WO2016091017A1

WO2016091017A1 - Extraction method for spectral feature cross-correlation vector in hyperspectral image classification

Info

Publication number: WO2016091017A1
Application number: PCT/CN2015/092591
Authority: WO
Inventors: 刘治; 唐波; 张海霞; 肖晓燕; 聂明钰; 孙育霖
Original assignee: 山东大学
Priority date: 2014-12-09
Filing date: 2015-10-23
Publication date: 2016-06-16
Also published as: CN104463247B; CN104463247A

Abstract

An extraction method for a spectral feature cross-correlation vector in a hyperspectral image classification, comprising hyperspectral image data preprocessing──normalization, denoising, and dimension reduction; boostrap sampling and weighted averaging to acquire a reference sample set; spectrum signal random process hypotheses──hypothesis one, spectrum signals are random trials at any one moment of a stationary random process, hypothesis two, probabilities are equal for every random trial value, then, the spectrum signals are abstracted on the basis of a random process autocorrelation theory to arrive at an autocorrelation coefficient calculation formula, and then finally combined into a autocorrelation feature vector; and, a method of optimal directions (MOD) is employed for sparse decomposition of the correlation feature vector. The present invention sets forth the feature extraction method in hyperspectral classification from a random process correlation perspective, thus providing improved noise immunity and high stability, and allowing increased precision for hyperspectral classification.

Description

A method for extracting spectral vector cross-correlation features in hyperspectral image classification

Technical field

The invention belongs to the field of hyperspectral image processing, and particularly relates to a method for extracting spectral vector cross-correlation features in hyperspectral image classification.

Background technique

The hyperspectral image organically combines the traditional spatial dimension information and the spectral dimension information. When acquiring the scene space image, the continuous spectrum of all objects in the scene is obtained, thereby realizing the classification and recognition target according to the spectral features of the object. Compared with traditional full-color and multi-spectral remote sensing, due to its high spectral resolution and spatial resolution, it combines spectral information and spatial information effectively, and the data is rich, the data model is easy to describe, and the hyperspectral image is recognized and accurate. There are outstanding advantages in classification. With the development and maturity of hyperspectral imaging technology, hyperspectral image processing technology has been widely used in medical diagnosis, agricultural detection, mineral detection, environmental monitoring and other fields.

Hyperspectral image classification is still a major problem for hyperspectral image analysis and processing. Hyperspectral images themselves have many drawbacks, such as excessive data redundancy caused by massive amounts, spectral mixing due to high spatial resolution, and the influence of noise, which greatly increases the difficulty of fine classification. The traditional hyperspectral feature matching classification method requires a large amount of prior knowledge, and the dependence on the spectral feature database is too high, while the statistical classification method is slow in operation and the accuracy is greatly affected by the training samples. Existing feature extraction and classification methods are often limited by the defects of the hyperspectral image itself, which is manifested by insufficient stability and robustness of the algorithm.

Summary of the invention

The object of the present invention is to solve the above problems, and to provide a method for extracting spectral vector cross-correlation features in hyperspectral image classification, which has the problem of effectively dealing with small sample learning classification, has good anti-noise performance, and can effectively improve the whole. The classification accuracy advantage of the classification system.

In order to achieve the above object, the present invention adopts the following technical solutions:

A method for extracting spectral vector cross-correlation features in hyperspectral image classification, comprising the following steps:

Step (1): preprocessing the raw data of the hyperspectral image to obtain a training sample set; the preprocessing comprises: converting the three-dimensional hyperspectral data into a two-dimensional eigenvector matrix, data normalization, and descending by principal component analysis The peacekeeping random extraction part of the marked samples constitutes a training sample set;

Step (2): taking a boostrap sampling method to obtain a reference sample set from the training sample set, calculating a correlation coefficient between the training sample in the training sample set and the reference sample of the reference sample set, and constructing a cross-correlation feature vector;

Step (3): Feature selection: reducing the computational complexity from the cross-correlation feature vector constructed in step (2) From the perspective, feature re-selection is needed. The selection method is to sparsely represent the correlation coefficient feature vector constructed in step (2), and sparse the correlation coefficient feature vector to obtain the sparse feature vector.

The step (1) includes:

Step (1-1): Convert the 3D hyperspectral image into a 2D feature vector form:

Where I is a three-dimensional hyperspectral image, M is the number of image lines, N is the number of image columns, K is the feature number, I ₁ is the transformed two-dimensional feature matrix, each row of I ₁ corresponds to one sample, and each column corresponds to one Feature, L is the total number of pixels, Lable is the label matrix corresponding to I, and Lable ₁ is the label matrix corresponding to I ₁ ;

Step (1-2): data normalization: according to the feature dimension, that is, the column of I ₁ in the step (1-1), searching for the minimum value x _min and the maximum value x _max of the feature values in each column, The original eigenvalues between [x _min , x _max ] are mapped to [-1, 1], and the mapping relationship is as shown in equation (3):

Where y _max =1, y _min =-1, x is the original eigenvalue in I ₁ , y is the eigenvalue after mapping to [-1, 1], and x is replaced by y, and I ₁ is normalized. Obtaining normalized image data I ₂ ;

Step (1-3): Principal component analysis is performed on the normalized image data I ₂ by principal component analysis (PCA) to reduce image noise and feature dimensions:

The principal component analysis process is as follows: the normalized image data I _{2 is} expressed as I ₂ = (x ₁ , x ₂ , ..., x _i , ..., x _L ) ^T , where x _i is a K × 1 column vector , indicating a sample.

The center of the sample, i.e., I ₂ of the central operation of all samples, the specific method is all I ₂ vectors by subtracting the global mean vector

Calculate the covariance matrix of I ₂ after centrifugation

Then, the feature matrix of the covariance matrix is decomposed to obtain the eigenvalue matrix Λ and the eigenvector matrix ω, and the principal component transformation is performed on I ₂ :

I ₃ =I ₂ *ω (4)

I ₃ is image data after dimensional reduction by PCA, and the characteristic dimension K1 (number of columns) of I ₃ is much smaller than the feature dimension K of I ₂ .

Step (1-4): Randomly extract the training sample set:

The extraction method adopts a random number method, that is, randomly generates a random number a = (a ₁ , a ₂ , ..., a _l ) between 1 and L, the random number is not repeated, and l is the number of random numbers.

The generated random number is taken as a row label, and the corresponding row is extracted from I ₃ in step (1-3) to form a training sample set train_matrix _l×K1 .

The generated random number is taken as a row label, and the corresponding label is extracted from the Lable _{1 of the} step (1-1) to form a training sample category label set trian_label _l×1 .

Each row of train_matrix _l×K1 represents a training sample corresponding to the class label of the corresponding row in trian_label _l×1 .

The step (2) includes:

Step (2-1): Construct a collection of reference samples of each category by boostrap sampling:

Suppose there are class c, and the number of reference samples per class N _i . For the category i, according to the training sample category label set trian_label _l×1 and the training sample set train_matrix _{l×K1 in} the step (1-4), the sample of the category label _{i is} first extracted from the training sample set train_matrix _l×K1 . sub _i constituting the subset of samples, followed by extracting the subset of samples a weighted average of 80% of the total number of samples from sub _i sub _i has returned to the weighted average formula is as follows:

Where ref _i is a weighted new sample, is a 1×K1 row vector, x _i is the extracted sample, and l1 is the extracted sample number.

After extracting N _i times, a new reference sample matrix will be obtained.

A summary of all class reference sample matrices results in an overall reference sample matrix Ref = (Ref ₁ , Ref ₂ , ..., Ref _c ) ^T .

Step (2-2): Assume that any two spectral feature vectors x ₁ = (x ₁₁ , x ₁₂ , ..., x _1K1 ) ^T , x ₂ = (x ₂₁ , x ₂₂ , ..., x _2K1 ) ^T , respectively For a random experiment of two different times t, t + τ of the stochastic process X(t, ω), then X(t, ω) = x ₁ , X(t + τ, ω) = x ₂ for discrete stationary stochastic processes, The cross-correlation of random trials at different times in the same stationary stochastic process is:

R _XY (τ) represents the cross-correlation of two random experiments X(t, ω), X(t+τ, ω), τ is the time interval, Ω={ω ₁ , ω ₂ ,..., ω _i ,..., ω _N } denotes the random test sample space, N is the number of all possible outcomes of the random test, ω _i is the result of a random test, and P(ω _i ) is the probability that the random test obtains ω _i .

When t is fixed, let X(ω)=X(t,ω), Y(ω)=X(t+τ,ω), then formula (6) is rewritten as:

R _XY (τ) represents the cross-correlation of two random experiments, x _1i =X(t,ω _i ), x _2i =X(t+τ,ω _i ) respectively corresponding to the spectral feature vector x ₁ , x ₂ i Eigenvalues, ω _i ∈ Ω = {ω ₁ , ω ₂ , ..., ω _i , ..., ω _N }, Ω represents the random test sample space, N is the number of all possible outcomes of the random test, and N = K1, ω _i is the result of a random test, and P(ω _i ) is the probability that a random test obtains ω _i ;

Step (2-3): Conditions for step (2-2) will be further assumed that - assume that the probability values of all features of the P (ω _i) are equal, it is possible to remove the equation (7) is P (ω _i), with R _XY replaces R _XY (τ) and further turns into:

Step (2-4): According to the form of the formula (8), in combination with the kernel method, the original data is mapped to the high-dimensional space, and the RBF kernel function is introduced, and the structure thereof is as follows:

Where k(x, r) represents the RBF kernel function, x represents the test sample column vector, r is the reference sample column vector, σ is the parameter of the RBF kernel function, and the parameter σ is adjustable.

Equation (8) can be replaced by equation (9):

R _xr =k(x,r) (10)

R _xr is the correlation coefficient between the vector x and r, x is the test sample column vector, and r is the reference sample column vector;

Step (2-5): obtaining a training sample matrix composed of correlation coefficient feature vectors by calculation;

Step (2-6): A test sample matrix composed of correlation coefficient feature vectors is obtained by calculation.

The steps of the step (2-5) are:

Step (2-5-1): for each training sample x _i in the training sample set train_matrix _l×K1 in the step (1-4), that is, a row of data in the train_matrix _l×K1 , according to the formula (9) and the formula Ref general reference sample matrix (10) training sample x _i is calculated in step (2-1) in each of the cross-correlation of the reference sample, the correlation coefficient for all reference samples to obtain a feature vector x _i with the total training sample set reference sample Ref cor _i, with x _i cor _i substituents as training samples;

Step (2-5-2): Perform steps (2-5-1) for all training samples x _i , i=1, 2, . . . , l in the training sample set train_matrix _l×K1 in the step (1-4). Operation, obtaining a training sample matrix train=(cor ₁ , cor ₂ ,...,cor _l ) ^T composed of correlation coefficient feature vectors;

The steps of the step (2-6) are as follows:

Step (2-6-1): I ₃ a test any sample x ^*, i.e., data of the line I step (1-3) of _3, according to the equation (9) and equation (10) calculates a test sample x ^* , with the correlation coefficient of each reference sample in the overall reference sample matrix Ref of step (2-1), obtain the test coefficient x ^* , and the correlation coefficient feature vector cor ^{* of} all the reference samples of the total reference sample set Ref, with cor ^* Replace x ^* as a test sample;

Step (2-6-2) for all the test samples in the step (1-3) I ₃

i=1,2,...,L,L is the number of samples in I ₃ , and step (2-6-1) is performed to obtain a test sample matrix composed of correlation coefficient feature vectors.

The method for selecting the feature of the step (3) is:

Step (3-1): Sparsely decompose the correlation coefficient training samples in the training sample matrix train of step (2-5-2) by using the sparse decomposition method, and obtain the sparse dictionary Φ at the same time, and after sparse decomposition, corresponding to each in the train A sample cor _i will get a sparse coefficient feature vector α _i , and replace cor _i with α _{i to} obtain a sparse coefficient training sample set Train.

Step (3-2): performing a sparse decomposition on the sparse dictionary Φ of the step (3-1) on the correlation coefficient test sample matrix test of the step (2-6-2), and corresponding to each of the tests after the sparse decomposition sample

Will get a sparse coefficient eigenvector

use

Replace

, get the sparse coefficient test sample set Test.

The sparse decomposition method is as follows, and the basic model of sparse decomposition is:

y is a feature vector that is not thinned out, Φ is a sparse dictionary, α is a sparse coefficient that the feature vector y decomposes on the sparse dictionary Φ, and λ is a parameter that controls the variation of the sparse coefficient α.

The sparse dictionary Φ is obtained by using the optimal direction method (MOD):

Using the correlation coefficient of the train in step (2-5-2) to train the sample as the input sample X of the MOD algorithm, find a sparse dictionary Φ:

X is the input sample, A sparse coefficient matrix, and β _i represents the ith column in A.

The beneficial effects of the invention:

From the perspective of stochastic process analysis theory, the present invention regards the spectral vector as the random test result at different times of the same stochastic process, and transforms the hyperspectral classification problem into a cross-correlation problem of two random experiments. The larger the mutual relation, the two The greater the probability that the spectral vectors belong to the same class. By introducing the kernel method, the spectral vector is nonlinearly mapped to the high-dimensional space, and the cross-correlation coefficient between the sample to be classified and the reference sample is calculated, and combined into the correlation coefficient feature vector, and through the sparse The solution method sparses the correlation coefficient feature vector and completes the hyperspectral classification feature extraction process.

The method has the advantages of good noise resistance, stability, low computational complexity and high classification accuracy.

DRAWINGS

1 is a flow chart of an autocorrelation feature extraction method in the over-spectral classification of the present invention;

Figure 2 is a schematic diagram of the correlation feature vector construction.

detailed description

The invention will be further described below in conjunction with the drawings and embodiments.

As shown in Figure 1, the process of the autocorrelation feature extraction method is:

1) Data conversion. The original three-dimensional hyperspectral data I is converted into two-dimensional feature data I ₁ , one for each sample and one for each feature.

2) Data normalization. Data normalization is done by a data mapping process. In step 1), the I ₁ data is projected onto the [-1, 1] interval. The process is to search for the minimum value x _min and the maximum value x _max of the feature values in each column, and map [x _min , x _max ] between [-1, 1].

3) Using the principal component analysis method (PCA) to reduce the noise of the data. Dimensioning the data on the basis of 2) requires that as much original image information as possible be represented with as few dimensions as possible. The PCA is reduced in dimension to obtain a preprocessed data set.

4) Acquisition of the sample set. The sample set includes a training sample set and a reference sample set. The invention firstly acquires a training sample set by using a random number method, randomly generates a sample number to be extracted, and then extracts corresponding samples from the preprocessed data set in step 3) to form a training sample set. Then, the boost sample sampling method is used to obtain the reference sample set, and the training samples are back-sampled according to the category, and a reference sample is calculated by weighting the average number of samples each time. On the one hand, the randomness of the sample selection is fully guaranteed. On the other hand, the representativeness of the reference sample is increased by the sample weighted average.

5) Calculation of correlation coefficient. The spectral signal is treated as a stationary stochastic process, and each sample corresponds to a random experimental result of time t. Smooth stochastic process cross-correlation formula:

In the hyperspectral image, ω _i corresponds to the number of bands, then it can be assumed that the probability P(ω _i ) of each ω _i is equal, so the influence of P(ω _i ) on the correlation coefficient in (13) can be eliminated. .

The correlation coefficient calculation is divided into two processes:

(1) For the training sample set, any one of the training samples is calculated separately and all of the reference sample sets in 4) The correlation coefficient R of the samples is then combined into a correlation feature vector.

(2) For the original data sample, for any unlabeled sample, 4) the cross-correlation number R of all samples in the reference sample set, and then combine them into a correlation feature vector.

As shown in Figure 2, the construction process of the correlation feature vector.

6) A sparse representation of the correlation feature vector. Necessity: The number of samples in the newly constructed reference sample set in (1) 4) may be large. According to the calculation in 5), the obtained correlation feature vector dimension will be very high; (2) the coefficient representation can simplify the calculation and reduce the data volume. Improve classification efficiency.

The selected method is sparse representation. Based on the training sample correlation feature vector in 5), a sparse dictionary is trained. The training method is the optimal direction method (MOD), and then the original correlation feature vector is sparsely decomposed.

7) Classifier testing, mainly to evaluate the feature extraction algorithm.

The above description of the specific embodiments of the present invention has been described with reference to the accompanying drawings, but it is not intended to limit the scope of the present invention. Those skilled in the art should understand that the skilled in the art does not require the creative work on the basis of the technical solutions of the present invention. Various modifications or variations that can be made are still within the scope of the invention.

Claims

A method for extracting spectral vector cross-correlation features in hyperspectral image classification, characterized in that it comprises the following steps:

Step (1): preprocessing the raw data of the hyperspectral image to obtain a training sample set; the preprocessing comprises: converting the three-dimensional hyperspectral data into a two-dimensional eigenvector matrix, data normalization, and descending by principal component analysis The peacekeeping random extraction part of the marked samples constitutes a training sample set;

Step (2): taking a boostrap sampling method to obtain a reference sample set from the training sample set, calculating a correlation coefficient between the training sample in the training sample set and the reference sample of the reference sample set, and constructing a cross-correlation feature vector;

Step (3): Feature selection: According to the cross-correlation feature vector constructed in step (2), feature re-selection is needed from the perspective of reducing computational complexity. The selection method is the correlation coefficient feature vector constructed in step (2). A sparse representation is performed, and the sparse correlation coefficient feature vector is obtained to obtain a sparse feature vector.
The method for extracting spectral vector cross-correlation features in hyperspectral image classification according to claim 1, wherein the step of converting the three-dimensional hyperspectral data into a two-dimensional feature vector matrix is:

Step (1-1): Convert the 3D hyperspectral image into a 2D feature vector form:

Where I is a three-dimensional hyperspectral image, M is the number of image lines, N is the number of image columns, K is the feature number, I 1 is the transformed two-dimensional feature matrix, each row of I 1 corresponds to one sample, and each column corresponds to one wherein, L is the total number of cells, Lable I corresponding to the tag matrix, Lable 1 is a matrix corresponding to the tag is I 1.
The method for extracting spectral vector cross-correlation features in hyperspectral image classification according to claim 1 or 2, wherein the step of normalizing the data is:

Step (1-2): searching for the minimum value x min and the maximum value x max of the feature values in each column according to the feature dimension, that is, the column of I 1 in the step (1-1), [x min , x The original eigenvalues between max ] are mapped to [-1, 1], and the mapping relationship is as shown in equation (3):

Where y max =1, y min =-1, x is the original eigenvalue in I 1 , y is the eigenvalue after mapping to [-1, 1], and x is replaced by y, and I 1 is normalized. The normalized image data I 2 is obtained .
The method for extracting spectral vector cross-correlation features in hyperspectral image classification according to claim 1, wherein the step of performing dimensionality reduction by principal component analysis is:

Step (1-3): Principal component analysis is used to perform principal component analysis on the normalized image data I 2 to reduce image noise and feature dimensions:

The principal component analysis process is as follows: the normalized image data I 2 is expressed as I 2 = (x 1 , x 2 , ..., x i , ..., x L ) T , where x i is a K × 1 column vector , indicating a sample;

The center of the sample, i.e., I 2 of the central operation of all samples, the specific method is all I 2 vectors by subtracting the global mean vector

The covariance matrix I=I 2 T *I 2 of the centralized I 2 is calculated, and then the covariance matrix Σ is decomposed into features to obtain the eigenvalue matrix Λ and the eigenvector matrix ω, and the principal component transformation is performed on I 2 :

I 3 =I 2 *ω (4)

I 3 is image data after dimensional reduction by PCA, and the feature dimension K1 of I 3 is much smaller than the feature dimension K of I 2 .
The method for extracting spectral vector cross-correlation features in hyperspectral image classification according to claim 1, wherein the step of randomly extracting the partially labeled samples to form the training sample set is:

Step (1-4): Randomly extract the training sample set:

The extraction method adopts a random number method, that is, randomly generates a random number a=(a 1 , a 2 , . . . , a l ) between 1 and L, the random number is not repeated, and l is the number of random numbers;

The generated random number is taken as a row label, and the corresponding row is extracted from I 3 in step (1-3) to form a training sample set train_matrix l×K1 .

The generated random number is used as a row label, and the corresponding label is extracted from the Lable 1 of the step (1-1) to form a training sample category label set trian_label l×1 ;

Each row of train_matrix l×K1 represents a training sample corresponding to the class label of the corresponding row in trian_label l×1 .
The method for extracting spectral vector cross-correlation features in hyperspectral image classification according to claim 1, wherein the step (2) comprises:

Step (2-1): Construct a collection of reference samples of each category by boostrap sampling:

Step (2-2): Assume that any two spectral feature vectors x 1 = (x 11 , x 12 , ..., x 1K1 ) T , x 2 = (x 21 , x 22 , ..., x 2K1 ) T , respectively For a random experiment of two different times t, t + τ of the stochastic process X(t, ω), then X(t, ω) = x 1 , X(t + τ, ω) = x 2 for discrete stationary stochastic processes The cross-correlation of random trials at different times in the same stationary stochastic process is:

R XY (τ) represents the cross-correlation of two random experiments X(t, ω), X(t+τ, ω), τ is the time interval, Ω={ω 1 , ω 2 ,..., ω i ,..., ω N } represents the random test sample space, N is the number of all possible outcomes of the random test, ω i is the result of a random test, and P(ω i ) is the probability that the random test obtains ω i ;

When t is fixed, let X(ω)=X(t,ω), Y(ω)=X(t+τ,ω), then formula (6) is rewritten as:

R XY (τ) represents the cross-correlation of two random experiments, x 1i =X(t,ω i ), x 2i =X(t+τ,ω i ) respectively corresponding to the spectral feature vector x 1 , x 2 i Eigenvalues, ω i ∈ Ω = {ω 1 , ω 2 , ..., ω i , ..., ω N }, Ω represents the random test sample space, N is the number of all possible outcomes of the random test, and N = K1, ω i is the result of a random test, and P(ω i ) is the probability that a random test obtains ω i ;

Step (2-3): Conditions for step (2-2) will be further assumed that - assume that the probability values of all features of the P (ω i) are equal, it is possible to remove the equation (7) is P (ω i), with R XY replaces R XY (τ) and further turns into:

Step (2-4): According to the form of the formula (8), in combination with the kernel method, the original data is mapped to the high-dimensional space, and the RBF kernel function is introduced, and the structure thereof is as follows:

Where k(x, r) represents the RBF kernel function, x represents the test sample column vector, r is the reference sample column vector, σ is the parameter of the RBF kernel function, and the parameter σ is adjustable;

Equation (8) can be replaced by equation (9):

R xr =k(x,r) (10)

R xr is the correlation coefficient between the vector x and r, x is the test sample column vector, and r is the reference sample column vector;

Step (2-5): obtaining a training sample matrix composed of correlation coefficient feature vectors by calculation;

Step (2-6): A test sample matrix composed of correlation coefficient feature vectors is obtained by calculation.
The method for extracting spectral vector cross-correlation features in hyperspectral image classification according to claim 6, wherein the steps of the step (2-5) are:

Step (2-5-1): for each training sample x i in the training sample set train_matrix l×K1 in the step (1-4), that is, a row of data in the train_matrix l×K1 , according to the formula (9) and the formula Ref general reference sample matrix (10) training sample x i is calculated in step (2-1) in each of the cross-correlation of the reference sample, the correlation coefficient for all reference samples to obtain a feature vector x i with the total training sample set reference sample Ref cor i, with x i cor i substituents as training samples;

Step (2-5-2): Perform steps (2-5-1) for all training samples x i , i=1, 2, . . . , l in the training sample set train_matrix l×K1 in the step (1-4). Operation, obtaining a training sample matrix train=(cor 1 , cor 2 ,...,cor l ) T composed of correlation coefficient feature vectors;

The steps of the step (2-6) are as follows:

Step (2-6-1): I 3 a test any sample x *, i.e., data of the line I step (1-3) of 3, according to the equation (9) and equation (10) calculates a test sample x * , with the correlation coefficient of each reference sample in the overall reference sample matrix Ref of step (2-1), obtain the test coefficient x * , and the correlation coefficient feature vector cor * of all the reference samples of the total reference sample set Ref, with cor * Replace x * as a test sample;

Step (2-6-2) for all the test samples in the step (1-3) I 3
i=1,2,...,L,L is the number of samples in I 3 , and step (2-6-1) is performed to obtain a test sample matrix composed of correlation coefficient feature vectors.
The method for extracting spectral vector cross-correlation features in hyperspectral image classification according to claim 1 or 7, wherein the method for selecting features in the step (3) is:

Step (3-1): Sparsely decompose the correlation coefficient training samples in the training sample matrix train of step (2-5-2) by using the sparse decomposition method, and obtain the sparse dictionary Φ at the same time, and after sparse decomposition, corresponding to each in the train A sample cor i will get a sparse coefficient eigenvector α i , and replace cor i with α i to obtain a sparse coefficient training sample set Train;

Step (3-2): performing a sparse decomposition on the sparse dictionary Φ of the step (3-1) on the correlation coefficient test sample matrix test of the step (2-6-2), and corresponding to each of the tests after the sparse decomposition sample
Will get a sparse coefficient eigenvector
use
Replace
Get the sparse coefficient test sample set Test.
The method for extracting spectral vector cross-correlation features in hyperspectral image classification according to claim 6, wherein the step of the step (2-1) is:

Suppose there are class c, and the number of reference samples per class N i ; for class i, according to the training sample class label set trian_label l×1 and the training sample set train_matrix l×K1 in the step (1-4), first from the training sample train_matrix set in l × K1 decimation class label of a sample i, i constituting the sample subsets sub, sub i secondly from the subset of samples extracted with replacement of 80% of the total weighted average sub i samples, the weighted average of the following formula:

Where ref i is a weighted new sample, is a 1×K1 row vector, x i is the extracted sample, and l1 is the extracted sample number;

After extracting N i times, a new reference sample matrix will be obtained.

A summary of all class reference sample matrices results in an overall reference sample matrix Ref = (Ref 1 , Ref 2 , ..., Ref c ) T .
A method for extracting spectral vector cross-correlation features in hyperspectral image classification according to claim 8, wherein:

The sparse decomposition method is as follows, and the basic model of sparse decomposition is:

y is a feature vector that is not thinned out, Φ is a sparse dictionary, α is a sparse coefficient of the feature vector y decomposed on the sparse dictionary Φ, and λ is a parameter that controls the variation of the sparse coefficient α;

The sparse dictionary Φ is learned by the optimal direction method:

The sample is trained as the input sample X of the optimal direction method with the correlation coefficient of the train in step (2-5-2), and a sparse dictionary Φ is found:

X is the input sample, A sparse coefficient matrix, and β i represents the ith column in A.