CN104200217A - Hyperspectrum classification method based on composite kernel function - Google Patents

Abstract

The invention provides a hyperspectral classification method based on a composite kernel function. Input a hyperspectral image, the number of categories is N; take the support vector machine as the base classifier, and at the same time randomly select s samples from each category of the hyperspectral image to form a training set, and the remaining samples form a test set, and determine each The variation range of the parameters, and then combined with K times of cross-validation to determine the optimal performance parameters of the support vector machine, including penalty factors and kernel parameters; use the composite kernel construction strategy to construct a composite kernel function, and train the support vector machine; use the training process to get The parameters of the decision function of the support vector machine, loop N times, and then get the decision function value of the test set belonging to each category, and form a matrix Determine the multi-classifier strategy, i.e. find the matrix The maximum value for each column. The invention has the characteristics of better describing the distribution characteristics of the data set, and relatively high classification accuracy. Compared with traditional multi-core learning methods, the time consumed by its parameter optimization is relatively short.

Description

A Hyperspectral Classification Method Based on a Composite Kernel Function

technical field

The present invention relates to a hyperspectral image classification method, in particular to a hyperspectral image classification method based on a new composite kernel function (Hyperspectral Image Classification Based on A New Composite Kernel). the

Background technique

In recent years, the development of satellite sensors has improved the spatial resolution and spectral resolution, and also shortened the access time, which in turn created the conditions for the development of hyperspectral classification methods. Neural network classifiers, K-nearest neighbor classifiers, Bayesian classifiers, decision tree classifiers, and kernel-based classifiers have been widely used in the hyperspectral field, among which kernel function-based classification methods have received more and more attention. . Support Vector Machine (SVM) is the most typical classification method based on kernel function. When dealing with limited high-dimensional training samples, it can still achieve good classification performance, making it occupy a certain position in hyperspectral classification. . In view of the multiple sources of actual data and the diversity of expression methods, the traditional single-core is not enough to meet the actual needs. In the past ten years, the technology of Multiple Kernel Learning (MKL) has received attention and development, and it has been mainly improved in four aspects: 1) The improvement of multi-kernel learning strategy; the simple multi-kernel learning (Simple-MKL) proposed by Rakotomamonjy et al. ), which uses gradient descent to solve multi-core learning problems. 2) Improvement of parameter optimization; Li et al. proposed a parameter optimization method that does not use any convex constraint items. 3) Improvement of the combination of multiple kernel functions; for example, Xia et al. used the combination of boosting. 4) Improvement of fast strategies; for example, Gu et al. proposed representative multi-kernel learning for hyperspectral classification. The most common multi-kernel learning method is Mixture Kernels (MKs), which is the linear weighted sum of multiple kernel functions. The effectiveness of this kernel function has been verified and is widely used in the field of hyperspectral classification. However, during the training process, the time consumed by parameter optimization is also huge. the

Contents of the invention

The purpose of the present invention is to provide a hyperspectral classification method based on a compound kernel function that can better describe the distribution characteristics of the data set, has relatively high classification accuracy, and consumes relatively short time for parameter optimization. the

The purpose of the present invention is achieved like this:

Step 1: Input a hyperspectral image, the number of categories is N;

Step 2: Using the support vector machine as the base classifier, randomly select s samples from each category of the hyperspectral image to form a training set, and the remaining samples form a test set, determine the variation range of each parameter, and then combine K times Cross-validation determines the optimal performance parameters of the support vector machine, including penalty factors and kernel parameters;

Step 3: Use the composite kernel construction strategy to construct a composite kernel function and train the support vector machine;

Step 4: Use the parameters of the decision function of the support vector machine obtained in the training process, loop N times, and then obtain the value of the decision function of each category in the test set, and form a matrix Where n _test represents the number of test samples;

Step 5: Determine the multi-classifier strategy, i.e. find the matrix the maximum value of each column whose row number corresponds to the predicted label for each test sample, i=1,...,n _test .

The present invention is characterized in that:

1. In the selection of the base classifier in the parameter setting process, other kernel-based classifiers can be used instead of the support vector machine. the

2. The determination of the optimal performance parameters in the parameter setting process is a combination of grid search and K times cross-validation. the

3. The construction of the composite kernel function in the training process is obtained through multiple nonlinear mappings, and the calculation is as follows:

K ₁ (x, z) = φ ₁ (x) · φ ₁ (z)

K ₂ (x,z)＝φ ₂ [φ ₁ (x)]·φ ₂ [φ ₁ (z)]

...

K _M (x,z)＝φ _M [φ _M-1 (x)]·φ _M [φ _M-1 (z)]

Among them, K _M (x, z) represents the kernel function of samples x and z, and φ _M represents the Mth nonlinear mapping function. φ _M can be Gaussian mapping, polynomial mapping or other nonlinear mapping, and when M=2, the classification performance of the classifier based on the compound kernel function can reach a convergent state.

4. When M=2, the parameter optimization time of the composite kernel function is much shorter than that of the traditional hybrid kernel function. the

Based on the above characteristics, the hyperspectral classification method based on a new composite kernel function proposed by the present invention has the characteristics of better describing the distribution characteristics of the data set and relatively high classification accuracy. At the same time, the time consumed by its parameter optimization is relatively short compared with traditional multi-core learning methods. the

Description of drawings

FIG. 1 is a flowchart of a hyperspectral classification method based on a compound kernel function of the present invention. the

Figures 2a-2b are hyperspectral images of Indian Pines, where: Figure 2a is a grayscale image, and Figure 2b is a reference object category. the

Figure 3a-Figure 3b is a visual map of the classification boundary of the double crescent-shaped simulated data set, wherein: Figure 3a is a visual map of the classification boundary of the Gaussian kernel function Gauss, and Figure 3b is a visual map of the classification boundary of the continuous Gaussian mapping composite kernel function G(G). Hollow symbols in the figure represent test samples, and solid symbols represent training samples. the

Figure 4 is a comparison table 1 of the classification performance of different kernel functions for the hyperspectral dataset Indian Pines. In the table, Linear represents a single linear kernel function, Polynomial represents a single polynomial kernel function, Gauss represents a single Gaussian kernel function, MKs represents a mixed kernel function, G(G) represents a continuous Gaussian mapping composite kernel function, and G(P) represents a polynomial-Gaussian mapping Composite kernel function, P(G) represents Gaussian-polynomial mapping composite kernel function, P(P) represents continuous polynomial mapping composite kernel function. the

Figure 5 is a comparison table 2 of the optimization time of different kernel function parameters. the

Figure 6 shows the impact of different numbers of training samples on the classification performance of different kernel functions in Table 3. the

Detailed ways

The present invention will be described in more detail below with reference to the accompanying drawings. the

The present invention is a hyperspectral classification method based on a new compound kernel function, including five steps of input process, parameter setting, training process, classification process and output process. The input process is to input a hyperspectral image; the parameter setting is the process of initialization and parameter optimization; the training process is the process of training the base classifier model based on the support vector machine; the classification process is the model parameters obtained by using the above process , so as to give the decision function value process that the test set belongs to each category; the output process is to determine the multi-classifier strategy and give the test sample prediction label process. The detailed process is given below:

The specific analysis steps are as follows:

Step S1: Input process. Input a hyperspectral image, N represents the number of categories. the

Step S2: parameter setting. Randomly select s samples from each category in the image to form the training set, and the remaining samples form the test set. Determine the variation range of each parameter, and then determine the optimal performance parameters of the support vector machine combined with K times of cross-validation. This step further includes the following steps:

Step S2.1: Determination of the sample set; randomly select s samples from each of the N categories of hyperspectral images to form a labeled training sample set D={(x ₁ ,y ₁ ),...,(x _n , y _n )}, where _xi represents the spectral feature of the i-th labeled sample, y _i represents the label of sample _xi , and n=16s represents the number of training samples. The remaining samples form the test set n _test represents the number of test samples.

Step S2.2: Determination of parameter ranges; polynomial kernel functions and Gaussian kernel functions are used as basic kernel functions to synthesize composite kernel functions. The polynomial kernel is K _P (x, z) = φ _P (x) φ _P (z) = [(x z) + 1] ^d , its parameter is d, φ _P (x) represents the polynomial mapping of sample x function. The Gaussian kernel is K _G (x, z) = φ _G (x) φ _G (z) = exp(-||xz|| ² /σ ² ), its Gaussian radius is σ, and φ _G (x) represents the sample The Gaussian map kernel function of x. The penalty factor C of the support vector machine, the Gaussian radius σ and the variation range of d are: {2 ⁰ ,2 ¹ ,...,2 ⁸ }, {2 ^-5 ,2 ^-4 ,...,2 ¹ } and {2 ^-2 ,2 ^-1 ,...,2 ⁴ }.

Step S2.3: K cross-validation; in the case of each parameter combination, the training sample set D is divided into K subsets, one subset is reserved for testing, and the remaining K-1 subsets are used for training SVM classification device model. The cross-validation is repeated K times, each subset is validated once, and the average of K classification accuracy is calculated. When the average precision reaches the maximum value, it indicates that the parameter combination is optimal. the

Step S3: training process. Using the composite kernel construction strategy, construct the composite kernel function, and train the support vector machine. This step further includes the following steps:

Step S3.1: Construction of the composite kernel function; the composite kernel function is obtained through multiple nonlinear mappings, and the calculation is as follows:

K ₁ (x, z) = φ ₁ (x) · φ ₁ (z)

K ₂ (x,z)＝φ ₂ [φ ₁ (x)]·φ ₂ [φ ₁ (z)]

...

K _M (x,z)＝φ _M [φ _M-1 (x)]·φ _M [φ _M-1 (z)]

Among them, K _M (x, z) represents the kernel function of samples x and z, and φ _M represents the Mth nonlinear mapping function. φ _M can be Gaussian mapping, polynomial mapping or other nonlinear mapping, and when M=2, the classification performance of the classifier based on the compound kernel function can reach a convergent state. When M=2, K _G(G) , K _G(P) , K _P(G) and K _P(P) are continuous Gaussian mapping composite kernel function, polynomial-Gaussian mapping composite kernel function, Gaussian-polynomial mapping respectively Composite kernel functions and continuous polynomial mapping composite kernel functions. The expression is as follows:

K _G(G) (x,z)＝φ _G [φ _G (x)]·φ _G [φ _G (z)]

＝exp[-||φ _G (x)-φ _G (z)|| ² /σ ₂ ² ]

(1)

＝exp[-[φ _G (x) φ _G (x)+φ _G (z) φ _G (z)-2φ _G (x) φ _G (z)]/σ ₂ ² ]

=exp[-[K _G (x,x)+K _G (z,z)-2K _G (x,z)]/σ ₂ ² ]

K _G(P) (x,z)＝φ _G [φ _P (x)]·φ _G [φ _P (z)]

＝exp[-||φ _P (x)-φ _P (z)|| ² /σ ₂ ² ]

(2)

＝exp[-[φ _P (x) φ _P (x)+φ _P (z) φ _P (z)-2φ _P (x) φ _P (z)]/σ ₂ ² ]

=exp[-[K _P (x,x)+K _P (z,z)-2K _P (x,z)]/σ ₂ ² ]

$\begin{array}{c}{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; G</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; \&CenterDot;</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; G</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; G</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CenterDot;</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; G</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ]</span>}^{{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}\\ <span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; G</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ]</span>}^{{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}\end{array}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

$\begin{array}{c}{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; \&CenterDot;</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CenterDot;</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ]</span>}^{{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}\\ <span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ]</span>}^{{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}\end{array}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

Where x and z represent two pixels, _σ2 represents the Gaussian radius of the second-order Gaussian mapping, and _d2 represents the polynomial parameter of the second-order polynomial mapping. Note that the Gaussian radius of the first Gaussian mapping in equations (1)(3) is σ ₁ , similarly, the polynomial parameter of the first polynomial mapping in equations (2)(4) is d ₁ .

Step S3.2: Training of the support vector machine; train the support vector machine to obtain the weight vector and thresholds α ^* and b ^* of the decision function of the support vector machine.

Step S4: classification process. Using the parameters of the decision function of the support vector machine obtained during the training process, loop N times, and then obtain the value of the decision function of the test set belonging to each category, and form a matrix Where n _test represents the number of test samples.

Step S5: Output process. Determine the multi-classifier strategy, i.e. find the matrix the maximum value of each column whose row number corresponds to the predicted label for each test sample, i=1,...,n _test .

Step S5.1: Determination of the multi-classifier strategy; the multi-classification problem is transformed into multiple binary classification problems by using the "one-to-remainder" multi-classifier strategy. the

Step S5.2: Calculation of classification accuracy; respectively calculate the overall classification accuracy (Overall Accuracy, OA) of the support vector machine under the situation of different composite kernel functions, the Kappa coefficient (Kappa Statistic, Kappa), the average classification accuracy (Average Accuracy, AA ) and the classification accuracy for each category. the

In order to illustrate the effectiveness of the present invention, the following experimental demonstration is specially carried out. The experimental data comes from the two-moon (Two Moons) simulation dataset and the real hyperspectral dataset (Indian Pines). the

1) Two Moons: This simulated data set contains two categories with 96 and 104 pixels respectively, and each pixel is represented by two features. the

2) Indian Pines: This hyperspectral remote sensing image comes from Indian agriculture and forestry in northwestern Indiana, USA acquired in 1992. It contains 144×144 pixels, 16 categories, and 220 bands. Due to noise and other factors, 20 bands were removed. . The image contains supervised data of 16 types of vegetation except the background. the

First set up some parameters of the present invention: first randomly select s samples from each class to form D={(x ₁ ,y ₁ ),...,(x _n ,y _n )}, and the remaining samples form the test set Choose the method of combining ten times of cross-validation (K=10) and grid search to estimate the optimal parameters. The range of penalty factor C, Gaussian radius σ and d of the support vector machine are: {2 ⁰ ,2 ¹ , ...,2 ⁸ }, {2 ^-5 ,2 ^-4 ,...,2 ¹ } and {2 ^-2 ,2 ^-1 ,...,2 ⁴ }. For the proposed compound kernel function, the nonlinear mapping times M=2, and the classification performance converges at this time.

The first group of experiments focused on using SVM as the base classifier, respectively using Gauss Kernel (Gauss Kernel) and composite kernel function (continuous Gaussian nonlinear mapping G(G) Composite Kernel), and the distribution of classification boundaries. As shown in Figure 2, a) Gaussian kernel function Gauss classification boundary b) compound kernel function G(G) classification boundary. Through this group comparison, it can be seen that when the composite kernel function is used, the classification boundary of SVM can better describe the distribution of the data set, and the classification accuracy is relatively high. the

The second set of experiments focuses on the impact of different kernel functions on classification performance. The parameters of this group of experiments were set to s=10. The experimental results are shown in Table 1, which can be summarized as follows: 1. For a single kernel function, the classification accuracy obtained by the Gaussian kernel function is higher than that obtained by the linear kernel (Linear Kernel) and the polynomial kernel (Polynomial Kernel); 2. When using Mixture Kernels (Mixture Kernels, MKs), that is, linear weighted summation mixed kernels, can effectively improve the classification accuracy of a single kernel function; 3. Compared with other kernel forms, when using a composite kernel function, the classification accuracy of SVM is higher, and G (G) Composite core has the best performance. the

The third group of experiments focuses on the influence of different kernel functions on the parameter optimization time. All experiments in this group adopt ten times of cross-validation, that is, K=10. The experimental results are shown in Table 2. It can be seen that when the number of nonlinear mappings is M=2, the optimization time of the composite kernel function parameters is longer than that of the single kernel, and at the same time, the optimization time is much shorter than the time consumed by the hybrid kernel function. the

The fourth group of experiments focuses on the influence of different numbers of training samples on the performance of different kernel functions. This group of experiments made s change in the set {5, 10, 50, 100}. The experimental results are shown in Table 3, which can be summarized as follows: 1. With the increase of s, the classification ability of different kernel functions is obviously improved, but different kernel functions 2. In the case of small samples, the advantages of the kernel function can be better reflected; 3. Each kernel function is not constant and effective, and its effectiveness is related to the actual distribution of the data set. If a kernel function can achieve good performance under certain circumstances, it means that the kernel function is effective. Based on this, the composite kernel function in the present invention is an effective kernel function. the

Claims (3)

1. Based on a kind of hyperspectral classification method of composite kernel function, it is characterized in that: Step 1: Input a hyperspectral image, the number of categories is N; Step 2: Using the support vector machine as the base classifier, randomly select s samples from each category of the hyperspectral image to form a training set, and the remaining samples form a test set, determine the variation range of each parameter, and then combine K times Cross-validation determines the optimal performance parameters of the support vector machine, including penalty factors and kernel parameters; Step 3: Use the composite kernel construction strategy to construct a composite kernel function and train the support vector machine; Step 4: Use the parameters of the decision function of the support vector machine obtained in the training process, loop N times, and then obtain the value of the decision function of each category in the test set, and form a matrix Where n _test represents the number of test samples; Step 5: Determine the multi-classifier strategy, i.e. find the matrix the maximum value of each column whose row number corresponds to the predicted label for each test sample, ${\widehat{\mathrm{the\; y}}}_i\&Element;\mathrm{the\; y},i=1,...,{\mathrm{no}}_{\mathrm{test}}.$ 2. the hyperspectral classification method based on a kind of composite kernel function according to claim 1, is characterized in that: described step 2 specifically comprises: Step 2.1: Randomly select s samples from each of the N categories of hyperspectral images to form a labeled training sample set D={(x ₁ ,y ₁ ),...,(x _n ,y _n )}, Among them, x _i represents the spectral feature of the i-th labeled sample, y _i represents the label of sample x _i , n=16s represents the number of training samples, and the remaining samples form the test set n _test represents the number of test samples; Step 2.2: Use the polynomial kernel function and the Gaussian kernel function as the basic kernel function to synthesize the composite kernel function. The polynomial kernel is K _P (x, z) = φ _P (x) φ _P (z) = [(x z )+1] ^d , whose parameter is d, φ _P (x) represents the polynomial mapping function of the sample x; the Gaussian kernel is K _G (x, z) = φ _G (x)·φ _G (z) = exp(- ||xz|| ² /σ ² ), its Gaussian radius is σ, φ _G (x) represents the Gaussian mapping kernel function of the sample x; the penalty factor C of the support vector machine, the variation range of the Gaussian radius σ and d are: {2 ⁰ ,2 ¹ ,...,2 ⁸ }, {2 ^-5 ,2 ^-4 ,...,2 ¹ } and {2 ^-2 ,2 ^-1 ,...,2 ⁴ }; Step 2.3: In the case of each parameter combination, the training sample set D is divided into K subsets, and one subset is reserved for testing in turn, and the remaining K-1 subsets are used for training the SVM classifier model, and the cross-validation is repeated for K times, each subset is verified once, and the average value of classification accuracy is calculated for K times. When the average accuracy reaches the maximum value, the parameter combination is optimal. 3. the hyperspectral classification method based on a kind of composite kernel function according to claim 1 or 2, is characterized in that: described step 3 specifically comprises: Step 3.1: Obtain the composite kernel function through multiple nonlinear mappings, and the calculation is as follows: K ₁ (x, z) = φ ₁ (x) · φ ₁ (z) K ₂ (x,z)＝φ ₂ [φ ₁ (x)]·φ ₂ [φ ₁ (z)] ... K _M (x,z)＝φ _M [φ _M-1 (x)]·φ _M [φ _M-1 (z)] Where K _M (x, z) represents the kernel function of samples x and z, φ _M represents the Mth nonlinear mapping function; φ _M is Gaussian mapping, polynomial mapping or other nonlinear mapping, and when M=2, The classification performance of the classifier based on the compound kernel function reaches a state of convergence; when M=2, K _G(G) , K _G(P) , K _P(G) and K _P(P) are continuous Gaussian mapping compound kernels respectively Function, polynomial-Gaussian mapping composite kernel function, Gaussian-polynomial mapping composite kernel function and continuous polynomial mapping composite kernel function; the expressions are as follows: K _G(G) (x,z)＝φ _G [φ _G (x)]·φ _G [φ _G (z)]＝exp[-||φ _G (x)-φ _G (z)|| ² /σ ₂ ² ] (1)＝exp[-[φ _G (x) φ _G (x)+φ _G (z) φ _G (z)-2φ _G (x) φ _G (z)]/ σ ₂ ² ]=exp[-[K _G (x,x)+K _G (z,z)-2K _G (x,z)]/σ ₂ ² ] K _G(P) (x,z)＝φ _G [φ _P (x)]·φ _G [φ _P (z)]＝exp[-||φ _P (x)-φ _P (z)|| ² /σ ₂ ² ] (2)＝exp[-[φ _P (x) φ _P (x)+φ _P (z) φ _P (z)-2φ _P (x) φ _P (z)]/ σ ₂ ² ]=exp[-[K _P (x,x)+K _P (z,z)-2K _P (x,z)]/σ ₂ ² ] $\begin{array}{c}{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; G</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; G</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; G</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; G</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ]</span>}^{{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}\\ <span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; G</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ]</span>}^{{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}\end{array}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$ $\begin{array}{c}{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ]</span>}^{{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}\\ <span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ]</span>}^{{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}\end{array}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$ Where x and z represent two pixels, σ ₂ represents the Gaussian radius of the second Gaussian mapping, d ₂ represents the polynomial parameter of the second polynomial mapping; the Gaussian radius of the first Gaussian mapping in formula (1)(3) is σ ₁ , and the polynomial parameter of the first polynomial mapping in formula (2)(4) is d ₁ ; Step 3.2: Train the support vector machine to obtain the weight vector and thresholds α ^* and b ^* of the decision function of the support vector machine.