CN104751176B - A kind of target in hyperspectral remotely sensed image band selection method - Google Patents

Abstract

The invention discloses a hyperspectral remote sensing image band selection method based on mixed binary particle swarm differential evolution. Firstly, the original hyperspectral remote sensing image is preprocessed, dual population individuals and algorithm parameters are initialized, and then mixed binary particle swarm differential evolution is applied. (HBPSODE) method, let the two populations iterate in parallel and transmit the optimal solution information between the populations, and use the SVM classifier to calculate the classification accuracy as the fitness value, and update the evolution until it reaches the specified number of evolutions or reaches the maximum accuracy.

Description

A method for band selection of hyperspectral remote sensing images

technical field

The present invention relates to a hyperspectral remote sensing image band selection method, specifically a hyperspectral remote sensing image band selection method based on a packaged HBPSODE-SVM algorithm based on mixed binary particle swarm differential evolution and aiming at obtaining the optimal band combination , belonging to the technical field of hyperspectral remote sensing image processing.

Background technique

Remote Sensing (Remote Sensing) is a technology that uses the principle of electromagnetic waves to obtain distant signals and image them, and can feel and perceive distant things remotely. It is a new science. With the improvement of computer technology and optical technology, remote sensing technology has also been developed rapidly. In recent years, a variety of remote sensing satellites have been successfully launched, which has promoted the remote sensing data acquisition technology towards three high (high spatial resolution, high spectral resolution and high temporal resolution) and three multi (multi-platform, multi-sensor, multi-angle) development direction.

Hyperspectral remote sensing has the characteristics of high spectral resolution. By carrying hyperspectral sensors on different space platforms, it can measure the surface of the earth in continuous spectral bands in the visible, near-infrared, mid-infrared and thermal infrared bands of the electromagnetic spectrum. The area is imaged at the same time, the number of bands can reach dozens or even hundreds, and continuous spectral information of ground objects is obtained, thus realizing the simultaneous acquisition of space, radiation and spectral information of ground objects. Compared with conventional remote sensing, the main difference is that hyperspectral remote sensing is narrow-band imaging, and in addition to two-dimensional spatial information, one-dimensional spectral information is added, which expands the application field of remote sensing technology.

Hyperspectral remote sensing can detect finer spectral characteristics. Hyperspectral images have spectral information that cannot be obtained by conventional remote sensing. However, while the amount of spectral information in hyperspectral images increases, the dimensionality of data also increases, which makes the amount of image data surge. Its high dimensionality and inter-band correlation will not only complicate the calculation and greatly reduce the processing speed, but also may lead to a decrease in classification accuracy in the case of limited samples. When the imaging spectrometer obtains the hyperspectral image data, the band selection is particularly important.

Hyperspectral remote sensing image band selection can be regarded as an NP-hard combinatorial optimization problem. Using the intelligent suboptimal search algorithm according to the evaluation criterion function, several bands can be selected from all bands to form a band combination with better performance for classification accuracy. The common intelligent algorithm has been successfully applied to the band selection, but its defects are quickly exposed. The genetic algorithm cannot effectively converge within a limited time. The ant colony algorithm generally needs more Long search time, and prone to premature phenomenon. Although the particle swarm optimization algorithm has a fast convergence speed, its precision is not high and it is prone to premature. Both the PSO algorithm and the DE algorithm are new heuristic algorithms based on swarm intelligence, and both can be used independently in the band selection of hyperspectral remote sensing images. However, both algorithms have some defects. In the initial stage of population iteration, the individuals maintain a high diversity. With the increase of the number of iterations, the individuals in the PSO algorithm gradually approach the optimal particle in the population, and the DE algorithm is performing selection. The greedy strategy is adopted in the operation, that is, only when the mutant individual has a better fitness function value than the current individual can it participate in the next iterative evolution. Although these evolutionary mechanisms can speed up the convergence speed of the algorithm, they also make the differences between individuals in the population gradually narrow, and the diversity of the population also decreases. no change. Aiming at the defects of the two algorithms, the method of the present invention adopts a dual-population evolutionary strategy, allowing the two algorithms to iteratively evolve simultaneously to find the optimal band combination solution, and to help each other's populations escape from the local optimal solution through an information exchange mechanism.

Contents of the invention

Purpose of the invention: In order to overcome the deficiencies in the existing hyperspectral remote sensing image band selection technology and improve the classification accuracy, the present invention provides a hyperspectral band selection method, which is a package-based hybrid method aimed at obtaining the optimal band combination Band Selection Method for Binary Particle Swarm Differential Evolution (HBPSODE-SVM) Algorithm.

Technical solution: A hyperspectral remote sensing image band selection method. First, the original hyperspectral remote sensing image is preprocessed, the dual population individuals and algorithm parameters are initialized, and then the hybrid binary particle swarm differential evolution (HBPSODE) method is applied to make the dual populations parallel Iterate and pass the optimal solution information between the populations, and use the SVM classifier to calculate the classification accuracy as the fitness value, and update the evolution until the specified number of evolutions is reached or the maximum accuracy is reached.

Make corresponding modifications to the PSO algorithm and the DE algorithm, and propose a binary differential evolution algorithm with mixed coding, so that it can be extended to the discrete domain; firstly, define the auxiliary search space S'=[-a,a] ^d , where a is Positive integer, solution space S={0,1} ^d , d is the dimension of the problem; then the auxiliary search space D-dimensional real number vector X plus the solution space binary D-dimensional vector B (X, B) as the individual (or variation body) mixed encoding representation; the real vector X still performs mutation operations and crossover operations according to the differential evolution algorithm. Before performing the selection operation, it is necessary to map the real vector X into a binary vector B through full homomorphic evolution. Function definition:

Among them, h _ij (t+1) is each component value of the variant after the crossover operation, is a fuzzy function, b _ij (t+1) is each component value of the binary vector B, the adjustment factor μ can control the probability that b _ij (t+1) is set to 1, and μ=0.5.

Let (X _i (t),B _i (t)) and (X _i (t+1),B _i (t+1)) denote individual i of the tth generation and the t+1th generation of the population respectively, (H _i (t+1), E _i (t+1)) represents the variant of individual i in the t+1th generation, and f(x) represents the fitness function. The new selection operation is defined as follows:

The hyperspectral band selection method specifically includes the following steps:

Step 1: Preprocessing of original hyperspectral remote sensing images. Eliminate interference bands, pre-select the type of ground objects, set the dimension D of the search space, and the maximum number of iterative evolutions of the algorithm MaxDT.

Step 2: Initialize the population Ppso and related parameters evolved according to the HBPSO algorithm. Set the population number as Np, set the learning factor c ₁ , learning factor c ₂ , the maximum inertia weight coefficient w _max , the minimum inertia weight coefficient w _min and so on. In order to improve the performance of the particle swarm optimization (PSO) algorithm, the inertia weight w is updated according to the following formula, and i represents the ith iteration.

Step 3: Initialize the population Pde and related parameters evolved according to the HBDE algorithm. Set the number of populations as Nd, scaling factor F, hybridization parameter CR, etc. In order to improve the performance of the differential evolution (DE) algorithm, the scaling factor F is updated according to the following formula, F0 is a constant, and i represents the ith iteration.

Step 4: Set evolution iteration counter t=0.

Step 5: The Ppso population performs a position and velocity update according to the HBPSO algorithm, uses the SVM classifier to classify the updated band combination, and calculates the classification accuracy as the fitness value, and records the best fitness value and band combination of the tth generation .

Step 6: The Pde population performs mutation, crossover and selection operations on all individuals according to the HBDE algorithm. Use the SVM classifier to calculate the fitness value, and record the best fitness value and band combination of the tth generation.

Step 7: Compare the best fitness values selected in the tth generation of Ppso and Pde, and adjust the optimal solutions of the respective populations.

Step 8: Update the evolution algebra counter t=t+1. If the evolution algebra reaches the maximum number of evolutions or meets the precision requirements, the algorithm is terminated, otherwise, go back to step 5.

In order to better understand the technology and method involved in the present invention, the theory involved in the present invention is introduced here.

1. Particle Swarm Optimization (PSO) Algorithm

The implementation method of the PSO algorithm is to let the particle i be any individual in the population, the position _Xi (t) of the tth iteration=[x ₁ ,x ₂ ,…,x _D ], the velocity V _i (t)=[v ₁ ,v ₂ ,…,v _D ], where D represents the dimension of the problem, pBest _i represents the historical optimal solution position of particle i, and gBest(t) represents the global optimal solution position in the t-th iteration. Particle i updates its velocity and position according to the following formula:

V _i (t+1)=w·V _i (t)+c ₁ ·rand(0,1)·(pBest _i -X _i (t))

+c ₂ rand(0,1)(gBest(t)-X _i (t)) (3)

X _i (t+1)=X _i (t)+V _i (t+1) (4)

Among them, c ₁ and c ₂ are acceleration constants, which indicate the degree to which particles are affected by social cognition and individual cognition, rand(0,1) represents a random number that obeys the [0,1] distribution, and w is the inertia weight, which can be The iterative process dynamically adjusts the particle velocity.

2. Differential evolution (DE) algorithm

The implementation method of the differential evolution algorithm is to first randomly generate an initial population, and let _Xi (t) represent the i-th individual in the t-th generation population. The mutation operation is to generate a new variant for _Xi (t). First, randomly select three different individuals X _r1 (t), X _r2 (t), and X _r3 (t) from the current population, and then generate mutant individuals X _i (t)' according to formula (2.3):

X _i (t)'=X _r1 (t)+F·(X _r2 (t)-X _r3 (t)) (5)

Where F is the scaling factor, and its value range is [0.1,1].

The purpose of the crossover operation is to crossover the current individual _Xi (t) and the mutant individual _Xi (t)', thereby introducing the diversity of individuals in the population. The specific operation process is as follows:

First randomly generate an integer r ∈ [1, D], D represents the dimension of the problem, and then operate according to formula (6) for each dimension of the individual vector:

Among them, CR represents the cross rate, and its value range is [0,1]. The integer r can guarantee that the value of at least one component of the new individual after crossover is different from the current individual.

The selection operation determines whether the new individual after the crossover operation can enter the next round of iterative evolution. Through the evaluation criterion function, it can be judged whether the fitness value of the new individual is better than the current individual, and the selection operation is operated according to (2.5):

In the DE algorithm, the mutation mode can generally be expressed as: DE/x/y/z, where x represents the selection method of the base vector in the mutation operation, y represents the number of difference vectors, and z represents the crossover mode. The variation pattern introduced above is expressed as: DE/rand/1/bin, and bin means that the binomial crossover pattern is used. In addition, there are some other classic mutation modes:

DE/best/1/bin. This mode indicates that the mutant individual is a combination of two random individuals selected from the current population and the optimal individual, and its expression is as follows:

X _i (t)'＝X _best (t)+F·(X _r1 (t)-X _r2 (t)) (8)

DE/best/2/bin. This mode indicates that four random individuals need to be selected from the current population to generate two difference vectors to form mutant individuals, and the expression is as follows:

X _i (t)'＝X _best (t)+F·(X _r1 (t)+X _r2 (t)-X _r3 (t)-X _r4 (t)) (9)

DE/rang-to-best/1/bin. This mode means that the optimal vector in the current population is added to the difference vector to form a mutant individual, and its expression is as follows:

X _i (t)'＝X _i (t)+λ·(X _best (t)-X _r1 (t))+F·(X _r2 (t)-X _r3 (t)) (10)

1. Overview of support vector machine (SVM) classifier

Support vector machine (support vector machine, SVM) is a new machine learning method proposed by Vapnik et al. on the basis of statistical learning theory. It is based on the linear classifier and evolved by introducing the principle of structural risk minimization, optimization theory and kernel method. It is essentially a convex quadratic programming problem, which has great advantages in solving small sample, nonlinear and high-dimensional pattern recognition problems and has been successfully applied to many problems such as regression problems, classification recognition, and discriminant analysis.

The mechanism of SVM is to find an optimal classification hyperplane that meets the classification requirements, so that the hyperplane can maximize the blank area on both sides of the plane while ensuring the classification accuracy. To illustrate with the classification of two types of data, it is assumed that the training sample set is (xi _, y _i ), i=1,2,...,l, x _i ∈R ⁿ , y _i ∈{±1}. The linear discriminant function in a general d-dimensional space can be expressed as:

g(x)=ω ^T x+b (11)

Then the classification hyperplane equation is expressed as:

ω ^T x + b = 0 (12)

All sample points that are correctly classified and have classification intervals must satisfy:

y _i ·(ω ^T x _i +b)≥1,i=1,2,...,l,y _i ∈{±1} (13)

Among them, y _i is the category label of the i-th sample, and ω is the weight coefficient. As shown in Figure 2, those points that happen to fall on the classification hyperplane are called support vectors, and the classification interval between the two types of samples is

Expressed as:

At this point, the problem of finding the optimal hyperplane is transformed into solving the function under the constraints of (13):

Introducing the Lagrange factor α _i , and solving equation (15), we can get:

Taking partial derivatives for ω, b, we can get:

Substituting formula (17) into formula (16), we get:

Solve for the optimal solution make The training samples of are the support vectors, and the weight coefficient vector ω ^* of the optimal classification hyperplane is a linear combination of support vectors. b ^* can be obtained by solving the constraint condition y _i (ω ^T x _i +b)-1=0.

The above description is a linear classification hyperplane, but in many practical problems, the classification surface between categories is often a nonlinear surface. In solving linear inseparable problems, SVM uses kernel functions to map nonlinear classification in low-dimensional space to high-dimensional space, and constructs a linear classification hyperplane in high-dimensional space. At present, there are mainly four types of typical kernel functions:

_Linear (linear) kernel function: K(x _i , x _j )=(xi x _j );

Polynomial (polynomial) kernel function: K( _xi , x _j )=[( _xi x _j )+1] ^q ;

Radial basis (RBF) kernel function: K(x _i ,x _j )=exp{-| _xi -x _j | ² /σ ₂ };

S-shaped (sigmoid) kernel function: K(x _i , x _j )=tanh(υ(x _i ·x _j )+c).

Description of drawings

Fig. 1 is the method flowchart of the embodiment of the present invention;

Fig. 2 is SVM classification hyperplane diagram;

Figure 3 is an image synthesized by 50, 27, and 17 bands;

Figure 4 is the calibration map of AVIRIS original ground objects;

Figure 5 is a curve diagram of the optimal individual fitness value of the algorithm;

Figure 6 is the classification results of the algorithm, where (a) is the classification results of BPSO-SVM, (b) is the classification results of HBPSO-SVM, (c) is the classification results of HBDE-SVM, and (d) is the classification of HBPSODE-SVM Result graph.

Detailed ways

Below in conjunction with specific embodiment, further illustrate the present invention, should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention, after having read the present invention, those skilled in the art will understand various equivalent forms of the present invention All modifications fall within the scope defined by the appended claims of the present application.

The hyperspectral remote sensing image band selection method first preprocesses the original hyperspectral remote sensing image, initializes the dual population individuals and algorithm parameters, and then applies the hybrid binary particle swarm differential evolution (HBPSODE) method to allow the dual populations to iterate in parallel and pass through the two populations. The information of the optimal solution is transmitted among them, and the SVM classifier is used to calculate the classification accuracy as the fitness value, and the evolution is updated until the specified number of evolutions is reached or the maximum accuracy is reached.

Make corresponding modifications to the PSO algorithm and the DE algorithm, and propose a binary differential evolution algorithm with mixed coding, so that it can be extended to the discrete domain; firstly, define the auxiliary search space S'=[-a,a] ^d , where a is Positive integer, solution space S={0,1} ^d , d is the dimension of the problem; then the auxiliary search space D-dimensional real number vector X plus the solution space binary D-dimensional vector B (X, B) as the individual (or variation body) mixed encoding representation; the real vector X still performs mutation operations and crossover operations according to the differential evolution algorithm. Before performing the selection operation, it is necessary to map the real vector X into a binary vector B through full homomorphic evolution. Function definition:

Among them, h _ij (t+1) is each component value of the variant after the crossover operation, is a fuzzy function, b _ij (t+1) is each component value of the binary vector B, the adjustment factor μ can control the probability that b _ij (t+1) is set to 1, and μ=0.5.

Let (X _i (t),B _i (t)) and (X _i (t+1),B _i (t+1)) denote individual i of the tth generation and the t+1th generation of the population respectively, (H _i (t+1), E _i (t+1)) represents the variant of individual i in the t+1th generation, and f(x) represents the fitness function. The new selection operation is defined as follows:

As shown in Figure 1, it specifically includes the following steps:

Step 1: Preprocessing of original hyperspectral remote sensing images. Eliminate interference bands, pre-select the type of ground objects, set the dimension D of the search space, and the maximum number of iterative evolutions of the algorithm MaxDT.

Step 2: Initialize the population Ppso and related parameters evolved according to the HBPSO algorithm. Set the population number as Np, set learning factor c1, learning factor c2, maximum inertia weight coefficient w _max , minimum inertia weight coefficient w _min , etc. In order to improve the performance of the particle swarm optimization (PSO) algorithm, the inertia weight w is updated according to the following formula, and i represents the ith iteration.

Step 3: Initialize the population Pde and related parameters evolved according to the HBDE algorithm. Set the number of populations as Nd, scaling factor F, hybridization parameter CR, etc. In order to improve the performance of the differential evolution (DE) algorithm, the scaling factor F is updated according to the following formula, F0 is a constant, and i represents the ith iteration.

Step 4: Set evolution iteration counter t=0.

Step 5: The Ppso population performs a position and velocity update according to the HBPSO algorithm, uses the SVM classifier to classify the updated band combination, and calculates the classification accuracy as the fitness value, and records the best fitness value and band combination of the tth generation .

Step 6: The Pde population performs mutation, crossover and selection operations on all individuals according to the HBDE algorithm. Use the SVM classifier to calculate the fitness value, and record the best fitness value and band combination of the tth generation.

Step 7: Compare the best fitness values selected in the tth generation of Ppso and Pde, and adjust the optimal solutions of the respective populations.

Step 8: Update the evolution algebra counter t=t+1. If the evolution algebra reaches the maximum number of evolutions or meets the precision requirements, the algorithm is terminated, otherwise, go back to step 5.

Simulation experiment result analysis

1. Experimental image

The performance of the algorithm is analyzed and evaluated through simulation experiments. In order to verify the effectiveness of the HBPSODE-SVM algorithm, this paper conducts an experiment on a standard hyperspectral remote sensing image. The remote sensing image used is a part of the hyperspectral remote sensing image acquired by the AVIRIS sensor in June 1992 of an agricultural and forestry mixed experimental area in northwestern Indiana, USA. The wavelength range is 0.4-2.5μm, the image size is 145×145pixel, and the spatial resolution is 25m. The bands seriously polluted by water vapor and noise (bands 1-4, 78, 80-86, 103-110, 149-165, 217-224) were removed from the original bands, and the remaining 179 bands were reserved for testing. Figure 3 is the synthetic R, G, B false color images of the 50th, 27th, and 17th bands selected for the test. Figure 4 is the original map of AVIRIS.

There are 17 types of ground objects in the image, and 7 types of ground objects are selected to participate in the classification experiment. The training samples and test samples are uniformly selected at a ratio of 1:1. Table-1- shows the number, name, number of training and testing samples of the 7 types of ground objects. The experimental program is implemented by Matlab (R2009b), and the SVM classifier is implemented by libsvm toolbox.

Table-1-Training samples and testing samples

2. Experimental method and related parameter settings

In order to verify the superiority of the HBPSODE-SVM method, design A, B, C, D, 4 groups of comparative simulation experiments. Considering the fairness of algorithm-related parameter settings, group A uses the binary particle swarm optimization (BPSO) algorithm, group B uses the hybrid binary particle swarm optimization (HBPSO) algorithm, group C uses the hybrid binary differential evolution (HBDE) algorithm, and group D uses the The band selection method of Hybrid Binary Particle Swarm Differential Evolution (HBPSODE) algorithm is compared. The four groups of experiments all use support vector machine (SVM) as the classifier and RBF kernel function. The relevant parameters of the algorithm and classifier are set as shown in Table -2-:

Table-2-Related parameter settings

3. Comparison of experimental results

After confirming the training samples and test samples, we analyze the (BPSO) algorithm from the following two aspects, the hybrid binary particle swarm optimization (HBPSO) algorithm, the hybrid binary differential evolution (HBDE) algorithm and the hybrid binary particle swarm differential evolution (HBPSODE ) algorithm to compare the experimental results.

① Error matrix

The error matrix obtained from the best band combinations obtained from the 4 groups of experiments participating in the ground object classification experiment is shown in Table 3-6.

Table-3-Group A BPSO-SVM error matrix

Table-4-Group B HBPSO-SVM error matrix

Table-5-Group C HBDE-SVM error matrix

Table-6-D group HBPSODE-SVM error matrix

It can be seen from Tables 3 to 6 that HBPSODE-SVM is generally superior to the other three algorithms in terms of producer accuracy and user accuracy, which means that the omission error and multi-point error generated by the HBPSODE-SVM algorithm are relatively small. Comparing the error matrices of HBPSO-SVM and BPSO-SVM algorithms, although the overall classification accuracy of the two is only 0.1%, but comparing the producer accuracy and user accuracy of each type of ground object, it can be seen that HBPSO-SVM is better than BPSO-SVM. SVMs. Comparing the error matrix of HBPSO-SVM and HBDE-SVM algorithm, DE algorithm is better than PSO algorithm in overall classification accuracy, producer accuracy and user accuracy. According to the differential evolution (DE) algorithm mentioned in the literature 39, the overall performance is better than the particle swarm optimization (PSO) algorithm.

②Kappa analysis and overall accuracy

Kappa analysis can combine the overall precision, producer precision and user precision of the error matrix to quantitatively evaluate the consistency between classification results and ground reference information. Table 7 shows the overall accuracy and Kappa value of the four experimental algorithms.

Table-7- Overall classification accuracy and Kappa value

Through the Kappa value in Table-7-, it can be quantitatively evaluated that the classification performance of the HBPSODE-SVM algorithm is better than that obtained by using the HBDE-SVM or HBPSO-SVM algorithm alone. HBDE-SVM performance is better than HBPSO-SVM, and HBPSO-SVM is better than general BPSO-SVM algorithm.

Figure-5- is the change curve of the optimal individual fitness value (classification accuracy) of the 4 groups of experimental algorithms in the whole iterative process.

As can be seen from Figure-5-, the length of the "platform" in the curve of Figure d is the shortest, indicating that the HBPSODE-SVM algorithm maintains a good population diversity during the iteration process, and can escape within a short number of iterations when evolutionary stagnation occurs local optimal solution. However, the other three algorithms all have the situation that the evolutionary stagnation cannot escape from the local optimal solution in a short period of time. Although the final classification accuracy of HBPSO-SVM is higher than that of BPSO-SVM, the HBPSO-SVM algorithm is more likely to fall into a local optimal solution and cannot escape in a short time during the entire evolution iteration process. have high diversity, but quickly fall into precocity until reaching the maximum number of evolution iterations. Compared with the above two algorithms, HBDE-SVM maintains better population diversity. Although it is also trapped in a local optimal solution, it can escape before the evolution iteration terminates. The classification results of the four groups of experimental algorithms are shown in Figure 6.

From Figure-6-(a)~(d), it can be seen intuitively that the classification accuracy of the optimal band combination of the HBPSODE-SVM algorithm is the best, and the classification accuracy of HBPSO-SVM is better than that of BPSO-SVM, and the classification accuracy of HBDE-SVM is better than that of BPSO-SVM. It is also better than HBPSO-SVM.

Claims (2)

1. A hyperspectral remote sensing image band selection method, characterized in that: firstly, the original hyperspectral remote sensing image is preprocessed, the double-population individuals and algorithm parameters are initialized, and then the hybrid binary particle swarm differential evolution (HBPSODE) method is applied to allow The dual populations iterate in parallel and transmit the optimal solution information between the populations, and use the SVM classifier to calculate the classification accuracy as the fitness value, and update the evolution until it reaches the specified number of evolutions or reaches the maximum accuracy; In the initialization process of population individuals and algorithm parameters, a dual-population parallel search strategy is adopted to initialize two populations, and the dual-population iterative evolution evolves in parallel; after each iteration, the populations share their respective searched optimal solutions to realize information exchange. Let the population not only refer to the optimal solution of its own population but also consider the optimal solution of the other population in the next iterative evolution, and guide the population to break away from the local optimal solution; Make corresponding modifications to the PSO algorithm and the DE algorithm, and propose a binary differential evolution algorithm with mixed coding, so that it can be extended to the discrete domain; firstly, define the auxiliary search space S'=[-a,a] ^d , where a is Positive integer, solution space S={0,1} ^d , d is the dimension of the problem; then the auxiliary search space D-dimensional real number vector X plus the solution space binary D-dimensional vector B (X, B) as an individual or variant The mixed encoding representation of the real number vector X still performs mutation operations and crossover operations according to the differential evolution algorithm. Before performing the selection operation, the real number vector X needs to be mapped into a binary vector B through full homomorphic evolution. The full homomorphic evolution mapping function is defined : Among them, h _ij (t+1) is each component value of the variant after the crossover operation, is a fuzzy function, b _ij (t+1) is each component value of the binary vector B, the adjustment factor μ can control the probability that b _ij (t+1) is set to 1, and μ=0.5. Let (X _i (t),B _i (t)) and (X _i (t+1),B _i (t+1)) denote individual i of the tth generation and the t+1th generation of the population respectively, (H _i (t+1), E _i (t+1)) represents the variant of individual i in the t+1th generation, and f(x) represents the fitness function. The new selection operation is defined as follows: The preprocessing of the original hyperspectral remote sensing images, and the initialization of dual population individuals and algorithm parameters include: Preprocessing of original hyperspectral remote sensing images: removing interference bands, pre-selecting object types, and setting the dimension D of the search space and the maximum iterative evolution of the algorithm MaxDT; Initialize the population Ppso and related parameters evolved according to the HBPSO algorithm: set the number of populations to Np, set the learning factor c ₁ , learning factor c ₂ , the maximum inertia weight coefficient w _max , the minimum inertia weight coefficient w _min , etc.; in order to improve the particle swarm optimization algorithm The performance of , where the inertia weight w is updated according to the following formula, i represents the ith iteration; <mrow><mi>w</mi><mo>=</mo><msub><mi>w</mi><mrow><mi>m</mi><mi>a</mi><mi>x</mi></mrow></msub><mo>-</mo><mfrac><mrow><msub><mi>w</mi><mrow><mi>m</mi><mi>a</mi><mi>x</mi></mrow></msub><mo>-</mo><msub><mi>w</mi><mi>min</mi></msub></mrow><mrow><mi>M</mi><mi>a</mi><mi>x</mi><mi>D</mi><mi>T</mi></mrow></mfrac><mo>&amp;CenterDot;</mo><mi>i</mi><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>1</mn><mo>)</mo></mrow></mrow> Initialize the population Pde and related parameters evolved according to the HBDE algorithm: set the number of populations to Nd, the scaling factor F, the hybridization parameter CR, etc.; in order to improve the performance of the differential evolution (DE) algorithm, the scaling factor F is updated according to the following formula, and F0 is a Constant, i represents the ith iteration; <mrow><mi>F</mi><mo>=</mo><mi>F</mi><mn>0</mn><mo>&amp;CenterDot;</mo><msup><mn>2</mn><mrow><mi>exp</mi><mrow><mo>(</mo><mfrac><mrow><mn>1</mn><mo>-</mo><mi>M</mi><mi>a</mi><mi>x</mi><mi>D</mi><mi>T</mi></mrow><mrow><mi>M</mi><mi>a</mi><mi>x</mi><mi>D</mi><mi>T</mi><mo>+</mo><mn>1</mn><mo>-</mo><mi>i</mi></mrow></mfrac><mo>)</mo></mrow></mrow></msup><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>2</mn><mo>)</mo></mrow><mo>.</mo></mrow> 2. The hyperspectral remote sensing image band selection method as claimed in claim 1, characterized in that, using the mixed binary particle swarm differential evolution method, allowing two populations to iterate in parallel and passing optimal solution information between the populations, and using SVM classification The classifier calculates the classification accuracy as the fitness value, and updates the evolution until it reaches the specified number of evolutions or reaches the maximum accuracy, specifically: Set evolution iteration counter t=0. The Ppso population performs a position and velocity update according to the HBPSO algorithm, uses the SVM classifier to classify the updated band combination, and calculates the classification accuracy as the fitness value, and records the best fitness value and band combination of the tth generation; The Pde population performs mutation, crossover, and selection operations on all individuals according to the HBDE algorithm; uses the SVM classifier to calculate the fitness value, and records the best fitness value and band combination of the tth generation; Compare the best fitness values selected by the tth generation of Ppso and Pde, and adjust the optimal solutions of the respective populations; Update the evolutionary algebra counter t=t+1; if the evolutionary algebra reaches the maximum number of evolutions or meets the precision requirement.