;

[0010] Step 3: Calculate the positional deviation between feature point pairs to determine the geometric positional differences between the pre- and post-images;

[0011] Step 4: Based on the positional deviation information, perform local positional correction on the later patch image to eliminate geometric positional deviations and align it with the earlier patch image;

[0012] Step 5: Perform dense matching on the corrected pre- and post-processed patch images to obtain pixel-by-pixel matching results;

[0013] Step 6: Based on the feature point matching results obtained in Step 2 and the dense matching results obtained in Step 5, analyze the invariance of the current patch and determine whether the patch is still changing.

[0014] In the above technical solution, step 1 specifically includes:

[0015] For each changed patch in the detection results, a buffer range of a certain size is set, and the buffer vector is used to crop the images before and after to obtain the corresponding images of the patches before and after.

[0016] The purpose of setting up a buffer is to ensure that information about surrounding features is included in subsequent analysis.

[0017] In the above technical solution, step 2 specifically includes:

[0018] (1) Feature detection: The scale-invariant feature transformation algorithm is used to extract stable local feature points from the previous and later patch images;

[0019] (2) Feature matching: The brute-force matching searcher is used to match the feature point descriptors in the images before and after the process to obtain the corresponding feature point pairs in the images before and after the process. During the matching process, the Euclidean distance is used to measure the similarity between feature vectors, and the random sampling consensus algorithm is used to eliminate incorrect matches.

[0020] In the above technical solution, step 3 specifically includes:

[0021] Feature point matching describes the correspondence between features in the preceding and following periods. The positional deviation between the preceding and following image patches is obtained by utilizing the coordinate differences of the matching point pairs. The calculation formula is as follows:

[0022]

[0023]

[0024] Where, x diff With y diff This indicates the displacement deviation of the image patches in the horizontal and vertical directions between the earlier and later stages. and This represents the coordinates of the i-th pair of matching points on the previous image. and This represents the coordinates of the i-th pair of matching points on the later image.

[0025] In the above technical solution, step 4 specifically includes:

[0026] The displacement deviation between the two images is regarded as the translation relationship between the images; the displacement deviation obtained by step 3 is used as the basis for the previous patch image to correct the later image to achieve position matching;

[0027] Based on the sign of the displacement deviation, consider the following cases:

[0028] ①If x diff Greater than 0 and y diff If the value is greater than 0, the later image will shift upwards relative to the earlier image by y. diff 1 pixel, move x to the right diff 1 pixel;

[0029] ②If x diff Greater than 0 and y diff If the value is less than 0, the later image will shift upwards relative to the earlier image by y. diff Pixels, move x to the left diff 1 pixel;

[0030] ③If x diff Less than 0 and y diff If the value is greater than 0, the later image is shifted downwards relative to the earlier image by y. diff 1 pixel, move x to the right diff 1 pixel;

[0031] ④If x diff Less than 0 and y diff If the value is less than 0, the later image is shifted downwards relative to the earlier image by y. diff Pixels, move x to the left diff 1 pixel.

[0032] In the above technical solution, step 5 specifically includes:

[0033] For the pre- and post-image patches after eliminating geometric displacement, the invariance between the images is recalculated; using the corresponding points obtained in step 2, regions with invariant structural features are directly marked; a dense matching algorithm is used to generate pixel-by-pixel matching results, describing the invariance of each pixel in the entire image patch, to supplement the deficiencies of feature point matching; the specific process steps include:

[0034] (1) Calculation of matching cost function: Census transform is selected as the matching cost function;

[0035] (2) Global energy function minimization: Based on the matching cost matrix of two images, multi-path cost aggregation is used to obtain the global energy function, and the optimal disparity is obtained by minimizing the global energy function;

[0036] (3) Multi-scale dense matching: The pyramid strategy is used to carry out multi-scale dense matching of dual-temporal satellite images to obtain the dense matching results of each pixel at different pyramid layers.

[0037] In the above technical solution, step 6 specifically includes:

[0038] Based on the calculation results of steps 2 and 5, and combining the feature structure invariance described by the feature point matching results and the multi-scale pixel-by-pixel invariance described by the dense matching results, the invariance within the current changing patch range is calculated to determine whether the patch is still changing; the invariance feature calculation and judgment method is as follows:

[0039]

[0040] Among them, MSF i The invariant feature; i represents the number of the changed patch; k∈K, which is the pyramid level of the multi-scale dense matching species; P mk P represents the number of densely matched pixels at the k-th scale; uk N represents the number of unmatched pixels at the k-th scale; sift λ represents the number of points successfully matched by SIFT within the polygon; λ is the weight coefficient; K represents the number of pyramid levels.

[0041] When MSF i When the value exceeds the threshold, it indicates that the current changed patch has a large area of ​​pixel-level matching results and is marked as a structurally invariant region.

[0042] The present invention has the following beneficial effects:

[0043] The satellite image change detection post-processing method based on invariance analysis of the present invention effectively improves the accuracy of change detection results by constructing modules such as deviation calculation, position correction, and invariance measurement to eliminate spurious changes caused by geometric displacement in the detection results.

[0044] The satellite image change detection post-processing method based on invariance analysis of the present invention, from the perspective of post-processing, has less computation and higher operating efficiency compared with the traditional geometric fine correction method before detection. Compared with the introduction of context features by deep learning networks, the present invention shows stronger adaptability and stability when dealing with larger offsets, making it more widely applicable in practical applications. Attached Figure Description

[0045] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0046] Figure 1 This is a flowchart illustrating the satellite image change detection post-processing method based on invariance analysis according to the present invention.

[0047] Figure 2 This is an example of the post-processing results of the satellite image change detection post-processing method based on invariance analysis according to the present invention. Detailed Implementation

[0048] The inventive concept of this invention is as follows:

[0049] This invention starts with post-processing methods to analyze the changed patches obtained by the algorithm. By constructing modules such as deviation calculation, position correction, and invariance measurement, it eliminates false changes caused by geometric displacement in the detection results and improves the detection accuracy.

[0050] The present invention will now be described in detail with reference to the accompanying drawings.

[0051] See Figure 1 As shown in the figure, an embodiment of the present invention provides a satellite image change detection post-processing method based on invariance analysis, which includes the following steps:

[0052] Step 1: For each change detection result patch, crop the corresponding patch image from the previous and later images;

[0053] Step 2: Extract and match feature points for each pair of image patches to obtain corresponding feature point pairs in the previous and later images;

[0054] Step 3: Calculate the positional deviation between feature point pairs to determine the geometric positional differences between the pre- and post-images;

[0055] Step 4: Based on the positional deviation information, perform local positional correction on the later patch image to eliminate geometric positional deviations and align it with the earlier patch image;

[0056] Step 5: Perform dense matching on the corrected pre- and post-processed patch images to obtain pixel-by-pixel matching results;

[0057] Step 6: Based on the feature point matching results and dense matching results, analyze the invariance of the current patch and determine whether the patch is still changing.

[0058] The satellite imagery used in this embodiment can be high-resolution satellite imagery with a resolution better than 1 meter, which can clearly depict the detailed information of ground features. However, due to the geometric displacement deviation between the two time phases, inconsistencies in information in the comparison areas can be introduced, leading to spurious changes.

[0059] In step 1, for each changed patch in the detection results, a buffer range of a certain size is set, and the buffer vector is used to crop the images before and after the changes to obtain the corresponding images of the patches before and after the changes (e.g., ...). Figure 2 (a) and (b)). The purpose of setting up a buffer is to ensure that surrounding ground feature information can be covered in subsequent analysis, avoiding insufficient usable information due to the small size of the patch, and also solving the problem that ground features at the patch boundary cannot be effectively matched due to positional deviations between previous and later images. Specifically, in this embodiment, the buffer size is set to 30 meters to ensure that the information in the patch image is rich enough to obtain sufficient matching point pairs.

[0060] In step 2, feature points are extracted and matched for each pair of image patches to obtain corresponding point pairs from the preceding and following images (e.g., ...). Figure 2 (c)). The main process is as follows:

[0061] (1) Feature detection: The Scale Invariant Feature Transform (SIFT) algorithm is used to extract stable local feature points from the pre- and post-image patches. The SIFT algorithm is robust to changes in spatial scale, rotation and illumination, and can accurately detect key points in the image.

[0062] (2) Feature matching: A brute-force matching searcher is used to match feature point descriptors in the images before and after the process to obtain corresponding feature point pairs. During the matching process, Euclidean distance is used to measure the similarity between feature vectors, and the Random Sample Consensus (RANSAC) algorithm is used to remove incorrect matches to improve the accuracy and reliability of the matching.

[0063] In step 3, feature point matching describes the correspondence between features in the preceding and following phases. By utilizing the coordinate differences between the matched point pairs, the positional deviation between the preceding and following image patches can be obtained. The calculation formula is shown below:

[0064]

[0065]

[0066] Where, x diff With y diff This indicates the displacement deviation of the image patches in the horizontal and vertical directions between the earlier and later stages. and This represents the coordinates of the i-th pair of matching points on the previous image. and This represents the coordinates of the i-th pair of matching points on the later image.

[0067] In step 4, considering the small area of ​​the patch image, the displacement deviation between the two images is considered as a translation relationship between the images. Using the displacement deviation calculated in the previous step, the later image is corrected based on the earlier patch image to achieve position matching (e.g., ...). Figure 2 (d) Based on the sign relationship of the displacement deviation, consider the following cases:

[0068] ①If x diff Greater than 0 and y diff If the value is greater than 0, the later image will shift upwards relative to the earlier image by y. diff 1 pixel, move x to the right diff 1 pixel;

[0069] ②If x diff Greater than 0 and y diff If the value is less than 0, the later image will shift upwards relative to the earlier image by y. diff Pixels, move x to the left diff 1 pixel;

[0070] ③If x diff Less than 0 and y diff If the value is greater than 0, the later image is shifted downwards relative to the earlier image by y. diff 1 pixel, move x to the right diff 1 pixel;

[0071] ④If x diff Less than 0 and y diff If the value is less than 0, the later image is shifted downwards relative to the earlier image by y. diff Pixels, move x to the left diff 1 pixel.

[0072] In step 5, the invariance between the images is recalculated for both the pre- and post-image patches after geometric displacement removal. Using the corresponding points obtained in step 2, regions with invariant structural features can be directly marked. A dense matching algorithm can generate pixel-by-pixel matching results, describing the invariance of each pixel in the entire image patch, which can supplement the deficiencies of feature point matching. The specific steps are as follows:

[0073] (1) Matching Cost Function Calculation: The Census transform is selected as the matching cost function. This method defines a rectangular window, compares the gray values ​​of each pixel within the window with the gray value of the center pixel, and concatenates them bit by bit to form a binary code stream, which can provide an effective description of the local structure of pixels in the image. Specifically, in this embodiment, the Census window size is set to 9*7.

[0074] (2) Global Energy Function Minimization: Based on the obtained matching cost matrices of the two images, multi-path cost aggregation is used to obtain the global energy function. Optimal disparity is achieved by minimizing the global energy function. A dynamic programming strategy is introduced into the multi-path cost aggregation to comprehensively consider pixel matching in multiple directions to calculate the matching cost in the image, thereby achieving pixel-level registration. Specifically, eight paths are used to aggregate the cost function. After obtaining the total matching cost of the current pixel, a winner-takes-all strategy is used to obtain the final disparity value. Post-processing operations such as left-right consistency checks and median filtering are used to eliminate erroneous matching results, resulting in high-precision pixel-by-pixel matching.

[0075] (3) Multi-scale dense matching: Due to the rich detail information of ground features in high-resolution images, using a single matching scale or a fixed-size matching window during dense matching can result in fragmented results. Furthermore, in some large areas of homogeneous ground features, matching information from multiple paths is difficult to transmit into the interior of the ground features. A pyramid strategy is used to perform multi-scale dense matching on dual-temporal satellite images to obtain dense matching results for each pixel at different pyramid layers, thereby comprehensively reflecting the matching attributes of ground features at different scales (e.g., ...). Figure 2 (e) and (f)). Specifically, the pyramid is set to 5 layers. The first layer uses the original resolution image for pixel-by-pixel matching. The second layer downsamples the image to 4 / 5 the size of the original image for pixel-by-pixel matching. The third layer downsamples the image to 3 / 5 the size of the original image for pixel-by-pixel matching. The fourth layer downsamples the image to 2 / 5 the size of the original image for pixel-by-pixel matching. The fifth layer downsamples the image to 1 / 5 the size of the original image for pixel-by-pixel matching.

[0076] In step 6, based on the calculation results of steps 2 and 5, and combining the feature structure invariance described by the feature point matching results and the multi-scale pixel-by-pixel invariance described by the dense matching results, the invariance within the current changing patch range is calculated to determine whether the patch is still considered changing. The invariance feature calculation and judgment method is as follows:

[0077]

[0078] Among them, MSF iFor invariant features, i represents the number of the changed patch, k∈K belongs to the pyramid level of the multi-scale dense matching species, and P mk P represents the number of densely matched pixels at the k-th scale. uk N represents the number of unmatched pixels at the k-th scale. sift λ represents the number of points successfully matched by SIFT within the patch, λ is the weight coefficient, and K represents the number of pyramid levels.

[0079] When MSF i When the value is greater than the threshold (usually set to 1), it indicates that the current changed patch has a large area of ​​pixel-level matching results and is marked as a structurally invariant region. Such patches will be removed from the detection results as invariant patches.

[0080] The satellite image change detection post-processing method based on invariance analysis of the present invention effectively improves the accuracy of change detection results by constructing modules such as deviation calculation, position correction, and invariance measurement to eliminate spurious changes caused by geometric displacement in the detection results.

[0081] The satellite image change detection post-processing method based on invariance analysis of the present invention, from the perspective of post-processing, has less computation and higher operating efficiency compared with the traditional geometric fine correction method before detection. Compared with the introduction of context features by deep learning networks, the present invention shows stronger adaptability and stability when dealing with larger offsets, making it more widely applicable in practical applications.

[0082] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A post-processing method for satellite image change detection based on invariance analysis, characterized in that, Includes the following steps: Step 1: For each change detection result patch, crop the corresponding patch image from the previous and later images; Step 2: Extract and match feature points for each pair of image patches to obtain corresponding feature point pairs in the previous and later image patches; Step 3: Calculate the positional deviation between feature point pairs to determine the geometric positional differences between the pre- and post-images; Step 4: Based on the positional deviation information, perform local positional correction on the later patch image to eliminate geometric positional deviations and align it with the earlier patch image; Step 5: Perform dense matching on the corrected pre- and post-processed patch images to obtain pixel-by-pixel matching results; Step 6: Based on the feature point matching results obtained in Step 2 and the dense matching results obtained in Step 5, analyze the invariance of the current patch and determine whether the patch is still changing; Step 5 specifically involves: For the pre- and post-image patches after eliminating geometric displacement, the invariance between the images is recalculated; using the corresponding points obtained in step 2, regions with invariant structural features are directly marked; and a dense matching algorithm is used to generate pixel-by-pixel matching results to describe the invariance of each pixel in the entire image patch, in order to supplement the deficiencies of feature point matching. The specific process steps include: (1) Calculation of matching cost function: Census transform is selected as the matching cost function; (2) Global energy function minimization: Based on the matching cost matrix of two images, multi-path cost aggregation is used to obtain the global energy function, and the optimal disparity is obtained by minimizing the global energy function; (3) Multi-scale dense matching: The pyramid strategy is used to carry out multi-scale dense matching of dual-temporal satellite images to obtain the dense matching results of each pixel at different pyramid layers; Step 6 specifically involves: Based on the calculation results of steps 2 and 5, and combining the feature structure invariance described by the feature point matching results and the multi-scale pixel-by-pixel invariance described by the dense matching results, the invariance within the current changing patch range is calculated to determine whether the patch is still changing; the invariance feature calculation and judgment method is as follows: Among them, MSF i The invariant feature is represented by 'i', which indicates the number of the changed patch. It is a pyramid hierarchy of multi-scale dense matching species; P mk P represents the number of densely matched pixels at the k-th scale; uk N represents the number of unmatched pixels at the k-th scale; sift This represents the number of point pairs that were successfully matched using SIFT within the image patch. These are the weighting coefficients; Indicates the number of pyramid levels; When MSF i greater than the threshold value, it indicates that the current change patch has a large area of pixel-level matching results and is marked as a structural invariant region.

2. The satellite image change detection post-processing method based on invariance analysis according to claim 1, characterized in that, Step 1 specifically involves: For each changed patch in the detection results, a buffer range of a certain size is set, and the buffer vector is used to crop the images before and after to obtain the corresponding images of the patches before and after. The purpose of setting up a buffer is to ensure that information about surrounding features is included in subsequent analysis.

3. The satellite image change detection post-processing method based on invariance analysis according to claim 1, characterized in that, Step 2 specifically involves: (1) Feature detection: The scale-invariant feature transformation algorithm is used to extract stable local feature points from the previous and later patch images; (2) Feature matching: The brute-force matching searcher is used to match the feature point descriptors in the images before and after the process to obtain the corresponding feature point pairs in the images before and after the process. During the matching process, the Euclidean distance is used to measure the similarity between feature vectors, and the random sampling consensus algorithm is used to eliminate incorrect matches.

4. The satellite image change detection post-processing method based on invariance analysis according to claim 1, characterized in that, Step 3 specifically involves: Feature point matching describes the correspondence between features in the preceding and following periods. The positional deviation between the preceding and following image patches is obtained by utilizing the coordinate differences of the matching point pairs. The calculation formula is as follows: in, and This indicates the displacement deviation of the image patches in the horizontal and vertical directions between the earlier and later stages. and This represents the coordinates of the i-th pair of matching points on the previous image. and This represents the coordinates of the i-th pair of matching points on the later image.

5. The satellite image change detection post-processing method based on invariance analysis according to claim 1, characterized in that, Step 4 specifically involves: The displacement deviation between the two images is regarded as the translation relationship between the images; the displacement deviation obtained by step 3 is used as the basis for the previous patch image to correct the later image to achieve position matching; Based on the sign of the displacement deviation, consider the following cases: ①If x diff Greater than 0 and y diff If the value is greater than 0, the later image will shift upwards relative to the earlier image by y. diff 1 pixel, move x to the right diff 1 pixel; ②If x diff Greater than 0 and y diff If the value is less than 0, the later image will shift upwards relative to the earlier image by y. diff Pixels, move x to the left diff 1 pixel; ③If x diff Less than 0 and y diff If the value is greater than 0, the later image is shifted downwards relative to the earlier image by y. diff 1 pixel, move x to the right diff 1 pixel; ④If x diff Less than 0 and y diff If the value is less than 0, the later image is shifted downwards relative to the earlier image by y. diff Pixels, move x to the left diff 1 pixel.

Images