CN103020978A - SAR (synthetic aperture radar) image change detection method combining multi-threshold segmentation with fuzzy clustering - Google Patents

Abstract

The invention discloses an SAR (synthetic aperture radar) image change detection method combining multi-threshold segmentation with fuzzy clustering. The method mainly aims at overcoming the defect of existing fuzzy clustering algorithms and is used for SAR image change detection by combining multi-threshold segmentation with fuzzy clustering. The implementation steps of the method include: (1), subjecting two SAR images to median filtering; (2), calculating to obtain a logarithmic ratio differential image after normalization; (3), adopting the Otsu method based on standard particle swarm optimization to perform multi-threshold segmentation to the logarithmic ratio differential image after normalization; (4), initializing membership matrixes U0 and U1; (5), adopting the FLICM (fuzzy local information C-means) algorithm to perform fuzzy clustering to pixels which cannot be determined whether changes occur or not after multi-threshold segmentation; (6), deblurring; and (7), outputting change detection results. The multi-threshold segmentation and fuzzy clustering are combined for SAR image change detection, so that change detection time is reduced, and change detection accuracy is improved.

Description

SAR image change detection method combining multi-threshold segmentation and fuzzy clustering

Technical Field

The invention belongs to the technical field of computers, and further relates to an SAR image change detection method combining multi-threshold segmentation and fuzzy clustering in the technical field of image processing. The SAR image change detection method based on the fuzzy clustering obtains the difference image from the two SAR images with different time phases, then performs multi-threshold segmentation and fuzzy clustering on the difference image, realizes SAR image change detection, and can be used in the fields of ground object coverage and utilization, natural disaster monitoring and evaluation, city planning, map updating and the like.

Background

Synthetic Aperture Radar (SAR) technology has become more and more widely used in recent years. Compared with the common optical remote sensing image, the SAR image has the characteristic of all-weather acquisition, and along with the continuous improvement of the resolution of the SAR image, more and more image processing technologies are provided based on the SAR image. The image change detection technology is used for obtaining the landform change information of the area in two periods by comparing and analyzing two remote sensing images of the same area at different times. At present, a change detection technology based on SAR images has been applied more, such as forest coverage change, natural disaster monitoring and evaluation and the like.

There are roughly two approaches to how to find the varying regions from the two SAR images: one is a comparison after classification method, namely, different SAR images are classified respectively, and then the classified results are compared to find out a changed area; the other method is to obtain the difference images of different SAR images from different SAR images and analyze the difference images to find the change area, and the latter method is applied to a plurality of methods at present. The processing of the difference image is actually to perform a two-class division on the difference image, so that a clustering algorithm can be used for solving the problem of the classification of the difference image.

The Fuzzy C-Means (FCM) algorithm is one of the most popular clustering algorithms, and is characterized in that uncertainty description of sample classes is established by adopting a Fuzzy theory, membership degrees of the samples to all clustering centers are obtained by optimizing an objective function, and then the class of each sample is judged. However, since the FCM does not consider the spatial information of the samples, it is sensitive to noise, and the clustering effect is affected. The Fuzzy Local information C-Means (FLICM) algorithm is a relatively new clustering algorithm, and is improved aiming at the defect that FCM is relatively sensitive to noise, and the Local spatial information of samples is introduced into a clustering objective function, so that a better clustering effect than FCM is achieved, but the calculation of neighborhood information is increased for all samples, and the calculation amount is larger compared with FCM.

The patent of the seian university of electronic technology "method for detecting SAR image change based on quantum immune cloning" (patent application No. 201010230980.6, publication No. CN101908213A), which defines a cluster center by quantum bits, searches for an optimal cluster center and obtains a global threshold, but has a disadvantage that only an objective function of FCM is used to construct an antibody affinity function, and neighborhood information of pixels is not considered.

Disclosure of Invention

The invention aims to overcome the defects of the prior art, and combines multi-threshold segmentation and fuzzy clustering for SAR image change detection so as to reduce the change detection time and improve the change detection precision.

In order to achieve the purpose, when the change detection is carried out, the invention firstly uses multi-threshold segmentation to judge the category of the simple pixels, and then the rest difficult pixels are reserved for the FLICM fuzzy clustering algorithm to be judged.

The invention comprises the following steps:

(1) median filtering

Selecting a common 3 x 3 median filter to preprocess the two SAR images to be detected to obtain two median-filtered images;

(2) obtaining a normalized logarithmic ratio difference image

2a) Calculating a logarithmic ratio difference image from the two median-filtered images by adopting a logarithmic ratio difference formula;

2b) carrying out normalization processing on the log ratio difference image by adopting a normalization formula to obtain a normalized log ratio difference image;

(3) multi-threshold segmentation

And performing multi-threshold segmentation on the normalized logarithmic ratio difference image by adopting an Otsu method based on standard particle swarm optimization, wherein pixels in the segmented image are divided into three types: pixels that have not changed, pixels that have changed, and pixels that cannot determine whether a change has occurred;

(4) initializing membership matrix

4a) Initializing membership matrix U₀，U₀Storing membership degree of pixel under non-variable class, wherein the membership degree is in interval [0, 1 ]]The real number of the up value is used for representing that the pixel which is not changed after multi-threshold segmentation is in U₀The membership degree in (1) is selected, and after multi-threshold segmentation, the membership degree belongs to the changed pixels in U₀The membership degree in (1) is 0, and after multi-threshold segmentation, the pixel which can not determine whether change occurs is in U₀The membership degree in (1) is randomly generated;

4b) initializing membership matrix U₁，U₁Storing the membership degree of the pixel under the variation class, wherein the membership degree is in the interval [0, 1 ]]The real number of the up value is used for representing that the pixel which is not changed after multi-threshold segmentation is in U₁The membership degree in (1) is 0, and after multi-threshold segmentation, the membership degree belongs to the changed pixels in U₁The membership degree in (1) is 1, and after multi-threshold segmentation, the pixel which can not determine whether the change occurs belongs to U₁The degree of membership in (1) is then equal to 1 minus the pixel in U₀Degree of membership in (1);

(5) fuzzy clustering

Performing fuzzy clustering on pixels which cannot determine whether change occurs after multi-threshold segmentation by using an FLICM (flash memory integration control algorithm), and performing iterative updating until a preset termination condition is reached, namely the maximum change of membership degrees before and after iteration is less than 0.00001;

(6) deblurring

According to the membership value of each pixel in the normalized log ratio difference image, which is subordinate to a non-change class and a change class, the pixel is judged to be the class with a larger membership value, so that all the pixels in the normalized log ratio difference image are divided into two classes: a non-change class and a change class;

(7) and outputting a change detection result.

Compared with the prior art, the invention has the following advantages:

1. the method firstly judges the category of the simple pixels by utilizing multi-threshold segmentation, and then reserves the rest difficult pixels for the FLICM algorithm to judge, and compared with the prior FLICM algorithm which directly clusters and judges the pixels, the method reduces the operation amount and the running time of change detection.

2. When the invention uses FLICM algorithm to carry out fuzzy clustering, the local window N_iThe classes of some non-central pixels in the image are determined in a multi-threshold segmentation stage, and the determined neighborhood information improves the fuzzy clustering effect in the prior art, so that the change detection precision is further improved.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a result of change detection performed on a floods SAR image in an Ottawa area in a simulation experiment;

FIG. 3 shows the result of SAR image variation detection of fire in Mexico area in simulation experiment;

fig. 4 is a result of change detection performed on the Bern urban flood SAR image in the simulation experiment.

Detailed Description

Referring to fig. 1, the method for detecting the change of the SAR image by combining multi-threshold segmentation and fuzzy clustering in the invention comprises the following steps:

(1) median filtering

Selecting a common 3 x 3 median filter to preprocess the two SAR images to be detected to obtain two median-filtered images;

(2) obtaining a normalized logarithmic ratio difference image

2a) And solving a log ratio difference image from the two median-filtered images by adopting a log ratio difference formula as follows:

I₃＝|log(I₁+1)-log(I₂+1)|

wherein, I₃Representing pixel gray values of a logarithmic ratio difference image, I₁And I₂Respectively representing pixel gray values of the two images after median filtering;

2b) carrying out normalization processing on the log ratio difference image by adopting a normalization formula as follows to obtain a normalized log ratio difference image:

$I_D=\frac{255\×(I_3-I_{\min })}{(I_{\max }-I_{\min })}$

wherein, I_DTo representNormalized pixel gray value, I, of log ratio difference image₃Representing pixel gray values of a logarithmic ratio difference image, I_maxIs represented by₃Middle maximum gray value, I_minIs represented by₃The smallest gray value;

(3) multi-threshold segmentation

Performing multi-threshold segmentation on the normalized logarithm ratio difference image by adopting an Otsu method based on standard particle swarm optimization, wherein pixels in the segmented image are divided into three types: pixels that have not changed, pixels that have changed, and pixels that cannot determine whether a change has occurred;

the specific steps of multi-threshold segmentation are as follows:

3a) encoding n thresholds for segmenting the image as one n-dimensional particle in a standard particle swarm optimization;

3b) using a standard particle swarm optimization search to maximize the objective function of Otsu's method by n thresholds, the Otsu's objective function is as follows:

Maximize σ²＝ω₀(μ₀-μ_T)²+ω₁(μ₁-μ_T)²+...+ω_n(μ_n-μ_T)²

the normalized logarithmic ratio difference image may be divided into n +1 parts by n thresholds, in the above formula, σ²Represents the variance, ω, between the n +1 parts₀，ω₁，...，ω_nRespectively representing the probability of occurrence of the respective portions, mu₀，μ₁，…，μ_nRespectively representing the mean value of the gray levels, mu, of the respective parts_TA gray level average value representing the entire image;

the updating formula of the d-dimension flight speed of the particles in the standard particle swarm optimization is as follows:

$v_i^d=\mathrm{\&omega;}{v}_i^d+{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="HSA00000839307100022.png"\; state="\{\{state\}\}">c</figure-callout>}}_1{\mathrm{rand}}_1^d({\mathrm{pBest}}_i^d-x_i^d)+c_2{\mathrm{rand}}_2^d({\mathrm{gBest}}^d-x_i^d)$

wherein,

representing the d-dimensional flight velocity of the ith particle in the population, omega representing the inertial weight, c₁And c₂Which represents a factor of learning that is,

representing the d-dimensional position of the ith particle in the population,and

is two independent random numbers, pBest, uniformly distributed between 0 and 1_iIs the historical optimum position of the ith particle, and gBest is the historical optimum position of the whole population;

the updating formula of the d-dimension position of the particle in the standard particle swarm optimization is as follows:

$x_i^d=x_i^d+v_i^d$

wherein,

representing the d-dimensional position of the ith particle in the population,

representing the d-dimension flight speed of the ith particle in the population;

3c) through the search of standard particle swarm optimization, n thresholds which enable the Otsu method objective function to be maximum are found, the n thresholds divide the logarithmic ratio difference image after normalization into n +1 parts, and the parts are recorded as follows according to the sequence of gray values from small to large: c₀，C₁，...，C_n. Handle C₀The pixels in (A) are classified as unchanged pixels, and C is set_nThe pixels in (A) are classified as changed pixels, and C is set₁To C_n-1The pixels in the image are classified into pixels which can not determine whether the change occurs, so that the pixels in the normalized logarithmic ratio difference image are divided into three types;

(4) initializing membership matrix

4a) Initializing membership matrix U₀，U₀Storing membership degree of pixel under non-variable class, wherein the membership degree is in interval [0, 1 ]]The real number of the up value is used for representing that the pixel which is not changed after multi-threshold segmentation is in U₀The membership degree in (1) is selected, and after multi-threshold segmentation, the membership degree belongs to the changed pixels in U₀The membership degree in (1) is 0, and after multi-threshold segmentation, the membership degree cannot be determined whether to send or notChanged pixel is in U₀The membership degree in (1) is randomly generated;

4b) initializing membership matrix U₁，U₁Storing the membership degree of the pixel under the variation class, wherein the membership degree is in the interval [0, 1 ]]The real number of the up value is used for representing that the pixel which is not changed after multi-threshold segmentation is in U₁The membership degree in (1) is 0, and after multi-threshold segmentation, the membership degree belongs to the changed pixels in U₁The membership degree in (1) is 1, and after multi-threshold segmentation, the pixel which can not determine whether the change occurs belongs to U₁The degree of membership in (1) is then equal to 1 minus the pixel in U₀Degree of membership in (1);

(5) fuzzy clustering

Performing fuzzy clustering on pixels which cannot determine whether change occurs after multi-threshold segmentation by using an FLICM (flash memory integration control algorithm), and performing iterative updating until a preset termination condition is reached, namely the maximum change of membership degrees before and after iteration is less than 0.00001;

5a) the FLICM algorithm uses membership degree to represent the degree of possible pixel belonging to a certain class, and simultaneously considers image gray information and spatial neighborhood information in the clustering process, and the target function of the FLICM is as follows:

$J=\underset{i\&Element;N}{\mathrm{\&Sigma;}}\underset{k=1}{\overset{C}{\mathrm{\&Sigma;}}}[u_{\mathrm{ki}}^m{d}^2(x_i,v_k)+G_{\mathrm{ki}}]$

wherein J represents an objective function value, N represents a set of pixels which cannot be determined whether change occurs after multi-threshold segmentation, C is the number of classes with a value of 2, and u is the number of classes with a value of 2_kiRepresenting degree of membership, x, of the ith pixel in the kth class_iRepresenting the gray value of the ith pixel, v_kDenotes the cluster center of the kth class, d²(x_i，v_k) Represents the distance measure from the ith pixel to the cluster center of the kth class, m represents a fuzzy weighting index with a value of 2, G_kiRepresents the blurring coefficient between the ith pixel and the cluster center of the kth class, which is defined as follows:

$G_{\mathrm{ki}}=\underset{\underset{i\&NotEqual;j}{j\&Element;N_i}}{\mathrm{\&Sigma;}}\frac{1}{e_{\mathrm{ij}}+1}{(1-u_{\mathrm{kj}})}^m{d}^2(x_j,v_k)$

wherein G is_kiRepresenting the blurring coefficient between the i-th pixel and the cluster center of the k-th class, N_iRepresenting a local window of size 3 x 3, x, centered on the ith pixel_jThe gray value of the jth pixel is shown and falls into the local window N_iV is a non-central pixel of_kDenotes the cluster center of the kth class, u_kjRepresenting degree of membership of the jth pixel in the kth class, e_ijRepresenting the spatial Euclidean distance, d, between the ith and jth pixels²(x_j，v_k) Representing the distance measurement from the jth pixel to the kth class of clustering center, and m representing a fuzzy weighting index with a value of 2;

5b) the FLICM algorithm adopts an iterative calculation mode to obtain a group of membership values which enable the objective function to be minimum, and the iterative calculation formula of the membership in the FLICM algorithm is as follows:

$u_{\mathrm{ki}}=\frac{1}{\underset{t=1}{\overset{C}{\mathrm{\&Sigma;}}}{\left(\frac{d^2(x_i,v_k)+G_{\mathrm{ki}}}{d^2(x_i,v_t)+G_{\mathrm{ti}}}\right)}^{\frac{1}{m-1}}}$

wherein u is_kiRepresenting the degree of membership of the ith pixel in the kth class, C being the number of classes taking the value 2, x_iRepresenting the gray value of the ith pixel, v_kRepresenting the cluster center of the kth class, v_tDenotes the cluster center of the t-th class, G_kiRepresenting the blurring coefficient between the i-th pixel and the cluster center of the k-th class, G_tiRepresenting the blurring coefficient between the ith pixel and the cluster center of the t-th class, d²(x_i，v_k) Representing a distance measure of the ith pixel to the cluster center of the kth class, d²(x_i，v_t) Representing the distance measurement from the ith pixel to the clustering center of the t-th class, wherein m represents a fuzzy weighting index with the value of 2;

5c) the iterative calculation formula of the clustering center in the FLICM algorithm is as follows:

$v_k=\frac{\underset{i\&Element;N}{\mathrm{\&Sigma;}}{u}_{\mathrm{ki}}^m{x}_i}{\underset{i\&Element;N}{\mathrm{\&Sigma;}}{u}_{\mathrm{ki}}^m}$

wherein v is_kRepresenting the cluster center of the kth class, N representing a set of pixels that cannot be determined whether a change has occurred after multi-threshold segmentation, u_kiRepresenting degree of membership, x, of the ith pixel in the kth class_iExpressing the gray value of the ith pixel, and m represents a fuzzy weighting index with the value of 2;

(6) deblurring

According to the membership value of each pixel in the normalized log ratio difference image, which is subordinate to a non-change class and a change class, the pixel is judged to be the class with a larger membership value, so that all the pixels in the normalized log ratio difference image are divided into two classes: a non-change class and a change class;

(7) and outputting a change detection result.

The effect of the invention can be further illustrated by the following simulation experiment:

1. simulation experiment conditions

Simulation experiment environment: the operating system is Windows XP, the CPU is AMD Athlon1.60GHz, the memory is 1.75GB, and the programming platform is Visual C + + 6.0.

The simulation experiment is that the SAR image change detection of flood in Ottawa area is 290 × 350, the image size is shown in FIG. 2(a) as SAR image in 5 months 1997, FIG. 2(b) as SAR image in 8 months 1997, FIG. 2(c) as the corresponding change detection reference image, FIG. 2(d) as the change detection result of FCM algorithm, FIG. 2(e) as the change detection result of FLICM algorithm, and FIG. 2(f) as the change detection result of the invention.

The second simulation experiment is the change detection of the SAR image of a fire in the mexico region, the image size is 512 × 512, fig. 3(a) is the SAR image of 4 months in 2000, fig. 3(b) is the SAR image of 5 months in 2002, fig. 3(c) is a change detection reference map, fig. 3(d) is the change detection result of the FCM algorithm, fig. 3(e) is the change detection result of the FLICM algorithm, and fig. 3(f) is the change detection result of the present invention.

The third simulation experiment is the change detection of the SAR image in Bern urban flood, the image size is 301 × 301, fig. 4(a) is the SAR image in 4 months 1999, fig. 4(b) is the SAR image in 5 months 1999, fig. 4(c) is the change detection reference map, fig. 4(d) is the change detection result of the FCM algorithm, fig. 4(e) is the change detection result of the FLICM algorithm, and fig. 4(f) is the change detection result of the present invention.

2. Contents and results of simulation experiments

The parameters in the simulation experiment were set as follows: the threshold number n of the multi-threshold segmentation is set to be 3 (the selected basis is given in the parameter analysis simulation experiment of the threshold number later); in the invention, two learning factors of the standard particle swarm optimization are set to be 2.0, the inertia weight is linearly reduced from 0.9 to 0.4 along with the searching process, and the termination condition of the standard particle swarm optimization is that the evaluation frequency of the objective function reaches 5000 times; in the simulation experiment, the termination conditions of the fuzzy clustering process in the three methods are set to be that the maximum change of membership degrees before and after iteration is less than 0.00001. The data in the following simulation experiments are the average results of 20 independent runs.

Simulation experiment I: the results of the three methods for detecting the change in this problem are shown in FIGS. 2(d) to (f), and the specific comparative analysis of the results is shown in Table 1. As can be seen from FIG. 2, the detection results of the present invention are most similar to the reference image. As can be seen from Table 1, the number of missed pixels is 1246 and 1309 less than FCM and FLICM, respectively, and the total number of error pixels is 1003 and 372 less than the two, respectively. Meanwhile, as can be seen from the aspect of running time, although the multi-threshold segmentation process is added, the task amount of fuzzy clustering later is reduced, so that the running time is still much faster than that of clustering by directly using the FLICM algorithm.

TABLE 1Ottawa regional flood disaster change detection results

And (2) simulation experiment II: the results of the three methods for detecting the change in this problem are shown in FIGS. 3(d) to (f), and the specific comparative analysis of the results is shown in Table 2. As can be seen from Table 2, the number of missing pixels is 623 and 530 less than that of FCM and FLICM, respectively, and the total number of error pixels is 149 and 383 less than that of FCM and FLICM, respectively. The invention organically combines multi-threshold segmentation only considering gray information and an FLICM algorithm mainly considering neighborhood information, improves the anti-noise performance, greatly reduces the number of missed detections and ensures that the overall detection performance is superior to that of the two compared methods. Also from Table 2, it can be seen that the runtime of the present invention is about 30% of the FLICM algorithm.

TABLE 2 Mexico area fire change detection results

And (3) simulation experiment III: the results of the three methods for detecting the change in this problem are shown in FIGS. 4(d) to (f), and the specific comparative analysis of the results is shown in Table 3. As can be seen from table 3, for the change detection problem, the present invention still has a significant advantage in the number of missed detections, so the total number of error pixels is also small, only 112, the error detection rate is only 0.12%, and the running time is also short, and is less than 25% of the flimc algorithm.

TABLE 3Bern urban flood change test results

The average run time reduction of the present invention in the above three sets of simulation experiments was calculated to be about 70% lower than the FLICM algorithm using the following formula:

$T=\frac{T_2-T_1}{T_2}$

where T represents the reduction ratio of the runtime of the present invention compared to the FLICM algorithm, T₁Represents the running time, T, of the invention₂Representing the runtime of the FLICM algorithm.

3. Threshold number of parametric analysis simulation experiments

The selection of the number n of thresholds when performing multi-threshold segmentation in the present invention has a large influence on the effect of change detection. In order to select the size of the threshold number n reasonably, the threshold number n is subjected to parameter analysis, the method of the invention is independently operated for 10 times under different threshold numbers, and the detection result after averaging is shown in table 4. As can be seen from table 4, the threshold number n being too small will significantly increase the false detection rate of change detection; an excessive threshold number n causes an increase in the operation time. Taking the threshold number n to be 3 is a reasonable choice after comprehensive consideration.

TABLE 4 results of parameter analysis simulation experiments for threshold number n in the present invention

Claims (5)

1. A SAR image change detection method combining multi-threshold segmentation and fuzzy clustering comprises the following steps:

(1) median filtering

Selecting a common 3 x 3 median filter to preprocess the two SAR images to be detected to obtain two median-filtered images;

(2) obtaining a normalized logarithmic ratio difference image

2a) Calculating a logarithmic ratio difference image from the two median-filtered images by adopting a logarithmic ratio difference formula;

2b) carrying out normalization processing on the log ratio difference image by adopting a normalization formula to obtain a normalized log ratio difference image;

(3) multi-threshold segmentation

And performing multi-threshold segmentation on the normalized logarithmic ratio difference image by adopting an Otsu method based on standard particle swarm optimization, wherein pixels in the segmented image are divided into three types: pixels that have not changed, pixels that have changed, and pixels that cannot determine whether a change has occurred;

(4) initializing membership matrix

4a) Initializing membership matrix U₀，U₀Storing membership degree of pixel under non-variable class, wherein the membership degree is in interval [0, 1 ]]The real number of the up value is used for representing that the pixel which is not changed after multi-threshold segmentation is in U₀The membership degree in (1) is selected, and after multi-threshold segmentation, the membership degree belongs to the changed pixels in U₀The membership degree in (1) is 0, and after multi-threshold segmentation, the pixel which can not determine whether change occurs is in U₀The membership degree in (1) is randomly generated;

4b) initializing membership matrix U₁，U₁Storing the membership degree of the pixel under the variation class, wherein the membership degree is in the interval [0, 1 ]]The real number of the up value is used for representing that the pixel which is not changed after multi-threshold segmentation is in U₁The membership degree in (1) is 0, and after multi-threshold segmentation, the membership degree belongs to the changed pixels in U₁The membership degree in (1) is 1, and after multi-threshold segmentation, the pixel which can not determine whether the change occurs belongs to U₁The degree of membership in (1) is then equal to 1 minus the pixel in U₀Degree of membership in (1);

(5) fuzzy clustering

Performing fuzzy clustering on pixels which cannot determine whether change occurs after multi-threshold segmentation by using an FLICM (flash memory integration control algorithm), and performing iterative updating until a preset termination condition is reached, namely the maximum change of membership degrees before and after iteration is less than 0.00001;

(6) deblurring

According to the membership value of each pixel in the normalized log ratio difference image, which is subordinate to a non-change class and a change class, the pixel is judged to be the class with a larger membership value, so that all the pixels in the normalized log ratio difference image are divided into two classes: a non-change class and a change class;

(7) and outputting a change detection result.

2. The method for detecting changes in SAR images combining multi-threshold segmentation and fuzzy clustering according to claim 1, wherein the logarithmic ratio difference formula in step 2a) is as follows:

I₃＝|log(I₁+1)-log(I₂+1)|

wherein, I₃Representing pixel gray values of a logarithmic ratio difference image, I₁And I₂Respectively representing the pixel gray values of the two images after median filtering.

3. The method for detecting changes in SAR images combining multi-threshold segmentation and fuzzy clustering according to claim 1, wherein the normalization formula in step 2b) is as follows:

$I_D=\frac{255\&times;(I_3-I_{\min })}{(I_{\max }-I_{\min })}$

wherein, I_DRepresenting pixel gray values, I, of the normalized log ratio difference image₃Representing pixel gray values of a logarithmic ratio difference image, I_maxIs represented by₃Zhongji (Chinese character of 'Zhongji')Large gray value, I_minIs represented by₃The smallest gray value.

4. The SAR image change detection method combining multi-threshold segmentation and fuzzy clustering according to claim 1, characterized in that the step (3) of performing multi-threshold segmentation on the normalized logarithmic ratio difference image based on the Otsu method of standard particle swarm optimization comprises the following specific steps:

the method comprises the steps of firstly, encoding n thresholds used for segmenting an image into one n-dimensional particle in standard particle swarm optimization;

secondly, using standard particle swarm optimization search to make the maximum n thresholds of the Otsu method objective function, wherein the Otsu method objective function is as follows:

Maximize σ²＝ω₀(μ₀-μ_T)²+ω₁(μ₁-μ_T)²+...+ω_n(μ_n-μ_T)²

the normalized logarithmic ratio difference image may be divided into n +1 parts by n thresholds, in the above formula, σ²Represents the variance, ω, between the n +1 parts₀，ω₁，...，ω_nRespectively representing the probability of occurrence of the respective portions, mu₀，μ₁，...，μ_nRespectively representing the mean value of the gray levels, mu, of the respective parts_TA gray level average value representing the entire image;

the updating formula of the d-dimension flight speed of the particles in the standard particle swarm optimization is as follows:

$v_i^d=\mathrm{\&omega;}{v}_i^d+c_1{\mathrm{rand}}_1^d({\mathrm{pBest}}_i^d-x_i^d)+c_2{\mathrm{rand}}_2^d({\mathrm{gBest}}^d-x_i^d)$

wherein,

representing the d-dimensional flight velocity of the ith particle in the population, omega representing the inertial weight, c₁And c₂Which represents a factor of learning that is,

representing the d-dimensional position of the ith particle in the population,

and

is two independent random numbers, pBest, uniformly distributed between 0 and 1_iIs the historical optimum position of the ith particle, and gBest is the historical optimum position of the whole population;

the updating formula of the d-dimension position of the particle in the standard particle swarm optimization is as follows:

$x_i^d=x_i^d+v_i^d$

wherein,

representing the d-dimensional position of the ith particle in the population,representing the d-dimension flight speed of the ith particle in the population;

thirdly, finding n thresholds which enable the Otsu method target function to be maximum through searching of standard particle swarm optimization, dividing the normalized logarithmic ratio difference image into n +1 parts by the n thresholds, and recording the parts as follows according to the sequence of gray values from small to large: c₀，C₁，...，C_n(ii) a Handle C₀The pixels in (A) are classified as unchanged pixels, and C is set_nThe pixels in (A) are classified as changed pixels, and C is set₁To C_n-1The pixels in (1) are classified as pixels that cannot be determined whether a change has occurred, and thus the pixels in the normalized logarithmic ratio difference image are classified into three types.

5. The SAR image change detection method combining multi-threshold segmentation and fuzzy clustering according to claim 1, characterized in that the implementation of the FLICM fuzzy clustering algorithm in step (5) is as follows:

the FLICM algorithm uses membership degree to represent the degree of possible pixel belonging to a certain class, and simultaneously considers image gray information and spatial neighborhood information in the clustering process, and the target function of the FLICM is as follows:

$J=\underset{i\&Element;N}{\mathrm{\&Sigma;}}\underset{k=1}{\overset{C}{\mathrm{\&Sigma;}}}[u_{\mathrm{ki}}^m{d}^2(x_i,v_k)+G_{\mathrm{ki}}]$

wherein J represents an objective function value, N represents a set of pixels which cannot be determined whether change occurs after multi-threshold segmentation, C is the number of classes with a value of 2, and u is the number of classes with a value of 2_kiRepresenting degree of membership, x, of the ith pixel in the kth class_iRepresenting the gray value of the ith pixel, v_kDenotes the cluster center of the kth class, d²(x_i，v_k) Represents the distance measure from the ith pixel to the cluster center of the kth class, m represents a fuzzy weighting index with a value of 2, G_kiRepresents the blurring coefficient between the ith pixel and the cluster center of the kth class, which is defined as follows:

$G_{\mathrm{ki}}=\underset{\underset{i\&NotEqual;j}{j\&Element;N_i}}{\mathrm{\&Sigma;}}\frac{1}{e_{\mathrm{ij}}+1}{(1-u_{\mathrm{kj}})}^m{d}^2(x_j,v_k)$

wherein G is_kiRepresenting the blurring coefficient between the i-th pixel and the cluster center of the k-th class, N_iRepresenting a local window of size 3 x 3, x, centered on the ith pixel_jThe gray value of the jth pixel is shown and falls into the local window N_iV is a non-central pixel of_kDenotes the cluster center of the kth class, u_kjRepresenting degree of membership of the jth pixel in the kth class, e_ijRepresenting the spatial Euclidean distance, d, between the ith and jth pixels²(x_j，v_k) Representing the distance measurement from the jth pixel to the kth class of clustering center, and m representing a fuzzy weighting index with a value of 2;

the FLICM algorithm adopts an iterative calculation mode to obtain a group of membership values which enable the objective function to be minimum, and the iterative calculation formula of the membership in the FLICM algorithm is as follows:

$u_{\mathrm{ki}}=\frac{1}{\underset{t=1}{\overset{C}{\mathrm{\&Sigma;}}}{\left(\frac{d^2(x_i,v_k)+G_{\mathrm{ki}}}{d^2(x_i,v_t)+G_{\mathrm{ti}}}\right)}^{\frac{1}{m-1}}}$

wherein u is_kiRepresenting the degree of membership of the ith pixel in the kth class, C being the number of classes taking the value 2, x_iRepresenting the gray value of the ith pixel, v_kRepresenting the cluster center of the kth class, v_tDenotes the cluster center of the t-th class, G_kiRepresenting the blurring coefficient between the i-th pixel and the cluster center of the k-th class, G_tiRepresenting the blurring coefficient between the ith pixel and the cluster center of the t-th class, d²(x_i，v_k) Representing a distance measure of the ith pixel to the cluster center of the kth class, d²(x_i，v_t) Representing the distance measurement from the ith pixel to the clustering center of the t-th class, wherein m represents a fuzzy weighting index with the value of 2;

the iterative calculation formula of the clustering center in the FLICM algorithm is as follows:

$v_k=\frac{\underset{i\&Element;N}{\mathrm{\&Sigma;}}{u}_{\mathrm{ki}}^m{x}_i}{\underset{i\&Element;N}{\mathrm{\&Sigma;}}{u}_{\mathrm{ki}}^m}$

wherein v is_kRepresenting the cluster center of the kth class, N representing a set of pixels that cannot be determined whether a change has occurred after multi-threshold segmentation, u_kiRepresenting degree of membership, x, of the ith pixel in the kth class_iExpressing the gray value of the ith pixel, and m represents a fuzzy weighting index with the value of 2;

and carrying out fuzzy clustering on pixels which cannot determine whether change occurs after multi-threshold segmentation by using an FLICM (flash memory integration) algorithm until a preset termination condition is reached, namely the maximum change of membership degrees before and after iteration is less than 0.00001.

Images