Swarm intelligence optimization band selection method for hyperspectral image target detection

Abstract

The application discloses a swarm intelligence optimization wave band selection method for hyperspectral image target detection, and comprises the following steps: S1, reading hyperspectral image data X, target signal d, and the number K of wave bands needed to be selected; S2, establishing a multi-objective optimization model based on information entropy, joint spectrum-space similarity and CEM error; S3, starting iteration, determining a global optimal solution gbest from a non-inferior solution set rep by adopting a roulette wheel operator strategy, calculating an individual optimal solution pbest, and updating the position and speed of each individual in the population; S4, performing crossover and mutation operations on each individual in the population; S5, calculating the fitness of each individual in the population P t+1 , updating the non-inferior solution set; and S6, performing evaluation by adopting an optimal / suboptimal ratio criterion, and selecting a best wave band subset. The application overcomes the one-sidedness of wave band evaluation of a single criterion method, and can realize effective dimension reduction of a hyperspectral image.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of remote sensing image processing, and particularly relates to a swarm intelligence optimization band selection method for hyperspectral image target detection. BACKGROUND

[0002] Hyperspectral remote sensing image usually contains hundreds of continuous narrow bands of ultraviolet, visible light, near infrared and mid infrared in the electromagnetic spectrum, providing rich spatial and spectral resolution, so that it can be used to accurately determine the ground object category. However, the high dimension and redundancy characteristics of hyperspectral image also cause difficulties for subsequent processing and analysis. As one of the effective means of dimensionality reduction of hyperspectral data, band selection aims to select a small number of bands with rich information and strong representativeness from the original data, and retains the original physical meaning of the data, so it has been widely concerned.

[0003] At present, the research on the band selection method of hyperspectral image usually has the following problems: first, most of the methods are based on a single evaluation criterion to measure the band, which leads to incomplete evaluation of the band, and weak generalization ability for different scenes or different data sets in subsequent application. Secondly, the existing research usually evaluates the band quality based on the digital statistical characteristics of single-band image, such as the method based on entropy and signal-to-noise ratio. However, for the target detection task, the above method ignores the characteristics of the target to be detected, so that the selected band lacks pertinence. In addition, the evaluation of band similarity is also a key step to reduce the redundancy of hyperspectral data. The traditional method usually focuses on the spectral similarity between bands and ignores the similarity of their spatial position distribution, and does not fully utilize the local neighborhood information of the band. Therefore, how to consider multiple evaluation indexes and introduce the measurement of target representation ability, how to comprehensively evaluate the spectral-spatial similarity between bands, so as to get a comprehensive band evaluation and improve the pertinence of the band selection method for subsequent target detection task, are the key problems to be solved in the research of band selection method. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a swarm intelligence optimization band selection method for hyperspectral image target detection. The present application overcomes the one-sidedness of the band evaluation of the single criterion method, and can realize effective dimensionality reduction of hyperspectral image. The use of joint spectral-spatial similarity helps to fully exploit the spectral similarity degree and spatial proximity relationship between bands. The design of CEM error metric and the introduction of optimal / suboptimal ratio criterion are both conducive to improving the target representation ability of the selected band subset, solving the problem that the traditional band selection method lacks pertinence for the target detection task of hyperspectral image, and having important application value.

[0005] The technical means adopted by the present application are as follows:

[0006] A swarm intelligence optimization band selection method for hyperspectral image target detection, comprising the following steps:

[0007] S1: reading hyperspectral image data X, target signal d, and the number of wave bands K required to be selected;

[0008] S2: a multi-objective optimization model is established based on information entropy, joint spectral-spatial similarity and CEM error to comprehensively measure the information amount, redundancy and target representation ability of the wave band; the wave band index code is used to represent the wave band, the initial population P1 containing K wave bands is initialized and the fitness of each individual is calculated, the non-dominated individuals in the current population are selected according to the Pareto dominance relationship and put into the non-inferior solution set;

[0009] S3: starting iteration, in the tth iteration, the global optimal solution gbest is determined from the non-inferior solution set rep according to the objective function value of the multi-objective optimization model, the individual optimal solution pbest is calculated according to the dominance relationship, and the position and speed of each individual in the population are updated under the guidance of the global and individual optimal solutions;

[0010] S4: the crossover and mutation operations are performed on each individual in the population, and a new population P t+1 is evolved;

[0011] S5: the fitness of each individual in the population P t+1 is calculated, the non-inferior solution set is updated according to the dominance relationship, if the number of non-inferior solution set reaches the maximum value, the redundant solutions are deleted by using the roulette wheel selection operator, if the maximum iteration number is reached, S6 is turned to continue execution, otherwise S3 is returned, and the iteration number t=t+1;

[0012] S6: the non-inferior solution set rep is regarded as the Pareto front of the optimization problem, for the n wave band subsets corresponding to the non-inferior solution set rep, the optimal / suboptimal ratio criterion is used for evaluation, and the best wave band subset is selected.

[0013] The application provides a swarm intelligence optimization band selection method for hyperspectral image target detection, which simultaneously considers three wave band evaluation indexes of information amount, redundancy and target representation ability, converts the hyperspectral wave band selection problem into a multi-objective optimization problem, abstracts the wave band into a population individual in the multi-objective optimization problem, finds the non-inferior solution set of the optimization problem under the guidance of the global optimal individual and the local optimal individual, designs genetic operations such as crossover and mutation to avoid falling into local optimum, selects the optimal solution on the Pareto front by using the optimal / suboptimal ratio criterion, and finally obtains the best wave band subset with high information amount, low redundancy and strong target representation ability.

[0014] Compared with the prior art, the application has the advantages and positive effects that: the CEM error is proposed as an index for evaluating the band target representation capability, the band selection method is improved to be more targeted for the target detection task, and the selected band subset is better adapted to the target detection task; the joint spectral space similarity is used to replace the traditional Euclidean distance, the spectral feature and the spatial feature are considered at the same time, and the band redundancy is more effectively reduced; the improved swarm intelligence optimization algorithm is proposed, the hyperspectral band selection problem is converted into a multi-objective optimization problem, a more comprehensive band evaluation is obtained, and the generalization ability of the selected band for different scenes is improved; and the genetic operations such as crossover and mutation are designed, and the optimization algorithm is effectively prevented from falling into local optimum.

[0015] The application solves the problems of single evaluation standard, poor adaptability of general criteria to the target detection task and easy falling into local optimum in the optimization process of the traditional hyperspectral image band selection method to some extent, and has important application value in the target detection of hyperspectral images. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor under the premise of the drawings.

[0017] Figure 1 The technical flow chart of the swarm intelligence optimization band selection method for the target detection of hyperspectral images in the application.

[0018] Fig. 2 (a) and (b) are pseudo-color images and ground truth of the HYDICE Urban hyperspectral data set used in the embodiments of the application.

[0019] Fig. 3 (a) and (b) are pseudo-color images and ground truth of the San Diego hyperspectral data set used in the embodiments of the application.

[0020] Figure 4 The quantitative analysis of the detection results of the HYDICE Urban hyperspectral data in the embodiments of the application.

[0021] Figure 5 The quantitative analysis of the detection results of the San Diego hyperspectral data in the embodiments of the application. DETAILED DESCRIPTION

[0022] In the following, the technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative effort should belong to the protection scope of the present application.

[0023] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device including a series of steps or units does not necessarily have to be limited to the clearly listed steps or units, but can include other steps or units not clearly listed or inherent to the process, method, product, or device.

[0024] As shown in Figure 1 The present application provides a swarm intelligence optimization band selection method for hyperspectral image target detection. Specifically, the method comprises the following steps:

[0025] Step S1: inputting hyperspectral image data X and target signal d, and setting the number K of bands to be selected;

[0026] Step S2: establishing a multi-objective optimization model and initializing a population P1.

[0027] (a) establishing a multi-objective optimization model: based on information entropy, joint spectral-spatial similarity, and CEM error, a multi-objective optimization model is established:

[0028]

[0029] wherein, represents any individual in the population, i.e., a band subset containing K bands, H(·), Similarity ss (·,·), and cemError(·) represent information entropy, joint spectral-spatial similarity, and CEM error measurement, respectively.

[0030] In related research on band selection, entropy is often used to measure the information richness of a band. Specifically, the band x information entropy H(x) is defined as follows:

[0031]

[0032] where Φ denotes the set of all gray values in band x, ω denotes the gray value of a pixel, p(ω) denotes the corresponding probability density, which can be calculated from the gray histogram, if the gray value of a pixel is considered as a random variable.

[0033] The spectral-spatial similarity between the i-th band x i and the j-th band x j is calculated by using the joint spectral-spatial similarity method, which is specifically defined as:

[0034] Similarity ss (x i ,x j )=S spectral (x i ,x j )×S spatial (x i ,x j ) (12)

[0035] Here, S spectral (x i ,x j ) and S spatial (x i ,x j ) denote the spectral similarity and spatial distribution similarity between band x i and band x j , respectively, which are specifically defined as:

[0036] S spectral (x i ,x j )=exp(-||x i -x j || 2 / 2σ 2 )

[0037]

[0038] where and are kernel functions, denotes the index value of band x i . According to the physical meaning of the hyperspectral image, the closer the wavelengths of band x i and band x j , the higher the degree of similarity.

[0039] The CEM error is used to measure the target representation ability of the hyperspectral band to the object to be detected, which is specifically defined as:

[0040]

[0041] for each subset of wavebands d represents the target signal, which is the prior information; R represents its covariance matrix. The concept of CEM error comes from the constrained target minimization waveband selection, which is considered as the variance produced by the signal of interest passing through the CEM linear filter. The smaller the variance, the more suitable the subset of wavebands can be considered for separating the target to be detected, and the stronger the ability to represent the target.

[0042] (b) initialize the population P1: initialize the initial population P1 containing K wavebands, and calculate the fitness of each individual in the population according to formula (1) where m represents the number of sub-target functions, which is set to 3 in this embodiment.

[0043] Then, the non-dominated individuals in the current population are selected according to the Pareto dominance relationship and put into the non-inferior solution set.

[0044] The dominance relationship of any two solutions and is defined as follows:

[0045]

[0046] where “>” represents the dominance relationship, i.e. the solution dominates the solution Therefore, all the non-dominated solutions can be stored in the non-inferior solution set rep.

[0047] Step S3: update the individual position and speed according to the global optimal solution and individual optimal solution: start iteration, determine the global optimal solution gbest from the non-inferior solution set rep using the roulette operator strategy, and calculate the individual optimal solution pbest according to the dominance relationship, and update the position and speed of each individual in the population as follows:

[0048]

[0049]

[0050] for each individual in the population v t represents the speed in the tthiteration, r1and r2are random numbers between 0 and 1, and δ, c1and c2represent the inertia index, the individual learning factor and the global learning factor, respectively.

[0051] Step S4: perform crossover and mutation operations. First, set the adaptive crossover rate pcand mutation rate pmrespectively as:

[0052] pc=(1-(t-1) / (T-1))^(1 / cu)​

[0053] pm = (1 - (t - 1) / (T - 1))A(1 / mu) (16)

[0054] where t is the current iteration number, T is the maximum iteration number, and cu and mu represent the crossover factor and mutation factor, respectively. According to equation (7), both the crossover rate and mutation rate decrease with the increase of iteration number, which makes the population active in the early stage of evolution to avoid falling into local optimum, and tends to be stable in the later stage to accelerate the optimization process.

[0055] In the crossover process, the crossover pool is set as the non-inferior solution set of each generation, from which two individuals are randomly selected, and the position and velocity of the offspring are updated as follows:

[0056]

[0057] childv = ((v1 + v2) * ||v1||) / (||v1|| + ||v2||) (17)

[0058] where and childv are the position and velocity of the offspring, respectively, and v1 and v2 represent the position and velocity of the two parents, respectively, and a is a random number between 0 and 1.

[0059] The mutation process can occur on any individual in the evolution process. If the generated random number is less than the mutation rate pm, one of the wave band indexes in the individual is randomly changed.

[0060] Step S5: Update the non-inferior solution set rep and determine whether the maximum iteration number is reached. The fitness of each individual in the new population is calculated according to equation (1) and the non-inferior solution set is updated. If the number of non-inferior solutions reaches the maximum value, the redundant solutions are deleted using the roulette wheel selection operator. If the maximum iteration number is reached, go to S6 for further execution, otherwise return to S3 and iteration number t = t + 1.

[0061] Step S6: Select the best wave band subset from the non-inferior solution set rep using the optimal / suboptimal ratio criterion. For all n candidate wave band subsets on the Pareto front, first use the CEM or ACE detector to generate the detection map, and perform a normalization operation on the gray image. For the detection result of the i-th candidate solution , select the pixel position with the highest gray value and its four nearest neighbor pixel positions, a total of 5 pixel positions, and calculate the average gray value as Then, calculate the average gray value of the top z pixel positions with the highest gray value other than the above 5 pixel positions, and denote it as z is generally set to 1% of the total number of pixel points. Thus, the optimal / suboptimal ratio criterion is defined as:​

[0062]

[0063] The scheme and effect of the present application are further described below through real hyperspectral data experiments.

[0064] In order to verify the effectiveness of the method described in the present embodiment, the present application is tested and described on two groups of public real hyperspectral data sets according to the above method steps, and the application effect of the one kind of swarm intelligence optimization band selection algorithm for hyperspectral image target detection is analyzed and evaluated by comparison with other five advanced band selection algorithms.

[0065] 1. Data set introduction

[0066] The first group of data sets used in the present embodiment is HYDICE Urban hyperspectral data, which is obtained by a Hyperspectral Digital Imagery Collection Experiment airborne hyperspectral sensor, has 162 spectral bands with a size of 80*100 pixels, the imaging spectral range is 400 to 2500 nanometers, and the pseudo-color image and ground truth are as shown in Fig. 2.

[0067] The second group of data sets used in the present embodiment is San Diego hyperspectral data, which is obtained by an Airborne Visible Infra-Red Imaging Spectrometer airborne hyperspectral sensor, contains 189 bands, the spectral range is 370 to 2510 nanometers, and a region with a size of 100*100 pixels is selected for experiment, and the pseudo-color image and ground truth are as shown in Fig. 3.

[0068] 2. Comparison method and parameter setting

[0069] Six advanced band selection methods, maximum variance principal component analysis (MVPCA), uniform band selection (UBS), fast density peak clustering (E-FDPC), adaptive subspace partitioning (ASPS) and optimal neighborhood reconstruction (ONR), are selected to verify the effectiveness of the proposed method.

[0070] The parameter settings of the experiment are shown in Table 1. The number of selected bands is determined by the virtual dimension method, wherein it is set to 11 on the HYDICE Urban data set and 10 on the San Diego data set.

[0071] Table 1 Experimental parameters

[0072]

[0073] 3. Experimental results

[0074] The band subset selected by the proposed method and the comparative methods were detected by the constrained energy minimization detector. The detection results were evaluated qualitatively and quantitatively by visual comparison of the detection maps and the area under curve (AUC) of the receiver operating characteristic (ROC) curves.

[0075] Figure 4 and 5 are the detection maps on the HYDICE Urban and San Diego hyperspectral data, respectively.

[0076] Among them, (a)-(f) are the detection results of the band subset selected by the MVPCA, UBS, E-FDPC, ASPS, ONR and the proposed method, respectively. It can be seen that the band subset selected by the proposed method can effectively detect the target pixels and suppress the background area, and outperforms other methods in the comparison.

[0077] Table 2 and Table 3 are the quantitative comparison of the detection results on the HYDICE Urban and San Diego hyperspectral data, respectively.

[0078] Table 2 Quantitative comparison of the detection results on the HYDICE Urban hyperspectral data

[0079]

[0080] Table 3 Quantitative comparison of the detection results on the San Diego hyperspectral data

[0081]

[0082] In the comparison with the five commonly used band selection methods, the proposed method performs superior performance on the three evaluation indexes and two public data sets, obtains the band subset with strong target representation ability, and verifies the feasibility and effectiveness of the proposed method.

[0083] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit it; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A swarm intelligence optimization band selection method for target detection in hyperspectral images, characterized in that, Includes the following steps: S1: Read the hyperspectral image data X, the target signal d, and the number of bands K to be selected; S2: A multi-objective optimization model is established based on information entropy, joint spectrum-space similarity and CEM error to comprehensively measure the information content, redundancy and target representation ability of the bands; the bands are represented by band index encoding, an initial population P1 containing K bands is initialized and the fitness of each individual is calculated, and non-dominant individuals in the current population are selected according to Pareto dominance and placed into the non-dominated solution set. S3: Start iteration. In the t-th iteration, based on the objective function value of the multi-objective optimization model, the global optimal solution gbest is determined from the non-dominated solution set rep using the roulette wheel operator strategy. The individual optimal solution pbest is calculated based on the dominance relationship. Under the guidance of the global and individual optimal solutions, the position and velocity of each individual in the population are updated. S4: Perform crossover and mutation operations on each individual in the population to evolve and generate a new population P. t+1 ; S5: Calculate population P t+1 The fitness of each individual is used to update the non-dominated solution set based on the dominance relationship. If the number of non-dominated solutions reaches the maximum value, the roulette wheel operator strategy is used to delete redundant solutions. If the maximum number of iterations is reached, the process is moved to S6 to continue execution; otherwise, it returns to S3. The number of iterations is t = t + 1. S6: The non-dominated solution set rep is regarded as the Pareto front of the optimization problem. For the n band subsets corresponding to the non-dominated solution set rep, the best / second best ratio criterion is used for evaluation to select the best band subset.

2. The swarm intelligence optimization band selection method for target detection in hyperspectral images according to claim 1, characterized in that, The multi-objective optimization model based on information entropy, joint spectrum-space similarity, and CEM error is as follows: in, Let H represent any individual in the population, K represent the number of band subsets, and H(·) represent Similarity. ss (·,·) and cemError(·) represent information entropy, joint spectral-space similarity and CEM error measure, respectively; The information entropy H(x) of band x is defined as follows: Where Φ represents the set of all gray values ​​on band x, ω represents the gray value of a pixel as a random variable, and p(ω) represents the probability density corresponding to the random variable. The joint spectral-space similarity method is used to measure the distribution difference of hyperspectral image bands, and the x-th band is calculated. i and the j-th band x j The spectral-space similarity is in the following form: Similarity ss (x i ,x j )=S spectral (x i ,x j )×S spatial (x i ,x j ) (3) Among them, S spectral (x i ,x j ) and S spatial (x i ,x j ) represent band x i and band x j The spectral similarity and spatial distribution similarity are specifically defined as follows: S spectral (x i ,x j )=exp(-||x i -x j || 2 / 2σ 2 ) in, and For kernel function, I xi Indicates band x i The index value; CEM error is used to measure the target characterization capability of the hyperspectral band for the analyte, specifically in the following form: For each band subset d represents the target signal, which is prior information; R represents its covariance matrix.

3. The swarm intelligence optimization band selection method for target detection in hyperspectral images according to claim 1, characterized in that, Selecting non-dominant individuals from the current population and placing them into the non-dominant solution set based on Pareto dominance relations includes defining dominance relations as follows: For any two solutions: and The dominance relationship is defined as follows: In this context, ">" indicates a dominance relationship.

4. The swarm intelligence optimization band selection method for target detection in hyperspectral images according to claim 1, characterized in that, Guided by global and individual optimal solutions, update the position and velocity of each individual in the population, including: The method for updating the velocity and position of an individual is defined as follows: For individuals in the population v t Let r1 and r2 represent the velocity in the t-th iteration, where r1 and r2 are random numbers between (0,1), and δ, c1, and c2 represent the inertia index, individual learning factor, and global learning factor, respectively.

5. The swarm intelligence optimization band selection method for target detection in hyperspectral images according to claim 1, characterized in that, Based on adaptive crossover and mutation rates, genetic operations involving crossover and mutation are performed on individuals in the population to co-evolve and generate a new population P. t+1 ,include: First, set the adaptive crossover rate pc and mutation rate pm as follows: Where t is the current iteration number, T is the maximum iteration number, and cu and mu represent the crossover factor and mutation factor, respectively; Secondly, during the crossover process, the crossover pool is set as the set of non-dominated solutions for each generation. Two individuals are randomly selected from this pool, and the position and velocity of the offspring are updated as follows: in, `childv` represents the position and velocity of the child generation, respectively. and v1 and v2 represent the positions and velocities of the two parents, respectively. α is a random number between 0 and 1. Mutation can occur in any individual during the evolutionary process. If the generated random number is less than the mutation rate pm, the individual is randomly modified. One of the band indexes.

6. The swarm intelligence optimization band selection method for target detection in hyperspectral images according to claim 1, characterized in that, For the n band subsets corresponding to the non-dominated solution set rep, the best / second best ratio criterion is used for evaluation to select the best band subset, including: For all n candidate band subsets on the Pareto front, a detection map is first generated using a CEM or ACE detector, and the grayscale image is then normalized. For the i-th candidate solution Based on the detection results, the pixel with the highest grayscale value and its four nearest neighbors (a total of 5 pixels) are selected, and the average grayscale value is calculated and denoted as . Then, calculate the average gray value of the z pixels with the highest gray values ​​(excluding the five pixels mentioned above), denoted as . z represents 1% of the total number of pixels; Therefore, the optimal / second-best ratio criterion is defined as:

Images