Filter evolution wave band selection method based on wrapper guidance

Abstract

The invention discloses a filter evolution wave band selection method based on wrapper guidance, relates to the technical field of hyperspectral images, and solves the technical problem that a traditional hyperspectral wave band selection algorithm is only optimized once to obtain a selected wave band subset with a given size and does not consider the relationship between the selected wave band subsets with different sizes. The method comprises the following steps: 1, selecting a hyperspectral band; 2, solving a multi-objective optimization problem; step 3, a filter evolution waveband selection method based on wrapper guidance: 1) dual-population evolution multi-objective definition, 2) a dual-population coding mode, 3) a dual-population evolution strategy, 4) a dual-population interaction process, and 5) environment selection; compared with the most advanced semi-supervised and unsupervised wave band selection algorithms, for several standard hyperspectral data sets, the experimental result shows that the wave band subset obtained by the WFBS method provided by the invention has better performance.

Description

Technical Field

[0001] The invention belongs to the field of hyperspectral images, and in particular is a filter evolution band selection method based on wrapper guidance. Background Art

[0002] In recent years, with the continuous development and maturity of spectral imaging technology, remote sensing technology has been greatly improved in both theory and application, among which hyperspectral imaging technology is one of them. Due to the different absorption of the spectrum by different material components and the difference in reflection degree at different wavelengths, hyperspectral images can fully reflect the differences in the physical structure and chemical composition of the sample, so they are widely used in many fields such as marine water quality monitoring.

[0003] However, the large amount of information in hyperspectral images also brings challenges to data transmission, processing and storage. The huge data volume of hyperspectral images will greatly increase the computational cost, and a large number of redundant bands will also affect the classification accuracy of ground object categories and the accuracy of target detection. Therefore, it is very necessary to perform data dimensionality reduction on hyperspectral images. As a method widely used in dimensionality reduction of hyperspectral image data, band selection (BS) can select a certain number of band images from the original image data and remove redundant and noisy bands. BS can achieve the purpose of dimensionality reduction while maintaining the original physical structure.

[0004] Depending on whether label information is used, hyperspectral image band selection can be divided into supervised, semi-supervised and unsupervised. Supervised band selection requires a large number of labeled samples to calculate the correlation between bands and labels to determine the bands to be selected, but it is difficult to obtain enough labeled samples in practical applications, so unsupervised and semi-supervised methods are more practical. Unsupervised band selection uses a large number of unlabeled samples to calculate the amount of information contained in the bands, the correlation between bands, band noise, etc. to select meaningful bands. On the basis of the unsupervised method, the semi-supervised method uses a small number of labeled samples to further measure the band's ability to distinguish the ground object category in order to select the band that is most conducive to classification.

[0005] According to different search strategies, there are methods based on evolutionary algorithms and methods based on non-evolutionary algorithms. Traditional mathematical programming methods can quickly locate a single optimal solution within the search range, but their limitation is that they may fall into local optimality. Relatively speaking, population-based evolutionary algorithms have a higher probability of discovering the global optimal solution, thereby effectively avoiding the problem of local optimality. Furthermore, multi-objective evolutionary algorithms can find a balance between multiple conflicting objectives and provide a set of optimal solutions. However, evolutionary algorithms are designed around optimization objectives, so when dealing with practical problems, they take more time to evaluate and reach convergence than traditional algorithms. Therefore, for hyperspectral images, it is particularly important to develop evolutionary operators that are both representative and can converge quickly.

[0006] However, these semi-supervised band selection methods based on evolutionary algorithms mainly focus on using different evolutionary algorithm frameworks to design hyperspectral band selection algorithms, and have the following problems: First, these existing works mainly optimize the statistical and spatial information contained in the selected bands, which often have high time complexity. Second, these methods only perform optimization once to obtain a selected band subset of a given size, and rarely consider the relationship between selected band subsets of different sizes. Therefore, the present invention provides a band selection method based on wrapper-guided filter evolution. Summary of the invention

[0007] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a filter evolution band selection method based on wrapper guidance, which is used to solve the technical problem that the traditional hyperspectral band selection algorithm only performs optimization once to obtain a selected band subset of a given size, and does not consider the relationship between selected band subsets of different sizes.

[0008] To achieve the above object, a first aspect of the present invention provides a filter evolution band selection method based on wrapper guidance, comprising the following steps:

[0009] Step 1: Selection of hyperspectral bands;

[0010] Select a band subset S from the original band space R, and the band subset S optimizes the performance criterion J to the minimum;

[0011] Step 2: Multi-objective optimization;

[0012] Establish a minimization objective function and solve a set of equilibrium solutions;

[0013] Step 3: Band selection based on wrapper-guided filter evolution.

[0014] The wrapper-guided filter evolution band selection includes:

[0015] Step S1: For a given band subset, a multi-objective function of dual population evolution is established under labeled samples and unlabeled samples;

[0016] Step S2: Encode the double population;

[0017] The two populations of the dual population adopt different encoding methods according to their respective evolutionary characteristics, and design the Wrapper population and the Filter population;

[0018] Step S3: dual population evolution;

[0019] The wrapper population uses the traditional NSGA-II framework. The individual update operator uses single-point crossover and bitwise mutation, and the new population is obtained by selecting the frontier level and crowding distance.

[0020] The Filter population contains several sub-populations. For each sub-population, it is combined with the sub-population corresponding to the offspring for environmental selection, and the results of environmental selection of each sub-population are returned to form a new population; the number of bands of each sub-population is fixed;

[0021] Step S4: two-population interaction;

[0022] In each generation of the Wrapper population, high-precision solutions are selected and added to the archive, and these high-precision individuals are used to guide the evolution of the Filter population;

[0023] In each generation of the dual population, solutions with classification accuracy greater than the accuracy threshold are selected and added to the archive, and the solutions in the archive guide the evolution of the Filter population.

[0024] Step S5: environment selection;

[0025] According to the interactive evolution iteration of the two populations, the new Wrapper population and Filter population of the current generation are obtained, and the excellent individuals of each generation are selected from the parent and offspring through environmental selection. If the number of iterations is less than the iteration number threshold, the iteration continues, and the best individual is selected according to the filter population evolution maintenance structure goal until the number of iterations reaches the iteration number threshold.

[0026] Preferably, a band subset S is selected from the original band space R, and the band subset S optimizes the performance criterion J to a minimum;

[0027] The hyperspectral image band S is as follows:

[0028]

[0029] Among them, R′ is the candidate band subset; D is the data set, D={x i ∈ZM |1≤i≤N}; where xi is the i-th band in the band subset, M is the number of bands in the band space R, Z is the valid sample value, i is the band number, i∈[1,N], and N is a positive integer representing the sample size;

[0030] If R′ is determined, the corresponding data set D′ is obtained; the performance criterion J depends on the data set D and the candidate band subset R′.

[0031] Preferably, the establishing of a minimization objective function and solving a set of equilibrium solutions comprises:

[0032] Minimize the objective function: Min.F(X)=(F1(X),F2(X),...,F On (X)); where X is the band subset; Min.F(X) is the minimization objective function; F On (X) is the objective function, O n is the number of objective functions;

[0033] If X is not dominated by any other decision vector, then X is a Pareto optimal solution, and the Pareto solution set is the set of all Pareto optimal solutions.

[0034] Preferably, the multi-objective function of the labeled samples is: Minimize.(f1(X), f2(X)); wherein f1(X) is the objective function for optimizing the number of bands, and f2(X) is the objective function for optimizing the classification accuracy;

[0035] Band number optimization objective function: Where xi is the i-th band in the band subset; X is the band subset; i is the band number, i∈[1,N], N is a positive integer;

[0036] Classification accuracy optimization objective function: f2(X) = ErrorRate(X, classifier); where ErrorRate(X) is the classification error rate that can be achieved by extracting the currently selected band subset X from the hyperspectral image for classification; the classifier is the KNN classifier.

[0037] Preferably, the multi-objective function of the unlabeled sample is: Minimize.(g1(Y), g2(Y), g3(Y)); wherein g1(Y) is a similarity objective function, g2(Y) is a structure maintenance objective function, and g3(Y) is an information amount objective function between bands; Y is a solution vector of the selected band subset;

[0038] It should be noted that Y is an integer-coded decision vector. Each value in Y represents the index of a selected band. The reason for using integer coding is to maintain the same number of selected bands in each subpopulation of the Filter.

[0039] Similar objective function:

[0040]

[0041] Among them, MI(x i ,x j ) is the similarity measure between the i-th and j-th selected bands, p(y1) and p(y2) are two probability distribution functions, p(y1,y2) is the joint probability distribution function; Ω is all the pixels in an image;

[0042] It should be noted that the smaller the MI value, the smaller the redundancy of the band subset information.

[0043] Structure maintenance objective function:

[0044]

[0045] Among them, y λ and z is the pixel value of the λth and kth pixels in the hyperspectral image in the selected band; K l is the number of actual adjacent pixels around the current pixel; function dist(y i ,y j ) is y λ and z The Euclidean distance between z and y′ z are the pixel values ​​of the z-th and z-th pixels in the hyperspectral image in the selected band; Szz′ is the neighbor relationship in all bands; z is the number of any pixel in the image, z∈[1,n]; λ is the pixel adjacent to the z-th pixel; considering that i and i′ are usually located in different regions, the non-local consistency target actually reflects the global structure in space.

[0046] The objective function of the amount of information between bands is:

[0047]

[0048] Among them, BN is the number of selected bands; R and C represent the number of rows and columns of the hyperspectral remote sensing image, respectively; is the pixel value of the rth row and cth column in the characteristic image of the i-th band in the hyperspectral remote sensing image; is the average pixel value of the i-th band image in the hyperspectral image.

[0049] The present invention calculates the variance value of each pixel point on the selected band and the average pixel value, reflecting the degree of dispersion on the band, that is, the amount of information contained in the band.

[0050] Preferably, the encoding of the dual populations comprises:

[0051] For the encoding of the Wrapper population, each individual is encoded using binary 0-1 encoding, where 1 indicates selecting the current band and 0 indicates not selecting it;

[0052] For the encoding of the Filter population, each individual is encoded using an integer, where the integer value represents the index of the selected band.

[0053] Preferably, the updating process of the Filter population includes updating the Filter population and updating the individual Filters;

[0054] The update of the Filter population includes the following steps:

[0055] Step A1: Initialize the Leader matrix to store the leader individuals of each subpopulation; the number of bands of each subpopulation is fixed;

[0056] Step A2: For each subpopulation, based on the offset density estimation, calculate the similarity between the current population and each individual in its adjacent population, mark it as the SDE value, and select the individual with the smallest SDE value as the leader of the current subpopulation;

[0057] Step A3: For each solution in the subpopulation, learn from the leader through the learning strategy to obtain a new offspring of the subpopulation. For the update of the leader, randomly select an individual from the external archive, and update the individual learning strategy through the leader, and finally integrate all the updated solutions together to form a new population;

[0058] For the update of individuals in the Filter population, the following steps are included:

[0059] Step B1: Extract two individuals, label them LeaderA and LeaderT respectively, and select their common band subsets, label them common;

[0060] Step B2: Remove common from LeaderT to obtain the band set LeaderTR, then enter the loop to generate a random;

[0061] Step B3: If the random number is greater than 0.5, a band is randomly selected from LeaderA; if the random number is less than 0.5, a band is randomly selected from LeaderTR; when the cycle ends, the randomly selected band is combined with common to form a new individual, marked as Ynew.

[0062] Preferably, the dual population interaction comprises the following steps:

[0063] Step C1: Input the external archive and the Wrapper population Pw, sort Pw in ascending order according to the error rate, sort the external archive in descending order, and enter a loop;

[0064] Step C2: After entering the loop, select the individual p with the lowest error rate from Pw, remove p from Pw, and select the individual q with the highest error rate from the external archive;

[0065] Step C3: Compare p and q. If the error rate of p is less than that of q or the error rate of p does not exceed 0.001 of q and the number of bands of p is at least 5 less than that of q, then add p to the external archive and remove q from the external archive.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] 1. The present invention designs a solution with high preservation accuracy for the external archive strategy, which serves as an intermediary for information exchange between the two populations. The corresponding update operator is designed for the external archive, so that effective information exchange can be achieved between the two populations.

[0068] 2. The experiment of the present invention can obtain band subsets with different numbers of bands through one run, providing multiple options for decision makers to choose from. The experimental results show that on three common hyperspectral image test sets, compared with the four latest hyperspectral band selection algorithms and three feature selection algorithms, this method has obvious performance advantages;

[0069] 3. The present invention defines the hyperspectral band selection problem as a multi-objective optimization problem, and simultaneously optimizes the uniformity of the band distribution of the selected band subset and the inter-class separation distance from a small number of labeled samples. In order to obtain a band subset set with low redundancy, high information content and high classification accuracy, the present invention will use a small number of labeled samples and a large number of unlabeled samples, and use a method based on the combination of wrappers (Wrapper) and filters (Filter) to design a band selection algorithm (WFBS) guided by wrapper filter evolution. In order to implement WFBS, the present invention proposes a dual population interaction strategy based on external archiving for information interaction between the Wrapper population and the Filter population to obtain high-quality individuals with a fixed number of bands. Compared with the most advanced semi-supervised and unsupervised band selection algorithms, for different standard hyperspectral data sets, experimental results show that the band subsets obtained by the WFBS method proposed in the present invention have better performance, and the time complexity is smaller than that of other evolutionary band selection methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0071] Figure 1 It is the algorithm framework diagram of the present invention;

[0072] Figure 2 This is an example of the dual population encoding strategy of the present invention;

[0073] Figure 3 This is the Filter population evolution process of the present invention;

[0074] Figure 4 This is an example of the Leader-based population learning strategy of the present invention. DETAILED DESCRIPTION

[0075] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0076] See also Figure 1 The first aspect of the present invention provides a filter t-band selection method based on wrapper guidance, comprising the following steps:

[0077] Step 1. Selection of hyperspectral bands;

[0078] Hyperspectral image band selection aims to select a band subset S from the original band space R, which should optimize the performance criterion J. Assuming that the performance criterion J is optimized to the minimum and the candidate band subset is denoted by R', the problem of hyperspectral image band selection can be described as follows:

[0079]

[0080] Among them, R′ is the candidate band subset; D is the data set, D={x i ∈Z M |1≤i≤N}; where xi is the i-th band in the band subset, M is the number of bands in the band space R, Z is the valid sample value, i is the band number, i∈[1,N], and N is a positive integer representing the sample size;

[0081] If R′ is determined, the corresponding data set D′ is formed, marked as {y i∈Z M |1≤i≤N}; the performance criterion J depends on the dataset D and the candidate band subset R′.

[0082] Step 2. Multi-objective optimization;

[0083] Multi-objective optimization problems aim to maximize or minimize multiple objectives at the same time to find a set of equilibrium solutions. A multi-objective optimization framework for minimization problems can be defined as follows:

[0084]

[0085] Among them, X is the band subset; Min.F(X) is the minimization objective function; F On (X) is the objective function, O n is the number of objective functions;

[0086] Assume that given two decision vectors X1 and X2, if all objectives satisfy F i (X1)≤F i (X2), for all i, then X1 dominates X2 or X2 is dominated by X1 (expressed as X1>X2); if a decision vector X is not dominated by any other decision vector, then X is a Pareto optimal solution; the Pareto solution set is the set of all Pareto optimal solutions, which is represented as the Pareto frontier in the target space.

[0087] Step 3. Wrapper-guided filter evolution band selection method;

[0088] Since most feature selection methods do not consider the advantages of both the Wrapper method and the Filter method, or do not effectively combine the two methods, in view of this, this paper proposes a hyperspectral image band selection method that uses Wrapper to guide the evolution of Filter to improve the classification accuracy of hyperspectral images. The excellent solutions in the Wrapper population saved in the external archive are used to guide the evolution of the Filter population. At the same time, in order to enhance the information interaction between the two populations, individuals in the Filter population are selected to participate in the evolution of the Wrapper, thereby obtaining a better band subset.

[0089] Step S1: Multi-objective function of dual population evolution

[0090] For the dual-population evolution, two objective functions are defined, and minimizing the number of bands becomes one of the optimization objectives. By reducing the number of bands, the data can be simplified and key information can be retained.

[0091] The objective function for band number optimization is defined as:

[0092]

[0093] Where xi is the i-th band in the band subset; X is the band subset; i is the band number, i∈[1,N], and N is the dimension of the binary vector X.

[0094] In addition, due to the existence of a small number of labeled samples, the classification accuracy can be directly used as one of the optimization objectives of the band subset performance, so that the time complexity is acceptable while finding a band subset with better quality.

[0095] The band classification accuracy optimization objective function is defined as:

[0096] f2(X)=ErrorRate(X,classifier);

[0097] Among them, ErrorRate(X) is the classification error rate that can be achieved by extracting the currently selected band subset X from the hyperspectral image for classification; the classifier classifier is the KNN classifier.

[0098] For a given band subset, the number of bands and the classification accuracy of the selected bands are optimized at the same time, and a solution with as few bands as possible and as high classification accuracy as possible is searched. The problem can be stated as:

[0099] Minimize.(f1(X),f2(X));

[0100] Traditional filter optimization targets redundancy and information content, finding a band subset with low redundancy and high information content, and rarely considering structural features at the same time. Therefore, mutual information is used to measure the redundancy between bands, variance is used to measure the information content between bands, and the structure maintenance target measures the information content of the band space.

[0101] First, MI (mutual information) is an effective indicator in information theory, which is used to measure the redundancy of band subsets. Generally speaking, the smaller the MI value, the smaller the redundancy of band subset information. For two features Y1 and Y2, the corresponding MI is defined as follows:

[0102]

[0103] Among them, p(y1) and p(y2) are two probability distribution functions, p(y1,y2) is the joint probability distribution function; Ω is all the pixels in an image;

[0104] Similar objective function:

[0105]

[0106] Among them, MI(y i ,y j) is a measure of the similarity between the i-th and j-th selected bands; Y is the solution vector of the selected band subset.

[0107] In order to take into account both local and global structures, the structure-maintaining objective can be defined as:

[0108]

[0109] Among them, y λ and z is the pixel value of the λth and kth pixels in the hyperspectral image in the selected band; K l is the number of actual adjacent pixels around the current pixel; function dist(y i ,y j ) is y λ and z The Euclidean distance between z and y′ z is the pixel value of the zth and z′th pixels in the hyperspectral image in the selected band; S zz ′ is the neighbor relationship in all bands; z is the number of any pixel in the image, z∈[1,n]; λ is the pixel adjacent to the zth pixel; considering that z and z′ are usually located in different regions, the non-local consistency target actually reflects the global structure in space.

[0110] The objective function of the amount of information between bands is:

[0111]

[0112] Among them, BN is the number of selected bands; R and C represent the number of rows and columns of the hyperspectral remote sensing image, respectively; is the pixel value of the rth row and cth column in the characteristic image of the i-th band in the hyperspectral remote sensing image; is the average pixel value of the i-th band image in the hyperspectral image.

[0113] The present invention calculates the variance value of each pixel point on the selected band and the average pixel value, reflecting the degree of dispersion on the band, that is, the amount of information contained in the band.

[0114] For a given band subset, without the participation of labeled samples, the optimized objective function is as follows:

[0115] Minimize.(g1(Y),g2(Y),g3(Y));

[0116] Among them, Y is the solution vector of the selected band subset, which is an integer-coded decision vector, and each value in Y represents the index of the selected band. Integer coding is used because it is necessary to maintain the same number of selected bands in each subpopulation of the Filter.

[0117] Step S2: encoding method of dual populations;

[0118] See also Figure 2 ,The band selection problem of hyperspectral images is a classic discrete ,combinatorial optimization problem. For the coding of discrete ,problems, binary 0-1 coding and integer coding are usually used.

[0119] Since the population evolved based on the Wrapper method needs to optimize the number of bands and classification accuracy, this paper uses binary 0-1 coding to design the Wrapper population. In order to ensure that a solution with a fixed number of bands can be generated in one experiment, integer coding is used to design the Filter population, and the corresponding evolutionary operator is designed so that the length of the code remains unchanged during the evolution process to achieve the purpose of fixing the number of bands.

[0120] Specifically, for the encoding of the Wrapper population, each individual is encoded using binary 0-1 encoding; where 1 indicates selecting the current band and 0 indicates not selecting;

[0121] For the encoding of the Filter population, each individual is encoded using an integer, where the integer value represents the index of the selected band.

[0122] See also Figure 3 , Step S3: Dual population evolution strategy, the specific process steps are as follows:

[0123] For the Wrapper population, the traditional NSGA-II framework is used, and the individual update operator uses single-point crossover and bitwise mutation, and a new population is obtained by selecting the frontier level and crowding distance. In the present invention, a small amount of labeled data is used for classification, and the classification results are obtained using the KNN classifier.

[0124] The update of the Filter population includes the following steps:

[0125] Step A1: Initialize the Leader matrix to store the leader individuals of each subpopulation; the number of bands of each subpopulation is fixed;

[0126] Step A2: For each subpopulation, based on the offset density estimation, calculate the similarity between the current population and each individual in its adjacent population, mark it as the SDE value, and select the individual with the smallest SDE value as the leader of the current subpopulation;

[0127] Step A3: For each solution in the subpopulation, learn from the leader through the learning strategy to obtain a new offspring of the subpopulation. For the update of the leader, randomly select an individual from the external archive, and update the individual learning strategy through the leader, and finally integrate all the updated solutions together to form a new population;

[0128] That is, for each sub-population, it is combined with the sub-population corresponding to the offspring for environmental selection, and the selection is still based on the frontier level and crowding distance. Finally, the results of each sub-population environmental selection are returned to form a new O f .

[0129] See also Figure 4 , for the update of individuals in the Filter population, the following steps are included:

[0130] Step B1: Extract two individuals, label them LeaderA and LeaderT respectively, and select their common band subsets, label them common;

[0131] Step B2: Remove common from LeaderT to obtain the band set LeaderTR, then enter the loop to generate a random;

[0132] Step B3: If the random number is greater than 0.5, a band is randomly selected from LeaderA; if the random number is less than 0.5, a band is randomly selected from LeaderTR; when the cycle ends, the randomly selected band is combined with common to form a new individual, marked as Ynew.

[0133] Step S4: two-population interaction process;

[0134] The classification task of hyperspectral images focuses on classification and has high requirements for classification accuracy, while the number of bands is acceptable within a certain range. Therefore, an external archiving strategy based on accuracy priority is proposed. In each generation of the Wrapper population, high-precision solutions are selected and added to the archive to avoid losing the original excellent individuals due to optimizing the number of bands. For external archives of fixed size, these high-precision individuals are then used to guide the evolution of the Filter population.

[0135] Among them, in each generation of the dual population, solutions with classification accuracy greater than the accuracy threshold are selected and added to the archive, and the solutions in the archive guide the evolution of the Filter population.

[0136] The specific algorithm flow is as follows:

[0137] Step C1: Input the external archive and the Wrapper population Pw, sort Pw in ascending order according to the error rate, sort the external archive in descending order, and enter a loop;

[0138] Step C2: After entering the loop, select the individual p with the lowest error rate from Pw, remove p from Pw, and select the individual q with the highest error rate from the external archive;

[0139] Step C3: Compare p and q. If the error rate of p is less than that of q or the error rate of p does not exceed 0.001 of q and the number of bands of p is at least 5 less than that of q, then add p to the external archive and remove q from the external archive.

[0140] Because the amount of hyperspectral image data is huge, there may be some individuals with very different numbers of bands but very little difference in accuracy during classification. Therefore, it is believed that individuals with slightly lower accuracy but a small number of bands are also important.

[0141] Step S5: environment selection;

[0142] According to the interactive evolution of the two populations, the new Wrapper population and Filter population of the current generation are obtained, and the excellent individuals of each generation are selected from the parent generation and the offspring generation through environmental selection. If the number of iterations is less than the iteration number threshold, the iteration continues, and the best individual is selected according to the filter population evolution maintenance structure goal until the number of iterations reaches the iteration number threshold;

[0143] Among them, for the output stage, the solution of the first frontier in the Wrapper population and the solution with the maximum precision for each band number in the external archive are first selected as part of the optimal solution set. If there is no solution for the current number of bands, the solution with the largest amount of information is selected from the subpopulation corresponding to the Filter population as a supplement, and several sets of solutions with solutions for each band number are obtained.

[0144] Part of the data in the above formula is calculated by removing the dimension and taking its numerical value. The formula is a formula closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0145] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A filter evolution band selection method based on wrapper guidance, characterized in that: The following steps are involved: Step 1: Selection of hyperspectral bands; Select a band subset S from the original band space R, and the band subset S optimizes the performance criterion J to the minimum; Step 2: Multi-objective optimization; Establish a minimization objective function and solve a set of equilibrium solutions; Step 3: Band selection based on wrapper-guided filter evolution.

2. The method for band selection of filter evolution based on wrapper guidance according to claim 1, characterized in that: The wrapper-guided filter evolution band selection includes: Step S1: For a given band subset, a multi-objective function of dual population evolution is established under labeled samples and unlabeled samples; Step S2: Encode the double population; The two populations of the dual population adopt different encoding methods according to their respective evolutionary characteristics, and design the Wrapper population and the Filter population; Step S3: dual population evolution; The wrapper population uses the traditional NSGA-II framework. The individual update operator uses single-point crossover and bitwise mutation, and the new population is obtained by selecting the frontier level and crowding distance. The Filter population contains several sub-populations. For each sub-population, it is combined with the sub-population corresponding to the offspring for environmental selection, and the results of each sub-population environmental selection are returned to form a new population. Step S4: two-population interaction; In each generation of the Wrapper population, high-precision solutions are selected and added to the archive, and these high-precision individuals are used to guide the evolution of the Filter population; Step S5: environment selection; According to the interactive evolution iteration of the two populations, the new Wrapper population and Filter population of the current generation are obtained, and the excellent individuals of each generation are selected from the parent and offspring through environmental selection; if the number of iterations is less than the iteration number threshold, the iteration continues, and the best individual is selected according to the filter population evolution maintenance structure goal until the number of iterations reaches the iteration number threshold.

3. The method for band selection of filter evolution based on wrapper guidance according to claim 1, characterized in that: The step of selecting a band subset S from the original band space R, wherein the band subset S optimizes the performance criterion J to a minimum; The hyperspectral image band S is as follows: Among them, R′ is the candidate band subset; D is the data set, D={x i ∈Z M |1≤i≤N}; where xi is the i-th band in the band subset, M is the number of bands in the band space R, Z is the valid sample value, i is the band number, i∈[1,N], and N is a positive integer representing the sample size; If R′ is determined, the corresponding data set D′ is obtained; the performance criterion J depends on the data set D and the candidate band subset R′.

4. The method for selecting a band of filter evolution based on wrapper guidance according to claim 2, characterized in that: The minimization objective function is established and a set of equilibrium solutions are solved, including: Minimize the objective function: Where X is the band subset; Min.F(X) is the minimization objective function; is the objective function, O n is the number of objective functions; If X is not dominated by any other decision vector, then X is a Pareto optimal solution, and the Pareto solution set is the set of all Pareto optimal solutions.

5. The method for selecting a band of filter evolution based on wrapper guidance according to claim 4, characterized in that: The multi-objective function of labeled samples: Minimize.(f1(X),f2(X)); wherein f1(X) is the objective function for optimizing the number of bands, and f2(X) is the objective function for optimizing the classification accuracy; Band number optimization objective function: Where xi is the i-th band in the band subset; X is the band subset; i is the band number, i∈[1,N], N is a positive integer; Classification accuracy optimization objective function: f2(X) = ErrorRate(X, classifier); where ErrorRate(X) is the classification error rate that can be achieved by extracting the currently selected band subset X from the hyperspectral image for classification; the classifier is the KNN classifier.

6. The method for band selection of filter evolution based on wrapper guidance according to claim 4, characterized in that: The multi-objective function of the unlabeled sample: Minimize.(g1(Y), g2(Y), g3(Y)); wherein g1(Y) is a similarity objective function, g2(Y) is a structure maintenance objective function, and g3(Y) is an information amount objective function between bands; Y is a solution vector of the selected band subset; Similar objective function: Among them, MI(x i ,x j ) is the similarity measure between the i-th and j-th selected bands, p(y1) and p(y2) are two probability distribution functions, p(y1,y2) is the joint probability distribution function; Ω is all the pixels in an image; Structure maintenance objective function: Among them, y λ and z is the pixel value of the λth and kth pixels in the hyperspectral image in the selected band; K l is the number of actual adjacent pixels around the current pixel; function dist(y i ,y j ) is y λ and z The Euclidean distance between z and y′ z is the pixel value of the zth and zth′ pixels in the hyperspectral image in the selected band; Szz′ is the neighbor relationship in all bands; z is the number of any pixel in the image, z∈[1,n]; λ is the pixel adjacent to the zth pixel; The information content objective function between bands: Among them, BN is the number of selected bands; R and C represent the number of rows and columns of the hyperspectral remote sensing image, respectively; is the pixel value of the rth row and cth column in the characteristic image of the i-th band in the hyperspectral remote sensing image; is the average pixel value of the i-th band image in the hyperspectral image.

7. The method for band selection of filter evolution based on wrapper guidance according to claim 2, characterized in that: The encoding of the dual populations comprises: For the encoding of the Wrapper population, each individual is encoded using binary 0-1 encoding, where 1 indicates selecting the current band and 0 indicates not selecting it; For the encoding of the Filter population, each individual is encoded using an integer, where the integer value represents the index of the selected band.

8. The method for band selection of filter evolution based on wrapper guidance according to claim 2, characterized in that: The updating process of the Filter population includes: updating the Filter population and updating the individual Filters; The update of the Filter population includes the following steps: Step A1: Initialize the Leader matrix to store the leader individuals of each subpopulation; the number of bands of each subpopulation is fixed; Step A2: For each subpopulation, based on the offset density estimation, calculate the similarity between the current population and each individual in its adjacent population, mark it as the SDE value, and select the individual with the smallest SDE value as the leader of the current subpopulation; Step A3: For each solution in the subpopulation, learn from the leader through the learning strategy to obtain a new offspring of the subpopulation. For the update of the leader, randomly select an individual from the external archive, and update the individual learning strategy through the leader, and finally integrate all the updated solutions together to form a new population; For the update of individuals in the Filter population, the following steps are included: Step B1: Extract two individuals, label them LeaderA and LeaderT respectively, and select their common band subsets, label them common; Step B2: Remove common from LeaderT to obtain the band set LeaderTR, then enter the loop to generate a random; Step B3: If the random number is greater than 0.5, a band is randomly selected from LeaderA; if the random number is less than 0.5, a band is randomly selected from LeaderTR; when the cycle ends, the randomly selected band is combined with common to form a new individual, marked as Ynew.

9. The method for band selection of filter evolution based on wrapper guidance according to claim 2, characterized in that: The dual population interaction comprises the following steps: Step C1: Input the external archive and the Wrapper population Pw, sort Pw in ascending order according to the error rate, sort the external archive in descending order, and enter a loop; Step C2: After entering the loop, select the individual p with the lowest error rate from Pw, remove p from Pw, and select the individual q with the highest error rate from the external archive; Step C3: Compare p and q. If the error rate of p is less than that of q or the error rate of p does not exceed 0.001 of q and the number of bands of p is at least 5 less than that of q, then add p to the external archive and remove q from the external archive.

10. The method for selecting a band of filter evolution based on wrapper guidance according to claim 2, characterized in that: The new Wrapper population and Filter population of the current generation are obtained by iteratively evolving the two populations interactively, including: For the output stage, the solution of the first frontier in the Wrapper population and the solution with the maximum precision for each band number in the external archive are selected as part of the optimal solution set. If there is no solution for the current number of bands, the solution with the largest amount of information is selected from the subpopulation corresponding to the Filter population as a supplement, and several sets of solutions are obtained for which each band number has a solution.

Images