A two-target fuzzy change detection method based on decomposition

Abstract

The present invention discloses a two-target fuzzy change detection method based on decomposition, which belongs to the technical field of remote sensing image detection. When the method performs change detection on SAR images, the change detection task is decomposed into two sub-problems of retaining details and removing noise. By analyzing the two goals of retaining details and removing noise, a higher quality detail-retaining differential image is obtained using the log-mean operator. A more effective denoised differential image is obtained by using homomorphic filtering and saliency detection. Different target models are constructed using FLICM, and the two targets are optimized simultaneously using the MOEA / D method. The new membership update method fully considers the balance between different targets and generates a better final change detection map. Experiments on four different data sets have proved that the method of the present application has more powerful detection performance in SAR images with complex noise interference.

Description

Technical Field

[0001] The invention relates to a two-target fuzzy change detection method based on decomposition, and belongs to the technical field of remote sensing image detection. Background Art

[0002] SAR (Synthetic Aperture Radar) images play an important role in the field of remote sensing monitoring because they are not affected by light and climate. However, due to the existence of speckle noise, change detection and analysis of SAR images is still a difficult task.

[0003] The main analysis methods for change detection tasks include supervised learning, semi-supervised learning, and unsupervised learning. In supervised learning and semi-supervised learning methods, it is necessary to obtain labeled data sets in advance to train classifiers to distinguish change areas. However, due to many complex factors such as the resolution, angle, and atmospheric environment of remote sensing satellites, it is difficult to obtain enough reliable labeled data, so it is time-consuming and costly. Unsupervised learning classifies samples based on the similarity between samples and does not require any prior knowledge, so it is more widely used in change detection tasks.

[0004] Unsupervised change detection can be regarded as a binary classification problem, and post-classification comparison is a popular unsupervised change monitoring method. Its main steps are as follows: First, perform image preprocessing operations on remote sensing images of the same area acquired at different times, mainly including geometric registration, radiation correction, denoising and other operations. Then, use certain means to obtain differential images, and finally analyze the obtained differential images to obtain the final change detection results. The analysis methods for differential images are mainly threshold method and clustering method.

[0005] The classic threshold algorithm analyzes the pixel values ​​of the difference image to find the optimal threshold, and then divides the pixels of the difference image into two categories according to the optimal threshold; common threshold methods include the OSTU method, the KI algorithm and the expectation maximization (EM) algorithm. For example, Hu and Ban used different algorithms to automatically determine the selected optimal threshold by analyzing the histogram pattern (Hu, H., and Ban, Y. Unsupervised change detection in multitemporal sar images overlarge urban areas. IEEE Journal of Selected Topics in Applied EarthObservations and Remote Sensing 7, 8 (2014), 3248–3261.); Bruzzone et al. combined the Markov random field with the expectation maximization algorithm for automatic threshold selection, thereby reducing the overall error probability (Bruzzone, L., and Prieto, DF Automatic analysis of the difference image for unsupervised change detection. IEEE Transactions on Geoscience and Remote Sensing 38, 3 (2000), 1171–1182.). Sumaiya et al. combined the Gabor filter with the KI algorithm to generate a binary change map, which improved the accuracy of the detection results (Sumaiya, MN, and Kumari, RSS Gabor filterbased change detection in sar images by ki thresholding. Optik-International Journal for Light and Electron Optics (2017).). In summary, the threshold method is relatively simple to operate and has a low time complexity, but it is sensitive to noise and has relatively weak processing capabilities for complex images with unclear grayscale differences.

[0006] Among the clustering-based change detection methods, fuzzy c-clustering (FCM) is one of the most popular methods. It introduces the concept of fuzzy sets into the determination of class centers and generates a membership matrix. It flexibly divides pixels through membership to detect change information and has better classification capabilities. Chen and Zhang proposed FCM_S1 and FCM_S2 to reduce time consumption (Songcan Chen, and Daoqiang Zhang. Robust image segmentation using fcm with spatial constraints based on new kernel-induced distance measure. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 34, 4 (2004), 1907–1916.), SIFCM introduced the concept of hesitation based on spatial information to further enhance the robustness to noise (thy, BK, Basu, A., and Govel, S. Image segmentation using spatial intuitionistic fuzzy c means clustering. In 2014IEEE International Conference on Computational Intelligence and Computing Research (2014), pp. 1–5.), and Krinidis et al. proposed FLICM (Krinidis, S., and Chatzis, VA robust fuzzy local information c-means clustering algorithm. IEEE Transactions on ImageProcessing 19,5(2010),1328–1337.), this method adds local spatial information so that the neighborhood pixels automatically adjust their influence on the target pixel according to the distance from the target pixel, making the algorithm more robust to noise.Furthermore, Mu et al. proposed an algorithm combined with saliency detection (Mu, C., Li, C., Liu, Y., Sun, M., Jiao, L., and Qu, R. Change detection in sar images based on the salient map guidance and an accelerated genetic algorithm. In 2017 IEEE Congress on Evolutionary Computation (CEC) (2017).), using the FCM algorithm to pre-classify the salient regions into three categories: change, invariant, and uncertain, and then using the accelerated GA algorithm to classify the labels of the uncertain classes to obtain the final change detection results. However, the above method only uses a single target as the cost function, and it is necessary to manually adjust the key parameters to adjust the balance between the ability to retain details and the ability to eliminate noise. However, without prior knowledge, it is very difficult to find suitable parameters, and it is not possible to achieve ideal classification results for images with severe noise.

[0007] In summary, the detection accuracy of existing analysis methods for change detection needs to be further improved. Summary of the invention

[0008] In order to further improve the detection accuracy of the change detection task, the present invention provides a two-target fuzzy change detection model and method based on decomposition, which decomposes and models the change detection problem from different target perspectives, and is used to perform index analysis and prediction on the changes occurring on the surface of the same area. The method comprises:

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

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

[0011] Step 3: Analyze the information contained in the detail-retaining difference image and the denoised difference image, build a two-target change detection model from the two perspectives of removing noise and retaining details, and use the fuzzy clustering objective function based on local spatial information as the cost function of the two targets respectively;

[0012] Step 4: By allocating different weights, weigh the impact of the objectives on the final analysis results, and use the decomposition-based multi-objective evolutionary calculation method to optimize the two objectives;

[0013] Step 5: Combined with the decomposition-based multi-objective weight allocation method, the new fuzzy membership update formula is used to autonomously adjust the membership value of each pixel through population iteration, so as to determine the classification attribute of each pixel and obtain the change detection result image.

[0014] Optionally, the two preprocessed SAR images obtained in step 1 are represented 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), and H and W represent the height and width of the original SAR image, respectively;

[0015] The step 2 uses the log-mean operator to obtain a detail-retaining difference image, including:

[0016] DI detail =mean(|log(X2+1)-log(X1+1)|)

[0017] Among them, mean() represents the mean filtering operation, and log represents the logarithmic ratio operator.

[0018] Optionally, in step 2, the denoised difference image is obtained by combining homomorphic filtering with saliency detection, including:

[0019] Homomorphic filtering converts multiplicative noise into additive noise through the logarithmic ratio operator, and then converts the image signal from the time domain to the frequency domain for processing; after filtering out the noise, the signal is reconstructed and its exponent is used to obtain the initial denoised image;

[0020] Perform a differential operation on the initial denoised image using the logarithmic ratio operator to obtain an initial denoised differential image DI(x);

[0021] Use saliency detection to find salient areas in an image:

[0022] Gamma correction is used to enhance the image DI(x), and then the enhanced initial denoised difference image DI(x) is Fourier transformed to obtain its amplitude spectrum F(f) and logarithmic spectrum L(f):

[0023] L(f)=log(F(f))

[0024] Use a 3*3 local mean filter h(f) and L(f) to obtain the mean log spectrum A(f) of the image:

[0025] A(f)=h(f)*L(f)

[0026] Then the spectral residual R(f) is obtained according to the following formula:

[0027] R(f)=L(f)-A(f)

[0028] Then perform inverse Fourier transform to obtain the salient region;

[0029] Choose whether to retain the pixel value in the area at position x based on the salient area:

[0030]

[0031] Thus, the noise-removed differential image DI is obtained denoise .

[0032] Optionally, in step 3, a two-target change detection model is constructed, and a fuzzy clustering objective function based on local spatial information is used as the cost function of the two targets, f(v):

[0033]

[0034] Where N is the number of pixels in the image; C is the number of categories; x i is the value of the i-th pixel; v k is the cluster center of the kth class; G ki is the fuzzy factor; m represents the fuzzy coefficient; u ki represents the membership of the i-th pixel to the k-th class;

[0035]

[0036] in, and Represent the noise-removed difference image DI denoise and detail-preserving difference image DI detail The value of the i-th pixel in , and is the corresponding calculated fuzzy factor; w1 is the noise-removed difference image DI denoise The weight of w2 is the difference image DI that retains the details. detail The weight of v j is the cluster center of the jth class.

[0037] Optionally, when optimizing the two objectives in step 4, the cost function ming ws (v|w i ) is defined as follows:

[0038]

[0039] w t Represents the weight given to the tth goal, t = 1, 2; the first goal is to remove noise, and the second goal is to retain details.

[0040] Optionally, when optimizing two objectives, the method uses the MOEA / D algorithm to update the cluster center v in the cost function.

[0041] Optionally, the preprocessing of the SAR image in step 1 includes:

[0042] The SAR images are geometrically registered and radiometrically corrected to obtain two co-registered SAR remote sensing satellite images of the same size.

[0043] The beneficial effects of the present invention are:

[0044] When performing change detection on SAR images, the present invention decomposes the change detection task into two sub-problems: detail preservation and noise removal. The idea of ​​this decomposition enables us to consider the change detection problem in more detail. By analyzing the two goals of detail preservation and noise removal, we use the log-mean operator to obtain a higher quality detail-preserving differential image. By using homomorphic filtering and saliency detection, a more effective denoised differential image can be obtained. Different target models are constructed using FLICM, and the two targets are optimized simultaneously using the MOEA / D method. The new membership update method fully considers the balance between different targets and generates a better final change detection map. Experimental results on four different data sets demonstrate the adaptability and effectiveness of DBFLICM. In SAR images with complex noise interference, DBFLICM has more powerful detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] 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.

[0046] Figure 1 This is a flow chart of the decomposition-based two-target fuzzy change detection method proposed in this application.

[0047] Figure 2 Two SAR images acquired at different times in the Ottawa dataset and the corresponding reference change detection maps.

[0048] Figure 3 A detail-retained differential image and a noise-removed differential image generated by the method proposed in the present application in one embodiment of the present application.

[0049] Figure 4 In order to adopt the present application method and the existing method Figure 2The figure shows a comparison of the change detection results of two SAR images acquired at different times. DETAILED DESCRIPTION

[0050] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0051] Embodiment 1:

[0052] This embodiment provides a two-target fuzzy change detection method based on decomposition, the method is used to predict surface change related index data in SAR (Synthetic Aperture Radar) images, the method includes:

[0053] S1 obtains SAR images of the same geographical area at two different times, and performs a series of preprocessing operations such as geometric registration and radiation correction on the images to obtain two co-registered SAR remote sensing satellite images of the same size.

[0054] S2 transforms the change detection task into two sub-problems: noise removal and detail preservation, and generates different matching differential images for different target focuses. The detail-preserving differential image is obtained by using the log-mean operator, and the denoised differential image is obtained by combining homomorphic filtering with saliency detection.

[0055] S3 analyzes the information contained in the two differential images, constructs a two-target change detection model from the two perspectives of removing noise and retaining details, and uses the fuzzy clustering objective function based on local spatial information as the two-target cost function.

[0056] S4 weighs the impact of the objectives on the final analysis results by allocating different weights, and optimizes the two objectives using a decomposition-based multi-objective evolutionary calculation method.

[0057] S5 combines the decomposition-based multi-objective weight allocation method with a new fuzzy membership update formula to autonomously adjust the membership value of each pixel through population iteration, thereby determining the classification attribute of each pixel.

[0058] According to the description, the size of the two co-registered SAR images obtained by S1 is H×W, which 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 represent the height and width of the original SAR image, respectively.

[0059] In S2 we generate the noise-removed difference image DI denoise and the detail-preserving difference image DIdetail . Where DI detail Obtained by the log-mean operator:

[0060] DI detail =mean(|log(X2+1)-log(X1+1)|)

[0061] Among them, mean() represents the mean filtering operation, and log represents the logarithmic ratio operator.

[0062] The noise-removed difference image DI denoise It is obtained by combining homomorphic filtering and saliency detection. The main type of noise in SAR images is multiplicative speckle noise, which will have a serious impact on the recognition accuracy of change detection. The method provided in this application uses homomorphic filtering to obtain an initial denoised SAR image.

[0063] Homomorphic filtering converts multiplicative noise into additive noise through the logarithmic ratio operator, and then converts the image signal from the time domain to the frequency domain for processing. Image information and noise show different characteristics in the frequency domain. Larger frequency coefficients contain low-frequency image information, and noise exists in high-frequency subbands. After filtering out the noise, the signal is reconstructed and its exponential is used to obtain the initial denoised image. The initial denoised differential image DI(x) is obtained by performing a differential operation using the logarithmic ratio operator.

[0064] Then, saliency detection is used to find salient areas in the image. First, gamma correction is used to enhance the image, and then the initial denoised difference image DI(x) is Fourier transformed to obtain its amplitude spectrum F(f) and logarithmic spectrum L(f):

[0065] L(f)=log(F(f))

[0066] Use a 3*3 local mean filter h(f) and L(f) to obtain the mean log spectrum A(f) of the image:

[0067] A(f)=h(f)*L(f)

[0068] Then, the spectral residual R(f) is obtained and inverse Fourier transform is performed to obtain the significant region. According to the significant region, it is determined whether to retain the pixel value in the region at position x:

[0069]

[0070] After obtaining the difference images for analyzing different targets, the change detection task is decomposed into a biological target optimization problem with the goal of preserving details and removing noise.

[0071] Since the fuzzy clustering method based on local neighborhood information (FLICM) has strong analytical capabilities, the method of the present application uses the cost function of FLICM to model these two objectives. When trying to maintain the maximum number of original details, it is possible to leave more noise information in the changed area at the same time. On the contrary, if you want to obtain a cleaner denoising result, you may ignore small change areas in the original image, resulting in ignoring change information. Therefore, this application uses a decomposition-based multi-objective evolutionary computing algorithm (MOEA / D) algorithm to optimize these two objectives simultaneously. Since the objective functions of detail preservation and noise elimination are both convex functions, the weighted method is used in the method of the present application to solve the problem, and the final cost function ming ws (v|w i ) is defined as follows:

[0072]

[0073] Among them, f(v) is the cost function of each target:

[0074]

[0075] where w t represents the weight given to the tth target (t = 1, 2), v represents the clustering center of the proposed algorithm, and f t (v) represents the cost function of each objective. In f(v), u ki is the membership of the i-th pixel to the k-th class, m represents the fuzzy coefficient, and G represents its fuzzy factor. The calculation method is as follows:

[0076]

[0077] Among them, N i represents the neighboring pixels centered on pixel i, d ij is the Euclidean distance between pixel i and pixel j.

[0078] Then, this application uses the MOEA / D algorithm to update the cluster center v in the cost function. First, construct 50 groups of uniform weight vectors and record the neighbor vector population of each group of vectors. Initialize the corresponding 50 groups of cluster centers, and update the positions of their cluster centers through crossover mutation selection between populations. Then update the pixel membership in each population through the obtained cluster center. The MOEA / D algorithm can be referred to the introduction of "A.Trivedi, D.Srinivasan, K.Sanyal, and A.Ghosh, "A survey of multiobjective evolutionary algorithms based on decomposition," IEEE Transactions on Evolutionary Computation, vol.21, no.3, pp.440–462, 2017." This application will not repeat it here.

[0079] In the membership update formula, the same idea of ​​decomposition is used, and a certain trade-off is made between the analysis results obtained using the membership update formula of FLICM. The designed membership update formula is as follows:

[0080]

[0081] in, and Represent the value of the i-th pixel in the noise-removed difference image and the detail-retained difference image, respectively. and is the corresponding calculated membership factor, which considers the trade-off between the two objectives by decomposing the change detection problem.

[0082] Embodiment 2:

[0083] This example provides a two-target fuzzy change detection method based on decomposition. This example takes the detection results of the Ottawa dataset as an example. The dataset contains two SAR images taken by the Radarsat-1 SAR sensor in July 1997 and August 1997 respectively. Figure 2 Figures (a) and (b) in the middle. This dataset records the changes caused by floods near Ottawa, Canada. Figure 2 Figure (c) is a reference change detection image obtained by using Figures (a) and (b) as detection objects.

[0084] In the present invention, a total of 5 indicators are adopted, including false positive number (FP), false negative number (FN), overall error (OE), correct classification percentage (PCC) and Kappa coefficient. FP and FN represent the number of pixels that are incorrectly identified as changed and unchanged, respectively. OE is the sum of FP and FN. PCC refers to the proportion of pixels that are correctly identified as changed and unchanged, and the Kappa coefficient is a classification accuracy measure based on the confusion matrix. Its purpose is to reduce the OE value as much as possible and increase the values ​​of PCC and Kappa. OE, PCC and Kappa are defined as follows:

[0085] OE=FN+FP

[0086]

[0087]

[0088]

[0089] Among them, TP and TN represent the total number of pixels accurately classified as changed and unchanged, respectively. AC and AU represent the number of pixels that actually changed and the number of pixels that actually did not change, respectively.

[0090] In terms of experimental results, this application conducted a series of experiments to prove the superiority of the method of this application.

[0091] 1. This application compares the prediction effects of the model under the combination of differential images generated in different ways. Table 1 summarizes the experimental results. log Represents the difference image obtained using the log operator, DI log-mean Represents the difference image generated using the log-mean operator, DI denoise It represents the denoised differential image obtained using the method in step S2 of the method of the present application.

[0092] Experimental results show that the logarithmic ratio operator can convert multiplicative noise into additive noise and has good detail information retention ability. However, its noise suppression ability is too weak. There is a lot of noise interference in the differential image, which seriously affects the analysis of detail information. The Log-mean operator suppresses the noise in the image to a certain extent while retaining the original details, and effectively improves the model's ability to extract original details. However, its denoising ability is limited. When it is used as a denoised differential image, it cannot completely remove noise. The denoising differential image generation method proposed in this application does not pay attention to the retention of original details, and can remove noise more effectively. Therefore, by making a compromise between denoising and detail retention, a more effective change detection result can be obtained. As can be seen from Table 1, the differential image combination method proposed in this application has a better recognition effect.

[0093] Table 1 Detection results obtained by three differential image combination methods

[0094]

[0095]

[0096] 2. To evaluate the performance, this embodiment compares the proposed method with six existing algorithm models in the Ottawa dataset; these methods include FCM and SIFCM, FLICM, RFLICM, MOFCM and NRCRM.

[0097] Among them, the FCM method can refer 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.”;

[0098] 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.”

[0099] The FLICM method can be found in “BKTripathy, A.Basu, and S.Govel, “Image segmentation using spatial intuitionistic fuzzy c means clustering,” in 2014IEEE International Conference on Computational Intelligence and Computing Research, 2014, pp. 1–5.”

[0100] The RFLICM method can be found in “M. Gong, Z. Zhou, and J. Ma, “Change Detection in Synthetic Aperture Radar Images based on Image Fusion and Fuzzy Clustering,” IEEE Transactions on Image Processing, vol. 21, no. 4, pp. 2141–2151, APR 2012”.

[0101] The MOFCM method can be found 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.”

[0102] The NRCRM method can be found in “Y.Gao, G.Feng, J.Dong, and S.Wang, “Sea ice change detection in sar images based on collaborative representation,” in IGARSS2018-2018IEEE International Geoscience and Remote Sensing Symposium, 2018.”

[0103] The qualitative experimental results are shown in Table 2. The difference image generated by the method of this application is Figure 3 Display in; Figure 3 The left image in the middle is the difference image with details preserved, and the right image is the difference image with noise removed. Figure 4 The final generated binary change map is shown, where Figure 4 (a)-(f) are the binary change result diagrams obtained by FCM, SIFCM, FLICM, RFLICM, MOFCM and NRCRM respectively. Figure 4 (g) is a graph of the change detection results obtained in this application, Figure 4 (h) is the reference change map. The distribution of the change regions in the Ottawa dataset is very complex, which increases the difficulty of distinguishing noise from the change regions. It can be seen that compared with other methods, the method proposed in this application generates the result closest to the reference change detection image.

[0104] contrast Figure 3 The denoised difference image and Figure 2 The reference change detection map given in Figure 3 The right picture and Figure 2 As can be seen from Figure (c), although the generated denoised difference image has effectively removed most of the noise interference, it can be seen that there is still some noise compared to the reference change detection image. Through further analysis of the image, the DBFLICM algorithm proposed in this application effectively eliminates the interference information. As shown in Table 2, this method obtains the lowest OE value, which is 589 less than NRCRM, and the PCC is 0.9818. The Kappa coefficient is 0.9317, which is 2% higher than the best result obtained by other algorithms. Therefore, DBFLICM has the best detection performance. Experiments have confirmed that the new fuzzy membership update method in DBFLICM effectively improves the detection accuracy compared with other algorithms.

[0105] Table 2. Results of quantitative experiments on change detection

[0106]

[0107] 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.

[0108] 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 principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A two-target fuzzy change detection method based on decomposition, characterized in that: The method comprises: Step 1: Obtain SAR images of the same geographical area at two different times and preprocess the images; Step 2: Use the log-mean operator to obtain detail-retaining difference images for the preprocessed images, and use a combination of homomorphic filtering and saliency detection to obtain denoised difference images; Step 3: Analyze the information contained in the detail-retaining difference image and the denoised difference image, build a two-target change detection model from the two perspectives of removing noise and retaining details, and use the fuzzy clustering objective function based on local spatial information as the cost function of the two targets respectively; Step 4: By allocating different weights, weigh the impact of the objectives on the final analysis results, and use the decomposition-based multi-objective evolutionary calculation method to optimize the two objectives; Step 5: Combined with the decomposition-based multi-objective weight allocation method, a new fuzzy membership update formula is used to autonomously adjust the membership value of each pixel through population iteration, thereby determining the classification attribute of each pixel point and obtaining the change detection result image; The two preprocessed SAR images obtained in step 1 are represented 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), and H and W represent the height and width of the original SAR image, respectively; The step 2 uses the log-mean operator to obtain a detail-retaining difference image, including: DI detail =mean(|log(X2+1)-log(X1+1)|) Among them, mean() represents the mean filter operation, and log represents the logarithmic ratio operator; In the step 2, a denoised difference image is obtained by combining homomorphic filtering with saliency detection, including: Homomorphic filtering converts multiplicative noise into additive noise through the logarithmic ratio operator, and then converts the image signal from the time domain to the frequency domain for processing; after filtering out the noise, the signal is reconstructed and its exponent is used to obtain the initial denoised image; Perform a differential operation on the initial denoised image using the logarithmic ratio operator to obtain an initial denoised differential image DI(x); Use saliency detection to find salient areas in an image: Gamma correction is used to enhance the image DI(x), and then the enhanced initial denoised difference image DI(x) is Fourier transformed to obtain its amplitude spectrum F(f) and logarithmic spectrum L(f): L(f)=log(F(f)) Use a 3*3 local mean filter hh(f) and L(f) to obtain the mean log spectrum A(f) of the image: A(f)=h(f)*L(f) Then the spectral residual R(f) is obtained according to the following formula: R(f)=L(f)-A(f) Then perform inverse Fourier transform to obtain the salient region; Choose whether to retain the pixel value in the area at position x based on the salient area: Thus, the noise-removed differential image DI is obtained denoise .

2. The method according to claim 1, characterized in that: In the step 3, a two-target change detection model is constructed, and the fuzzy clustering objective function based on local spatial information is used as the cost function of the two targets, which is f(v): Where N is the number of pixels in the image; C is the number of categories; x i is the value of the i-th pixel; v k is the cluster center of the kth class; G ki is the fuzzy factor; m represents the fuzzy coefficient; u ki represents the membership of the i-th pixel to the k-th class; in, and Represent the noise-removed difference image DI denoise and detail-preserving difference image DI detail The value of the i-th pixel in , and is the corresponding calculated fuzzy factor; w1 is the noise-removed difference image DI denoise The weight of w2 is the difference image DI that retains the details. detail The weight of v j is the cluster center of the jth class.

3. The method according to claim 2, characterized in that When optimizing the two objectives in step 4, the cost function ming ws (v|w i ) is defined as follows: w t Represents the weight given to the tth goal, t = 1, 2; the first goal is to remove noise, and the second goal is to retain details.

4. The method according to claim 3, characterized in that: The method uses the MOEA / D algorithm to update the cluster center v in the cost function when optimizing the two objectives.

5. The method according to claim 1, characterized in that The preprocessing of SAR images in step 1 includes: The SAR images are geometrically registered and radiometrically corrected to obtain two co-registered SAR remote sensing satellite images of the same size.

Images