A Decomposition-Based Multi-Target Fuzzy Change Detection Method for SAR Images

Abstract

The present invention discloses a method for multi-target fuzzy change detection of SAR images based on decomposition, belonging to the technical field of remote sensing monitoring. Starting from two objective points of retaining details and removing noise, different difference images are generated for different targets, and different objective functions are constructed by using the FCM and FLICM fuzzy clustering cost functions respectively. The multi-objective task is optimized by means of MOEA / D. Finally, a new membership update formula is adopted, and through the population iteration method, the membership values of each pixel point with respect to different categories are calculated with different weight distributions. Experiments prove that the method of the present application has better detection effect on SAR images. At the same time, the present application conducts a more detailed analysis on the two objectives of retaining details and removing noise, and selects two conflicting objective functions for the two objectives to further improve its discrimination ability and the robustness to speckle noise.

Description

Technical Field

[0001] The invention relates to a decomposition-based SAR image multi-target fuzzy change detection method, belonging to the technical field of remote sensing monitoring. Background Art

[0002] Change detection analyzes remote sensing images of the same area over time to identify changes in the Earth's surface caused by human intervention or natural causes. Change detection has been applied to many important areas, such as glacier melt monitoring, disaster management, and land cover monitoring.

[0003] With the continuous advancement of Earth observation technology, it is now easier to obtain satellite remote sensing images of various resolutions and categories. Synthetic Aperture Radar (SAR) imagery has attracted widespread attention due to its all-weather, all-day operation and strong penetrating capabilities. However, due to its imaging principle, SAR images are often subject to scattered noise, making their analysis a difficult task.

[0004] From a machine learning perspective, the main steps of currently popular unsupervised change detection methods are as follows: First, remote sensing images of the same area acquired at different times undergo image preprocessing, primarily including geometric registration, radiometric correction, and denoising. Then, differential images are obtained using certain techniques, and finally, these differential images are analyzed to obtain the final change detection results.

[0005] Li and Gong et al. proposed a multiobjective fuzzy clustering (MOFCM) algorithm (Li, H., Gong, M., Wang, Q., Liu, J., & Su, L. (2016). A multiobjective fuzzy clustering method for change detection in SAR images. Applied Soft Computing, 46, 767-777.). This algorithm transforms the change detection problem into a multi-objective optimization problem (MOP). A difference image generated by a log-ratio operator is used to preserve details. This difference image is then subjected to mean filtering to remove noise, resulting in a denoised difference image. The FCM cost function is then used directly as the objective function for both objectives to construct a multi-objective fuzzy clustering model. A weighted sum decomposition approach is used to weigh different objectives, and an evolutionary algorithm is used to update the cluster centers. Finally, information from the log-ratio image and the mean-filtered log-ratio image is used to determine the membership degree of each pixel, thereby obtaining the final change detection result.

[0006] However, its noise processing is too crude; using only mean filtering cannot effectively handle speckle noise. Furthermore, because it directly uses the FCM cost function as the objective function for both targets, and FCM does not consider the spatial location of pixels and is sensitive to noise, it is not suitable for analyzing denoised images. Therefore, the prediction accuracy of its change detection results needs further improvement. Summary of the Invention

[0007] In order to enhance the analysis capability of SAR images and improve the detection accuracy of change detection tasks, the present invention provides a decomposition-based multi-target fuzzy change detection method. The method addresses the problem that the mean filter used in the MOFCM algorithm cannot effectively process speckle noise, and proposes a new method for generating denoised differential images, namely, wavelet filtering + saliency detection to obtain denoised differential images, which effectively removes speckle noise; and in response to the problem that the MOFCM algorithm uses FCM, which is sensitive to noise and does not consider spatial information, as the denoising objective function, resulting in weak analysis capability of denoised images, the application proposes a new multi-target fuzzy clustering model, which uses different membership update formulas to analyze the membership of each pixel point to different categories, thereby improving the detection accuracy of change detection tasks.

[0008] A multi-target fuzzy change detection method based on decomposition, the method comprising:

[0009] Step 1: Acquire SAR images of the same geographical area at two different times and preprocess the images;

[0010] Step 2: Use the log operator to obtain detail-retaining difference images on the preprocessed images, and use a combination of wavelet filtering and frequency-domain saliency detection to obtain denoised difference images;

[0011] Step 3: Two conflicting objective functions are used for detail-preserving difference images and denoising difference images, and a decomposition-based multi-objective evolutionary computing algorithm is used to simultaneously optimize the two objectives.

[0012] Step 4: Construct an initial uniform weight vector, assign different weight ratios to the two objectives of retaining details and removing noise, and use a multi-objective evolutionary calculation method to update the fuzzy cluster center value;

[0013] Step 5: Combining the idea of decomposition, the newly designed fuzzy membership update formula is used to update the membership value of the pixel to obtain the final classification result and the change detection image.

[0014] Optionally, in step 3, for retaining the detail difference image, the objective function is as follows:

[0015]

[0016] Among them, N is the number of pixel points in the image; C is the number of classifications, C = {1, 2}, representing the classes of changed and unchanged respectively; is to retain the value of the i-th pixel point in the detailed difference image; v k is the clustering center of the k-th class; G ki is the fuzzy factor; m represents the fuzzy coefficient; u ki represents the membership degree of the i-th pixel in the detailed difference image to the k-th class.

[0017] Optionally, in the third step, for the denoised difference image, the objective function is as follows:

[0018]

[0019] Among them, is the value of the i-th pixel point in the noise-removed difference image; is the corresponding calculated membership degree factor.

[0020] Optionally, the membership degree factor The calculation formula of is:

[0021]

[0022] Among them, d ij is the Euclidean distance between pixel point i and pixel point j; N i is the adjacent pixel points of the noise-removed difference image centered on pixel point i that do not include itself; is the value of the j-th pixel point in the noise-removed difference image.

[0023] Optionally, the newly designed fuzzy membership degree update formula in the fifth step is:

[0024]

[0025] Among them, w1 represents the weight assigned to the goal of retaining details, and w2 represents the weight assigned to the goal of removing noise.

[0026] Optionally, the two preprocessed SAR images at different times in the first step are respectively denoted as X1 = {P1(a, b)|1 ≤ a ≤ H, 1 ≤ b ≤ W} and X2 = {P2(a, b)|1 ≤ a ≤ H, 1 ≤ b ≤ W};

[0027] Among them, P(a, b) represents the pixel value at the coordinate (a, b); H and W respectively represent the height and width of the SAR image;

[0028] In the second step, the log operator is used on the preprocessed image to obtain the detail-preserved difference image DI detail , including:

[0029] DI detail = |log(X2 + 1) - log(X1 + 1)|

[0030] where log represents the logarithmic ratio operator.

[0031] Optionally, in the second step, a method combining wavelet filtering and frequency-domain based saliency detection is used on the preprocessed image to obtain the denoised difference image, including:

[0032] Perform wavelet transform on the preprocessed image to transform the image signal from the time domain to the wavelet domain;

[0033] Filter out the noise through the Bayesian threshold method, and then reconstruct the threshold-processed wavelet coefficients to obtain the initial denoised images corresponding to two SAR images at different times respectively;

[0034] Differentiate the initial denoised images corresponding to two SAR images at different times respectively to obtain the initial denoised difference image DI(x);

[0035] Use gamma transform to enhance the initial denoised difference image DI(x);

[0036] Perform Fourier transform on the enhanced initial denoised difference image DI(x) to obtain its amplitude spectrum F(f), and thus obtain its logarithmic spectrum L(f):

[0037] L(f) = log(F(f))

[0038] Use a 3*3 local average filter h(f) to convolve with the logarithmic spectrum L(f) to obtain the image mean logarithmic spectrum A(f):

[0039] A(f) = h(f) * L(f)

[0040] Obtain the spectral residual R(f) from L(f) and A(f) according to the following formula:

[0041] R(f) = L(f) - A(f)

[0042] Then perform inverse Fourier transform on R(f) to obtain the saliency map S(x) in the spatial domain:

[0043] S(x) = g(x) * F -1 [exp(R(f) + P(f))] 2

[0044] Where g(f) represents a Gaussian filter used to smooth the saliency image, and P(f) represents the phase spectrum of the image;

[0045] Finally, the OTSU algorithm is used to obtain the salient area, retaining the pixel values in the salient area and setting the values of the non-salient area to zero, thus obtaining the final denoised difference image DI denoise .

[0046] Optionally, when the decomposition-based multi-objective evolutionary computing algorithm is used to optimize two objectives simultaneously in step 3, the cost function min F(v|w t ) is defined as follows:

[0047] min F(v|w t )=w1f1(v)+w2f2(v)

[0048] Among them, w1 represents the weight given to the goal of preserving details, and w2 represents the weight given to the goal of removing noise.

[0049] Optionally, preprocessing the SAR image in step 1 includes:

[0050] The SAR images are geometrically registered and radiometrically corrected to obtain two co-registered SAR remote sensing satellite images of the same size.

[0051] The beneficial effects of the present invention are:

[0052] This application proposes a decomposition-based multi-target fuzzy change detection method for SAR images. Starting from the two objectives of retaining details and removing noise, this method generates different differential images for different targets, and uses FCM and FLICM fuzzy clustering cost functions to construct different objective functions. The multi-target task is optimized through the MOEA / D method. Finally, a new membership update formula is proposed. Through population iteration, different weight distributions are used to calculate the final membership value of each pixel point to different categories. Experiments have shown that compared with other algorithms, the method proposed in this application has better detection effect on SAR images. At the same time, this application conducts a more detailed analysis of the two objectives of retaining details and removing noise, and selects more suitable objective functions for different targets to further improve its discrimination ability and enhance its robustness to speckle noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0054] Figure 1 This is a flow chart of the decomposition-based multi-target fuzzy change detection method for SAR images proposed in this application.

[0055] Figure 2 These are two SAR images acquired at different times in the Yellow River dataset and the corresponding reference change detection maps.

[0056] Figure 3 The difference images generated by this method; the left image is the detail-retaining difference image, and the right image is the noise-removing difference image.

[0057] Figure 4 In order to use the present application method and the existing method Figure 2 The comparison chart shows the change detection results of two SAR images acquired at different times. DETAILED DESCRIPTION

[0058] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0059] Embodiment 1:

[0060] This embodiment provides a SAR image multi-target fuzzy change detection method based on decomposition. Figure 1 , the method comprising:

[0061] S1 obtains two SAR images of the same area at different times through remote sensing satellites, and preprocesses the original images to obtain two SAR remote sensing satellite images with the same size.

[0062] S2 analyzes the conflicting perspectives of detail preservation and noise removal, constructing the corresponding target difference image. A logarithmic ratio operator is used to obtain a difference image that maximizes image information preservation, while wavelet filtering and a frequency-domain-based saliency detection method are used to obtain a denoised difference image.

[0063] S3 constructs its objective function in an adaptive manner according to different target characteristics, using FCM as the objective function for detail preservation and FLICM as the objective function for noise removal.

[0064] S4 constructs an initial uniform weight vector, assigns different weight ratios to the targets, and uses the multi-objective evolutionary calculation method (MOEA / D) to update the fuzzy cluster center value

[0065] S5 combines the idea of decomposition and uses a newly designed fuzzy membership update formula to update the membership value of the pixel to obtain the final classification result.

[0066] According to the above, the two co-registered SAR images obtained through S1 have a size of H×W and can be expressed as X1 = {P1(a, b)|1 ≤ a ≤ H, 1 ≤ b ≤ W}, X2 = {P2(a, b)|1 ≤ a ≤ H, 1 ≤ b ≤ W}. P(a, b) represents the pixel value at the coordinate (a, b), and H and W respectively represent the height and width of the original SAR image.

[0067] In S2, we generate a denoised difference image DI denoise and a detail-preserved difference image DI detail . Among them, DI detail is obtained through the log operator:

[0068] DI detail = |log(X2 + 1) - log(X1 + 1)|

[0069] where log represents the logarithmic ratio operator.

[0070] In order to obtain a better denoising effect, a frequency-domain-based method is used to obtain the denoised difference image DI denoise . A wavelet filtering and frequency-domain-based saliency detection method is used to obtain the denoised difference image. The noisy image is subjected to wavelet transform to transform the image signal from the time domain to the wavelet domain. In the wavelet domain, the coefficients corresponding to the effective signal are large, while the coefficients corresponding to the noise are small. The noise is filtered out through the Bayesian threshold method, and then the wavelet coefficients after threshold processing are reconstructed. The initial denoised image is obtained, and then the two denoised images are differentiated to obtain the initial denoised difference image.

[0071] Then, further denoising operations are performed on the obtained initial denoised difference image. In the difference image, the pixel points contaminated by speckle noise mostly show isolated and insignificant changes, while the regions where real changes occur are obvious. In order to remove these noise points, means of saliency detection are used to obtain a cleaner denoising effect. At the same time, in order to enhance the saliency change region and suppress the insignificant noise, the present application uses gamma transformation for image enhancement, and then uses a saliency detection method based on the spectral residual method. The specific steps are as follows:

[0072] First, gamma transformation is used for image enhancement, and then the enhanced initial denoised difference image DI(x) is subjected to Fourier transform to obtain its amplitude spectrum F(f), and thus its logarithmic spectrum L(f) is obtained:

[0073] L(f) = log(F(f))

[0074] A 3*3 local average filter h(f) is convolved with the logarithmic spectrum L(f) to obtain the image mean logarithmic frequency spectrum A(f):

[0075] A(f) = h(f) * L(f)

[0076] The spectral residual R(f) is obtained from L(f) and A(f):

[0077] R(f) = L(f) - A(f)

[0078] After that, the inverse Fourier transform is performed on R(f) to obtain the saliency map S(x) in the spatial domain:

[0079] S(x) = g(*) * F -1 [exp(R(f) + P(f))] 2

[0080] where g(f) represents a Gaussian filter used to smooth the saliency image, and P(f) represents the phase spectrum of the image.

[0081] Finally, the OTSU algorithm is used to obtain the salient region, and the pixel values within the salient region are retained, while the values in the non-salient region are set to zero, thereby obtaining the final denoised difference image.

[0082] After obtaining different difference images for the two objectives of retaining details and removing noise, the present application designs two conflicting objective functions for the two obtained images: for the objective of retaining details, since the classical FCM can maximize the retention of the detail information of the original image, the present application uses the classical FCM cost function as the objective function for retaining the original information for the difference image obtained for retaining details.

[0083] For the objective of removing noise, since FLICM has stronger robustness to noise, the present application uses the cost function of FLICM as the objective function for removing noise for the difference image obtained for removing noise.

[0084] The multi-objective evolutionary algorithm based on decomposition (MOEA / D) algorithm is used to optimize these two objectives simultaneously. Since the objective functions of detail preservation and noise elimination are both convex functions, the weighted method is adopted to solve the problem, and the final cost function min F(v|w t ) is defined as follows:

[0085] min F(v|w t ) = w1f1(v) + w2f2(v)

[0086]

[0087]

[0088] where w trepresents the weight assigned to the t-th target, v represents the clustering center of the proposed algorithm, f1(v) is the objective function for retaining details, and f2(v) is the objective function for removing noise. In f t (v), N represents the number of pixel points in the image, C = {1, 2}, representing the classes of changed and unchanged. u ki is the membership degree of the i-th pixel to the k-th class, m represents the fuzzy coefficient, and represent the values of the i-th pixel point in the differential image for retaining details and the differential image for removing noise respectively, is the corresponding calculated membership degree factor, The calculation method of

[0089]

[0090] where N i represents the adjacent pixel points that do not include itself and are centered on the pixel point i in the differential image, d ij is the Euclidean distance between the pixel point i and the pixel point j.

[0091] In the membership degree update formula, the same idea of decomposition is also used. A certain trade-off is made among the analysis results obtained by using the membership degree update formula of FLICM. The designed membership degree update formula is as follows:

[0092]

[0093] By decomposing the change detection problem, the result takes into account the compromise between the two objectives. Therefore, the results generated under different weight ratios can be more fully considered to obtain a more detailed analysis result.

[0094] Example Two:

[0095] This example provides a two-objective fuzzy change detection method based on decomposition. This example is illustrated by taking the detection results of the Yellow River dataset as an example, which contains two SAR images, respectively collected in June 2008 and June 2009, and are two SAR images taken of the change situation in the area near the Yellow River Estuary. This dataset records the surface changes in the area near the Yellow River Estuary due to cultivated land.

[0096] In the example, the results of the experiment are evaluated from two perspectives: one is the finally obtained binary change detection map, and the other is to use some quantitative means for analysis.

[0097] Five metrics are used to evaluate the effectiveness of the algorithm, including false positives (FP), false negatives (FN), total error number (OE), percentage of correct classification (PCC), and Kappa coefficient. Among them, assuming the total number of pixels in the image is N, the number of false positives (FP) refers to the number of pixels that are incorrectly identified as changed when they are not, while the number of false negatives (FN) refers to the number of pixels that are incorrectly identified as unchanged when they have changed. The total error number (OE) is the sum of FP and FN. At the same time, in the embodiment, the number of pixels correctly identified as changed is defined as TP (True positive), and the number of pixels correctly identified as unchanged is defined as TN (True negative). The actual number of changed pixels calibrated in the reference image is AC, and the actual number of pixels that have not changed is AU. The purpose of the actual application is to minimize the OE value and increase the values of PCC and Kappa.

[0098] PCC refers to the proportion of pixels correctly identified as changed and unchanged in the overall population, and its calculation formula is:

[0099]

[0100] The Kappa coefficient is a metric for measuring classification accuracy based on a confusion matrix, and its formula is:

[0101]

[0102] It can be seen from the definition and formula of the evaluation criteria that the smaller the value of OE, the better; while the higher the overall accuracy PCC and the kappa coefficient, the better the classification result.

[0103] To evaluate the performance, in the Yellow River dataset, this embodiment compares the method proposed in this application with four existing algorithm models; these methods include FCM and SIFCM, FLICM and MOFCM.

[0104] Among them, the FCM method can be referred to the introduction in "M. Ahmed, S. Yamany, N. Mohamed, A. Farag, and T. Moriarty, “A modified fuzzy C-means algorithm for bias field estimation and segmentation of MRI data,” IEEE Transactions on Medical Imaging, vol. 21, no. 3, pp. 193–199, MAR 2002.";

[0105] The SIFCM method can be found in “BK Tripathy, A. Basu, and S. Govel, “Image segmentation using spatial intuitionistic fuzzy c means clustering,” in 2014 IEEE International Conference on Computational Intelligence and Computing Research, 2014, pp. 1–5.”

[0106] The FLICM method can be found in “BK Tripathy, A. Basu, and S. Govel, “Image segmentation using spatial intuitionistic fuzzy c means clustering,” in 2014 IEEE International Conference on Computational Intelligence and Computing Research, 2014, pp. 1–5.”

[0107] The MOFCM method can be referred to in “Li, H., Gong, M., Wang, Q., Liu, J., & Su, L. (2016). A multiobjective fuzzy clustering method for change detection in SAR images. Applied Soft Computing, 46, 767-777.”

[0108] In this example, two SAR images were taken in June 2008 and June 2009 respectively to show the changes in the area near the Yellow River Estuary. Figure 2 (a) and 2(b), Figure 2 Figure (c) is the reference change detection image obtained by using Figures (a) and (b) as detection objects;

[0109] The difference image generated by the method of this application is shown in Figure 3 The left image is the detail-retaining difference image, and the right image is the noise-removing difference image. Figure 2 (c) and Figure 3 As can be seen from the left image, the detail-retaining difference image effectively retains the original change information in the image, but also introduces a lot of noise information; Figure 2 (c) and Figure 3As can be seen in the right figure, the difference image after removing the noise has achieved excellent denoising effect, but it also caused some loss of image detail information. The final binary change detection result is shown in Figure 4 Among them Figure 4 (a)-(d) are the binary change results obtained by FCM, SIFCM, FLICM and MOFCM respectively. Figure 4 (e) is the change detection result diagram obtained by this application, Figure 4 (f) is the reference change map. It can be seen that compared to other methods, this application uses saliency detection in the noise removal process, resulting in a cleaner background and significantly reduced speckle noise in the image. By comprehensively considering and weighing the analysis results obtained from the two objectives, the method proposed in this application generates a result that is closest to the reference change detection image. Its quantitative detection results can be shown in Table 1:

[0110] Table 1. Results of quantitative change detection experiments

[0111]

[0112] The method proposed in this application achieved optimal results across all five metrics, achieving not only the lowest OE value but also the lowest FP and FN values of 766 and 499, respectively. PCC improved by 1.15%, and the Kappa coefficient increased by nearly 10%. Experiments confirmed the effectiveness of the proposed method.

[0113] Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD or a hard disk.

[0114] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A decomposition-based multi-target fuzzy change detection method for SAR images, characterized in that, The method includes the following steps: Step 1: Obtain SAR images of two different periods in the same geographical area and preprocess the images; Step 2: Respectively use the log operator on the preprocessed images to obtain a detail-preserved difference image, and use a method combining wavelet filtering and frequency-domain-based saliency detection to obtain a denoised difference image; Step 3: For the detail-preserved difference image and the denoised difference image, construct two different objective functions using the FCM and FLICM fuzzy clustering cost functions, and use a decomposition-based multi-objective evolutionary algorithm to optimize the two objectives simultaneously; Step 4: Construct an initial uniform weight vector, assign different weight ratios to the two objectives of detail preservation and noise removal, and use the multi-objective evolutionary algorithm to update the fuzzy clustering center values; Step 5: Combining the idea of decomposition, use a newly designed fuzzy membership update formula to update the membership values of pixels, so as to obtain the final classification result and obtain a change detection image; In the above Step 2, using a method combining wavelet filtering and frequency-domain-based saliency detection on the preprocessed images to obtain a denoised difference image includes the following steps: Perform wavelet transform on the preprocessed images to transform the image signals from the time domain to the wavelet domain; Filter out noise through the Bayesian threshold method, and then reconstruct the wavelet coefficients after threshold processing to obtain the initial denoised images corresponding to the two SAR images of different periods respectively; Perform difference on the initial denoised images corresponding to the two SAR images of different periods respectively to obtain an initial denoised difference image DI(x); Use gamma transform to enhance the initial denoised difference image DI(x); Perform Fourier transform on the enhanced initial denoised difference image DI(x) to obtain its amplitude spectrum F(f), and thus obtain its logarithmic spectrum L(f): L(f) = log(F(f)) Use a 3*3 local average filter h(f) to convolve with the logarithmic spectrum L(f) to obtain the image mean logarithmic frequency spectrum A(f): A(f) = h(f) * L(f) Obtain the spectral residual R(f) according to the following formula through L(f) and A(f): R(f) = L(f) - A(f) Then perform inverse Fourier transform on R(f) to obtain the saliency map S(x) in the spatial domain: S(x) = g(x) * F -1 [exp(R(f) + P(f))] 2 where g(f) represents a Gaussian filter used to smooth the saliency image, and P(f) represents the phase spectrum of the image; Finally, the OTSU algorithm is used to obtain the salient region, the pixel values within the salient region are retained, while the values of the non-salient region are set to zero, thereby obtaining the final denoised difference image DI denoise .

2. The method according to claim 1, wherein In the above Step 3, for the detail-preserved difference image, the objective function is as follows: where N is the number of pixel points in the image; C is the number of classifications, C = {1, 2}, representing the changed and unchanged classes respectively; x i is the value of the i-th pixel point in the retained detail difference image; v k is the cluster center of the k-th class; m represents the fuzzy coefficient; u ki represents the membership degree of the i-th pixel in the retained detail difference image to the k-th class.

3. The method according to claim 2, wherein In the above Step 3, for the denoised difference image, the objective function is as follows: Among them, is the value of the i-th pixel point in the denoised difference image; is the corresponding calculated blur factor.

4. The method according to claim 3, characterized in that, The fuzzy factor has the following calculation formula: where d ij is the Euclidean distance between pixel point i and pixel point j; N i is the adjacent pixel points centered on pixel point i in the denoised difference image excluding itself; is the value of the j-th pixel point in the denoised difference image.

5. The method according to claim 4, wherein The newly designed fuzzy membership update formula in the above Step 5 is: where w1 represents the weight assigned to the objective of detail preservation, and w2 represents the weight assigned to the objective of noise removal.

6. The method according to claim 5, characterized in that, The two SAR images of different periods after preprocessing in the above Step 1 are respectively denoted as X1 = {P1(a, b)|1 ≤ a ≤ H, 1 ≤ b ≤ W} and X2 = {P2(a, b)|1 ≤ a ≤ H, 1 ≤ b ≤ W}; where P(a, b) represents the pixel value at the coordinate (a, b); H and W respectively represent the height and width of the original SAR image; In the second step, the preprocessed image is processed using a log operator to obtain a detail-preserved difference image DI detail , including: DI detail = |log(X2 + 1) - log(X1 + 1)| where log represents the logarithmic ratio operator.

7. The method according to claim 6, wherein When using the multi-objective evolutionary algorithm based on decomposition to optimize two objectives simultaneously in the third step, the cost function min F(v|w t ) is defined as follows: min F(v|w t ) = w1f1(v) + w2f2(v) Among them, w1 represents the weight assigned to the target of retaining details, and w2 represents the weight assigned to the target of removing noise.

8. The method according to claim 1, wherein The preprocessing of the SAR image in Step 1 includes: Performing geometric registration and radiometric correction on the SAR image to obtain two co-registered SAR remote sensing satellite images with the same size.

Images