CN105869161B - Hyperspectral image band selection method based on image quality evaluation - Google Patents

Abstract

The present invention proposes a hyperspectral image band selection method based on image quality evaluation, which is used to solve the problem of poor discrimination performance of the selected band subset existing in the unsupervised band selection method; it includes the following steps: input the high Spectral image; normalize the hyperspectral image; average the band images in the normalized hyperspectral image to obtain the average image; quantify the band image and the average image; use the average image as a reference to calculate each band The image quality evaluation value of the image; the mutual information between the band and the sample label in the MRMR band selection method is replaced by the image quality evaluation value; the band selection is performed according to the improved MRMR. By introducing image quality evaluation, the present invention can select high-quality wave bands, and the correlation between wave bands is small, and the selected wave band subset has good discrimination performance, and can be used to reduce the dimension of hyperspectral images.

Description

Band Selection Method for Hyperspectral Image Based on Image Quality Evaluation

technical field

The invention belongs to the technical field of image processing, and relates to a hyperspectral image band selection method, in particular to a hyperspectral image band selection method based on image quality evaluation, which can be used for dimensionality reduction of hyperspectral images.

Background technique

Remote sensing is a technology to detect objects in nature from a distance. All objects in nature have the characteristics of reflecting or emitting electromagnetic waves, and the objects are different, and the environments they are in are different. The electromagnetic waves emitted or reflected by objects are also different. Or the reflected electromagnetic wave contains the characteristic information of the object, which can be used as a tool to perceive different objects. Based on this principle, remote sensing technology uses physical equipment to collect electromagnetic waves emitted or reflected by objects to perceive and identify objects. Images obtained through remote sensing are remote sensing images.

Multispectral image is a remote sensing image that we usually contact with more, and it acquires information in several visible light bands in the electromagnetic spectrum at the same time. After a long period of development, remote sensing technology has made great progress. Hyperspectral remote sensing is the product of remote sensing technology developed to a certain stage. Hyperspectral remote sensing images simultaneously acquire information on tens to hundreds of very narrow and continuous spectral segments in the ultraviolet, visible, near-infrared, and mid-infrared regions of the electromagnetic spectrum. Compared with traditional remote sensing images such as multispectral images, hyperspectral images can provide rich information and have high spectral resolution. While hyperspectral images provide a large amount of information, there is also the phenomenon of information redundancy. A large amount of highly correlated information brings difficulties such as high computational complexity and memory consumption to the further processing of hyperspectral images. Therefore, dimensionality reduction is an important part of hyperspectral image processing on the basis of retaining hyperspectral rich information to the greatest extent.

Methods for dimensionality reduction of hyperspectral images include feature extraction and band selection. Feature extraction generally achieves dimensionality reduction by mapping a high-dimensional data space to a low-dimensional data space through a linear or nonlinear method. The data in the low-dimensional space obtained by the feature extraction method through mapping loses the physical meaning represented by the original data. Compared with the method of feature extraction, band selection is based on a non-transformation dimensionality reduction method. This method selects a subset of bands with good performance from the original bands. Compared with the method of feature extraction, band selection does not change the original Data, better retain the physical meaning and information of ground objects in the original data.

Hyperspectral image dimensionality reduction can be divided into supervised band selection and unsupervised band selection according to whether the participation of labeled samples is required. Supervised band selection requires the participation of labeled samples. The existing supervised band selection methods include methods based on inter-class separability, methods based on maximum correlation and minimum redundancy (MRMR), among which MRMR is a very classic supervised band The selection method has a simple principle and high computational efficiency, and can select a high-quality band subset with low redundancy between bands. Supervised band selection can select a subset of bands with better performance because of the participation of marked samples, but obtaining marked samples requires a lot of manpower and material resources, and some areas are not marked. This largely limits the scope of applications of supervised band selection. Unsupervised band selection does not involve labeled samples, and it selects a high-quality subset of bands based on the intrinsic structure between bands. Because it does not require the participation of labeled samples, the unsupervised band selection method has a wide range of applications.

The existing unsupervised band selection methods include information-based band selection methods, similarity-based band selection methods, and cluster-based band selection methods, among which information-based band selection methods specifically include principal component analysis based on maximum variance. method (MVPCA), MVPCA sorts the original bands according to the load factor matrix, so as to select bands with large variance. Although this method can select high-quality bands, the redundancy between bands is very high. The information in is largely repetitive, so the performance performance of the selected band set is relatively poor. The second is the band selection method based on similarity. Its idea is to select the band least similar to the selected band set from the unselected remaining bands. The disadvantage of this method is that only the selected bands and the selected bands are guaranteed. The redundancy between the bands is the lowest, but it may contain very little useful information, and the performance of the band set selected by this method has a lot to do with the two initial bands set. The third cluster-based band selection method is specifically based on the K-medoids band selection method (k-medoids). The implementation idea is to cluster the bands and then select the bands located in the middle. The disadvantages of this method are: The set of selected bands is susceptible to the influence of the initial center point, so the performance of the selected bands also needs to be improved.

Image is an important source of information for human perception and machine pattern recognition. However, the problem of image quality degradation occurs in the process of image acquisition, compression, processing, transmission and display. How to establish an effective image quality evaluation mechanism has been extensively studied by scholars. Image quality evaluation can be divided into subjective evaluation methods and objective evaluation methods. For subjective evaluation, objective evaluation has the advantages of easy operation and low cost, so objective evaluation has become the focus of current research. According to whether the original reference image participates in image quality evaluation, objective evaluation is divided into three types: full reference, semi-reference and no reference. The full-reference image quality assessment method needs to use the original reference image. Structural similarity-based image quality evaluation (SSIM) is a classic full-reference image quality evaluation criterion. SSIM is a method based on structural information to measure the similarity between the original image and the processed image. Its calculation is simple and subjective. Quality evaluation is highly correlated. In recent years, the quality evaluation mechanism has been widely used in video transmission, character recognition, security monitoring, medicine, etc., but image quality evaluation has not been applied in hyperspectral image dimensionality reduction, and each band of hyperspectral image is actually a The quality of images in each band is evaluated, and the obtained quality evaluation value may be instructive for hyperspectral dimensionality reduction.

Contents of the invention

The purpose of the present invention is to overcome the above-mentioned defects in the prior art, propose a hyperspectral image band selection method based on image quality evaluation, improve the MRMR band selection method by introducing image quality evaluation, and propose a new The unsupervised band selection method is used to solve the problem of poor recognition ability of selected bands existing in the existing unsupervised band selection method.

The technical idea of the present invention is: a new unsupervised band selection method is obtained by introducing a quality evaluation criterion. This method can select a band with a large quality evaluation value, and at the same time, this band is redundant with the selected band. The degree is low, so that a subset of bands containing more discriminant information can be obtained, and the purpose of dimensionality reduction of hyperspectral images can be achieved.

According to above-mentioned technical train of thought, realize the technical scheme that the present invention takes, comprise the steps:

Step 1, input hyperspectral image A;

Step 2, normalize the spectral value of the hyperspectral image A to between 0 and 1 to obtain the normalized hyperspectral image A';

Step 3, record each band image in the normalized hyperspectral image A′ as x _i , where x _i represents the i-th band image, i=1,2...N, for hyperspectral image A ' in each band image averaged to obtain the average image, denoted as y';

Step 4: Quantize each band image x _i and average image y' in the normalized hyperspectral image A' to obtain quantized band image x _i ' and quantized average image y';

Step 5, using the quality evaluation criterion, taking the quantized average band image as a reference, to obtain the quality evaluation value of the quantized band image x _i ', and the quality evaluation value of the ith quantized band image x _i ' , denoted as Q( _xi ′, y′);

Step 6, using the quality evaluation value Q( _xi ', y') to replace the mutual information between the band and the sample label in the formula used in the MRMR band selection method, to obtain the replaced formula:

Among them, m refers to the selected m-th band, which is a band in the XS _m-1 band, X is the full set of bands, S _m-1 is the set of selected m-1 bands, XS _{m- 1} is the set of bands to be selected, x _i ′ is a band to be selected, I( _xi ′, x _q ′) refers to the mutual information of band x _i ′ and band x _q ′, x _q ′ is the selected band, Q( _xi ′, y′) refers to the image quality evaluation value of the band _xi ′ and the band x _q ′, and the parameter α is used to adjust the image quality evaluation value and the redundancy between the bands the importance of

Step 7: Obtain an improved MRMR band selection method according to the replaced formula, and use this method to select bands.

Compared with the prior art, the present invention has the following advantages:

First: the present invention replaces the mutual information between the band and the sample label in the MRMR band selection method formula with the image quality evaluation value due to the use of the quality evaluation criterion, and then performs band selection according to the MRMR after replacement, which is different from the prior art In contrast, bands with rich information can be selected, and at the same time, the correlation between bands is guaranteed to be small, and the discrimination ability of band subsets is improved.

Second: the present invention replaces the mutual information between the band and the sample label in the formula of the MRMR band selection method with a quality evaluation value due to the use of the quality evaluation criterion, and does not require the participation of the sample label in the band selection process, which is different from the existing Compared to the MRMR band selection method, hyperspectral images of unlabeled samples can be processed.

Description of drawings

Fig. 1 is the realization flow diagram of the present invention;

Fig. 2 is the performance contrast curve figure on the indian pines hyperspectral image of the inventive method and other band selection methods;

Fig. 3 is a graph showing the performance comparison between the present invention and other band selection methods on paviaU hyperspectral images;

Detailed ways:

Below in conjunction with accompanying drawing and specific embodiment, the present invention is described in further detail:

Referring to Figure 1, the hyperspectral image used in this embodiment is a hyperspectral image of Indiana.

Step 1, input hyperspectral image A;

Step 2, normalize the spectral value of the hyperspectral image A to between 0 and 1 to obtain the normalized hyperspectral image A';

Step 3, record each band image in the normalized hyperspectral image A′ as x _i , where x _i represents the i-th band image, i=1,2...220, for hyperspectral image A ' in each band image averaged to obtain the average image, denoted as y';

Step 4: Quantize each band image x _i and average image y' in the normalized hyperspectral image A' to obtain quantized band image x _i ' and quantized average image y';

Step 5, using the quality evaluation criteria and taking the quantized average image as a reference, to obtain the quality evaluation value of the quantized band image x _i ', the quality evaluation value of the i-th quantized band image x _i ', Denoted as Q( _xi ′, y′);

The present invention adopts the structural similarity quality evaluation criterion SSIM (structural similarity) to evaluate the quality of the image. The SSIM is a method for measuring the similarity between images based on the structural information, and the calculation is simple and has strong correlation with the subjective quality evaluation. The specific steps for obtaining it are as follows:

(5a) Calculate the average intensity of the quantized band image x _i ', i=1,2...220 and the quantized average image y'

Indicates the average intensity of the quantized band image x _i ', μ _y' indicates the average intensity of the quantized average image y', M refers to the number of pixels in the band image x _i ' and the average image y', Indicates the jth pixel of the band image x _i ', y _j 'indicates the jth pixel of the average image y', j=1,2...21025;

(5b) Calculate the brightness contrast function I(x _i ′, y′) of the quantized band image x _i ′ and the quantized average image y′

Wherein C ₁ is a small non-zero constant added in order to avoid the denominator being zero. In the present invention, C ₁ =(K ₁ L) ² , parameter L can take any value in the range of pixel value variation, in this example The range of variation of the pixel value is 0-255, so L can take any value in 0-255, and in the present invention, L takes a value of 255, and parameter K is a constant less than ₁ , and among the present invention, K takes _a value of 0.01.

(5c) Calculate the standard deviation of the quantized band image x _i ' and the quantized mean image y'

in is the standard deviation of the quantized band image x _i ', and δ _y' is the standard deviation of the quantized average image y';

(5d) Calculate the contrast function C(x _i ′, y′) of the quantized band image x _i ′ and the quantized average image y′

Wherein C ₂ is a small non-zero constant added in order to avoid denominator being zero, order C ₂ =(K ₂ L) ² in the present invention, parameter K ₂ is a constant less than 1, K ₂ value among the present invention is 0.03.

(5e) Calculate the structure comparison function S(x _i ′, y′) of the quantized band image x _i ′ and the quantized average image y′

where C ₃ is a small non-zero constant added to avoid a zero denominator, so that is defined as follows

(5f) Calculate the value of SSIM

SSIM(x _i ′,y′)=[I(x _i ′,y′)] ^η [C(x _i ′,y′)] ^β [S(x _i ′,y′)] ^γ

η, β, and γ are constants greater than zero, controlling the importance of I( _xi ′, y′), C( _xi ′, y′), and S( _xi ′, y′) in calculating the SSIM value SSIM( _xi ′, y′) is the image quality evaluation value recorded as Q( _xi ′, y′) in step 5. In the present invention, let η=β=γ, the above-mentioned I( _xi ', y'), C( _xi ', y'), S( _xi ', y') and parameters η, β , γ into the SSIM( _xi ′, y′) formula to get:

Step 6, using the quality evaluation value SSIM( _xi ', y'), replace the mutual information between the band and the sample label in the formula used in the MRMR band selection method, and obtain the replaced formula:

Among them, m refers to the selected m-th band, which is a band in the XS _m-1 band, X is the full set of bands, S _m-1 is the set of selected m-1 bands, XS _{m- 1} is the set of bands to be selected, x _i ′ is a band to be selected, I( _xi ′, x _q ′) refers to the mutual information of band x _i ′ and band x _q ′, x _q ′ is the selected band, SSIM( _xi ′, y′) refers to the image quality evaluation value of the band x _i ′ and y′, the parameter α is used to adjust the image quality evaluation value and the redundancy between the bands the importance of;

Step 7 According to the formula after replacement, an improved MRMR band selection method is obtained, and this method is used for band selection. The specific steps are as follows:

(7a) Denote the band set composed of 220 bands in the hyperspectral image as X, and set an empty set S=[];

(7b) According to the obtained SSIM value, select a band from the band set X, add it to the empty set S, and obtain the selected band set S ₁ , remove the selected first band from the original band set X, and obtain Selected band set XS ₁ ;

Select the largest SSIM value from 220 SSIM values, take the corresponding band image as the first selected band, and record this band as p ₁ , add this band to the empty set S, and get the selected band set S ₁ =[p ₁ ], the band p ₁ is removed from the original band set X, and the band set to be selected is XS ₁ , and the number of bands is 219;

(7c) Select the second band from the band set XS ₁ to be selected according to the improved formula, add it to S ₁ to obtain the set S ₂ of the selected bands, remove the selected second band from the set XS ₁ to be selected, Get new candidate band set XS ₂

In the MRMR band selection method, the mutual information between the band and the sample label in the formula used is replaced by the image quality evaluation value SSIM, and the improved formula is as follows:

Select the second band according to the above formula, where x _k ′ is a band in the band set XS ₁ to be selected, k=1,2...219, x _q ′ is a band in the set S ₁ of the selected bands, At this time, there is only one band in the selected band set, that is, the band p ₁ selected in (7b). The purpose of this formula is to select the band with a large SSIM value and the redundancy of this band with the bands in the selected band set minimum. Record the second selected band as p ₂ , add this band to the band set S ₁ , and obtain the selected band set S ₂ =[p ₁ p ₂ ], S ₂ means that there are two bands in the selected band set, Remove the selected second band from the XS ₁ set to get a new candidate band set XS ₂ , the number of bands is 218, and the parameter α is used to adjust the image quality evaluation value SSIM and the redundancy between the bands The importance of , α is set to 1.573 on the Indian pines image;

(7d) Select the third band from the candidate band set XS ₂ according to the improved formula, add it to S ₂ to obtain the selected band set S ₃ , remove the selected third band from the candidate band set XS ₂ bands to get a new set of bands to be selected XS ₃ , the improved formula is as follows:

Select the third band according to the above formula, where x _k ′ is a band in the band set XS ₂ to be selected, k=1,2...218, x _q ′ is a band in the set S ₂ of the selected bands, At this time, there are two bands in the selected band set, that is, the band p ₁ selected in (7b) and the band p ₂ selected in (7c). The purpose of this formula is to select the band with a large SSIM Minimum redundancy with the bands in the Selected Bands Set. Record the selected third band as p ₃ , add this band to S ₂ , and get the selected band set S ₃ =[p ₁ p ₂ p ₃ ], S ₃ means that there are three bands in the selected band set, in Remove the selected third band from the XS ₂ set to obtain a new candidate band set XS ₃ , the number of which is 217;

(7e) Select the mth band from the candidate band set XS _m-1 according to the improved formula, add it to S _m-1 to obtain the selected band set S _m , and remove the selection from the candidate band set XS _m-1 out of the mth band, get a new set of bands to be selected XS _m

Select the mth band according to the above formula, where x _k ′ is a band in the candidate band set XS _m-1 , k=1,2...N-(m-1), x _q ′ is the selected band A band in the set S _m-1 . At this time, there are m-1 bands in the selected band set, that is, the band p ₁ p ₂ ...p _m-1 selected through the above steps. The purpose of this formula is A band with a large SSIM value is selected, and at the same time, the redundancy between this band and the bands in the selected band set is the smallest. Record the selected mth band as p _m , add this band to the selected band set S _m-1 to get the selected band set S _m =[p ₁ p ₂ p ₃ ...p _m-1 p _m ], S _m represents that there are m bands in the selected band set, remove the selected m band in the XS _m-1 set, and obtain a new candidate band set XS _m ;

(7f): Repeat step (7e) to gradually select bands until the required number of bands p is reached. At this time, the required bands are in the selected set S _p .

The effect of the present invention will be further described in conjunction with the simulation diagram below.

The simulation process includes:

1. Simulation conditions:

The hardware test platform adopted in the simulation experiment of the present invention is: the processor is Inter Core i7-4720HQ, the main frequency is 2.60GHz, and the internal memory is 8GB; the software platform is: Windows 10 Ultimate Edition 64-bit operating system, Matlab R2015a for simulation testing.

2. Simulation data:

What adopted in the present invention is the image of the University of Pavia recorded by the ROSIS-03 optical sensor, and the hyperspectral image of Indiana obtained by the airborne visible light and infrared imaging spectrometer AVIRIS to carry out the classification algorithm simulation to verify the effectiveness of the selected band of the present invention The hyperspectral image of the University of Pavia is recorded as paviaU, and the hyperspectral image of Indiana is recorded as Indian pines. These two data sets are industry-recognized experimental data for hyperspectral data processing.

The above hyperspectral data and the corresponding ground truth images are from:

http://www.ehu.es/ccwintco/index.php?html title=Hyperspectral_Remote_Sensing_Scenes

3. Simulation content:

In order to verify the effectiveness of the selected band set, generally after band selection, the hyperspectral data corresponding to the selected band will be used, and the category in the ground truth image will be used as the label to classify the hyperspectral image, and the classification accuracy will be used as Criteria for judging band selection performance.

(3.1) Simulation 1:

In the experiment, three existing representative unsupervised band selection methods are selected for comparison with the method mentioned in the present invention. These three methods are SBBS (similarity-based band selection method), MVPCA (maximum variance principal component analysis), k-medoids (K center point band selection method);

For the Indian pines image, use the method of the present invention and SBBS, MVPCA and k-medoids to select 100 bands, select 10% of the known labeled samples as training samples, and the remaining known labeled samples as testing samples, using support vectors machine (svm) classifier for classification experiments. All svm classifiers in this hyperspectral image experiment are implemented using libsvm-3.20, the parameters are set to c=1024, g= ^2-7 , and the rest parameters use default values;

The effectiveness of the present invention is set forth below by simulation result analysis; Result is as accompanying drawing 2;

Accompanying drawing 2 is the effect comparison curve diagram of the present invention method and existing three kinds of band selection methods obtained on Indian pines hyperspectral image, the more discriminant information that the band subset that the band selection method selects comprises, the classification accuracy that obtains is more high. The figure shows the accuracy obtained by selecting 10 to 100 bands for classification by the four methods respectively. It can be seen from the figure that the method of the present invention (SSIM) is slightly lower than k-medoids in band 10, but the effect is better than SBBS and MVPCA, and the effect in other bands is better than the three comparison methods, especially The effect obtained by selecting 40 bands in the present invention is already better than the performance of the 100 bands obtained by the existing three methods, which shows that the band selection performance of the present invention is very good, contains rich discriminant information, and reduces dimensionality. The effect is good.

(3.2) Simulation 2:

For the paviaU image, use the method of the present invention and SBBS, MVPCA and k-medoids to select 50 bands, select 1% of the known labeled samples as training samples, and use the remaining known labeled samples as testing samples, use the support vector machine (svm) classifier for classification experiments. The svm classifier used in this hyperspectral image experiment is implemented using libsvm-3.20, the parameters are set to c=1024, g=2, and the rest parameters use default values; the results are shown in Figure 3;

Accompanying drawing 3 is the comparison curve of the effect of the method of the present invention and the existing three band selection methods obtained on the paviaU hyperspectral image. The figure shows the accuracy obtained by selecting 5 to 50 bands for classification by the four methods respectively. As can be seen from the figure, the effect of the method of the present invention is obviously better than that of the existing three methods in the entire range of band variation, and when the present invention selects 10 bands, the recognition performance has tended to a stable level, and the classification effect is very good. It has reached the level when other methods select 50 bands, indicating that the method of the present invention can achieve the optimal effect when selecting less bands.

To sum up, the effect of the present invention is obviously better than several other methods. In the case of different numbers of bands, except for the number of individual bands, the recognition performance of other methods has been significantly improved, indicating that the band subset selected by the present invention contains abundant The discriminant information has better recognition performance.

Claims (3)

1. a hyperspectral image band selection method based on image quality evaluation, is characterized in that, comprises the steps: (1) input hyperspectral image α; (2) normalizing the spectral value of the hyperspectral image A to between 0-1 to obtain a normalized hyperspectral image A'; (3) Record each band image in the normalized hyperspectral image A' as x _i , where x _i represents the i-th band image, i=1, 2...N, for normalization Each band image in the hyperspectral image Α ' after being transformed is averaged, obtains average image, is denoted as y '; (4) Quantify each band image x _i and average image y' in the normalized hyperspectral image A' to obtain quantized band image x _i ' and quantized average image y'; (5) Using the quality evaluation criteria and taking the quantized average image y′ as a reference, obtain the quality evaluation value of the quantized band image x _i ′, and the quality evaluation value of the ith quantized band image x _i ′ Value, denoted as Q( _xi ′, y′); (6) Using the quality evaluation value Q( _xi ', y'), replace the mutual information between the band and the sample label in the formula used in the MRMR band selection method, and obtain the replaced formula: Among them, m refers to the selection of the mth band, which is a band in the XS _m-1 band set, X is the full set of bands, S _m-1 is the set of selected m-1 bands, and XS _{m- 1} is the set of bands to be selected, x _i ′ is a band to be selected, I( _xi ′, x _q ′) refers to the mutual information of band x _i ′ and band x _q ′, x _q ′ is the selected band, Q( _xi ′, y′) refers to the image quality evaluation value of the band _xi ′ and the band y′, and the parameter α is used to adjust the image quality evaluation value and the redundancy between the bands. the importance of (7) According to the formula after the replacement, an improved MRMR band selection method is obtained, and the band selection is performed using this method. 2. the hyperspectral image band selection method based on image quality evaluation according to claim 1, is characterized in that, the quality evaluation value described in step (5), adopts the quality evaluation criterion SSIM based on structural similarity to ask for, Follow the steps below to achieve: (5a) Calculate the average intensity of the quantized band image x _i ', i=1,2...N and the quantized average image y' Indicates the average intensity of the quantized band image x _i ', μ _y' indicates the average intensity of the quantized average image y', M refers to the number of pixels in the band image x _i ' and the average image y', Indicates the jth pixel of the band image x _i ', y _j 'indicates the jth pixel of the average image y', j=1,2...M; (5b) Calculate the brightness contrast function I(x _i ′, y′) of the quantized band image x _i ′ and the quantized average image y′ Wherein C ₁ =(K ₁ L) ² , the parameter L can take any value within the variation range of the pixel value, and the parameter K1 is a constant less than 1; (5c) Calculate the standard deviation of the quantized band image x _i ' and the quantized mean image y' in is the standard deviation of the quantized band image x _i ', and δ _y' is the standard deviation of the quantized average image y'; (5d) Calculate the contrast function C(x _i ′, y′) of the quantized band image x _i ′ and the quantized average image y′ Where C ₂ =(K ₂ L) ² , parameter K ₂ is a constant less than 1; (5e) Calculate the structure comparison function S(x _i ′, y′) of the quantized band image x _i ′ and the quantized average image y′ in is defined as follows (5f) Calculate the value of SSIM SSIM(x _i ′,y′)=[I(x _i ′,y′)] ^η [C(x _i ′,y′)] ^β [S(x _i ′,y′)] ^γ η, β, and γ are constants greater than zero, controlling the importance of I( _xi ′, y′), C( _xi ′, y′), and S( _xi ′, y′) in calculating the SSIM value SSIM( _xi ′, y′) is the image quality evaluation value recorded as Q( _xi ′, y′) in step (5). 3. the hyperspectral image band selection method based on image quality evaluation according to claim 2, is characterized in that, the band selection described in step (7), the band number of its selection is denoted as P, realizes as follows: (7a) Denote the band set composed of N bands in the hyperspectral image as X, and set an empty set S=[]; (7b) According to the SSIM value, select a band from the band set X, add it to the empty set S, and obtain the selected band set S ₁ , remove the selected first band from the original band set X, and obtain Selected band set XS ₁ ; Select the largest SSIM value from N SSIM values, use its corresponding band image as the first selected band, and record this band as p ₁ , add this band to the empty set S, and get the selected band set S ₁ =[p ₁ ], the band p ₁ is removed from the original band set X, and the band set to be selected is XS ₁ , and the number of bands is N-1; (7c) Select the second band from the candidate band set XS ₁ according to the improved formula, add it to S ₁ to obtain the selected band set S ₂ , remove the selected second band from the candidate band set XS ₁ Band, get the new band set XS ₂ to be selected; In the MRMR band selection method, the mutual information between the band and the sample label in the formula used is replaced by the quality evaluation value SSIM, and the improved formula is as follows: Select the second band according to the above formula, where x _k ′ is a band in the band set XS ₁ to be selected, k=1,2...N-1, x _q ′ is a band in the set S ₁ of the selected bands At this time, there is only one band in the selected band set, that is, the band p ₁ selected in (7b). The purpose of this formula is to select a band with a large SSIM The margin is the smallest; record the second selected band as p ₂ , add this band to the band set S ₁ , and get the selected band set S ₂ =[p ₁ p ₂ ], S ₂ means that there are Two bands, remove the selected second band in the XS ₁ set to get a new candidate band set XS ₂ , the number of bands is N-2, and the parameter α is used to adjust the image quality evaluation value SSIM and the band The importance of inter-redundancy; (7d) Select the third band from the candidate band set XS ₂ according to the improved formula, add it to S ₂ to obtain the selected band set S ₃ , remove the selected third band from the candidate band set XS ₂ bands to obtain a new band set XS ₃ to be selected; Select the third band according to the above formula, where the band image x _k ′ is a band in the candidate band set XS ₂ , k=1,2...N-2, and the band image x _q ′ is the selected band set S ₂ , there are two bands in the selected band set at this time, that is, the band p ₁ selected in (7b) and the band p ₂ selected in (7c). The purpose of this formula is to select At the same time, this band has the smallest redundancy with the bands in the selected band set; record the third selected band as p ₃ , add this band to S ₂ , and get the selected band set S ₃ =[p ₁ p ₂ p ₃ ], S ₃ means that there are three bands in the selected band set, remove the selected third band in the XS ₂ set, and get a new candidate band set XS ₃ , the number of bands is N-3 ; (7e) Select the mth band from the candidate band set XS _m-1 according to the improved formula, add it to S _m-1 to obtain the selected band set S _m , in the candidate band set XS _m-1 Remove the selected mth band to obtain a new set of bands to be selected XS _m ; Select the m-th band according to the above formula, where x _k ′ is a band in the band set XS _m-1 to be selected, k=1,2...N-(m-1), and the band image x _q ′ is Select a band in the band set S _m-1 . At this time, there are m-1 bands in the selected band set, that is, the band p ₁ p ₂ ...p _m-1 selected through the above steps. The formula The purpose is to select a band with a large SSIM value and at the same time, this band has the smallest redundancy with the bands in the selected band set; record the selected mth band as p _m , and add this band to the selected band set S _m- In ₁ , get the selected band set S _m = [p ₁ p ₂ p ₃ ...p _m-1 p _m ], S _m means that there are m bands in the selected band set, and remove the selection in the XS _m-1 set The m-th wave band out, obtains the new band set XS _m to be selected; (7f): Repeat step (7e) to gradually select bands until the required number of bands p is reached. At this time, the required bands are in the selected set S _p .