Hyperspectral endmember extraction method based on global-to-local evolutionary algorithm

Abstract

The present invention discloses a hyperspectral endmember extraction method based on a global-to-local evolutionary algorithm, comprising the following steps: 1. importing hyperspectral remote sensing image data and constructing a population of individuals; 2. performing global and multi-generational evolution on the resulting hyperspectral remote sensing image population to obtain a high-quality population; 3. performing local and multi-generational evolution on the high-quality population to obtain a set of non-dominated Pareto solutions, thereby obtaining a set of hyperspectral endmember extraction results. The present invention can measure the quality of endmembers from multiple perspectives, explore the entire image, resolve the endmember duplication problem, and further improve the endmember accuracy, thereby achieving a more perfect endmember extraction effect.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a hyperspectral endmember extraction method based on a global-to-local evolutionary algorithm. Background Art

[0002] Hyperspectral imaging is widely used in various applications due to its ability to capture images with multiple spectral bands, which can provide diagnostic spectral information to identify different land cover types. Due to resolution limitations, there are mixed pixels in hyperspectral remote sensing images, which poses a problem for the identification of ground objects and the accurate analysis of hyperspectral remote sensing images. In order to solve the problem of mixed pixels, hyperspectral unmixing is an effective method that can decompose the pixel spectrum into a set of pure pixel spectra (called endmembers) and the proportion of the corresponding endmembers (called abundance). Most mixed pixel decomposition techniques require the use of endmember extraction results for abundance inversion, so the accuracy of the endmember extraction results will affect the accuracy of the resolution of the distribution of ground objects. It can be seen that endmember extraction is an important task in hyperspectral unmixing.

[0003] In order to solve the problem of endmember extraction in hyperspectral images, domestic and foreign scholars have conducted in-depth research and proposed many methods, which are mainly divided into non-evolutionary algorithm methods and evolutionary optimization methods.

[0004] However, there are at least the following technical problems in the prior art:

[0005] (1) Traditional non-evolutionary computing methods cannot effectively measure the hyperspectral endmember extraction problem from multiple perspectives and are often interfered by noise.

[0006] (2) Existing intelligent optimization methods focus on using different evolutionary computing frameworks, such as particle swarm optimization and differential evolution for equal-endmember extraction, which cannot effectively search all pixels.

[0007] (3) The endmember results obtained by the existing methods are not accurate enough, and there are often cases where two or more extracted endmembers correspond to the same type of land feature. Summary of the Invention

[0008] In order to address the shortcomings of the above-mentioned existing technologies, the present invention proposes a hyperspectral endmember extraction method based on a global-to-local evolutionary algorithm, in order to measure the quality of endmembers from multiple angles and explore the entire image to solve the problem of endmember duplication and further improve the accuracy of endmembers, thereby achieving a more perfect endmember extraction effect.

[0009] In order to achieve the above-mentioned object, the present invention adopts the following technical solutions:

[0010] The hyperspectral endmember extraction method based on a global-to-local evolutionary algorithm of the present invention is characterized in that it includes the following steps:

[0011] S1: Get a hyperspectral remote sensing image matrix R with a dimension of L×N, where L represents the number of bands of the hyperspectral remote sensing image and N represents the number of pixels of the hyperspectral remote sensing image; let X represent the pixel vector in the hyperspectral remote sensing image matrix R, and X={x1,x2,...,x i ,...,x N}, x i Represents the i-th pixel, if x i =1, it means that the i-th pixel is an end member; if x i =0, it means that the i-th pixel is a mixed pixel; let P represent the number of end members of the hyperspectral remote sensing image matrix R;

[0012] S2: Use equations (1) to (3) to construct the multi-objective optimization function minF(X):

[0013] minF(X)=(f1(X),f2(X)) (1)

[0014]

[0015]

[0016] In formula (1) to formula (3), f1(X) is the first optimization objective function, f2(X) is the second optimization objective function, q1, q2, ..., q P is the P endmember value after dimension reduction in the hyperspectral remote sensing image matrix R; r i is the i-th pixel in the hyperspectral remote sensing image matrix R, is the reconstructed value of the i-th pixel;

[0017] S3: Define the current number of iterations as t and initialize t = 1; the maximum number of iterations is maxiter;

[0018] The distribution of end members and mixed pixels in N pixels constitutes an individual, thereby randomly initializing NP individuals and forming the t-th generation population in, represents the t-th generation population Q t The jth individual of ;

[0019] S4: population Q of generation t t The global evolution of

[0020] S4.1: From the tth generation population Q through two binary competitions tIn the jth order, two parent individuals are selected, and the fitness values ​​of the parent individuals selected for the jth order are calculated by the multi-objective optimization function minF(X), and the parent individual with the lower fitness value is recorded as The other parent individual is

[0021] S4.2: Generate the jth random number rand of the tth generation between [0,1] t,j ;

[0022] If rand t,j <0.5, then for the parent individual Perform an asymmetric mutation operation to obtain an offspring individual Otherwise, for the parent individual Perform a SAD-based patching operation to obtain a descendant individual

[0023] S4.3: Follow the process of S4.1-S4.2 for NP times, thus generating NP offspring individuals of the tth generation to form the offspring population of the tth generation

[0024] S4.4: Select operations from O by fast non-dominated sorting method and crowding distance calculation t and Q t Select the best NP individuals to form the t+1 generation population Q t+1 ;

[0025] S4.5: After assigning t+1 to t, return to step S4 and continue until t=4 / 5maxiter, thereby obtaining a high-quality population after global evolution;

[0026] S5: Local evolution of population:

[0027] S5.1: From the tth generation population Q through two binary competitions t In the jth order, two parent individuals are selected, and the fitness values ​​of the parent individuals selected for the jth order are calculated by the multi-objective optimization function minF(X), and the parent individual with the lower fitness value is recorded as The other parent individual is

[0028] S5.2: For parent individuals Perform local perturbation operation to obtain a descendant individual

[0029] S5.3: Follow the process of S5.1-S5.2 for NP times, thus generating NP offspring individuals of the tth generation to form the offspring population of the tth generation.

[0030] S5.4: Select operations from O by fast non-dominated sorting method and crowding distance calculation t and Q t Select the best NP individuals to form the t+1 generation population Q t+1 ;

[0031] S5.5: After assigning t+1 to t, return to step S5 and execute until t=maxiter, thereby obtaining the final population Q after local evolution. maxiter ;

[0032] S5.6: Calculate the final population Q based on the multi-objective optimization function minF(X) maxiter The fitness value of the final population Q maxiter The individual corresponding to the Pareto solution set with the smallest non-dominated level is selected as the end member of the spectral remote sensing image matrix R and the optimal distribution of mixed pixels, so as to obtain the position of the end member in the spectral remote sensing image matrix R.

[0033] The characteristic of the hyperspectral endmember extraction method of the present invention is that the asymmetric mutation operation of S4.2 is to change the values ​​of "0" and "1" of each pixel in the individual according to the probability p 0-1 and probability p 1-0 Flip to "1" and "0"; the probability of the value "0" mutating to "1" The probability of a value of "1" changing to "0"

[0034] If the number of pixels with the value "1" after the mutation is greater than P, the pixel value "1" is randomly changed to "0" so that the number of pixels with the value "1" is equal to P;

[0035] If the number of pixels with the value "1" after the mutation is less than P, then the pixel value "0" is randomly changed to "1" so that the number of pixels with the value "1" is equal to P.

[0036] The SAD-based patching operation in S4.2 is as follows:

[0037] Step a: Let Among them, ea k1 Represents the parent individual The k1th end member value in

[0038] make Among them, eb k2 Represents the parent individual The k2th end member value in;

[0039] Step b: Calculate the k1th end member value ea using formula (4) k1 and the k2th end member value ebk2 SAD value between SAD(ea k1 ,eb k2 ):

[0040]

[0041] Step c: Select the k2th end member value eb k2 and The minimum SAD value of each end member value is recorded as minSAD k2 , and minSAD k2 The corresponding end member value ea min,k2 As EB k2 The learning goal of The minimum value set MINSAD = (minSAD1, minSAD2, ..., minSAD k2 ,...,minSAD P ) and its corresponding learning objective set MINea=(ea min,1 ,ea min,2 ,...,ea min,k2 ,...,ea min,P );

[0042] Step d: If If there are no multiple endmember values ​​corresponding to the same learning objective, the maximum value is selected from MINSAD and recorded as minSAD max , minSAD max The corresponding set of end member values ​​ea min,max and eb max In the parent individual EB in max Replaced with ea min,max , thus obtaining the updated parent individual and as offspring individuals

[0043] like If there are multiple endmember values ​​corresponding to the same learning objective, the parent individuals Check each end member value in m k1 is the current end element value ea k1 The number of corresponding end member values;

[0044] If m k1 =0, then ea k1 join in

[0045] If m k1 ≠0, then mk1 The minimum value minSAD corresponding to the end member value in MINSAD min , minSAD min The corresponding EB min join in Thus, a progeny individual consisting of P end members is obtained

[0046] The present invention provides an electronic device, comprising a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the hyperspectral endmember extraction method, and the processor is configured to execute the program stored in the memory.

[0047] The present invention provides a computer-readable storage medium having a computer program stored thereon. The computer-readable storage medium is characterized in that the computer program executes the steps of the hyperspectral endmember extraction method when executed by a processor.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. The method of the present invention is different from the traditional one. It starts from the perspective of multi-objective evolutionary computing, measures two objectives, overcomes the interference of noise, and is therefore more suitable for endmember extraction of hyperspectral remote sensing images in a variety of complex scenes.

[0050] 2. The present invention regards the hyperspectral endmember extraction problem as a sparse large-scale optimization problem, designs an asymmetric mutation operator, and fully explores the sparse large-scale search space, making it more suitable for endmember extraction of hyperspectral remote sensing images of real big data.

[0051] 3. The present invention utilizes a SAD-based patching strategy, which ultimately greatly reduces the situation where multiple endmembers correspond to the same type of ground feature, thereby saving the computational cost of endmember extraction of hyperspectral remote sensing images.

[0052] 4. The present invention uses the local evolution method to refine the global evolution results, further improving the quality of endmembers, thereby further improving the accuracy of endmember extraction of hyperspectral remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The figure is a flow chart of the hyperspectral endmember extraction method of the present invention. DETAILED DESCRIPTION

[0054] In this embodiment, a method for solving the hyperspectral endmember extraction problem based on a global-to-local evolutionary algorithm is used to extract endmembers from hyperspectral remote sensing images. Specifically, Figure 1 As shown, the method is performed as follows:

[0055] Step 1: Obtain a hyperspectral remote sensing image matrix R with a dimension of L×N, where L represents the number of bands of the hyperspectral remote sensing image and N represents the number of pixels of the hyperspectral remote sensing image; let X represent the pixel vector in the hyperspectral remote sensing image matrix R, and X={x1,x2,...,x i ,...,x N}, x i Represents the i-th pixel, if x i =1, it means that the i-th pixel is an end member; if x i =0, it means that the i-th pixel is a mixed pixel; let P represent the number of end members of the hyperspectral remote sensing image matrix R;

[0056] Step 2: Use equations (1) to (3) to construct the multi-objective optimization function minF(X):

[0057] minF(X)=(f1(X),f2(X)) (1)

[0058]

[0059]

[0060] In formula (1) to formula (3), f1(X) is the first optimization objective function, f2(X) is the second optimization objective function, q1, q2, ..., q P is the P endmember value after dimension reduction in the hyperspectral remote sensing image matrix R; r i is the i-th pixel in the hyperspectral remote sensing image matrix R, is the reconstructed value of the i-th pixel;

[0061] Step 3: Define the current number of iterations as t and initialize t = 1; the maximum number of iterations is maxiter;

[0062] The distribution of end members and mixed pixels in N pixels constitutes an individual, thereby randomly initializing NP individuals and forming the t-th generation population in, represents the t-th generation population Q t The jth individual of ;

[0063] Step 4: The t-th generation population Q t The global evolution of

[0064] Step 4.1: Use two binary competitions to select the t-th generation population Q t In the jth order, two parent individuals are selected, and the fitness values ​​of the parent individuals selected for the jth order are calculated by the multi-objective optimization function minF(X), and the parent individuals with lower fitness values ​​are recorded as The other parent individual is

[0065] Step 4.2: Generate the jth random number rand of the tth generation between [0,1] t,j ;

[0066] If rand t,j <0.5, then for the parent individual Perform asymmetric mutation operation, specifically: The values ​​of each pixel in the matrix are “0” and “1” according to the probability p. 0-1 and probability p 1-0 Flip to "1" and "0"; the probability of the value "0" mutating to "1" The probability of a value of "1" mutating to "0"

[0067] If the number of pixels with the value "1" after the mutation is greater than P, the pixel value "1" is randomly changed to "0" so that the number of pixels with the value "1" is equal to P;

[0068] If the number of pixels with the value "1" after the mutation is less than P, then randomly change the pixel value "0" to "1" so that the number of pixels with the value "1" equals P. Finally, Assignment The asymmetric mutation operation makes all pixels equally likely to be selected as endmembers, which can fully explore the entire hyperspectral image.

[0069] If rand t,j Greater than or equal to 0.5, for the parent individual Perform SAD-based patching operations, specifically:

[0070] For a good individual, the endmembers it contains must be different endmembers corresponding to different land feature types, while for some bad individuals, the endmembers it contains may have multiple endmembers corresponding to the same land feature type. This is the endmember duplication phenomenon. Such bad individuals do not meet the requirements for endmember extraction, so we repair some bad individuals through the following steps.

[0071] Step a: Let Among them, ea k1 Represents the parent individual The k1th end member value in

[0072] make Among them, eb k2 Represents the parent individual The k2th end member value in the next bad individual Good individuals study.

[0073] Step b: Calculate the k1th end member value ea using formula (4) k1 and the k2th end member value eb k2 SAD value between SAD(ea k1 ,eb k2 ):

[0074]

[0075] If the SAD value between the two is smaller, it means that the two pixels are more similar, otherwise, the two pixels are less similar.

[0076] Step c: Select the k2th end member value eb k2 and The minimum SAD value is recorded as minSAD k2 , i.e. minSAD k2 =argmin SAD(eb k2 ,{ea1,ea2,...,ea k1 ,...,ea P}), and minSAD k2 The corresponding end member value ea min,k2 As EB k2 The learning goal of The minimum value set MINSAD = (minSAD1, minSAD2, ..., minSAD k2 ,...,minSAD P ) and its corresponding learning objectives set MINea=(ea min,1 ,ea min,2 ,...,ea min,k2 ,...,ea min,P );

[0077] Step d: At this point, there may be two corresponding relationships corresponding to two processing methods.

[0078] For correspondence 1: If There is no situation where multiple end member values ​​correspond to the same learning objective, that is, there is no ea in MINea. min,P1 =ea min,P2 , then, select the largest value in MINSAD and record it as minSAD max , then In, use the corresponding ea min,max End member replacement eb max , about to and The most dissimilar end members are replaced. After the change Assign to That's it.

[0079] If it is correspondence 2: There are multiple end-member values ​​corresponding to the same learning objective in MINea, that is, there are ea in MINea min,P1 =ea min,P2 At this time, Check each end member value in m k1 is the current end element value ea k1 The corresponding EB k2 Quantity, if m k1 =0, then ea k1 join in If m k1 ≠0, then m k1 The minimum value of MINSAD corresponding to the end member values ​​is recorded as minSAD min , and the corresponding eb min join in Finally, we get a system consisting of P end members.

[0080] Step 4.3: Execute the process of S4.1-S4.2 NP times, thereby generating NP offspring individuals of the tth generation to constitute the offspring population of the tth generation.

[0081] Step 4.4: Select the operation from O by fast non-dominated sorting method and crowding distance calculation. t and Q t Select the best NP individuals to form the t+1 generation population Q t+1 ;

[0082] Step 4.5: After assigning t+1 to t, return to step S4 and continue until t=4 / 5maxiter, thereby obtaining a high-quality population after global evolution;

[0083] Step 5: Local evolution of the population:

[0084] Step 5.1: Use two binary competitions to select the t-th generation population Q t In the jth order, two parent individuals are selected, and the fitness values ​​of the parent individuals selected for the jth order are calculated by the multi-objective optimization function minF(X), and the parent individuals with lower fitness values ​​are recorded as The other parent individual is

[0085] Step 5.2: For the parent individual Perform local perturbation operation. Since the search range of hyperspectral image is very large and the spatial correlation between pixels is relatively strong, we perform perturbation near the end member. Specifically: Randomly Select an endmember from the endmembers and change its current position within the range of the 5×5 image around it. The changed endmember and The unchanged end members in the

[0086] Step 5.3: Execute the process of S5.1-S5.2 NP times, thereby generating NP offspring individuals of the tth generation to constitute the offspring population of the tth generation.

[0087] Step 5.4: Select the operation from O by fast non-dominated sorting method and crowding distance calculation. t and Q t Select the best NP individuals to form the t+1 generation population Q t+1 ;

[0088] Step 5.5: After assigning t+1 to t, return to step S5 and execute until t=maxiter, thereby obtaining the final population Q after local evolution. maxiter ;

[0089] Step 5.6: Calculate the final population Q based on the multi-objective optimization function minF(X) maxiter The fitness value of the final population Q maxiter The individual corresponding to the Pareto solution set with the smallest non-dominated level is selected as the end member of the spectral remote sensing image matrix R and the optimal distribution of mixed pixels, so as to obtain the position of the end member in the spectral remote sensing image matrix R.

[0090] In this embodiment, an electronic device includes a memory and a processor. The memory is used to store a program that supports the processor to execute the above-mentioned hyperspectral endmember extraction method. The processor is configured to execute the program stored in the memory.

[0091] In this embodiment, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the hyperspectral endmember extraction method are executed.

[0092] The specific implementation method of the present invention is illustrated and the effect of the method of the present invention is verified by using examples of real data sets.

[0093] 1) Prepare real data sets;

[0094] The present invention uses the Samson dataset as a standard dataset to verify the effectiveness of the hyperspectral endmember extraction method based on the global to local evolutionary algorithm. The Samson dataset is a dataset widely used in endmember extraction problems. In the Samson dataset, the image size is 95 pixels × 95 pixels, with a total of 156 bands, and the image contains three endmembers: land, trees, and water. The method of this patent is GL_EA, which uses N-FINDR, DPSO, MODPSO, IMODPSO, and (μ+λ)MODE for endmember extraction. Table 1 shows the minimum SAD values ​​of the endmembers obtained by the embodiments of the present invention and the existing endmember extraction method and the true endmember.

[0095] 2) Evaluation indicators;

[0096] The SAD values ​​of the real end members are used to evaluate the performance of the end member extraction of different multi-objective algorithms and the present invention, as shown in Table 1:

[0097] Table 1

[0098]

[0099] As can be seen from Table 1, the accuracy of the endmembers extracted by the present invention is further evaluated by comparing the SAD values ​​with the real endmembers. In the three endmembers: land, tree and water, the SAD values ​​of the present invention can reach the minimum, indicating that the endmembers extracted by the present invention are closest to the real endmembers.

Claims

1. A hyperspectral endmember extraction method based on a global-to-local evolutionary algorithm, characterized in that: The steps include: S1: Get a hyperspectral remote sensing image matrix R with a dimension of L×N, where L represents the number of bands of the hyperspectral remote sensing image and N represents the number of pixels of the hyperspectral remote sensing image; let X represent the pixel vector in the hyperspectral remote sensing image matrix R, and X={x1,x2,...,x i ,...,x N }, x i Represents the i-th pixel, if x i =1, it means that the i-th pixel is an end member; if x i =0, it means that the i-th pixel is a mixed pixel; let P represent the number of end members of the hyperspectral remote sensing image matrix R; S2: Use equations (1) to (3) to construct the multi-objective optimization function minF(X): minF(X)=(f1(X),f2(X)) (1) In formula (1) to formula (3), f1(X) is the first optimization objective function, f2(X) is the second optimization objective function, q1, q2, ..., q P is the P endmember value after dimension reduction in the hyperspectral remote sensing image matrix R; r i is the i-th pixel in the hyperspectral remote sensing image matrix R, is the reconstructed value of the i-th pixel; S3: Define the current number of iterations as t and initialize t = 1; the maximum number of iterations is maxiter; The distribution of end members and mixed pixels in N pixels constitutes an individual, thereby randomly initializing NP individuals and forming the t-th generation population in, represents the t-th generation population Q t The jth individual of ; S4: population Q of generation t t The global evolution of S4.1: From the tth generation population Q through two binary competitions t In the jth order, two parent individuals are selected, and the fitness values ​​of the parent individuals selected for the jth order are calculated by the multi-objective optimization function minF(X), and the parent individual with the lower fitness value is recorded as The other parent individual is S4.2: Generate the jth random number rand of the tth generation between [0,1] t,j ; If rand t,j <0.5, then for the parent individual Perform an asymmetric mutation operation to obtain an offspring individual Otherwise, for the parent individual Perform a SAD-based patching operation to obtain a descendant individual S4.3: Follow the process of S4.1-S4.2 for NP times, thus generating NP offspring individuals of the tth generation to form the offspring population of the tth generation S4.4: Select operations from O by fast non-dominated sorting method and crowding distance calculation t and Q t Select the best NP individuals to form the t+1 generation population Q t+1 ; S4.5: After assigning t+1 to t, return to step S4 and continue until t=4 / 5maxiter, thereby obtaining a high-quality population after global evolution; S5: Local evolution of population: S5.1: From the tth generation population Q through two binary competitions t In the jth order, two parent individuals are selected, and the fitness values ​​of the parent individuals selected for the jth order are calculated by the multi-objective optimization function minF(X), and the parent individual with the lower fitness value is recorded as The other parent individual is S5.2: For parent individuals Perform local perturbation operation to obtain a descendant individual S5.3: Follow the process of S5.1-S5.2 for NP times, thus generating NP offspring individuals of the tth generation to form the offspring population of the tth generation. S5.4: Select operations from O by fast non-dominated sorting method and crowding distance calculation t and Q t Select the best NP individuals to form the t+1 generation population Q t+1 ; S5.5: After assigning t+1 to t, return to step S5 and execute until t=maxiter, thereby obtaining the final population Q after local evolution. maxiter ; S5.6: Calculate the final population Q based on the multi-objective optimization function minF(X) maxiter The fitness value of the final population Q maxiter The individual corresponding to the Pareto solution set with the smallest non-dominated level is selected as the end member of the spectral remote sensing image matrix R and the optimal distribution of mixed pixels, so as to obtain the position of the end member in the spectral remote sensing image matrix R.

2. The hyperspectral endmember extraction method according to claim 1, wherein: The asymmetric mutation operation in S4.2 is to change the values ​​of each pixel in the individual to "0" and "1" according to the probability p. 0-1 and probability p 1-0 Flip to "1" and "0"; the probability of the value "0" mutating to "1" The probability of a value of "1" changing to "0" If the number of pixels with the value "1" after the mutation is greater than P, the pixel value "1" is randomly changed to "0" so that the number of pixels with the value "1" is equal to P; If the number of pixels with the value "1" after the mutation is less than P, the pixel value "0" is randomly changed to "1" so that the number of pixels with the value "1" is equal to P.

3. The hyperspectral endmember extraction method according to claim 1, wherein: The SAD-based patching operation in S4.2 is as follows: Step a: Let Among them, ea k1 Represents the parent individual The k1th end member value in make Among them, eb k2 Represents the parent individual The k2th end member value in; Step b: Calculate the k1th end member value ea using formula (4) k1 and the k2th end member value eb k2 SAD value between SAD(ea k1 ,eb k2 ): Step c: Select the k2th end member value eb k2 and The minimum SAD value of each end member value is recorded as minSAD k2 , and minSAD k2 The corresponding end member value ea min,k2 As EB k2 The learning goal of The minimum value set selected by each end member value in MINSAD=(minSAD1,minSAD2,...,minSAD k2 ,...,minSAD P ) and its corresponding learning objective set MINea=(ea min,1 ,ea min,2 ,...,ea min,k2 ,...,ea min,P ); Step d: If If there are no multiple endmember values ​​corresponding to the same learning objective, the maximum value is selected from MINSAD and recorded as minSAD max , minSAD max The corresponding set of end member values ​​ea min,max and eb max In the parent individual EB in max Replaced with ea min,max , thus obtaining the updated parent individual and as offspring individuals like If there are multiple endmember values ​​corresponding to the same learning objective, the parent individuals Check each end member value in m k1 is the current end element value ea k1 The number of corresponding end member values; If m k1 =0, then ea k1 join in If m k1 ≠0, then m k1 The minimum value minSAD corresponding to the end member value in MINSAD min , minSAD min The corresponding EB min join in Thus, a progeny individual consisting of P end members is obtained 4. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the hyperspectral endmember extraction method according to any one of claims 1 to 3, and the processor is configured to execute the program stored in the memory.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the hyperspectral endmember extraction method according to any one of claims 1 to 3 are executed.

Images