An evolutionary multi-modal based hyperspectral endmember extraction method

Abstract

The application discloses a hyperspectral end member extraction method based on evolutionary multi-modal, comprising the following steps: 1, obtaining hyperspectral remote sensing image data and constructing population individuals; 2, using a population grouping evolution mechanism, grouping and respectively evolving the obtained hyperspectral remote sensing image population, and generating a population with higher quality; 3, performing front repair and multi-generation evolution on the high-quality population, obtaining a set of non-dominated Pareto solutions, and obtaining a set of hyperspectral end member extraction results. The application can measure the quality of the end member from multiple angles, improve the end member precision, obtain more excellent end member combinations in the decision space, and thus improve the end member extraction effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image end member extraction, in particular to a hyperspectral end member extraction method based on evolutionary multi-modal. BACKGROUND

[0002] Hyperspectral images contain rich radiation, spatial and spectral information, and these spectral bands can provide diagnostic spectral information to identify different land cover types. However, due to the limitation of resolution, there are mixed pixels in hyperspectral remote sensing images, which affects the identification of ground objects and the accuracy of hyperspectral remote sensing images. Hyperspectral unmixing can solve the problem of mixed pixels by searching for basic ground object types that make up the mixed pixels and calculating the proportion of each basic ground object type in the mixed pixels. End member extraction is responsible for finding pure pixels in the unmixing process. However, due to the large search space of end member extraction, for a scene containing tens of thousands of pixels and several end members, the number of combinations has reached a very high order of magnitude, and existing methods are difficult to balance exploration and utilization. There are cases in practical application of hyperspectral unmixing where the precision is the same but the end member combination is completely different, and mining such individuals can provide target space excellent and decision space more diverse solutions for different scenes.

[0003] At least the following technical problems exist in the prior art:

[0004] (1) In a sparse large-scale decision space, existing evolutionary computation-based methods are difficult to effectively balance global exploration and local search, resulting in decreased search performance and affecting the evolution efficiency of the population.

[0005] (2) Existing end member extraction methods mostly use evolutionary computation and its improved framework, such as particle swarm optimization, differential evolution, etc., which cannot effectively search all pixel positions.

[0006] (3) Existing methods only consider the individual fitness performance and population target space diversity, ignoring the diversity in the decision space. SUMMARY

[0007] The present application is to solve the above-mentioned deficiencies in the prior art, and proposes a hyperspectral end member extraction method based on evolutionary multi-modal, in order to measure the quality of end members from multiple angles, explore the entire image, and achieve better solutions in practical applications. The accuracy of the end member is further improved, so as to achieve a more perfect end member extraction effect.

[0008] In order to achieve the above-mentioned application purposes, the following technical solutions are adopted:

[0009] The hyperspectral end member extraction method based on evolutionary multi-modal has the following steps:

[0010] S1, obtain a hyperspectral remote sensing image matrix R of dimension L×N, where L represents the number of bands in the hyperspectral remote sensing image and N represents the number of pixels in the hyperspectral remote sensing image; let X represent the pixel vector in the hyperspectral remote sensing image matrix R, and , Represents the i-th pixel, if Then it represents the i-th pixel. For the end element; if Then it represents the i-th pixel. For mixed pixels;

[0011] S2, construct a multi-objective optimization function using equations (1)-(3). :

[0012] (1)

[0013] (2)

[0014] (3)

[0015] In equations (1)-(3), The first optimization objective function is... For the second optimization objective function, , , , The dimensionality-reduced hyperspectral remote sensing image matrix R Each end-member value; This represents the number of endmembers in the hyperspectral remote sensing image matrix R; For the i-th pixel The reconstructed value;

[0016] S3, define the current iteration number as t, and initialize t=1; define the maximum iteration number as t. max ;

[0017] An individual is formed by the distribution of endmembers and mixed pixels among N pixels, thus randomly initializing NP individuals and forming the t-th generation population. ;

[0018] S4. According to equations (2) and (3), calculate the non-dominance level of each individual in the t-th generation population, and divide the A non-dominant individuals with a non-dominance level of "1" in the t-th generation population into non-dominant subpopulations. ,in, Denotes the non-dominated subpopulation of generation t. The a-th individual, and the remaining B individuals in the t-th generation population are divided into the t-th generation dominated subpopulation. ,in, denotes the bth individual of the tth generation dominated sub-population ; ;

[0019] S5, the local search strategy based on the niche is adopted to process , and the ath sub-child individual is obtained, so that the tth generation non-dominated sub-child population is obtained; and the better A individuals are selected from and by the selection operation of the fast non-dominated sorting method and the crowding distance calculation to form the t+1th generation non-dominated sub-population ;

[0020] S6, the ith individual is randomly selected from as the guide vector of the current bth individual , and the global search strategy based on the SAD guide is adopted to process and , and the bth sub-child individual is obtained, so that the tth generation dominated sub-child population is obtained; and the better B individuals are selected from and by the selection operation of the fast non-dominated sorting method and the crowding distance calculation to form the t+1th generation dominated sub-population ;

[0021] S7, the combination of and is the t+1th generation combined population , and the front repair processing is performed on , and the t+1th generation complement set is obtained; wherein denotes the zth complement of the t+1th generation;

[0022] S8, the combination of and is the t+1th generation repair population , and the better NP individuals are selected from by the selection operation of the fast non-dominated sorting method and the endmember distance calculation to form the t+1th generation population ;

[0023] S9, t+1 is assigned to t, and then the step S4 is returned to be executed until t>t max , so that the evolved high-quality population is obtained, and the individual corresponding to the non-dominated level of "1" is selected from as the optimal endmember extraction result of the spectral remote sensing image.

[0024] The endmember extraction method based on evolutionary multimodality described in this invention is also characterized in that the niche-based local search strategy in S5 is performed according to the following steps:

[0025] Step a: For the current a-th individual The P endmembers in the equation are used to calculate the first P endmembers respectively using equation (4). individual terminal and the individual terminal spectral angle distance between This yields an endmember similarity matrix of dimension P×P. :

[0026] (4)

[0027] Step b: Calculate the first individual terminal With other The sum of the spectral angular distances of the endmembers Thus, the sum of the P spectral angle distances is obtained. and will Endmembers corresponding to the minimum value This is designated as a terminal element to be optimized.

[0028] Step c: Calculate the population of the current generation t. The sum of the indices corresponding to all endmembers in each individual, and combined with... The sum of the indices corresponding to all endmembers is compared, and the result is selected. The n closest individuals constitute Within the niche of the decision space, select the individual whose sum of the first and second objective functions is minimized. As Learning objectives;

[0029] Step d: Calculate the endmember to be optimized using equation (4) Individuals with learning targets The spectral angular distance between each endmember, and the minimum spectral angular distance in The corresponding endmember As a terminator to be optimized The evolutionary goal, in A new endmember is randomly selected within the range of the image and its radius r. and to End-members to be optimized Replacement is performed to obtain a child individual. .

[0030] Furthermore, the global search strategy based on SAD in S6 is implemented in the following steps:

[0031] Step a: Calculate the current b-th individual using equation (4). Each endmember in the vector is associated with the guiding vector. The spectral angle distance between each endmember is preserved. and The P / 2 endmembers with the smallest spectral angle distance in the middle spectrum, and Zhongyu The P / 2 endmembers with larger spectral angles are selected as endmembers to be updated.

[0032] Step b: One of the terminal elements awaiting update is in The matched endpoint is locked, and a dynamic radius is applied to the locked endpoint and its surrounding area. A new endmember is randomly selected within the image range to replace the corresponding endmember to be updated, thereby completing the process. Replace P / 2 terminal elements to be updated and obtain the b-th offspring individual. .

[0033] Furthermore, S7 includes:

[0034] S7.1: Will and Merge into the t+1 generation merged population and will After sorting the individuals within the first objective function in ascending order, they are uniformly divided into Z individual clusters, with each individual cluster containing NP / Z individuals;

[0035] Based on equations (2) and (3), calculate the non-dominated level of each individual in the z-th individual cluster, and select the individual with the smallest first objective function value from the individuals corresponding to the non-dominated level "1". And calculate The k individuals closest to each other in the target space by Euclidean distance constitute... nearest neighbor set ;Depend on All endmembers contained in all individuals constitute the z-th endmember set of the (t+1)th generation. ;

[0036] S7.2: Select the individual with the smallest second objective function value from the z-th individual cluster. and calculate respectively Middle-end element and end-end element set The spectral angular distance between the endmembers is used to select the P endmember pairs corresponding to the minimum spectral angular distance. The endmembers in the middle form the z-th complement in the (t+1)th generation. ; thereby obtaining Z complements of the t+1th generation and forming a complement set of the t+1th generation .

[0037] Further, S8 comprises:

[0038] S8.1: according to formula (2) and formula (3), calculating The non-dominated rank of each individual is 1, the neighborhood set of all individuals corresponding to the non-dominated rank of 1 is constructed, and the non-dominated rank of each individual in the neighborhood set is set to 1;

[0039] S8.2: using the end member distance competition method to Select the environment and keep NP individuals to form the t+1th generation population ; when the individuals are in the same non-dominated rank, the average Euclidean distance of the end members in the individual on the spectral image of R is calculated, so that the individual corresponding to the larger average Euclidean distance is kept.

[0040] The electronic device comprises a memory and a processor, and the memory is used to store a program supporting the processor to execute the hyperspectral end member extraction method, and the processor is configured to execute the program stored in the memory.

[0041] The computer readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the hyperspectral end member extraction method are executed.

[0042] Compared with the prior art, the present application has the following advantages:

[0043] 1、The method of the present application is different from the traditional one, and is derived from the actual application of hyperspectral unmixing, adopts an evolutionary multimodal method to mine excellent target space and more diverse end member combinations in the solution space, and is more suitable for end member extraction of hyperspectral remote sensing images in various complex scenes.

[0044] 2、The present application adopts a grouping evolution strategy for the population, and different strategies are used to update the convergence and diversity of the leading solution and the non-leading solution, so as to fully explore the sparse large-scale search space of hyperspectral end member extraction, and thus the present application is more suitable for end member extraction of real big data hyperspectral remote sensing images.

[0045] 3、The present application uses a leading repair mechanism to combine the advantages of excellent individuals to obtain better end member combinations, thereby further improving the accuracy of end member extraction.

[0046] 4. This invention also proposes an environmental selection competition mechanism based on endmember distance for the problem of hyperspectral endmember extraction. By widening the Pareto front, it retains potential multimodal individuals, fully explores the diversity of the decision space, and improves the endmember extraction effect. Attached Figure Description

[0047] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0048] In this embodiment, a hyperspectral endmember extraction method based on evolutionary multimodal approaches is used for endmember extraction from hyperspectral remote sensing images. Specifically, as... Figure 1 As shown, the method is performed according to the following steps:

[0049] Step 1: Obtain a hyperspectral remote sensing image matrix R of dimension L×N, where L represents the number of bands in the hyperspectral remote sensing image and N represents the number of pixels in the hyperspectral remote sensing image; let X represent the pixel vector in the hyperspectral remote sensing image matrix R, and , Represents the i-th pixel, if Then it represents the i-th pixel. For the end element; if Then it represents the i-th pixel. These are mixed pixels.

[0050] Step 2: Construct a multi-objective optimization function using equations (1)-(3) :

[0051] (1)

[0052] (2)

[0053] (3)

[0054] In equations (1)-(3), The first optimization objective function is... For the second optimization objective function, , , , The dimensionality-reduced hyperspectral remote sensing image matrix R Each end-member value; This represents the number of endmembers in the hyperspectral remote sensing image matrix R; For the i-th pixel The reconstructed value.

[0055] Step 3: Define the current iteration number as t, and initialize t=1; define the maximum iteration number as t. max ;

[0056] An individual is formed by the distribution of endmembers and mixed pixels in N pixels, so that NP individuals are randomly initialized and form the tth generation population .

[0057] Step 4: The individuals in the population often show different performance in the target space. For the individuals on the front with non-dominated level of "1", further search in their local neighborhood will probably find more potential excellent solutions. The remaining dominated individuals can be fully explored in the decision space, which can improve the population diversity and enhance the overall quality of the population.

[0058] Therefore, according to formula (2) and formula (3), the non-dominated level of each individual in the tth generation population is calculated, and A non-dominated individuals with non-dominated level of "1" in the tth generation population are divided into the tth generation non-dominated sub-population , wherein, represents the ath individual in the tth generation non-dominated sub-population , and the remaining B individuals in the tth generation population are divided into the tth generation dominated sub-population , wherein, represents the bth individual in the tth generation dominated sub-population , .

[0059] Step 5, evolution of the tth generation non-dominated sub-population ;

[0060] Step 5.1: for P endmembers in the current ath individual , the spectral angle distance between the first endmember and the second endmember is calculated respectively by using formula (4), so as to obtain the endmember similarity matrix with dimension of P × P :

[0061] (4)

[0062] The smaller the value of , the more similar the corresponding endmembers and .

[0063] Step 5.2: calculate the sum of spectral angle distances between the first endmember and the other endmembers, so as to obtain the sum of P spectral angle distances , , The smaller the value, the higher the similarity between the endmember and other endmembers, therefore... Endmembers corresponding to the minimum value Designated as end-members to be optimized; updating only some end-members of excellent solutions can improve the diversity of non-dominated individuals while ensuring the quality of offspring.

[0064] Step 5.3: The positions of the end-members to be updated within the individual have now been identified. Next, the decision space will be used to update the individual... Create a microenvironment, and select outstanding end-users from excellent individuals within the microenvironment as learning targets for end-users to be updated;

[0065] Calculate the population of the current generation t. The sum of the indices corresponding to all endmembers in each individual, and combined with... The sum of the indices corresponding to all endmembers is compared, and the result is selected. The n closest individuals constitute Within the niche of the decision space, select the individual whose sum of the first and second objective functions is minimized. As The learning objectives.

[0066] Step 5.4: Calculate the endmember to be optimized using equation (4). Individuals with learning targets The spectral angular distance between each endmember, and the minimum spectral angular distance in The corresponding endmember As a terminator to be optimized The evolutionary target is determined by the fact that the search range of hyperspectral images is very large and the spatial correlation between pixels is relatively strong. Therefore, in this embodiment, perturbation is performed near the target endmember. A new endmember is randomly selected within a 5×5 area of ​​the image and its surrounding area. and to End-members to be optimized The replacement is performed to obtain the a-th offspring individual. .

[0067] Thus, the non-dominated offspring population of generation t is obtained. Then, through the selection operation of the fast non-dominated sorting method and crowding distance calculation, from... and Select A individuals with better traits to form the (t+1)th generation non-dominated subpopulation. .

[0068] Step 6, the t-th generation offspring population The evolution;

[0069] Step 6.1: From Randomly select the ith individual from the population The guide vector of the current bth individual is calculated by formula (4) The guide vector of the current bth individual is calculated by formula (4) The spectral angle distance between each endmember in the current bth individual and the guide vector is calculated The spectral angle distance between each endmember in the current bth individual and the guide vector is calculated The P / 2 endmembers with the smallest spectral angle distance in the current bth individual are selected The P / 2 endmembers with the smallest spectral angle distance in the current bth individual are selected The P / 2 endmembers with the smallest spectral angle distance in the current bth individual are selected The P / 2 endmembers with the smallest spectral angle distance in the current bth individual are selected

[0070] Step 6.2: Lock the endmember matched by the to-be-updated endmember in the current bth individual in the current bth individual, and randomly select a new endmember within the image range of the dynamic radius Step 6.2: Lock the endmember matched by the to-be-updated endmember in the current bth individual in the current bth individual, and randomly select a new endmember within the image range of the dynamic radius Step 6.2: Lock the endmember matched by the to-be-updated endmember in the current bth individual in the current bth individual, and randomly select a new endmember within the image range of the dynamic radius Step 6.2: Lock the endmember matched by the to-be-updated endmember in the current bth individual in the current bth individual, and randomly select a new endmember within the image range of the dynamic radius Step 6.2: Lock the endmember matched by the to-be-updated endmember in the current bth individual in the current bth individual, and randomly select a new endmember within the image range of the dynamic radius Step 6.2: Lock the endmember matched by the to-be-updated endmember in the current bth individual in the current bth individual, and randomly select a new endmember within the image range of the dynamic radius Step 6.2: Lock the endmember matched by the to-be-updated endmember in the current bth individual in the current bth individual, and randomly select a new endmember within the image range of the dynamic radius Step 6.2: Lock the endmember matched by the to-be-updated endmember in the current bth individual in the current bth individual, and randomly select a new endmember within the image range of the dynamic radius Step 6.2: Lock the endmember matched by the to-be-updated endmember in the current bth individual in the current bth individual, and randomly select a new endmember within the image range of the dynamic radius Step 6.2: Lock the endmember matched by the to-be-updated endmember in the current bth individual in the current bth individual, and randomly select a new endmember within the image range of the dynamic radius

[0071] Step 7, the population grouping evolution strategy of the previous stage can produce excellent offspring in the local area, in order to further improve the accuracy of the endmember and enhance the diversity of the optimal solution, a repair mechanism for the front surface individual is adopted, the population individuals are first clustered, then new supplementary individuals are generated within each cluster by combining the excellent features of the excellent solutions of the two objective functions, and this strong combination can combine the advantages of the excellent individuals to obtain better endmember combination;

[0072] Merge the current bth individual and the (t+1)th individual to obtain the t+1th merged population Merge the current bth individual and the (t+1)th individual to obtain the t+1th merged population Merge the current bth individual and the (t+1)th individual to obtain the t+1th merged population Merge the current bth individual and the (t+1)th individual to obtain the t+1th merged population Merge the current bth individual and the (t+1)th individual to obtain the t+1th merged population Merge the current bth individual and the (t+1)th individual to obtain the t+1th merged population Merge the current bth individual and the (t+1)th individual to obtain the t+1th merged population

[0073] Step 7.1: Merge the current bth individual and the (t+1)th individual to obtain the t+1th merged population Step 7.1: Merge the current bth individual and the (t+1)th individual to obtain the t+1th merged population Step 7.1: Merge the current bth individual and the (t+1)th individual to obtain the t+1th merged population Step 7.1: Merge the current bth individual and the (t+1)th individual to obtain the t+1th merged population After sorting the individuals within the first objective function in ascending order, they are uniformly divided into Z individual clusters, with each individual cluster containing NP / Z individuals;

[0074] Based on equations (2) and (3), calculate the non-dominated level of each individual in the z-th individual cluster, and select the individual with the smallest first objective function value from the individuals corresponding to the non-dominated level "1". And calculate The k individuals closest to each other in the target space by Euclidean distance constitute... nearest neighbor set ;Depend on All endmembers contained in all individuals constitute the z-th endmember set of the (t+1)th generation. , Endmembers in the algorithm have the excellent property of making the first objective function value of an individual smaller;

[0075] Step 7.2: Select the individual with the smallest second objective function value from the z-th individual cluster. Its endmembers have the excellent property of making the individual's second objective function value smaller. (Calculate separately) Middle-end element and end-end element set The spectral angular distance between endmembers reflects the degree of matching between two excellent endmember pairs. Among the P endmember pairs corresponding to the smallest spectral angular distance, [the following is considered a suitable match]. The endmembers in the middle form the z-th complement in the (t+1)th generation. Thus, the Z complements of the (t+1)th generation are obtained and the complement set of the (t+1)th generation is formed. The complement generated by this method can take into account the excellent characteristic that both objective function values ​​are relatively small.

[0076] Step 8, and Merge into the repair population of generation t+1 And through the selection operation of fast non-dominated sorting method and endmember distance calculation, from Select NP of the better individuals to form the (t+1)th generation population. ;

[0077] First, the first frontier is widened, and non-frontier individuals from the nearest neighbor set of individuals on the first frontier are pulled to the first frontier to retain them for the next generation. At the same time, a new method for competition among individuals on the same frontier is proposed. When selecting the environment, endmember distance competition is used instead of the original crowding distance ranking method to enhance the diversity of non-dominated solutions in the decision space.

[0078] Step 8.1: Calculate according to equations (2) and (3). For each individual, the non-dominated level is determined. For all individuals with a non-dominated level of "1", a nearest neighbor set is constructed. The non-dominated level of each individual in the nearest neighbor set is set to "1". Individuals in the nearest neighbor set can be pulled to the first frontier to be retained to the next generation.

[0079] Step 8.2: When individuals are at the same frontier, traditional environmental selection uses crowding distance competition, which reflects the differences among individuals in the goal space. Here, we consider their diversity in the decision space and use the endmember distance competition method. Environmental selection is performed, and NP individuals are retained to form the population of generation t+1. When individuals compete within the same non-dominant rank, the average Euclidean distance of endmembers within an individual on the spectral image of R is calculated, thus retaining individuals with larger average Euclidean distances.

[0080] Step 9: After assigning t+1 to t, return to step S4 and continue until t>t. max This process aims to obtain a high-quality, evolved population. ,from Individuals with a non-dominant level of "1" are selected as the optimal endmember extraction results from the spectral remote sensing image.

[0081] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0082] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0083] The specific implementation method of the present invention and the effectiveness of the method of the present invention are illustrated by examples of real datasets.

[0084] 1) Prepare a real dataset;

[0085] This invention uses the Samson dataset as a standard dataset to verify the effectiveness of the hyperspectral endmember extraction method based on a global-to-local evolutionary algorithm. The Samson dataset is widely used for 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 types of endmembers: land, trees, and water. The method of this patent is EMO, using MODPSO, IMODPSO, (μ+λ)MODE, and GLEA for endmember extraction. Table 1 shows the minimum SAD values ​​of the endmembers obtained by the embodiments of this invention and existing endmember extraction methods compared to the true endmembers.

[0086] 2) evaluation index;

[0087] The SAD values of the real end members are used to evaluate the performance of different multi-target algorithms and the end member extraction of the present application, as shown in Table 1:

[0088]

[0089] As can be seen from Table 1, the accuracy of the extracted end members of the present application is further evaluated by the SAD values of the real end members, wherein the SAD values of the present application can reach the minimum on three end members: land, trees and water, indicating that the extracted end members of the present application are closest to the real end members.

Claims

1. An evolutionary multi-modal based hyperspectral endmember extraction method, characterized in that, The method comprises the following steps: S1, acquire a hyperspectral remote sensing image matrix R with a dimension of LxN, wherein L represents a band number of the hyperspectral remote sensing image, and N represents a pixel number of the hyperspectral remote sensing image; let X represent a pixel vector in the hyperspectral remote sensing image matrix R, and , represent the i-th pixel, if , the i-th pixel is an endmember; if , the i-th pixel is a mixed pixel; S2, constructing a multi-objective optimization function using formula (1) - formula (3) : (1) (2) (3) in formula (1) - formula (3), is a first optimization objective function, is a second optimization objective function, , , , is the reduced dimension of the hyperspectral remote sensing image matrix R end member values; represents the number of end members of the hyperspectral remote sensing image matrix R; is the reconstruction value of the i th pixel . S3, define the current iteration number as t, and initialize t = 1; define the maximum iteration number as t max ; An individual is formed by the distribution of endmembers and mixed pixels in N pixels, so that NP individuals are randomly initialized and form the tth generation population ; S4, calculating the non-dominated rank of each individual in the tth generation population according to formula (2) and formula (3), dividing A non-dominated individuals with the non-dominated rank of "1" in the tth generation population into a tth generation non-dominated sub-population wherein, denotes the a th individual in the tth generation non-dominated sub-population , and the remaining B individuals in the tth generation population are divided into a tth generation dominated sub-population wherein, denotes the b th individual in the tth generation dominated sub-population , ; S5, using local search strategy based on niche to process the individual in the population to obtain the a th offspring individual , thereby obtaining the t th non-dominated offspring population ; and selecting the better A individuals from the non-dominated offspring population and the non-dominated population to form the t+1 th non-dominated offspring population by using the fast non-dominated sorting method and the selection operation of the crowding distance calculation ;​​​ S6, from randomly select the ith individual as the current bth individual guide vector, and adopt the global search strategy based on SAD guide to process and , to obtain the bth offspring individual , so as to obtain the tth generation dominated offspring population ; and then through the selection operation of fast non-dominated sorting method and crowded distance calculation, select the better B individuals from and to form the t+1th generation dominated offspring population ; S7, will be combined into the t+1 generation combined population , and the leading edge repair processing is performed to obtain the t+1 generation complement set ; wherein, represents the zth complement of the t+1 generation ; wherein, represents the zth complement of the t+1 generation S8, will with merge into the t+1 generation of the repair population , and select the better NP individuals from to form the t+1 generation population by the fast non-dominated sorting method and the selection operation of end member distance calculation ; S9, after assigning t+1 to t, return to step S4 for execution until t>t max , thus obtaining the evolved high-quality population , from select the individual corresponding to the non-dominated rank of "1" as the optimal endmember extraction result of the spectral remote sensing image.

2. The method of claim 1, wherein the method is based on evolutionary multi-modal endmember extraction. The local search strategy based on niches in S5 is performed according to the following steps: Step a: For the current a-th individual , compute the spectral angle distance between the P endmembers and the P endmembers using equation (4) respectively, thus obtaining an endmember similarity matrix of size P x P: ​​​ (4) Step b: calculate the sum of spectral angle distance between the first endmember and the other endmembers , thus obtaining the sum of P spectral angle distances , and the endmember corresponding to the minimum value in is defined as the endmember to be optimized ​​​​ Step c: Calculate the population of the current generation t. The sum of the indices corresponding to all endmembers in each individual, and combined with... The sum of the indices corresponding to all endmembers is compared, and the result is selected. The n closest individuals constitute Within the niche of the decision space, select the individual whose sum of the first and second objective functions is minimized. As Learning objectives; Step d: Calculate the endmember to be optimized by using formula (4) The spectral angle distance between each endmember and the learning target individual The minimum spectral angle distance in The corresponding endmember is the evolutionary target of the endmember to be optimized A new endmember is randomly selected within the image range of and its radius r , and the endmember to be optimized in is replaced, thereby obtaining a child individual .

3. The method of claim 1, wherein the method is based on evolutionary multi-modal endmember extraction. The global search strategy based on SAD guidance in S6 is performed according to the following steps: Step a: Calculate the current b-th individual using formula (4) the spectral angle distance between each endmember in the spectral angle distance between each endmember in the spectral angle distance between each endmember in the P / 2 endmembers with the smallest spectral angle distance in the P / 2 endmembers with the largest spectral angle distance in the P / 2 endmembers with the largest spectral angle distance in Step b: the image of the selected new endomembrane is replaced by the image of the corresponding endomembrane in the bth sub-generation individual Step c: the image of the selected new endomembrane is replaced by the image of the corresponding endomembrane in the bth sub-generation individual Step d: the image of the selected new endomembrane is replaced by the image of the corresponding endomembrane in the bth sub-generation individual Step e: the image of the selected new endomembrane is replaced by the image of the corresponding endomembrane in the bth sub-generation individual Step f: the image of the selected new endomembrane is replaced by the image of the corresponding endomembrane in the bth sub-generation individual Step g: the image of the selected new endomembrane is replaced by the image of the corresponding endomembrane in the bth sub-generation individual 4. The method of claim 1, wherein the method is based on evolutionary multi-modal endmember extraction. S7 comprises: S7.1: combine the individuals in the tth generation population and the t+1th generation population into a (t+1)th generation combined population with S7.2: sort the individuals in the (t+1)th generation combined population in ascending order according to the first objective function, and divide the individuals into Z individual clusters, each of which contains NP / Z individuals S7.3: select the best individual in each individual cluster as the (t+1)th generation population S7.4: combine the individuals in the tth generation population and the (t+1)th generation population into a (t+1)th generation combined population According to formula (2) and formula (3), the non-dominated rank of each individual in the zth individual cluster is calculated, and the individual with the minimum first objective function value is selected from the individual corresponding to the non-dominated rank of "1" And the calculation of The k individuals closest in the Euclidean distance of the target space, thereby constituting The neighbor set of ; all endmembers contained in all individuals in The zth endmember set of the t+1th generation ; S7.2: selecting the individual with the minimum second objective function value from the zth individual cluster and respectively calculating the central endmember and the endmember set the spectral angle distance of the central endmember, thereby selecting the P endmember pairs with the minimum spectral angle distance, belonging to the endmember in the zth complement of the tth generation ; thereby obtaining the zth complement of the t+1th generation and constituting the complement set of the t+1th generation .

5. The method of claim 1, wherein S8 When the individuals are in the same non-dominated level, the average Euclidean distance of the end members in the individual on the spectral image of R is calculated, so as to reserve the individual corresponding to the larger average Euclidean distance. S8.1: Calculate, according to formula (2) and formula (3), For each individual, the non-dominated rank is constructed for all individuals corresponding to the non-dominated rank of "1", and the non-dominated rank of each individual in the neighborhood set is set to "1"; S8.2: Using end-member distance competition method to Perform environment selection, and reserve NP individuals to constitute the t+1 generation population ; The memory is used for storing a program supporting the processor to execute the hyperspectral end member extraction method in any one of claims 1-5, and the processor is configured to execute the program stored in the memory.

6. An electronic device comprising a memory and a processor, characterized in that The computer program is executed by the processor to perform the steps of the hyperspectral end member extraction method in any one of claims 1-5.

7. A computer-readable storage medium having stored thereon a computer program, characterized in that ​

Images