CN111340761B - Remote sensing image change detection method based on fractal attribute and decision fusion - Google Patents

Abstract

The invention discloses a remote sensing image change detection method based on fractal attribute and decision fusion, which comprises the following steps: collecting multi-temporal high-resolution remote sensing images; establishing an objective function based on the correlation minimum value between the average scales, determining a scale parameter set of each attribute in a self-adaptive manner through iterative calculation, and extracting a morphological attribute profile with self-adaptive scale parameters; and constructing a multi-feature decision fusion framework, calculating a change intensity index and an evidence confidence index to respectively describe change information and corresponding confidence degree, and fusing the change information of a morphological attribute section and an original spectrum from the adaptive scale parameter by using the multi-feature decision fusion framework to obtain a final change detection image. The method establishes the target function based on the correlation between the minimum average scales, adaptively obtains a group of scale parameters, constructs a multi-feature decision fusion framework on the basis, and improves the decision reliability by reducing the uncertainty of the change information of different sources.

Description

Remote sensing image change detection method based on fractal attributes and decision fusion

Technical Field

The invention belongs to the field of image processing, and in particular relates to a remote sensing image change detection method.

Background Art

With the continuous development of remote sensing systems, change detection (CD) has attracted widespread attention as one of the most important applications in the field of remote sensing. Accurate understanding of land cover change is an important issue in human activities, such as dynamic land use, vegetation health, and environmental monitoring. The widespread use of a new generation of high-resolution sensors (such as IKONOS, Quickbird, and GF2) has further expanded the application scope of CD technology. Compared with medium- and low-resolution remote sensing images, high-resolution remote sensing images (HRRS) contain more land cover spatial information and thematic information, making it feasible to identify different types of complex structures in the scene. However, since objects of various shapes are composed of many pixels and the spectral information is very limited, these characteristics of high-resolution remote sensing images make it difficult for traditional pixel change detection methods based on spectral differences to achieve ideal results.

In order to solve this problem, a large number of studies have introduced spatial structural information as a supplement. It turns out that this information is very effective in improving the recognition ability of CD in HRRS images. In the existing literature, supervised machine learning methods are the most widely used in CD. However, these methods require a large number of training samples to determine the model parameters to avoid overfitting. At the same time, scholars have proposed a variety of unsupervised spatial structural information extraction methods for CD in HRRS. These studies have adopted different strategies, such as object-based methods, linear transformation-based methods, Markov Random Field (MRF)-based methods, multi-scale analysis methods, and change intensity index methods. In recent years, in order to cope with the detail information that is meaningless or even unfavorable to CD due to the improvement of remote sensing image resolution, morphological attribute profiles (MAPs) have been introduced into the application of CD.

As one of the most effective methods in spatial modeling of HRRS images, the operators in MAPs can effectively achieve multi-scale representation of land cover through a tree structure. Compared with the traditional feature extraction strategy based on a given filter window, MAPs can expand the analysis unit to all connected pixels with similar attributes, which helps to accurately extract the spatial structure information of the object to which the pixel belongs. In addition, in CD applications, MAPs have also been proven to be effective in reducing image complexity and extracting spatial structure information. Even so, most MAPs-based CD methods still have some problems: (1) In order to highlight representative spatial structure information and reduce redundant information in a limited number of attribute profiles (APs), a set of reasonable scale parameters needs to be adaptively determined. However, MAPs theory does not give a clear standard, and most of the current scale parameters are manually determined based on experience. (2) Given the complexity of land cover changes in scenes, when combining change information from multiple APs and other features, existing studies rarely consider the uncertainty contained in change information from different sources.

Summary of the invention

In order to solve the technical problems mentioned in the above background technology, the present invention proposes a remote sensing image change detection method based on fractal attributes and decision fusion.

In order to achieve the above technical objectives, the technical solution of the present invention is:

The remote sensing image change detection method based on fractal attributes and decision fusion includes the following steps:

(1) Collect multi-temporal and high-resolution remote sensing images;

(2) Establish an objective function based on the minimum value of the average inter-scale correlation, adaptively determine the scale parameter set of each attribute through iterative calculation, and extract the morphological attribute profile with adaptive scale parameters;

(3) A multi-feature decision fusion framework is constructed to calculate the change intensity index and the evidence confidence index to describe the change information and the corresponding trust level respectively. The multi-feature decision fusion framework is used to fuse the change information of the morphological attribute profile and the original spectrum from the adaptive scale parameter to obtain the final change detection map.

Furthermore, in step (2), four morphological attributes, namely area, diagonal, standard deviation and normalized moment of inertia, are selected.

Furthermore, in step (2), the adaptive scale parameter extraction method is as follows:

(201) Set the total number of scales for each attribute to W, and the value interval of the scale parameter to [T _min , T _max ], where T _min and T _max are the minimum and maximum values that the scale parameter can take, respectively;

(202) Calculate the interval Sub _w . The w-th scale parameter should be within the interval Sub _w , w ∈ {1, 2, ..., W}:

(203)Define the objective function:

Iteratively calculate all combinations of scale parameters, and take the combination corresponding to the minimum GRSIM _sum as the optimal scale parameter set for extraction; where GRSIM _w,w+1 represents the gradient similarity of two adjacent attribute profiles:

和

表示两个影像的梯度幅度矩阵的方差，σ_M1,M2两个影像的梯度方向矩阵的协方差。In the above formula, σ _Z1 and σ _Z2 represent the standard deviations of the gradient magnitude matrices of the two images, σ _M1 and σ _M2 represent the standard deviations of the gradient direction matrices of the two images,

and

represents the variance of the gradient magnitude matrices of the two images, and σ _{M1, M2} represents the covariance of the gradient direction matrices of the two images.

Furthermore, in step (3), the method for calculating the change intensity index is as follows:

(301) extracting differential images between different time attribute sections under the same scale parameter through differential processing, and obtaining a differential image set of morphological attribute sections of each attribute based on the adaptive scale parameter;

(302) extracting the differential images between images of the same band at different times through differential processing, and obtaining a differential image set based on the original spectrum;

(303) In the differential image, the grayscale value of pixel i reflects the possibility of whether pixel i is a changed pixel. Therefore, it is normalized in the interval [0, 255] and used as the change intensity index corresponding to pixel i. Based on the differential image set obtained in steps (301) and (302), multiple groups of change intensity indexes corresponding to pixel i based on different attributes and original spectra are obtained.

Furthermore, in step (3), the evidence confidence index is calculated as follows:

In the above formula, CIE is the evidence confidence index.

Furthermore, in step (3), the method of constructing a multi-feature decision fusion framework is as follows:

The decision fusion framework is defined as Θ:{CT,NT}, where Θ represents the hypothesis space, CT and NT represent the changed pixels and unchanged pixels respectively. For each pixel i, the basic probability assignment formula is established by the following formula:

m _n ({CT}) = CII _n × CIE _n

m _n ({NT})=(1-CII _n )×CIE _n

m _n ({CT, NT}) = 1 - CIE _n

In the above formula, _CIIn and _CIEn represent the nth change intensity index and evidence confidence index corresponding to pixel i, _mn ({CT}), _mn ({NT}) and _mn ({CT,NT}) represent the basic probability distribution formulas corresponding to the nth group of evidence of the non-empty subsets {CT}, {NT} and {CT,NT};

The basic probability distribution formulas m({CT}), m({NT}) and m({CT,NT}) corresponding to the non-empty subsets {CT}, {NT} and {CT,NT} are calculated using the following formula:

In the above formula, A represents a non-empty subset, N represents the total number of evidences, m _n (F _n ) represents the basic probability distribution formula obtained by the nth group of evidence, and F _n ∈ 2 ^Θ ,

Establish the following judgment rules:

If pixel i satisfies the above judgment rule, pixel i is judged as a changed pixel, otherwise pixel i is judged as an unchanged pixel; all pixels are traversed to obtain the final transformation detection map.

The beneficial effects brought by adopting the above technical solution are:

By establishing an objective function based on the minimum average inter-scale correlation, the present invention can adaptively obtain a set of scale parameters to extract representative APs while reducing redundant information. On this basis, a multi-feature decision fusion framework based on D-S theory is constructed to improve the reliability of decision making by reducing the uncertainty of change information from different sources. The effectiveness of the method of the present invention is verified through experiments on multi-phase HRRS image datasets.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a flow chart of the method of the present invention;

Figure 2 shows the effect of different scale numbers W on the overall accuracy OA in the experiment.

DETAILED DESCRIPTION

The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

Change detection is crucial for accurately understanding surface changes using multi-temporal earth observation data. Due to its great advantages in spatial information modeling, morphological attribute profiles are increasingly favored in improving the accuracy of change detection. However, most change detection methods based on morphological attribute profiles are achieved by manually setting the scale parameters of the attribute profiles, and ignore the uncertainty of change information from different sources. In response to these problems, the present invention proposes a new high-resolution remote sensing image change detection method based on morphological attribute profiles and decision fusion. By establishing an objective function based on the minimum value of the average inter-scale correlation, a morphological attribute profile with adaptive scale parameters (Morphological Attribute Profiles with Adaptive Scale Parameters, ASP-MAPs) is proposed to mine spatial structural information. On this basis, a multi-feature decision fusion framework based on the Dempster-Shafer (D-S) theory is constructed to obtain change detection results. The method flow of the present invention is shown in Figure 1.

(1) MAPs theory

MAPs theory is developed on the basis of set theory. The theory uses spectral similarity and spatial connectivity as basic analysis units, extracts connected regions corresponding to pixels, and designs multi-scale operators with different attributes. The calculation process of MAPs is briefly described as follows: Let B be a grayscale image, i be a pixel in B, k represents grayscale, and then the binary image can be obtained.

Traverse all pixels in B to obtain the image sequence Th _k (B), and set Γ ⁱ (B) = max(k) as the result of the i-open operation. On this basis, using the symmetry of the attribute transformation, the i-closed operation result Φ ⁱ (B) = min(k) is obtained. Let T _w ∈ {T ₁ ,T ₂ ,...,T _W } be the w-th scale parameter, W be the total number of scales, and the open profile Ψ(Γ(B)) and the closed profile Ψ(Φ(B)) are expressed as follows:

Finally, by combining Ψ(Γ(B)) and Ψ(Φ(B)), MAPs can be obtained.

(2) Adopted attributes

Based on the research results related to MAPs, four attributes, namely area, diagonal, standard deviation, and normalized moment of inertia (NMI), are used in the present invention. The effectiveness of these attributes in HRRS image classification and CD applications has been proven.

For the connected area corresponding to pixel i: Area represents the area size; Diagonal represents the diagonal length of the minimum external matrix; Standard Deviation represents the degree of grayscale change; NMI reflects the shape and the position of the center of gravity.

(3) Construction of ASP-MAPs

As shown in Figure 1, in the process of constructing ASP-MAPs, the scale parameter is first determined: among a limited number of APs with different scale parameters, the constructed APs should highlight the representative spatial structural characteristics of typical objects in the scene, thereby improving the ability to recognize changes in such objects; in addition, reducing redundant information between APs also requires a reasonable set of scale parameters. On this basis, according to the principle that the smaller the average inter-scale correlation of APs, the stronger the representativeness of APs. The specific process of ASP-MAPs construction is as follows:

Gradient Similarity (GRSIM): In order to measure the scale correlation of APs, it is necessary to select an appropriate similarity metric. According to the MAPs theory, pixels that meet the attribute range determined by the corresponding scale parameter have the largest grayscale response, that is, they appear as newly generated edges (or objects). Therefore, the similarity metric used should be sensitive to edge changes. Based on the above analysis, the present invention proposes a similarity metric GRSIM based on gradient vectors: a third-order Sobel filter ^[31] is used to extract gradient information, and the GRSIM index between images B1 and B2 is defined as follows:

和σ_M1,M2分别表示标准差、方差和协方差。GRSIM_B1,B2的值越大，B1和B2之间的相关性越高。Where _Z1 and _Z2 represent the gradient magnitude matrices of B1 and B2 respectively; M1 and M2 represent the gradient direction matrices of _B1 and _B2 respectively.

and σ _M1,M2 represent the standard deviation, variance, and covariance, respectively. The larger the value of _{GRSIM B1,B2,} the higher the correlation between B1 and B2.

On this basis, the steps of the adaptive scale parameter extraction strategy are as follows:

Step 1: Set the interval [T _min , T _max ] and the number of scales W for each attribute, and adaptively search for the optimal scale parameter set. Set the area interval to [500, 28000], the diagonal interval to [10, 100], the standard deviation interval to [10, 70], the NMI interval to [0.2, 0.5], and W not more than 10. In addition, based on multiple sets of experimental results, it is recommended in the present invention that W be set to 6.

Step 2: To avoid falling into the local optimum, the Wth (w∈{1,2,...,W}) scale parameter should be within the interval Sub _w . Set Sub _w according to equation (4):

Step 3: The objective function is defined as follows:

Among them, GRSIM _w,w+1 represents the GRSIM of two adjacent APs. According to equations (3)-(5), all combinations of scale parameters are iteratively calculated, and the combination corresponding to the minimum GRSIM _sum is taken as the optimal scale parameter set extracted. On this basis, the ASP-MAPs of multi-temporal images are obtained according to the above equation (2).

(4) Description of change information based on the change intensity index CII

In order to uniformly describe the change information extracted from ASP-MAPs and original spectra, the calculation steps of the change intensity index CII are as follows:

Step 1: Through differential processing, the differential images between APs at different times under the same scale parameters are extracted to obtain a set of differential images based on ASP-MAPs for each attribute.

Step 2: Through differential processing, the differential images between images of the same band at different times are extracted to obtain a set of differential images based on the original spectrum.

Step 3: In the difference image, since the gray value of pixel i reflects the possibility of whether i is a changed pixel, it is normalized in the interval [0,255] and used as one of the CIIs corresponding to i. CIIs are calculated based on ASP-MAPs and all bands in the original image, and then five sets of CIIs can be obtained based on area, diagonal, standard deviation, NMI, and the original spectrum corresponding to i.

(5) Multi-feature decision fusion

D-S theory is a decision theory for multi-source evidence fusion. One of the significant advantages of D-S theory is that it has a strong quantitative evaluation capability for the uncertainty of multi-source evidence. Therefore, the present invention constructs a decision fusion framework for fusing the change information from ASP-MAPs and original spectra.

Basic Probability Assignment Formula (BPAF): According to DS theory, A is represented as a non-empty subset of ^2Θ , Θ is represented as the hypothesis space, and the BPFA of A is represented as m(A). m: ^2Θ →[0,1]BPAF should satisfy the following constraints:

Where m(A) represents the trust level of A, and m(A) is calculated as follows:

N represents the total number of evidences, m _n (F _n ) represents the basic probability distribution formula obtained from the nth group of evidence, and F _n ∈ 2 ^Θ ,

Calculation of CIE: In order to measure the trust level of CIIs from different sources (including area, diagonal, standard deviation, NMI, and original spectrum), a confidence index of evidence CIE is proposed. For each piece of evidence, CIE can be calculated using equation (8). For each CII, a larger CIE indicates that a higher degree of relief should be given in the decision fusion process.

Construction of decision fusion framework: The decision fusion framework is defined as Θ:{CT,NT}, where CT and NT represent changed pixels and unchanged pixels, respectively. Therefore, the non-empty subsets include {CT}, {NT} and {CT,NT}. For each pixel i, the BPAF can be established by the following formula:

m _n ({CT})=CII _n ×CIE _n (9)

m _n ({NT})=(1-CII _n )×CIE _n (10)

m _n ({CT, NT}) = 1 - CIE _n (11)

Where CII _n and CIE _n represent the nth CII and CIE corresponding to pixel i. On this basis, m({CT}), m({NT}) and m({CT,NT}) of pixel i are calculated by equation (7), and the judgment rules are as follows:

If i satisfies the above rules, it is determined to be a changed pixel, otherwise i is determined to be an unchanged pixel. Finally, all pixels are traversed according to the above decision process to obtain the CD map.

(6) Experiment and analysis

Dataset 1 is a set of aerial remote sensing images of Nanjing, China, including red, green and blue bands; the images were collected in March 2009 and February 2012, with a spatial resolution of 0.5 meters and an image size of 512×512 pixels. Dataset 2 is a set of Quickbird images of Chongqing, China, including red, green and blue bands; the images were collected in September 2007 and August 2011, with a spatial resolution of 2.4 meters and an image size of 512×512 pixels. Dataset 3 is a set of SPOT-5 panchromatic multispectral fusion images of Shanghai, China, including red, green and blue bands; the images were collected in June 2004 and July 2008, with a spatial resolution of 2.5 meters and an image size of 512×512 pixels. The reasons for selecting these three datasets for experiments are as follows: these datasets represent different urban scenes, mainly composed of buildings, roads, vegetation, wasteland, etc., which helps to verify the recognition ability of the proposed method for these typical landform changes and helps to evaluate the applicability and stability of the method in CD applications.

In order to comprehensively evaluate the performance of the method proposed in the present invention, five advanced CD methods were used for comparative experiments: improved change vector analysis (CVA), including CVA-Expectation Maximum (CVA-EM) (method 1), method based on spectral angle mapping (method 2), method based on spectral and texture features (method 3); method based on MAPs (method 4); method based on deep learning (DL) (method 5). The adaptive extraction scale parameter set of the method proposed in the present invention is shown in Table 1-3.

Table 1 Extracted dataset 1 Scale parameter set

Table 2 Extracted dataset 2 scale parameter set

Table 3 Extracted dataset 3 scale parameter sets

The quantitative evaluation results of different methods are shown in Tables 4-6. Among the three sets of data, the overall accuracy (OA) of the method proposed in the present invention can reach more than 83.9%, and the fluctuation range is less than 1.5%, which is significantly better than other comparison methods. Therefore, in the challenges brought by different data sources, the present invention has the advantages of high accuracy and good stability.

Among the three CVA-based CD methods, Method 1 and Method 2 only use spectral differences as the basis for CD, with false positive (FP) rates and false negative (FN) rates exceeding 30% and 20%, respectively. Due to the introduction of texture features as a supplement, the three evaluation indicators have been significantly improved in the experimental results of Method 3. Therefore, in order to generate more accurate CD maps, it is necessary to consider utilizing the spatial neighborhood information of pixels. Nevertheless, Method 3 uses an artificial setting to define a series of specified filter windows to extract texture features, which makes it difficult to keep consistent with the inherent shape of the corresponding object to which the current pixel belongs. In contrast, MAPs can extract more accurate spatial structure information based on non-fixed local areas composed of all connected pixels with similar attributes.

Compared with the method proposed in the present invention, although method 4 uses APs to extract change information, the results of OA are significantly reduced in the three data sets of the method proposed in the present invention, with a fluctuation range of more than 8%. This is mainly due to the manual setting of the scale parameters in method 4, which ignores the redundant information contained in APs and highlights the representative spatial structure information. In addition, this method uses a single threshold to obtain the final CD map based on the change information from different sources, thereby ignoring the uncertainty of multi-source change information.

Method 5 is a D-L based method. In the comparative experiments of the present invention, method 5 showed lower accuracy and poorer stability in all three data sets. Without sufficient training samples, DP based methods cannot obtain reliable results in CD applications. However, it is certain that the performance of method 5 will be significantly improved with the increase of training samples.

Table 4 Quantitative evaluation of CD accuracy in dataset 1. OA: overall accuracy; FP: false positive rate; FN: missed positive rate

Methods/Indicators OA (%) FP(%) FN(%) Evaluation Criteria The higher the better The lower the better The lower the better Method of the present invention 83.9 15.1 9.1 Method 1 57.2 40.4 39.1 Method 2 63.5 32.3 25.2 Method 3 79.8 19.3 11.9 Method 4 71.2 28.5 19.4 Method 5 77.1 21.4 15.3

Table 5 Quantitative evaluation of CD accuracy in dataset 2

Methods/Indicators OA (%) FP(%) FN(%) Evaluation Criteria The higher the better The lower the better The lower the better Method of the present invention 84.5 12.6 9.8 Method 1 68.4 39.1 34.9 Method 2 72.8 30.6 29.8 Method 3 81.5 15.3 11.4 Method 4 74.8 26.5 24.4 Method 5 51.1 46.6 42.8

Table 6 Quantitative evaluation of CD accuracy in dataset 3

Methods/Indicators OA (%) FP(%) FN(%) Evaluation Criteria The higher the better The lower the better The lower the better Method of the present invention 85.1 13.9 10.9 Method 1 59.4 40.2 39.7 Method 2 68.6 30.3 31.6 Method 3 78.1 21.9 17.4 Method 4 80.2 19.4 15.8 Method 5 71.4 26.4 27.8

In order to verify the effectiveness of the adaptive scale parameter extraction strategy and decision fusion framework proposed in this paper, the following two experimental schemes were carried out: (1) the scale parameters of area, diagonal, standard deviation and NMI were manually set to {100, 918, 1734, 2548, 3368, 4185, 5000}, {10, 25, 40, 55, 70, 85, 100}, {0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5} and {20, 25, 30, 35, 40, 45, 50}, respectively, and the rest of the steps were consistent with the proposed method (method 6); (2) the extracted CIIS corresponding to pixel i were averaged, and all pixels in the image were traversed, and the threshold for obtaining the CD map was determined using the EM method (method 7). Table 7 lists the OA of different methods.

Table 7 OA of the proposed method, method 6 and method 7

As shown above, the OA of the proposed method is significantly higher than that of the other two methods. Therefore, the adaptive scale parameter extraction strategy and decision fusion framework proposed in the present invention are necessary and effective for improving the accuracy of change detection: the former helps to highlight representative spatial structure information while reducing redundant information in APs; the latter can improve the reliability of decision making by reducing the uncertainty of change information from different sources.

In the adaptive scale parameter extraction process proposed in this paper, the scale number W is the only subordinate parameter that needs to be manually set. In order to clarify the basis for setting W, this chapter analyzes the impact of different OA. As shown in Figure 2, the horizontal axis is W and the vertical axis is OA. The results of the three data sets are represented by curves of different styles.

As shown in Figure 2, in the experiments of the three data sets, with the continuous increase of W, OA shows a similar trend of gradually rising first, stabilizing and then gradually decreasing. Among them, W = 6, W = 4, and W = 6 correspond to the peak values of the OA curve in the experiments of data sets 1, 2, and 3, which are 83.9%, 84.9%, and 85.1%, respectively. The detailed values are shown in Table 8.

Table 8 Detailed W-OA values in the experiments of three datasets

As shown in the table above, in the experiment of dataset 2, when W is set to 6, OA can reach 84.5%, and is only 0.4% lower than the corresponding highest OA. This means that setting W to 6 can get ideal results in all experiments of the three datasets. Therefore, considering automation and reliability, it is recommended to set W directly to 6 in CD applications.

The embodiments are only for illustrating the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims (4)

1. A remote sensing image change detection method based on fractal attributes and decision fusion, characterized in that it comprises the following steps: (1) Collect multi-temporal and high-resolution remote sensing images; (2) Establish an objective function based on the minimum value of the average inter-scale correlation, adaptively determine the scale parameter set of each attribute through iterative calculation, and extract the morphological attribute profile with adaptive scale parameters; (3) Construct a multi-feature decision fusion framework, calculate the change intensity index and the evidence confidence index to describe the change information and the corresponding trust level respectively, and use the multi-feature decision fusion framework to fuse the change information of the morphological attribute profile and the original spectrum from the adaptive scale parameter to obtain the final change detection map; In step (2), the adaptive scale parameter extraction method is as follows: (201) Set the total number of scales for each attribute to W, and the value interval of the scale parameter to [T _min , T _max ], where T _min and T _max are the minimum and maximum values that the scale parameter can take, respectively; (202) Calculate the interval Sub _w . The w-th scale parameter should be within the interval Sub _w , w ∈ {1, 2, ..., W}:

(203)Define the objective function:

Iteratively calculate all combinations of scale parameters, and take the combination corresponding to the minimum GRSIM _sum as the optimal scale parameter set for extraction; where GRSIM _w,w+1 represents the gradient similarity of two adjacent attribute profiles:

In the above formula, GRSIM _{B1, B1} is the gradient similarity between the two images B1 and B2; σ _Z1 and σ _Z2 represent the standard deviations of the gradient magnitude matrices of the two images, σ _M1 and σ _M2 represent the standard deviations of the gradient direction matrices of the two images,

and

represents the variance of the gradient magnitude matrices of the two images, σ _M1,M2 represents the covariance of the gradient direction matrices of the two images; In step (3), the method of constructing a multi-feature decision fusion framework is as follows: The decision fusion framework is defined as Θ:{CT,NT}, where Θ represents the hypothesis space, CT and NT represent the changed pixels and unchanged pixels respectively. For each pixel i, the basic probability assignment formula is established by the following formula: m _n ({CT}) = CII _n × CIE _n m _n ({NT})=(1-CII _n )×CIE _n m _n ({CT, NT}) = 1 - CIE _n In the above formula, _CIIn and _CIEn represent the nth change intensity index and evidence confidence index corresponding to pixel i, _mn ({CT}), _mn ({NT}) and _mn ({CT,NT}) represent the basic probability distribution formulas corresponding to the nth group of evidence of the non-empty subsets {CT}, {NT} and {CT,NT}; The basic probability distribution formulas m({CT}), m({NT}) and m({CT,NT}) corresponding to the non-empty subsets {CT}, {NT} and {CT,NT} are calculated using the following formula:

In the above formula, A represents a non-empty subset, N represents the total number of evidences, m _n (F _n ) represents the basic probability distribution formula obtained by the nth group of evidence, and F _n ∈ 2 ^Θ ,

Establish the following judgment rules:

If pixel i satisfies the above judgment rule, pixel i is judged as a changed pixel, otherwise pixel i is judged as an unchanged pixel; all pixels are traversed to obtain the final transformation detection map. 2. According to claim 1, the remote sensing image change detection method based on fractal attributes and decision fusion is characterized in that, in step (2), four morphological attributes of area, diagonal, standard deviation and normalized moment of inertia are selected. 3. The remote sensing image change detection method based on fractal attributes and decision fusion according to claim 1 is characterized in that, in step (3), the method for calculating the change intensity index is as follows: (301) extracting differential images between different time attribute sections under the same scale parameter through differential processing, and obtaining a differential image set of morphological attribute sections of each attribute based on the adaptive scale parameter; (302) extracting the differential images between images of the same band at different times through differential processing, and obtaining a differential image set based on the original spectrum; (303) In the differential image, the grayscale value of pixel i reflects the possibility of whether pixel i is a changed pixel. Therefore, it is normalized in the interval [0, 255] and used as the change intensity index corresponding to pixel i. Based on the differential image set obtained in steps (301) and (302), multiple groups of change intensity indexes corresponding to pixel i based on different attributes and original spectra are obtained. 4. The remote sensing image change detection method based on fractal attributes and decision fusion according to claim 1 is characterized in that, in step (3), the evidence confidence index is calculated as follows:

In the above formula, CIE is the evidence confidence index, and GRSIM _w,w+1 represents the gradient similarity of two adjacent attribute profiles.

Images