A multi-objective endmember identification method based on meme optimization

Abstract

The present invention discloses a multi-objective endmember identification method based on meme optimization, comprising: obtaining an original hyperspectral image of an endmember to be identified; constructing a self-dictionary based on the data of the original hyperspectral image and pruning the pixels of the hyperspectral image using vertex component analysis based on a pure pixel hypothesis; constructing a multi-objective optimization model for endmember identification of hyperspectral images based on the self-dictionary; obtaining a non-dominated solution set of the multi-objective optimization problem model based on a meme optimization algorithm; and determining an optimal endmember combination based on a knee method using the non-dominated solution set of the multi-objective optimization problem model. The present invention can integrate the global search capability of an evolutionary algorithm and the local optimization capability of a local search operator to improve the accuracy and efficiency of endmember identification, and can also determine an optimal endmember combination from a non-dominated solution set based on a knee method.

Description

Technical Field

[0001] The invention belongs to the technical field of hyperspectral image processing, and in particular relates to a multi-target endmember recognition method based on meme optimization. Background Art

[0002] As the culmination of the development of remote sensing technology, hyperspectral imaging is able to image with high resolution in the ultraviolet to infrared wavelength range of the electromagnetic spectrum. Therefore, hyperspectral images can not only provide spatial distribution information of objects, but also provide their spectral scattering information. Since each substance has unique spectral absorption and scattering characteristics, hyperspectral images have good differentiation and recognition capabilities for objects. Therefore, hyperspectral images have been widely used in tasks such as detection, recognition and classification of objects, and have successfully served in the fields of agricultural and forestry resource assessment, geological exploration, hydrological monitoring, military reconnaissance and medical testing.

[0003] Due to the insufficient spatial resolution of hyperspectral detectors and the complexity of ground object distribution, the pixel spectra of hyperspectral images are usually mixed by spectra of multiple basic substances. Common hyperspectral image mixing models include linear, bilinear and complex nonlinear mixing models. In order to more accurately quantify the basic material composition of hyperspectral images and improve the accuracy of hyperspectral image processing tasks, it is necessary to unmix the hyperspectral images and extract the spectra (endmembers) of the basic substances contained in the images and their proportion (abundance) in the image pixels. Therefore, hyperspectral image unmixing includes two subtasks: endmember extraction and abundance estimation. These two subtasks can be performed simultaneously or successively.

[0004] In the endmember identification task of hyperspectral images, it is usually necessary to predetermine the number of endmembers contained in the image. Traditional endmember number estimation methods include virtual dimension method, signal subspace estimation method and vertex component analysis method, but the estimation accuracy of such methods is easily affected by noise and outliers. Although regularization methods can introduce sparse regularization terms without estimating the number of endmembers, such methods need to solve the problem of the value of the regularization parameter.

[0005] Traditional endmember identification methods are usually based on the spatial distribution characteristics of hyperspectral data, using iterative algorithms or greedy algorithms to gradually optimize; or using regularization methods to iteratively optimize. However, these methods cannot fully search the solution space and are prone to fall into local optimality. In recent years, evolutionary optimization algorithms have been successively applied to the endmember identification task of hyperspectral images, and have achieved good global convergence effects, but they also highlight the disadvantage of low search efficiency. Summary of the invention

[0006] In order to solve the above problems existing in the prior art, the present invention provides a multi-target endmember identification method based on meme optimization. The technical problem to be solved by the present invention is achieved by the following technical solutions:

[0007] The present invention provides a multi-target endmember identification method based on meme optimization, comprising:

[0008] S1: Obtain the original hyperspectral image of the endmember to be identified;

[0009] S2: Based on the data of the original hyperspectral image, the pixels of the hyperspectral image are pruned using vertex component analysis based on the pure pixel hypothesis to construct a self-dictionary;

[0010] S3: constructing a multi-objective optimization model for endmember recognition of hyperspectral images based on the self-dictionary;

[0011] S4: Obtaining a non-dominated solution set of the multi-objective optimization problem model based on a memetic optimization algorithm;

[0012] S5: Using the non-dominated solution set of the multi-objective optimization problem model, determine the optimal end member combination based on the knee method.

[0013] In one embodiment of the present invention, the S2 includes:

[0014] S2.1: Generate k unit random vectors u1,u2,…,u with the same dimension as the original hyperspectral image band k ;

[0015] S2.2: All pixels a1, a2, ..., a in the original hyperspectral image are N Projecting onto the k unit random vectors respectively, wherein N represents the number of pixels in the original hyperspectral image;

[0016] S2.3: Count the cumulative number of times each pixel falls on the two ends of the k unit random vectors t1, t2, ..., t N ;

[0017] S2.4: Select the M pixels with the largest cumulative times to form a self-dictionary A = (a1, a2, ..., a M ).

[0018] In one embodiment of the present invention, the multi-objective optimization problem model is

[0019]

[0020] Where Y represents the L×N dimensional original hyperspectral image data, L represents the number of image bands, N represents the number of image pixels, E represents the L×L dimensional unit matrix, A represents the L×M dimensional self-dictionary, I represents the M dimensional 0-1 vector, AI represents the matrix consisting of the columns of A corresponding to the non-zero elements of I, A I The projection matrix on the column space of The calculation formula is: ||I||0 means counting the number of non-zero elements in I.

[0021] In one embodiment of the present invention, the S4 includes:

[0022] S4.1: Set the number of subproblems pop of the multi-objective optimization problem, and construct a set of single-objective subproblems, generating pop weight vectors uniformly between (0,1) and (1,0) Construct a set of single-objective subproblems

[0023]

[0024] in, and is the vector λ i Elements in

[0025] S4.2: According to the Euclidean distance between weight vectors, each subproblem Subp i Find T nearest neighbor subproblems and record their index set as B{i};

[0026] S4.3: Let I i Equal to the M-dimensional standard unit vector e i , initialize the population Calculate the target vector F(I i ), and record the set of target vectors as

[0027] S4.4: Take the individuals in the initial population as the parent individuals, and perform multi-point crossover recombination on each parent individual to obtain new offspring individuals

[0028] S4.5: For the new offspring individuals The elements in the bit-reversal mutation are performed to obtain the mutated individuals

[0029] S4.6: Based on the spatial distribution of hyperspectral image data, construct a heuristic local search operator to find the mutated individuals Make updates;

[0030] S4.7: For all l∈B{i}∪i, if Then And update the individual collection Set of target vectors with individuals

[0031] S4.8: Let the number of iterations Iter = Iter + 1, and repeat steps S4.4 to S4.7 until Iter reaches the set maximum number of iterations, and the iteration is terminated to obtain the final population PS and the target vector PF of the population;

[0032] S4.9: Remove the non-optimal solutions and their targets from the final population PS and the population target vector PF.

[0033] In one embodiment of the present invention, the S4.4 includes:

[0034] For each individual i , select two sub-problem indexes j, k from the set B{i}∪i, and use the multi-point crossover strategy to solve the problem of individual Ι j and k Recombination is performed to obtain new individuals of the offspring Right now

[0035]

[0036] Among them, r represents an M-dimensional random vector, and its d-th element r d The value is in the interval [0,1], cr is a constant, Respectively represent Ι j ,Ι k , The d-th element of .

[0037] In one embodiment of the present invention, the S4.5 includes:

[0038] With a probability of 1 / M, the new individuals of the next generation Each element of is bit-reversed and mutated to obtain the mutated individual

[0039]

[0040] in, represents an M-dimensional random vector, whose dth element The value is in the interval [0,1]. Represents the individual after mutation The d-th element of .

[0041] In one embodiment of the present invention, the S4.6 includes:

[0042] S4.61: Recording individuals The endmembers searched are Calculate the centroid of the end member

[0043] S4.62: Calculate the direction from the center of gravity c The direction vector of each search endmember in in,

[0044] S4.63: Calculate all candidate end elements a1, a2, ..., a in the dictionary A pointed to by the centroid c. M The vectors s1,s2,…,s M , where s k =a k -c,k=1,2,…,M;

[0045] S4.64: Calculate all vectors s1,s2,…,s M In the direction vector d j The projection ε on 1j ,ε 2j ,…,ε Mj , and find the maximum projection ε max =ε tj ,t∈{1,2,…,M};

[0046] S4.65: Determine ε max The size of ε max > 0, then use the candidate end member a from the dictionary t replace The search end member a in j , and will The tth element of is replaced by 1;

[0047] S4.66: Repeat steps S4.64 and S4.65 until all

[0048] In one embodiment of the present invention, the S4.9 includes:

[0049] S4.91: Initialize the non-dominated solution set and its target set

[0050] S4.92: For any Ι i ∈PS, if there is Make Then let PS = PS / Ι i If it does not exist Make Then let PS = P / S1 i , and At the same time, if there is Make Then and in If and only if

[0051]

[0052] S4.93: Repeat step S4.92 until Get the final Pareto optimal set and its target set

[0053] Another aspect of the present invention provides a storage medium storing a computer program for executing the steps of the multi-target endmember identification method based on meme optimization described in any one of the above embodiments.

[0054] Another aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor calls the computer program in the memory, it implements the steps of the multi-target end member identification method based on meme optimization as described in any one of the above embodiments.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] 1. The present invention can integrate the global search capability of the evolutionary algorithm and the local optimization capability of the local search operator to improve the accuracy and efficiency of endmember identification, and can also determine the optimal endmember combination from the non-dominated solution set according to the knee point method. According to the specific steps of the local search operator, it can directly locate the vertex position of the hyperspectral data manifold according to the size of the projection length, and according to the linear mixing mechanism, it can be known that such vertices are most likely to be endmembers contained in the image.

[0057] 2. The population-based evolutionary algorithm has extremely strong global search capabilities, while the local search operator based on the empty spectrum prior can guide the individual search directly to possible representative end members. The present invention organically combines the two and can quickly and accurately search for the optimal end member combination, saving the number of iterative searches and avoiding the waste of search resources.

[0058] 3. From the multi-objective decision-making theory and experimental results, it can be seen that the Pareto front obtained by the memetic optimization algorithm of the present invention has an obvious inflection point area. The knee point in this area usually corresponds to the optimal number and combination of end members. This feature can be used to adaptively determine the optimal end member, which is convenient for practical use.

[0059] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1is a flow chart of a multi-target endmember identification method based on meme optimization provided by an embodiment of the present invention;

[0061] Figure 2 is a hyperspectral image of an end member to be identified provided by an embodiment of the present invention;

[0062] Figure 3 It is a principle diagram of obtaining from a dictionary by vertex component analysis provided by an embodiment of the present invention;

[0063] Figure 4 It is a schematic diagram of a sub-problem construction strategy for a multi-objective optimization problem provided by an embodiment of the present invention;

[0064] Figure 5 This is a schematic diagram of a process of recombination of a parent individual to produce a new offspring individual provided by an embodiment of the present invention;

[0065] Figure 6 It is a schematic diagram of a process of a bit reversal variation operation provided by an embodiment of the present invention;

[0066] Figure 7 is a schematic diagram of updating an individual using local search provided by an embodiment of the present invention;

[0067] Figure 8 is the Pareto front obtained by using the endmember identification method of the embodiment of the present invention;

[0068] Fig. 9 It is a schematic diagram of four angles required for a knee point method provided in an embodiment of the present invention;

[0069] Fig.10 It is the end member and abundance obtained by using the end member identification method of the embodiment of the present invention. DETAILED DESCRIPTION

[0070] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the following is a detailed description of a multi-target endmember identification method based on meme optimization proposed in accordance with the present invention in combination with the accompanying drawings and specific implementation methods.

[0071] The above and other technical contents, features and effects of the present invention are clearly presented in the following detailed description of the specific implementation modes in conjunction with the accompanying drawings. Through the description of the specific implementation modes, the technical means and effects adopted by the present invention to achieve the predetermined purpose can be more deeply and specifically understood. However, the attached drawings are only for reference and explanation purposes and are not used to limit the technical solutions of the present invention.

[0072] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants are intended to cover non-exclusive inclusion, so that an article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed. In the absence of more restrictions, the elements defined by the statement "including one..." do not exclude the existence of other identical elements in the article or device including the elements.

[0073] This embodiment constructs a series of 0-1 vectors equal to the number of self-dictionaries to represent a search combination, and uses multi-point crossover, bit reversal mutation and local search strategies to generate various new search combinations, thereby completing the evaluation of various possible end member combinations and obtaining the optimal end member combination. For details, please refer to Figure 1 , Figure 1 1 is a flowchart of a multi-target endmember identification method based on meme optimization provided by an embodiment of the present invention, the method comprising the following steps:

[0074] S1: Obtain the original hyperspectral image of the endmember to be identified.

[0075] See also Figure 2 , Figure 2 : is a hyperspectral image of an end member to be identified provided by an embodiment of the present invention. The original hyperspectral image provided by this embodiment contains 162 bands, 307×307=94249 pixels.

[0076] S2: Construct a dictionary based on the data of the original hyperspectral image itself, which is called the self-dictionary.

[0077] In the embodiment, step S2 is based on the pure pixel hypothesis and uses a vertex component analysis method to roughly prune the pixels of the hyperspectral image to construct a self-dictionary.

[0078] Specifically, S2 is implemented by the following steps:

[0079] S2.1: Generate k unit random vectors with dimensions equal to the original hyperspectral image bands.

[0080] In this embodiment, k=10 is randomly generated. 4 162-dimensional unit vectors u1,u2,…,u k .

[0081] S2.2: If Figure 3 As shown, all pixels a1, a2, ..., a in the original hyperspectral image are NProject onto the k unit random vectors respectively, wherein N represents the number of pixels in the original hyperspectral image. In this embodiment, N=94249.

[0082] S2.3: Count the number of times each pixel falls on the two ends of k random vectors t1, t2, …, t N That is, count each pixel falling on the random vector u1,u2,…,u k The cumulative number of times at the two ends.

[0083] S2.4: Select the M pixels with the largest cumulative times to form a self-dictionary A = (a1, a2, ..., a M ). Preferably, M=150.

[0084] S3: Constructing a multi-objective optimization model for endmember recognition of hyperspectral images based on the self-dictionary.

[0085] In this embodiment, the goal of step S3 is to identify as few endmembers as possible from the self-dictionary to accurately characterize the hyperspectral image. The multi-objective optimization problem model constructed is:

[0086]

[0087] Where Y is L×N dimensional hyperspectral data, L=162 represents the number of bands, N=94249 represents the number of pixels, E is the L×L dimensional unit matrix, A is the L×M dimensional dictionary, I is the M dimensional 0-1 vector, A I is the matrix consisting of the columns of A corresponding to the non-zero elements of I, A I The projection matrix on the column space of The calculation formula is: ||I||0 means counting the number of non-zero elements in I.

[0088] S4: Obtaining a non-dominated solution set of the multi-objective optimization problem model based on a memetic optimization algorithm, mainly including multi-objective problem decomposition, population initialization, genetic strategy, local search strategy, population update strategy and non-dominated sorting operations.

[0089] Specifically, step S4 includes the following steps:

[0090] S4.1: Set the number of subproblems of the multi-objective optimization problem and construct a set of single-objective subproblems.

[0091] Specifically, see Figure 4 , Figure 4 Schematic diagram of a sub-problem construction strategy for a multi-objective optimization problem provided by an embodiment of the present invention. Set the number of sub-problems pop = 101, and generate pop weight vectors uniformly between (0,1) and (1,0) Construct a set of single-objective subproblems Right now

[0092]

[0093] in, and is the vector λ i The elements in

[0094] The sub-question parameters are set based on experience, and other appropriate values ​​may also be selected.

[0095] S4.2: Set the number of neighborhood subproblems T = 15, and assign each subproblem Subp to the Euclidean distance between weight vectors. i Find T nearest neighbor subproblems and record their index set as B{i};

[0096] S4.3: Let I i Equal to the M-dimensional standard unit vector e i (i.e. a vector whose i-th element is 1 and other elements are 0), initialize the population Calculate the target vector F(I i ), and record the set of target vectors as

[0097] Preferably, the initial number of iterations Iter=1, and the maximum number of iterations Max_Iter=50 are set;

[0098] S4.4: Take the individuals in the initial population as the parent individuals, and perform multi-point crossover recombination on each parent individual to obtain new offspring individuals

[0099] Specifically, for each individual i , choose any two subproblem indices j, k from the set B{i}∪i, such as Figure 5 As shown, a multi-point crossover strategy is used to j and k Recombination is performed to obtain new individuals of the offspring Right now

[0100]

[0101] Among them, r represents an M-dimensional random vector, and its d-th element r d The value is in the interval [0,1], the constant cr=0.5, Respectively represent Ι j ,Ι k , The d-th element of .

[0102] S4.5: For the new offspring individuals The elements in the bit-reversal mutation are performed to obtain the mutated individuals

[0103] Specifically, see Figure 6 , Figure 6 is a schematic diagram of a bit reversal mutation process provided by an embodiment of the present invention, in which the probability of 1 / M is used to perform a bit reversal mutation on the new individual of the offspring. Each element of the bit reversal mutation is performed to obtain the mutated individual Right now

[0104]

[0105] in, is an M-dimensional random vector, whose d-th element The value is in the interval [0,1]. Represents the individual after mutation The d-th element of .

[0106] S4.6: Based on the spatial distribution of hyperspectral image data, construct a heuristic local search operator to find the mutated individuals to update.

[0107] See also Figure 7 , Figure 7 is a schematic diagram of updating an individual using local search provided by an embodiment of the present invention. In the embodiment, step S4.6 includes the following steps:

[0108] S4.61: Recording individuals The endmembers searched are Calculate the centroids of these end members

[0109] S4.62: Calculate the direction from the center of gravity c The direction vector of each search endmember in in, in,

[0110] S4.63: Calculate all candidate end elements a1, a2, ..., a in the dictionary A pointed to by the centroid c. M The vectors s1,s2,…,s M , where s k =a k -c,k=1,2,…,M.

[0111] S4.64: Calculate all vectors s1,s2,…,s M In the direction vector d j The projection ε on 1j ,ε2j ,…,ε Mj , and find the maximum projection ε max =ε tj ,t∈{1,2,…,M};

[0112] S4.65: Determine ε max The size of ε max > 0, then use the candidate end member a from the dictionary t replace Search endmembers in Accordingly, The tth element of is replaced by 1;

[0113] S4.66: Repeat steps S4.64 and S4.65 until all

[0114] S4.7: For all l∈B{i}∪i, if Then And update the individual collection Set of target vectors with individuals

[0115] S4.8: Let the number of iterations Iter = Iter + 1, and repeat steps S4.4 to S4.7 until Iter>Max_Iter is satisfied and the iteration is terminated to obtain the final population PS and the target vector PF of the population;

[0116] S4.9: Remove the dominated solutions (non-optimal solutions) and their targets from the final population PS and the population's target vector PF.

[0117] In this embodiment, step S4.9 is implemented in the following manner:

[0118] S4.91: Initialize the non-dominated solution set and its target set

[0119] S4.92: For any Ι i ∈PS, if there is Make Then let PS = PS / Ι i If it does not exist Make Then let PS = PS / Ι i , and At the same time, if there is Make Then and in If and only if

[0120]

[0121] S4.93: Repeat step S4.92 until The final Pareto optimal set can be obtained and its target set like Figure 8 shown.

[0122] S5: Using the non-dominated solution set of the multi-objective optimization problem model, determine the optimal end member combination based on the knee method.

[0123] In this embodiment, step S5 is implemented in the following manner:

[0124] S5.1: For each non-dominated solution On its Pareto front Find the two nearest points F(Ι j ),F(Ι k ) and F(Ι l ),F(Ι u ), take one point from each of the two groups, and it can form an angle with point F(Ι); according to this method, a total of four angles can be formed, such as Fig. 9 As shown, calculate the average of the four angles;

[0125] S5.2: Compare the average angles of all non-dominated solutions and determine the solution with the smallest average angle as the knee point of the Pareto front, e.g. Figure 8 As shown by the * mark in , the searched end member is used as the final end member identification result, such as Fig.10 In order to better reflect the accuracy of the end member identification results, the abundance graph corresponding to each end member was obtained using the non-negative least squares method, as shown in the first row of graphs in . Fig.10 As shown in the second row of graphs in . Figure 8 It can be seen that the knee point position of the Pareto front can be used to determine the optimal number and combination of end members; Fig.10 It can be seen that the end members and abundances obtained by the method of this embodiment are similar to Figure 2 The material distribution in the sample is consistent, which verifies the correctness of the end member detection results.

[0126] In step S5.2, the four points at the two ends of the Pareto front cannot construct four corners according to step S5.1, so they can be ignored when comparing the average angles. However, this processing strategy does not affect the final endmember identification result when the number of endmembers is greater than 2.

[0127] The present invention can integrate the global search capability of the evolutionary algorithm and the local optimization capability of the local search operator to improve the accuracy and efficiency of endmember identification, and at the same time, the optimal endmember combination can be determined from the non-dominated solution set according to the knee method. According to the specific steps of the local search operator, it can be directly located to the vertex position of the hyperspectral data manifold according to the size of the projection length, and according to the linear mixing mechanism, it can be known that such vertices are most likely to be the endmembers contained in the image. The population-based evolutionary algorithm has a very strong global search capability, and the local search operator based on the space spectrum prior can guide the individual search to directly point to the possible representative endmembers. The present invention organically combines the two, and can quickly and accurately search for the optimal endmember combination, saving the number of iterative searches and avoiding the waste of search resources. From the multi-objective decision-making theory and experimental results, it can be seen that the Pareto front obtained by the meme optimization algorithm of the present invention has an obvious inflection point area, and the knee point in this area usually corresponds to the optimal number and combination of endmembers. This feature can be used to adaptively determine the optimal endmember, which is convenient for practical use.

[0128] Another embodiment of the present invention provides a storage medium, in which a computer program is stored, and the computer program is used to execute the steps of the multi-target end member identification method based on meme optimization described in the above embodiment. Another aspect of the present invention provides an electronic device, including a memory and a processor, in which a computer program is stored, and when the processor calls the computer program in the memory, the steps of the multi-target end member identification method based on meme optimization as described in the above embodiment are implemented. Specifically, the above-mentioned integrated module implemented in the form of a software function module can be stored in a computer-readable storage medium. The above-mentioned software function module is stored in a storage medium, including several instructions for an electronic device (which can be a personal computer, a server, or a network device, etc.) or a processor (processor) to execute some steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), disk or optical disk and other media that can store program codes.

[0129] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A multi-target endmember identification method based on meme optimization, characterized in that: include: S1: Obtain the original hyperspectral image of the endmember to be identified; S2: Based on the data of the original hyperspectral image, the pixels of the hyperspectral image are pruned using vertex component analysis based on the pure pixel hypothesis to construct a self-dictionary; S3: constructing a multi-objective optimization model for endmember recognition of hyperspectral images based on the self-dictionary; S4: Obtaining a non-dominated solution set of the multi-objective optimization problem model based on a memetic optimization algorithm; S5: using the non-dominated solution set of the multi-objective optimization problem model, determining the optimal end member combination based on the knee method; The multi-objective optimization problem model is: Where Y represents the L×N dimensional original hyperspectral image data, L represents the number of image bands, N represents the number of image pixels, E represents the L×L dimensional unit matrix, A represents the L×M dimensional self-dictionary, I represents the M dimensional 0-1 vector, A I represents the matrix consisting of the columns of A corresponding to the non-zero elements of I, A I The projection matrix on the column space of The calculation formula is: ||I||0 means counting the number of non-zero elements in I; The S4 includes: S4.1: Set the number of subproblems pop of the multi-objective optimization problem, and construct a set of single-objective subproblems, generating pop weight vectors uniformly between (0,1) and (1,0) Construct a set of single-objective subproblems Sub i : in, and is the vector λ i Elements in S4.2: According to the Euclidean distance between weight vectors, each subproblem Subp i Find T nearest neighbor subproblems and record their index set as B{i}; S4.3: Let I i Equal to the M-dimensional standard unit vector e i , initialize the population Calculate the target vector F(I i ), and record the set of target vectors as S4.4: Take the individuals in the initial population as the parent individuals, and perform multi-point crossover recombination on each parent individual to obtain new offspring individuals S4.5: For the new offspring individuals The elements in the bit-reversal mutation are performed to obtain the mutated individuals S4.6: Based on the spatial distribution of hyperspectral image data, construct a heuristic local search operator to find the mutated individuals Make updates; S4.7: For all l∈B{i}∪i, if Then And update the individual collection Set of target vectors with individuals S4.8: Let the number of iterations Iter = Iter + 1, and repeat steps S4.4 to S4.7 until Iter reaches the set maximum number of iterations, and the iteration is terminated to obtain the final population PS and the target vector PF of the population; S4.9: Remove the non-optimal solutions and their targets from the final population PS and the population target vector PF; Further, the S4.6 includes: S4.61: Recording individuals The endmembers searched are Calculate the centroid of the end member S4.62: Calculate the direction from the center of gravity c The direction vector of each search endmember in in, S4.63: Calculate all candidate end elements a1, a2, ..., a in the dictionary A pointed to by the centroid c. M The vectors s1,s2,…,s M , where s k =a k -c,k=1,2,…,M; S4.64: Calculate all vectors s1,s2,…,s M In the direction vector d j The projection ε on 1j ,ε 2j ,…,ε Mj , and find the maximum projection ε max =ε tj ,t∈{1,2,…,M}; S4.65: Determine ε max The size of ε max >0, then use the candidate end member a from the dictionary t replace Search endmembers in And will The tth element of is replaced by 1; S4.66: Repeat steps S4.64 and S4.65 until all 2. The multi-target endmember identification method based on meme optimization according to claim 1 is characterized in that: The S2 includes: S2.1: Generate k unit random vectors u1,u2,…,u with the same dimension as the original hyperspectral image band k ; S2.2: All pixels a1, a2, ..., a in the original hyperspectral image are N Projecting onto the k unit random vectors respectively, wherein N represents the number of pixels in the original hyperspectral image; S2.3: Count the cumulative number of times each pixel falls on the two ends of the k unit random vectors t1, t2, ..., t N ; S2.4: Select the M pixels with the largest cumulative times to form a self-dictionary A = (a1, a2, ..., a M ).

3. The multi-target endmember identification method based on meme optimization according to claim 2 is characterized in that: The S4.4 includes: For each individual i , select two sub-problem indexes j, k from the set B{i}∪i, and use the multi-point crossover strategy to solve the problem of individual Ι j and k Recombination is performed to obtain new individuals of the offspring Right now Among them, r represents an M-dimensional random vector, and its d-th element r d The value is in the interval [0,1], cr is a constant, Respectively represent Ι j ,Ι k , The d-th element of .

4. The multi-target endmember identification method based on meme optimization according to claim 3 is characterized in that: The S4.5 includes: With a probability of 1 / M, the new individuals of the next generation Each element of is bit-reversed and mutated to obtain the mutated individual in, represents an M-dimensional random vector, whose dth element The value is in the interval [0,1]. Represents the individual after mutation The d-th element of .

5. The multi-target endmember identification method based on meme optimization according to claim 4 is characterized in that: The S4.9 includes: S4.91: Initialize the non-dominated solution set and its target set S4.92: For any Ι i ∈PS, if there is Make Ι i <Ι, then let PS=PS / Ι i If it does not exist Make Ι i <Ι, then let PS=PS / I i , and At the same time, if there is So that Ι′<Ι i , then let and Where Ι2<Ι1, if and only if S4.93: Repeat step S4.92 until Get the final Pareto optimal set and its target set 6. A storage medium, characterized in that: The storage medium stores a computer program, which is used to execute the steps of the multi-target endmember identification method based on meme optimization as described in any one of claims 1 to 5.

7. An electronic device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program, and when the processor calls the computer program in the memory, it implements the steps of the multi-target end member identification method based on meme optimization as claimed in any one of claims 1 to 5.

Images