A method for automatic extraction of endmember bundles from multi-modal and multi-target hyperspectral images

By using bionic optimization algorithm and ring topology in the end element beam extraction of hyperspectral images, combined with relative spectral angular distance and quantum particle swarm update, the spectral changes and insufficient extraction accuracy of multimodal multi-objective problems in the existing technology are solved. The problem is to achieve efficient and accurate end element bundle extraction.

CN116543279BActive Publication Date: 2024-11-08深圳市规划和自然资源数据管理中心(深圳市空间地理信息中心) +1
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202310514059.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-05-08
Publication Date
2024-11-08
Estimated Expiration
2043-05-08

AI Technical Summary

Technical Problem

When the existing hyperspectral image end element beam extraction algorithm deals with multimodal multi-objective problems, it is difficult to take into account spectral changes, extraction accuracy and efficiency, and ignores the globality of the image and the distribution characteristics of the decision space, resulting in the extracted end element beam. Not comprehensive and accurate enough.

Method used

The multimodal multi-objective hyperspectral image end element beam automatic extraction method is adopted based on the bionic optimization algorithm, and particle information interaction is realized through the ring topology structure, the crowded distance is introduced to calculate the decision space of relative spectral angle distance, and the quantum particle swarm position update method is improved. Avoid local optimal traps, and judge iterative termination through distance information to improve extraction efficiency and accuracy.

Benefits of technology

More effective end element beams are extracted in hyperspectral images, which improves extraction accuracy and efficiency, can better handle multiple evaluation indicators, and has higher universality and accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116543279B_ABST
    Figure CN116543279B_ABST
Patent Text Reader

Abstract

The invention relates to the application field of hyperspectral remote sensing, and discloses a method for automatically extracting endmember bundles of multimodal multi-target hyperspectral images. The method comprises the following steps: step 1, encoding hyperspectral data; step 2, performing environment configuration; step 3, evaluating the fitness value of an initialized particle position P; step 4, updating a neighborhood optimal archive NBA; step 5, updating the particle position P; step 6, evaluating the fitness value of the updated particle position P; step 7, updating an individual optimal archive PBA; step 8, judging whether the update is successful; step 9, judging whether the program reaches a maximum number of iterations Max_Gen; and step 10, outputting a result of a neighborhood optimal archive NBA, which is a candidate endmember bundle. The present invention can take into account the change of each type of spectrum of the endmember bundle, the extraction accuracy and the efficiency, and obtain more effective endmember bundles.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the application field of hyperspectral remote sensing, and specifically to an automatic extraction method for multi-modal multi-target hyperspectral image endmember bundles, and more particularly to an automatic extraction method combining a bionic optimization algorithm and a multi-target multi-modal task with hyperspectral image endmember bundle extraction. Background Art

[0002] Hyperspectral image endmember extraction is the basis for obtaining rich spatial information of ground objects. The spectral signals obtained by the hyperspectral remote sensor are recorded in pixels. The pixel is a combination of the spectral signals of the corresponding surface materials. Affected by the resolution of the hyperspectral remote sensor, each pixel usually contains more than one type of ground object, that is, a mixed pixel. Endmembers are the main parameters to describe mixed pixels, representing a characteristic ground object with a relatively fixed spectrum. Endmembers only contain one type of ground object information, and pixels usually include multiple ground objects.

[0003] Affected by the environment, image resolution, etc., the endmembers of the same type of objects also have spectral variations. By extracting endmember bundles for each type of object, this phenomenon can be effectively avoided, which is of great significance for spectral unmixing.

[0004] At present, there are fewer algorithms for extracting endmember bundles than for extracting endmembers. Existing methods use block extraction, that is, extracting endmembers block by block through random (EBE), projection (Max-D), and clustering (SSEBE) methods, screening out singular values, and then merging them into endmember bundles. This type of method mostly relies on the accuracy of the block, and cannot guarantee that each block can have endmember information, ignoring the globality of the image, and the screening method has uncertainty. In addition, iterative methods, such as IQPSO and MODPSO, treat endmember extraction as a multi-objective optimization problem for solution. This method measures the effect of the solution only in the target space. Once a good convergence result is achieved, it is considered that the multi-objective optimization problem has been solved, ignoring the multi-modal problem, that is, the distribution characteristics of the decision space. Therefore, in the design of the endmember bundle extraction algorithm, it is necessary to take into account the changes in each type of spectrum of the endmember bundle, the extraction accuracy and efficiency, and obtain more effective endmember bundles. Summary of the invention

[0005] The present invention aims to provide a method for automatically extracting endmember bundles of multi-modal multi-target hyperspectral images, so as to take into account the variation of each type of spectrum of the endmember bundle, the extraction accuracy and efficiency, and obtain more effective endmember bundles.

[0006] To achieve the above object, the present invention adopts the following technical solution: a method for automatically extracting endmember bundles of multi-modal multi-target hyperspectral images, comprising:

[0007] Step 1, inputting hyperspectral data and performing row and column encoding according to the specifications of the hyperspectral data;

[0008] Step 2, perform environment configuration, the environment configuration includes setting relevant parameters, setting objective function and relevant initialization, the relevant parameters include particle population size M, maximum number of iterations Max_Gen, iteration early termination index Gen_Percent, particle weight update parameter CE, number of end members E_Num; the relevant initialization includes individual optimal archive PBA initialization, neighborhood optimal archive NBA initialization, random integer particle position P initialization within specification range, and construct random individual optimal archive PBA initialization and neighborhood optimal archive NBA initialization with row and column coded particles;

[0009] Step 3, two objective functions are used to evaluate the fitness value of the initialized particle position P;

[0010] Step 4: Update the neighborhood optimal archive NBA through the ring topology update rule;

[0011] Step 5, update the particle position P;

[0012] Step 6, using two objective functions to evaluate the fitness value of the updated particle position P;

[0013] Step 7, updating the individual optimal archive PBA;

[0014] Step 8, judging whether the neighborhood optimal archive NBA and the individual optimal archive PBA are updated successfully, if the update is successful, execute step 9, if the update is not successful, execute step 10;

[0015] Step 9, determine whether the program has reached the maximum number of iterations Max_Gen, if yes, execute step 10, if no, execute step 4;

[0016] Step 10, output the result of the neighborhood optimal archive NBA, which is the candidate endmember bundle.

[0017] The principles and advantages of this scheme are: in actual application, the hyperspectral data is first encoded, and then the environment is configured, and the encoded data is used for related initialization; then the neighborhood optimal archive NBA is updated through the ring topology update rule, and the particle position P is updated at the same time; the individual optimal archive PBA is updated; then it is determined whether the neighborhood optimal archive NBA and the individual optimal archive PBA are updated successfully, and then it is determined whether the neighborhood optimal archive NBA and the individual optimal archive PBA reach the maximum number of iterations Max_Gen. When the neighborhood optimal archive NBA and the individual optimal archive PBA reach the maximum number of iterations Max_Gen, the result in the neighborhood optimal archive NBA is output, which is the candidate end member beam.

[0018] Compared with the prior art, the beneficial effects of the present invention are:

[0019] 1. Use a ring-based topology structure to enable individuals in different neighborhoods to interact with each other, so that each particle only transmits information with the historical optimal position and neighbors, avoiding random transmission of particles;

[0020] 2. Replace the global optimal position with the optimal position in the neighborhood, and effectively avoid falling into the local optimal position through the improved quantum particle swarm position update method;

[0021] 3. Introduce relative spectral angular distance (RSAD) to calculate the crowding distance of the decision space, improve the diversity of particle decision space, and be more universal in different hyperspectral image endmember extraction problems;

[0022] 4. Use distance information update to automatically determine whether the particle iteration needs to be terminated early to avoid running to the maximum number of iterations and causing information redundancy;

[0023] 5. It can extract more effective endmember bundles in hyperspectral images and has higher accuracy in multiple evaluation indicators.

[0024] Preferably, as an improvement, the step 5 further comprises:

[0025] Step 5.1, calculate the distance information S_NBA of each row of the neighborhood optimal archive NBA, and re-sort the neighborhood optimal archive NBA from large to small according to the distance information S_NBA value;

[0026] Step 5.2, select the first row of the corresponding sorting of the neighborhood best archive NBA and the individual best archive PBA as the neighborhood best nbest and the individual best pbest, and use the neighborhood best nbest and the individual best pbest to update the particle position P.

[0027] Technical effect: The update of particle position P by the neighborhood optimal nbest and individual optimal pbest can enable the particle to be updated to a more optimal solution position.

[0028] Preferably, as an improvement, the distance information S_NBA includes a decision space distance CDDS and a target space distance CDOS; and the step 5.1 further includes:

[0029] Step 5.11, the decision space distance CDDS is calculated by the relative spectral angle distance (RSAD), and the specific calculation formula is:

[0030]

[0031] In the formula, e a 、e bare the end member curves extracted from two particles respectively. When the particle is a non-boundary point, the calculation formula of the decision space distance CDDS is:

[0032]

[0033] In the formula, a represents the spatial dimension of the particle, and j represents the particle index; when j = M, the calculation formula of the decision space distance CDDS is:

[0034]

[0035] When j=1, the calculation formula of the decision space distance CDDS is:

[0036]

[0037] In step 5.12, all decision space distances CDDS of the particles are added up, and then the average value is calculated, and the average value is compared and sorted with the particle's own value.

[0038] Technical effect: By sorting the decision space distance CDDS, the optimal particle can be found in the decision space for the next round of iteration. The larger the decision space distance CDDS ranking, the greater the contribution to particle update.

[0039] Preferably, as an improvement, the step 5.1 further comprises:

[0040] Step 5.13, the calculation formula of the target space distance CDOS is:

[0041]

[0042] In the formula, f(x a,j ) represents the target space distance value; if the mth target value of the jth particle is the smallest, CDOS j When the mth component is set to 1, when the mth target value of the jth particle is the largest, CDOS j The mth component is set to 0;

[0043] Step 5.14, add up all the target space distances CDOS of the particle, calculate the average value, and compare and sort with its own value.

[0044] Technical effect: By sorting the target space distance CDOS, the optimal particles can be found for the next round of iteration. The larger the target space distance CDOS, the greater the contribution to particle update.

[0045] Preferably, as an improvement, the step 7 further comprises:

[0046] Step 7.1, calculate the distance information \(S_{PBA}\) for each row of the individual optimal archive \(PBA\), and sort \(PBA\) in descending order according to the value of the distance information \(S_{PBA}\).

[0047] Step 7.2, replace the values of the original individual optimal archive \(PBA\) with those of the sorted individual optimal archive \(PBA\).

[0048] Technical effect: It can provide sorting information for particles.

[0049] Preferably, as an improvement, the ring topology update rule is: when \(1 < j < M\) for a particle, the particle \(j\) only conducts information interaction with the \((j - 1)\)-th and \((j + 1)\)-th particles; when \(j = 1\), the particle \(j\) conducts information interaction with the \(M\)-th and 2nd particles; when \(j = M\), the particle \(j\) conducts information interaction with the first and \((M - 1)\)-th particles.

[0050] Technical effect: The ring topology enables particles to only conduct information interaction with nearby particles, can control the information transfer speed, and achieve effective optimization.

[0051] Preferably, as an improvement, in the said Step 8, the judgment rule is: judge whether the update times of the distance information \(S_{PBA}\) are greater than the maximum number of iterations \(Max\_Gen\times\) iteration early termination index \(Gen\_Percent\). If so, the update is successful; if not, the update stops.

[0052] Technical effect: By adjusting the result of the maximum number of iterations \(Max\_Gen\times\) iteration early termination index \(Gen\_Percent\), it can be judged whether the iteration terminates early, preventing meaningless iterations from consuming running time.

[0053] Preferably, as an improvement, the range of the iteration early termination index \(Gen\_Percent\) is \([0, 1]\).

[0054] Technical effect: The termination index \(Gen\_Percent\) is any number within the range of \([0, 1]\). The larger this value is, the more times the distance information \(S_{PBA}\) is updated, and it is more inclined to more iterations. The update times of the distance information \(S_{PBA}\) can be controlled through the termination index.

[0055] Preferably, as an improvement, in the said Step 3, the objective function is the root mean square error of the difference between the back-mixed image remixed after extracting the endmember solution abundances and the original image. The calculation formula of the root mean square error is:

[0056]

[0057] In the formula, \(L\) represents the number of hyperspectral bands, \(N\) represents the number of pixels, where \(N = m\times n\), and \(r\) i represents a pixel, Represents the root mean square error RMSE_ucls and RMSE_fcls obtained by solving the abundance and re-synthesizing the remixed image using the unconstrained least squares method and the fully constrained least squares method.

[0058] Technical effect: The root mean square error can reflect the difference between the anti-mixed image and the original image.

[0059] Preferably, as an improvement, in step 5, the updating method used for location updating is:

[0060]

[0061] In the formula, It represents the jth individual optimal pbest in the current iteration, and then calculates the position update P of the jth particle through the individual optimal pbest and the neighborhood optimal nbest j and the particle's position update x j , the position of the jth particle is updated P j and the particle's position update x j The specific calculation formula is:

[0062] P j =φ·pbest j +(1-φ)nbest

[0063]

[0064] Where φ and u are the mean distribution values ​​on (0,1).

[0065] Technical effect: Through this update method, the global search performance is good, the control parameters are few, and the calculation speed is fast. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a flowchart of a method for automatic extraction of endmember bundles from multi-modal and multi-target hyperspectral images;

[0067] Figure 2 A schematic diagram of the ring topology of an automatic method for extracting endmember bundles from multi-modal and multi-target hyperspectral images;

[0068] Figure 3 A schematic diagram of the RSAD calculation method for an automatic endmember bundle extraction method for multi-modal multi-target hyperspectral images;

[0069] Figure 4 A schematic diagram of a target space CDOS calculation method for an automatic endmember bundle extraction method for multi-modal multi-target hyperspectral images;

[0070] Figure 5A multi-modal multi-target hyperspectral image endmember bundle automatic extraction method is applied to the actual extracted endmember bundles. DETAILED DESCRIPTION

[0071] The following is further described in detail through specific implementation methods:

[0072] The embodiment is basically as shown in the attached Figure 1 As shown:

[0073] A method for automatically extracting endmember bundles from multi-modal multi-target hyperspectral images, comprising:

[0074] Step 1, input hyperspectral data, and perform row and column encoding according to the specifications of the hyperspectral data; for example, input hyperspectral data R with a size of m×n×L in width×height×band, perform m×n row and column encoding according to the width and height of the hyperspectral data;

[0075] The hyperspectral data input in this embodiment is Samson data. The image includes 156 bands in total, covering wavelengths from 401nm to 889nm, with a spectral resolution of 3.13nm and an image size of 95×95 pixels, including three end members: "1# soil", "2# tree", and "3# water".

[0076] First, the particles are encoded in rows and columns. For the hyperspectral data image of size m×n×L, they are numbered in rows and columns and encoded by the pixel position in the hyperspectral data image. Each particle is represented as:

[0077] Z j ={[x1,x2,...,x E_Num ,y1,y2,...,y E_Num ] T |x k ∈S m ,y k ∈S n},j=1,2,...,M

[0078] In the formula, Z j represents a sequence of labels corresponding to j particle pixels, x k ,y k Respectively represent the row and column coordinates of the position corresponding to the kth end member; S m ,S n Represent the serial number sets of rows and columns respectively; through this encoding method, the position encoding of the particles can be restored to the image to form a set of end-member solutions.

[0079] Step 2: Perform environment configuration, which includes setting relevant parameters, setting the objective function and related initialization.

[0080] The relevant parameters include particle population size M, maximum number of iterations Max_Gen, iteration early termination index Gen_Percent, particle weight update parameter CE, and end member number E_Num; in this application, the range of iteration early termination index Gen_Percent is [0,1].

[0081] In order to demonstrate the effect of extracting endmember bundles, the particle population size M is set to 40, the maximum number of iterations Max_Gen is set to 600, the iteration early termination index Gen_Percent is set to 0.05, the particle weight update parameter CE is set to [0.5,0.5], the number of endmembers E_Num is set to 3, and the number of effective endmember bundles actually extracted for each endmember is 9.

[0082] It should be noted that in order to extract more effective endmember bundles, the operation is terminated when the maximum number of iterations is reached in this embodiment. However, when dealing with the endmember extraction problem of other hyperspectral images, different iteration early termination index Gen_Percent values ​​can be set to terminate the iteration early to prevent redundant calculations.

[0083] The related initialization includes individual optimal archive PBA initialization, neighborhood optimal archive NBA initialization, and random integer particle position P initialization within the specification range, where the specification range is m×n. Here, the row-column encoded particles are constructed to initialize the random individual optimal archive PBA and the neighborhood optimal archive NBA. The role of random initialization is that particles do not have a fixed position when they are initially updated, and the position of the particles is given in a random manner, which is random initialization.

[0084] Step 3, two objective functions are used to evaluate the fitness value of the initialized particle position P, and the fitness value is the calculation result of the two objective functions; the two objective functions are: root mean square error function RMSE_ucls and RMSE_fcls; their role is: the smaller the fitness value, the more optimized the particle is, so the fitness value is used for measurement;

[0085] In step 3, the objective function is the root mean square error of the difference between the remixed image and the original image after extracting the end member solution abundance. The calculation formula of the root mean square error is:

[0086]

[0087] In the formula, L represents the number of hyperspectral bands, N represents the number of pixels, where N = m × n, ri represents the pixel, The root mean square errors RMSE_ucls and RMSE_fcls obtained by solving the abundance and re-synthesizing the remixed image using the unconstrained least squares method and the fully constrained least squares method are the two objective functions.

[0088] Step 4: Update the neighborhood optimal archive NBA according to the ring topology update rule; the neighborhood optimal archive NBA is updated in each iteration, and its role is that the neighborhood optimal archive NBA provides values for sorting, and thus provides a priority order for the new iteration of the particles; as Figure 2 shown, the ring topology update rule is: when the particle 1 < j < M, the particle j only performs information interaction with the (j - 1)-th and (j + 1)-th particles; when j = 1, the particle j performs information interaction with the M-th and 2nd particles; when j = M, the particle j performs information interaction with the first and (M - 1)-th particles, where j represents the particle index.

[0089] Step 5: Update the particle position P according to a preset update method to perform optimization through the update; the update method used for position update is as follows:

[0090]

[0091] In the formula, represents the j-th individual optimal pbest in the current iteration, and then calculates the position update P of the j-th particle through the individual optimal pbest and the neighborhood optimal nbest j and the position update x of the particle j , the position update P of the j-th particle j and the position update x of the particle j The specific calculation formulas are as follows:

[0092] P j = φ·pbest j +(1 - φ)nbest

[0093]

[0094] In the formula, φ and u are numerical values with a uniform distribution on (0, 1), pbest is the individual optimal, nbest is the neighborhood optimal, and CE is the particle weight update parameter.

[0095] The coding method in this embodiment results in the particles being updated in a discrete form. Therefore, after the position of each particle is updated, rounding is required. At the same time, in order to ensure that the updated particle position is within the optimization range, the above formula is further optimized as follows:

[0096]

[0097] ceil()x r

[0098] In the formula, represents the rounding function, and x r represents a random number within the interval S m Similarly, for the column coordinate index y jThe same location update method is used and y r Take the interval S n Random numbers within.

[0099]

[0100] In the present application, the distance information is fused to determine the spatial distance, that is, the target spatial distance information, and the step 5 further includes:

[0101] Step 5.1, calculate the distance information S_NBA of each row of the neighborhood optimal archive NBA, and reorder the neighborhood optimal archive NBA from large to small according to the distance information S_NBA value; the distance information S_NBA includes the decision space distance CDDS and the target space distance CDOS; the distance information S_NBA calculation method is:

[0102]

[0103] For solving the decision space distance, the step 5.1 further includes:

[0104] Step 5.11, the decision space distance CDDS is calculated by the relative spectral angle distance (RSAD), and the specific calculation formula is:

[0105]

[0106] In the formula, e a 、e b are the end member curves extracted from two particles respectively. When the particle is a non-boundary point, the calculation formula of the decision space distance CDDS is:

[0107]

[0108] In the formula, a represents the spatial dimension of the particle, and j represents the particle index; when j = M, the calculation formula of the decision space distance CDDS is:

[0109]

[0110] When j=1, the calculation formula of the decision space distance CDDS is:

[0111]

[0112] In this embodiment, the hyperspectral image is a three-end member extraction problem. Taking particle ② as an example, its CDDS calculation method is:

[0113]

[0114] When calculating boundary particles, they are calculated by themselves and critical particles. For example, the calculation method of particle ④ is:

[0115]

[0116] In step 5.12, all decision space distances CDDS of the particles are added up, and then the average value is calculated, and the average value is compared and sorted with the particle's own value.

[0117] like Figure 3 As shown, considering the multimodal problem, the decision space distance needs to be sorted in the process of updating the particle individual optimal archive PBA, the neighborhood optimal archive NBA and the particle position P. In step 5.11, the decision space distance CDDS is calculated by the relative spectral angular distance (RSAD). The sorting of the decision space affects the distribution of the next particle. Therefore, the particle distribution with a distance greater than the average distance is selected first to improve the diversity of the space. However, the endmember beam extraction problem of hyperspectral images is similar to a discrete problem. When actually calculating the particle space, taking particle No. 1 as an example, the straight-line distance between particle No. 3 and particle No. 1 in the coordinate position is the farthest, which is theoretically better for improving spatial diversity. In fact, although point No. 4 is closer to point No. 1, it tends to be a pure endmember. From the perspective of diversity contribution: x 1,4 >x 1,3 =x 1,2 ,Therefore, the present invention adopts RSAD index to measure the ,crowding distance of the decision space, which ensures the optimization ,speed on one hand, and improves the spatial diversity on the ,other hand.

[0118] For the solution of the target space, such as Figure 4 As shown, RMSE_ucls and RMSE_fcls are selected as the objective functions of the two objective optimization problems, the dimension of the target space is 2, and the same objective function is used in steps 3 and 6.

[0119] Step 5.13, the calculation formula of the target space distance CDOS is:

[0120]

[0121] In the formula, f(x a,j ) represents the target space distance value; if the mth target value of the jth particle is the smallest, CDOS j When the mth component is set to 1, when the mth target value of the jth particle is the largest, CDOS j The mth component is set to 0;

[0122] In this embodiment, particle ② is taken as an example, and the CDOS calculation method is:

[0123]

[0124] If the mth target value of the jth particle is the smallest, CDOS jWhen the mth component is set to 1, when the mth target value of the jth particle is the largest, CDOS j The mth component is set to 0. For example, the distance of particle ④ in the target space is CDOS4=0+1=1.

[0125] Step 5.14, add up all the target space distances CDOS of the particle, calculate the average value, and compare and sort with its own value.

[0126] Step 5.2, select the first row of the corresponding sorting of the neighborhood best archive NBA and the individual best archive PBA as the neighborhood best nbest and the individual best pbest, and use the neighborhood best nbest and the individual best pbest to update the particle position P.

[0127] Step 6, using two objective functions to evaluate the fitness value of the updated particle position P;

[0128] Step 7, updating the individual optimal archive PBA to provide information for particle update; said step 7 also includes:

[0129] Step 7.1, calculate the distance information S_PBA of each row of the individual optimal archive PBA, and sort the PBA from large to small according to the distance information S_PBA value;

[0130] Step 7.2, replace the value of the original individual optimal archive PBA with the sorted individual optimal archive PBA.

[0131] Step 8, determine whether the neighborhood optimal archive NBA and the individual optimal archive PBA are updated successfully, if the update is successful, execute step 9, and if the update is terminated, execute step 10; in step 8, the judgment rule is: determine whether the number of updates of the distance information S_PBA is greater than the maximum number of iterations Max_Gen×iteration early termination index Gen_Percent, if so, the update is successful, otherwise the update is terminated.

[0132] Step 9, determine whether the neighborhood optimal archive NBA and the individual optimal archive PBA have reached the maximum number of iterations Max_Gen, if so, execute step 10, if not, execute step 4;

[0133] Step 10, output the result of the neighborhood optimal archive NBA, which is the candidate endmember bundle.

[0134] The SAD values ​​of the spectral angle information of the 9 groups of endmembers extracted are shown in Table 1. Each type of ground object contains a total of 9 groups of endmembers, and all 9 groups of endmembers can be used as candidate endmembers in the problem of deconvoluting abundance and demixing. The accuracy evaluation indicators of the extracted endmember bundles are shown in Table 2. The root mean square error RMSE_fcls of the deconvoluted and resynthesized remixed image using the fully constrained least squares method is 0.0056, and the mSAD of the 9 groups of endmembers is 0.03572. The visualization results of the endmember bundles of the three extracted ground objects are shown in Figure 5 As shown. Compared with other endmember bundle extraction methods (IQPSO, MODPSO), this method comprehensively considers the distribution rules of decision space and target space. The MODPSO method extracts only 1-2 groups of endmembers for each feature, while this method can obtain multiple groups of valid endmember bundles. Secondly, the IQPSO method controls the cumulative number of algorithm runs by setting a threshold, which has more iterative redundancy and low efficiency. This method is more efficient and obtains all endmember bundles when the iteration is optimal.

[0135] Table 1 SAD values ​​of each end member

[0136]

[0137] Table 2 Samson data accuracy results

[0138]

[0139] The above is only an embodiment of the present invention, and the common knowledge such as the known specific technical solutions and / or characteristics in the solution is not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several modifications and improvements can be made, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for automatically extracting endmember bundles from multi-modal multi-target hyperspectral images, characterized in that: include: Step 1, inputting hyperspectral data and performing row and column encoding according to the specifications of the hyperspectral data; Step 2, perform environment configuration, the environment configuration includes setting relevant parameters, setting objective function and relevant initialization, the relevant parameters include particle population size M, maximum number of iterations Max_Gen, iteration early termination index Gen_Percent, particle weight update parameter CE, number of end members E_Num; the relevant initialization includes individual optimal archive PBA initialization, neighborhood optimal archive NBA initialization, random integer particle position P initialization within specification range, and construct random individual optimal archive PBA initialization and neighborhood optimal archive NBA initialization with row and column coded particles; Step 3, two objective functions are used to evaluate the fitness value of the initialized particle position P; Step 4: Update the neighborhood optimal archive NBA through the ring topology update rule; Step 5, updating the particle position P; said step 5 also includes: Step 5.1, calculate the distance information S_NBA of each row of the neighborhood optimal archive NBA, and re-sort the neighborhood optimal archive NBA from large to small according to the distance information S_NBA value; Step 5.2, select the first row of the corresponding sorting of the neighborhood best archive NBA and the individual best archive PBA as the neighborhood best nbest and the individual best pbest, and use the neighborhood best nbest and the individual best pbest to update the particle position P; In step 5, the updating method used for location updating is: In the formula, It represents the jth individual optimal pbest in the current iteration, and then calculates the position update P of the jth particle through the individual optimal pbest and the neighborhood optimal nbest j and the particle's position update x j , the position of the jth particle is updated P j and the particle's position update x j The specific calculation formula is: P j =φ·pbest j +(1-φ)nbest Where φ and u are the mean distribution values ​​on (0,1); Step 6, using two objective functions to evaluate the fitness value of the updated particle position P; Step 7, updating the individual optimal archive PBA; Step 8, judging whether the neighborhood optimal archive NBA and the individual optimal archive PBA are updated successfully, if the update is successful, execute step 9, if the update is not successful, execute step 10; Step 9, determine whether the neighborhood optimal archive NBA and the individual optimal archive PBA have reached the maximum number of iterations Max_Gen, if so, execute step 10, if not, execute step 4; Step 10, output the result of the neighborhood optimal archive NBA, which is the candidate endmember bundle.

2. The method for automatically extracting endmember bundles from multi-modal multi-target hyperspectral images according to claim 1, characterized in that: The distance information S_NBA includes the decision space distance CDDS and the target space distance CDOS; the step 5.1 also includes: Step 5.11, the decision space distance CDDS is calculated by the relative spectral angle distance (RSAD), and the specific calculation formula is: In the formula, e a 、e b are the end member curves extracted from two particles respectively. When the particle is a non-boundary point, the calculation formula of the decision space distance CDDS is: In the formula, a represents the spatial dimension of the particle, and j represents the particle index; when j = M, the calculation formula of the decision space distance CDDS is: When j=1, the calculation formula of the decision space distance CDDS is: In step 5.12, all decision space distances CDDS of the particles are added up, and then the average value is calculated, and the average value is compared and sorted with the particle's own value.

3. The method for automatically extracting endmember bundles from multi-modal multi-target hyperspectral images according to claim 2, characterized in that: The step 5.1 also includes: Step 5.13, the calculation formula of the target space distance CDOS is: In the formula, f(x a,j ) represents the target space distance value; if the mth target value of the jth particle is the smallest, CDOS j When the mth component is set to 1, when the mth target value of the jth particle is the largest, CDOS j The mth component is set to 0; Step 5.14, add up all the target space distances CDOS of the particle, calculate the average value, and compare and sort with the own value.

4. The method for automatically extracting endmember bundles from multi-modal multi-target hyperspectral images according to claim 3, characterized in that: The said step 7 further includes: Step 7.1, calculate the distance information S_PBA of each row of the individual optimal archive PBA, and sort PBA in descending order according to the value of the distance information S_PBA; Step 7.2, replace the value of the original individual optimal archive PBA with the sorted individual optimal archive PBA.

5. The method for automatically extracting endmember bundles from multi-modal multi-target hyperspectral images according to claim 1, characterized in that: The said ring topology update rule is: when 1 < j < M for the particle, the particle j only conducts information interaction with the (j - 1)-th and (j + 1)-th particles; when j = 1, the particle j conducts information interaction with the M-th and 2nd particles; when j = M, the particle j conducts information interaction with the first and (M - 1)-th particles.

6. The method for automatically extracting endmember bundles from multi-modal multi-target hyperspectral images according to claim 1, characterized in that: In the said step 8, the judgment rule is: judge whether the update times of the distance information S_PBA are greater than the maximum iteration number Max_Gen × the iteration early termination index Gen_Percent. If so, the update is successful; otherwise, the update stops.

7. The method for automatically extracting endmember bundles from multi-modal multi-target hyperspectral images according to claim 1, characterized in that: The range of the said iteration early termination index Gen_Percent is [0, 1].

8. The method for automatically extracting endmember bundles from multi-modal multi-target hyperspectral images according to claim 1, characterized in that: In the said step 3, the objective function is the root mean square error between the remixed image after extracting the endmember solution abundances and the original image, and the calculation formula of the root mean square error is: In the formula, L represents the number of hyperspectral bands, N represents the number of pixels, where N = m × n, r i represents a pixel, Represents the root mean square error RMSE_ucls and RMSE_fcls obtained by solving the abundance and re-synthesizing the remixed image using the unconstrained least squares method and the fully constrained least squares method.

Citation Information

Patent Citations

  • A modified particle swarm intelligent optimization method for solving high-dimensional optimization problems of large oil and gas production systems

    AU2020103709A4

  • Automatic extraction method for hyperspectral remote sensing image end member beam

    CN113780168A

  • Method and apparatus for quantitative and depth resolved hyperspectral fluorescence and reflectance imaging for surgical guidance

    US20160278678A1

  • End-member extraction method based on segmented vertex component analysis (VCA)

    US20190392261A1

  • Multi-objective multimodal particle swarm optimization method based on bayesian adaptive resonance

    WO2022007376A1