End member extraction method based on large-scale sparse optimization algorithm

Through large-scale sparse optimization algorithm and multi-objective optimization function, combined with population evolution strategy, the problem of large search space and repeated calculations in end-member extraction of hyperspectral images is solved, achieving a more efficient and accurate end-member extraction effect.

CN118941810APending Publication Date: 2024-11-12ANHUI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411089327.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-09
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The existing hyperspectral image end-element extraction algorithm has huge search space, repeated calculations and low accuracy, and it is impossible to effectively search all pixels, resulting in inaccurate end-element results.

Method used

Large-scale sparse optimization algorithm is adopted to record the exploration of each pixel through multi-objective optimization functions and population evolution strategies, and use multiple descendant generation methods to reduce the search space, reduce repeated calculations, and improve the accuracy of end-member extraction.

Benefits of technology

Faster and more accurate end-member extraction is achieved, repeated calculations are avoided, diversity and accuracy of understanding is improved, and decision-making space can be searched more comprehensively.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941810A_ABST
    Figure CN118941810A_ABST
Patent Text Reader

Abstract

The invention discloses an end member extraction method based on a large-scale sparse optimization algorithm, and the method comprises the steps: 1, importing hyperspectral remote sensing image data, and initializing a population and a global record; 2, screening out an excellent end member composition solution through the global record, reducing an end member search space, and continuously updating the global record in an iteration process; 3, a multi-seed population generation mode is adopted, population diversity is increased, and end members are searched comprehensively; and 4, combining the newly generated sub-population with the original population, obtaining a Pareto solution set through non-dominated sorting and congestion distance, and obtaining a group of hyperspectral end member extraction results through multiple iterative evolution. According to the method, the quality of the end member can be measured from multiple angles, the problems of repeated and incomplete end member search and overlarge search space are solved, and the end member extraction result is more accurate and diversified.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of hyperspectral image processing, in particular to a hyperspectral endmember extraction method based on large-scale sparse concept. Background Art

[0002] Hyperspectral sensors can capture spectral information of continuous narrow bands in space to generate three-dimensional images, namely hyperspectral images. Compared with ordinary images, each pixel of hyperspectral images has higher dimensions, often hundreds of dimensions, and the information they contain is richer. However, due to the limitations of imaging space, only a very small number of pixels in the generated hyperspectral images are composed of one substance, namely endmembers, and the vast majority of pixels contain information of multiple substances at the same time. That is, mixed pixels. In order to conduct more accurate research and analysis on hyperspectral images, the concept of hyperspectral unmixing is proposed, that is, to solve the different endmembers contained in the mixed pixels and their ratios, namely abundance. The key to hyperspectral unmixing is endmember extraction.

[0003] With the continuous progress of genetic algorithms and single-objective and multi-objective evolutionary algorithms, more and more evolutionary algorithms are applied to endmember extraction, but existing endmember extraction algorithms often have the following problems: 1. Hyperspectral images usually contain tens of thousands of pixels and have a large search space. Some algorithms reduce the computing time and resource consumption by reducing the search space to obtain better solutions. The methods used are relatively simple, and the search space is reduced by clustering, multi-tasking, etc. 2. Most existing algorithms randomly search the image space without distinguishing between searched and unsearched pixels, resulting in the same pixel being selected for calculation multiple times, causing repeated solutions and a large number of repeated calculations, and not being able to effectively search all pixels. 3. The endmember results obtained by the existing methods are not accurate, ignoring the diversity of the decision space, and cannot effectively obtain richer and higher-quality solutions. Summary of the invention

[0004] In order to address the deficiencies in the prior art, the present invention proposes an endmember extraction method of a large-scale sparse optimization algorithm, which can record the exploration status of each pixel in the image, use a variety of offspring generation methods, in order to measure the quality of the endmember from multiple angles, quickly explore the entire image, to solve the problems of endmember duplication and repeated calculation, and further improve the accuracy of the endmember, so as to achieve a more perfect endmember extraction effect.

[0005] The present invention adopts the following technical solutions to solve the above technical problems:

[0006] The invention discloses a method for extracting endmembers from a hyperspectral spectrum based on a large-scale sparse optimization algorithm, which comprises the following steps:

[0007] S1 Let the number of pixel rows of the three-dimensional hyperspectral remote sensing image be , the number of pixel columns is , the number of bands is ;

[0008] Get the dimension Two-dimensional hyperspectral remote sensing image matrix ,in, is the number of pixels of the hyperspectral remote sensing image, and ;

[0009] Assume the number of endmembers of the hyperspectral remote sensing image is ;

[0010] S2 uses equations (1) to (3) to construct a multi-objective optimization function :

[0011] (1)

[0012] (2)

[0013] f 2 = ( N p − 1 ) ! | d e t [ 1 … 1 a 1 … a N p ] | (3)

[0014] In formula (1) to formula (3), represents the first optimization objective function, is the hyperspectral remote sensing image matrix The pixels, For the Pixels The reconstruction value of represents the second optimization objective function, represents the calculation of the determinant of the matrix, is the spectral remote sensing image matrix After dimensionality reduction end members, among which, express After dimensionality reduction end members;

[0015] S3 defines the current number of iterations as , and initialize , define the maximum number of iterations as ;

[0016] Assume the population size is , and initialize the dimensions to No. Generation population ,in, Indicates Individuals, and ,in, express Middle genes;

[0017] S4 generates Generation population Subpopulation or ;

[0018] S5 will subpopulation or With Generation population After the merger, the composition contains The merged population of individuals is then calculated using non-dominated sorting and crowding distance to eliminate half of the inferior individuals, and the remaining half of the high-quality individuals form the first Generation population ;

[0019] S6 Assign to After that, if , then the Generation population The endmembers corresponding to all non-dominated individuals in are taken as the optimal endmember extraction results; otherwise, return to S4 to continue sequential execution.

[0020] The hyperspectral endmember extraction method based on a large-scale sparse optimization algorithm described in the present invention is also characterized in that the first Generation population The steps include:

[0021] S3.1 Initialize global logging information is an all-zero matrix with dimension ;

[0022] S3.2 According to the number of rows , Number of columns , from the two-dimensional hyperspectral remote sensing image matrix Evenly selected The pixels constitute a location set ,in, express Middle pixels;

[0023] S3.3 Initialization Generation population The Individual ; thus completing the Generation population Initialization of all individuals in .

[0024] S3.3 includes the following steps:

[0025] S3.3.1 In Random selection Members make up A collection of pixel locations ,by is the center pixel, and the surrounding dimensions are The pixel positions constitute the Regions , thus obtaining the region set ;

[0026] S3.3.2 Update A collection of pixel locations :

[0027] Step a: Initialization ;

[0028] Step b: equal For each pixel position in , use equations (1) to (3) to calculate The corresponding objective function value is the minimum objective function value in The corresponding pixel position in the Genes in , thus obtaining The first Gene , and record information globally Update The element value at the corresponding position: assign the minimum objective function value to Middle elements, middle The corresponding element values ​​at other positions are set to -1;

[0029] Step c: Assign to ,like , it means that the Individual Otherwise, return to step b and execute sequentially.

[0030] S4 includes the following steps:

[0031] S4.1 Determine global record information Is there a 0 value? If so, execute S4.2-4.5; otherwise, execute S4.6-4.9;

[0032] S4.2 Obtaining global record information The positions of elements with median values ​​greater than 0 are randomly combined to generate individuals and constitute the first subpopulation ;

[0033] S4.3 Recording information from the global Select the position of the element with value 0 and generate according to the principle of S3.3 individuals and form the second subpopulation ;

[0034] S4.4 Using the binary bidding method Two parent individuals are selected to randomly exchange end members, and the number of exchanged end members is less than , to produce two offspring individuals; thus producing individuals and constitute the third subpopulation ;

[0035] S4.5 by , , Subpopulation ;

[0036] S4.6 is generated according to the process of S4.2 individuals and constitute the first subpopulation ;

[0037] S4.7 Using the binary bidding method Two parent individuals are selected from the endmembers to produce two new offspring individuals, thus producing individuals and form the second subpopulation ;

[0038] The endmember fine-tuning is randomly selected from the parent individual endmembers as the replaced endmembers, and then randomly select a pixel around the replaced endmember for replacement. ;

[0039] S4.8 is generated according to the process of S4.4 individuals and constitute the third subpopulation ;

[0040] S4.9 by , , Subpopulation .

[0041] The electronic device of the present invention includes 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.

[0042] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the characteristic feature is that the computer program executes the steps of the hyperspectral endmember extraction method when executed by a processor.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] 1. Apply the sparse optimization idea to endmember extraction and propose a global search record. While narrowing the search space and accelerating the algorithm execution, it prevents repeated calculations and ensures a comprehensive search of the decision space.

[0045] 2. Perform computational optimization on the two objectives, measure the two evaluation directions of the endmember solution, based on the entire decision space, and avoid falling into the local optimum; overcome noise interference and be suitable for endmember extraction of hyperspectral images in various scenarios.

[0046] 3. Use multiple sub-population generation methods, not limited to traditional offspring generation methods, to increase population diversity, improve solution diversity, and increase solution accuracy.

[0047] 4. A new method of narrowing the search space is used to apply the idea of ​​sparse problems to endmember extraction to more comprehensively search the decision space, reduce unnecessary repetitive calculations, and obtain a solution set with rich diversity and higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0049] Considering that the number of endmembers in a hyperspectral image is sparse compared to the number of mixed pixels, it meets the large-scale sparsity problem. In this embodiment, an endmember extraction method based on a large-scale sparse optimization algorithm is proposed. It is based on global records for evolution, uses the sparse idea to narrow the search space, and uses two population update strategies that are different from the traditional ones. It can reduce repeated calculations, improve efficiency, and can more comprehensively search the image space to obtain a diverse and accurate endmember solution set. The overall process is as follows Figure 1 As shown, the method is characterized by comprising the following steps:

[0050] Step 1: Let the number of pixel rows of the three-dimensional hyperspectral remote sensing image be , the number of pixel columns is , the number of bands is ;

[0051] Get the dimension Two-dimensional hyperspectral remote sensing image matrix ,in, is the number of pixels of the hyperspectral remote sensing image, and ;

[0052] Assume the number of endmembers of the hyperspectral remote sensing image is ;

[0053] Step 2: Use equations (1) to (3) to construct a multi-objective optimization function :

[0054] (1)

[0055] (2)

[0056] f 2 = ( N p − 1 ) ! | d e t [ 1 … 1 a 1 … a N p ] | (3)

[0057] In formula (1) to formula (3), represents the first optimization objective function, is the hyperspectral remote sensing image matrix The pixels, For the Pixels The reconstruction value of represents the second optimization objective function, represents the calculation of the determinant of the matrix, is the spectral remote sensing image matrix After dimensionality reduction end members, among which, express After dimensionality reduction end members;

[0058] Step 3: Define the current number of iterations as , and initialize ; Define the maximum number of iterations as , assuming the population size is , and initialize the dimensions to No. Generation population ,in, Indicates Individuals, and ,in, express Middle genes; using the position of the pixel in the image as the gene, , if the individual middle The value of 128 means The 128th pixel is Middle genes;

[0059] Step 3.1: Initialize global logging information is an all-zero matrix with dimension ;

[0060] Step 3.2: According to the number of rows , Number of columns , from the two-dimensional hyperspectral remote sensing image matrix Evenly selected The pixels constitute a location set ,in, express Middle pixels;

[0061] Step 3.3: Initialize Generation population The Individual :

[0062] Step 3.3.1 Random selection Members make up A collection of pixel locations ,by is the center pixel, and the surrounding dimensions are The pixel positions constitute the Regions , The same operation is performed on other pixels in the image to obtain the region set , whose dimensions are ;

[0063] in, is an integer, according to The number of pixels contained in , select an appropriate value; It is a plane with the band as the z-axis, the horizontal direction of the image as the x-axis, and the vertical direction as the y-axis. Figure area, set 3, not considered For the edge of the image, It can be expressed as: ;

[0064] S3.3.2 Update A collection of pixel locations :

[0065] Step a: Initialization ;

[0066] Step b: equal For each pixel position in , use equations (1) to (3) to calculate The corresponding objective function value is the minimum objective function value in The corresponding pixel position in the Genes in , thus obtaining The first Gene , and record information globally Update The element value at the corresponding position: assign the minimum objective function value to Middle elements, middle The corresponding element values ​​at other positions are set to -1;

[0067] Step c: Assign to ,like , it means that the Individual Otherwise, return to step b and execute sequentially.

[0068] Step 3.4: Follow the process in step 3.3 to complete the Generation population Initialization of all individuals in;

[0069] Step 4: Generate the Generation population subpopulations;

[0070] Step 4.1: Determine global record information Is there a 0 value? If so, execute steps 4.2-4.5; otherwise, execute steps 4.6-4.9;

[0071] The position of the element with a median value of 0 corresponds to the two-dimensional hyperspectral image matrix The pixel positions that are not selected as end members;

[0072] Step 4.2: Get global record information The positions of elements with median values ​​greater than 0 are randomly combined to generate individuals and constitute the first subpopulation ;

[0073] The location of the element with a median value greater than 0 indicates: in the two-dimensional hyperspectral image matrix Around In the region, the pixel is selected as the end member, and the objective function value is the smallest, that is, The best performing pixel in the region;

[0074] Step 4.3: Record information from the global Select the position of the element with value 0 and generate according to the principle of step 3.3 individuals and form the second subpopulation ;

[0075] Step 4.4: Use the Binary Bidding Method to Two parent individuals are selected to randomly exchange end members, and the number of exchanged end members is less than , to produce two offspring individuals; thus producing individuals and constitute the third subpopulation ;

[0076] Step 4.5: By , , Subpopulation ;

[0077] Step 4.6: Follow the process in step 4.2 to generate individuals and constitute the first subpopulation ;

[0078] Step 4.7: Use the Binary Bidding Method to Two parent individuals are selected from the endmembers to produce two new offspring individuals, thus producing individuals and form the second subpopulation ;

[0079] The endmember fine-tuning is randomly selected from the parent individual endmembers as the replaced endmembers, and then randomly select a pixel around the replaced endmember for replacement. ;

[0080] Step 4.8: Follow the process of step 4.4 to generate individuals and constitute the third subpopulation ;

[0081] Step 4.9: By , , Subpopulation ;

[0082] Step 5: Subpopulation or With Generation population After the merger, the composition contains The merged population of individuals is then calculated using non-dominated sorting and crowding distance to eliminate half of the inferior individuals, and the remaining half of the high-quality individuals form the first Generation population ;

[0083] Step 6: Assign to After that, if , then the Generation population The endmembers corresponding to all non-dominated individuals in are taken as the optimal endmember extraction results; otherwise, return to S4 to continue sequential execution.

[0084] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above-mentioned hyperspectral endmember extraction method, and the processor is configured to execute the program stored in the memory.

[0085] In this embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the hyperspectral endmember extraction method are executed.

Claims

1. A hyperspectral endmember extraction method based on a large-scale sparse optimization algorithm, characterized in that: The steps include: S1 Let the number of pixel rows of the three-dimensional hyperspectral remote sensing image be , the number of pixel columns is , the number of bands is ; Get the dimension Two-dimensional hyperspectral remote sensing image matrix ,in, is the number of pixels of the hyperspectral remote sensing image, and ; Assume the number of endmembers of the hyperspectral remote sensing image is ; S2 uses equations (1) to (3) to construct a multi-objective optimization function : (1) (2) (3) In formula (1) to formula (3), represents the first optimization objective function, is the hyperspectral remote sensing image matrix The pixels, For the Pixels The reconstruction value of represents the second optimization objective function, represents the calculation of the determinant of the matrix, is the spectral remote sensing image matrix After dimensionality reduction end members, among which, express After dimensionality reduction end members; S3 defines the current number of iterations as , and initialize , define the maximum number of iterations as ; Assume the population size is , and initialize the dimensions to No. Generation population ,in, Indicates Individuals, and ,in, express Middle genes; S4 generates Generation population Subpopulation or ; S5 will subpopulation or With Generation population After the merger, the composition contains The merged population of individuals is then calculated using non-dominated sorting and crowding distance to eliminate half of the inferior individuals, and the remaining half of the high-quality individuals form the first Generation population ; S6 Assign to After that, if , then the Generation population The endmembers corresponding to all non-dominated individuals in are taken as the optimal endmember extraction results; otherwise, return to S4 to continue sequential execution.

2. The method for extracting endmembers from a hyperspectral image based on a large-scale sparse optimization algorithm according to claim 1, characterized in that: Initialize the first Generation population The steps include: S3.1 Initialize global logging information is an all-zero matrix with dimension ; S3.2 According to the number of rows , Number of columns , from the two-dimensional hyperspectral remote sensing image matrix Evenly selected The pixels constitute a location set ,in, express Middle pixels; S3.3 Initialization Generation population The Individual ; thus completing the Generation population Initialization of all individuals in .

3. The hyperspectral endmember extraction method based on large-scale sparse optimization algorithm according to claim 2 is characterized in that: S3.3 includes the following steps: S3.3.1 In Random selection Members make up A collection of pixel locations ,by is the center pixel, and the surrounding dimensions are The pixel positions constitute the Regions , thus obtaining the region set ; S3.3.2 Update A collection of pixel locations : Step a: Initialization ; Step b: equal For each pixel position in , use equations (1) to (3) to calculate The corresponding objective function value is the minimum objective function value in The corresponding pixel position in the Genes in , thus obtaining The first Gene , and record information globally Update The element value at the corresponding position: assign the minimum objective function value to Middle elements, middle The corresponding element values ​​at other positions are set to -1; Step c: Assign to ,like , it means that the Individual Otherwise, return to step b and execute sequentially.

4. The hyperspectral endmember extraction method based on large-scale sparse optimization algorithm according to claim 3 is characterized in that: S4 includes the following steps: S4.1 Determine global record information Is there a 0 value? If so, execute S4.2-4.5; otherwise, execute S4.6-4.9; S4.2 Obtaining global record information The positions of elements with median values ​​greater than 0 are randomly combined to generate individuals and constitute the first subpopulation ; S4.3 Recording information from the global Select the position of the element with value 0 and generate according to the principle of S3.3 individuals and form the second subpopulation ; S4.4 Using the binary bidding method Two parent individuals are selected to randomly exchange end members, and the number of exchanged end members is less than , to produce two offspring individuals; thus producing individuals and constitute the third subpopulation ; S4.5 by , , Subpopulation ; S4.6 is generated according to the process of S4.2 individuals and constitute the first subpopulation ; S4.7 Using the binary bidding method Two parent individuals are selected from the endmembers to produce two new offspring individuals, thus producing individuals and form the second subpopulation ; The endmember fine-tuning is randomly selected from the parent individual endmembers as the replaced endmembers, and then randomly select a pixel around the replaced endmember for replacement. ; S4.8 is generated according to the process of S4.4 individuals and constitute the third subpopulation ; S4.9 by , , Subpopulation .

5. 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 4, and the processor is configured to execute the program stored in the memory.

6. 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 4 are executed.