CN104794497B - Multicenter approximating method in a kind of classification hyperspectral imagery - Google Patents

Abstract

The invention discloses multicenter approximating method in a kind of classification hyperspectral imagery, comprising: the sample of the label of Stochastic choice known class, build training sample set X and label matrix y; Initialization controling parameters: the maximum division number K that every class sample is set, deviation division threshold value σ _t, and smallest sample number N in the rear class of division _min; Calculate all kinds of matching center c _iwith all kinds of mean deviation division decision-making; Division efficiency assessment, smallest sample number N in class after sample number and division in the subclass comparing division _minsize, if sample number to be all greater than after division smallest sample number N in class in two subclasses _min, then division effectively, until classification convergence; If sample number has one to be less than smallest sample number N in the rear class of division in subclass _min, then illustrate that this classification restrains, division terminates.The present invention is directed to mixed pixel problem in hyperspectral classification, in multidimensional feature space, decision region can be divided more accurately.

Description

Multicenter approximating method in a kind of classification hyperspectral imagery

Technical field

The present invention relates to Hyperspectral imagery processing field, particularly relate to multicenter approximating method in a kind of hyperspectral classification.

Background technology

High light spectrum image-forming, based on multispectral imaging, from ultraviolet near infrared spectral range, adopts imaging spectrometer, to target object continuous imaging on the tens of of spectral coverage or hundreds of spectral bands.While acquisition object space characteristic imaging, also obtain the spectral information of testee.Spectral imaging technology has super multiband, high spectral resolution, wave band is narrow, spectral range is wide and the feature of collection of illustrative plates unification.It is advantageous that the image information that collects is abundant, resolution is compared with high and data descriptive model is many.Because it is in the excellent performance of field of detecting, be widely used in reality.

In recent years, along with the widespread use of high light spectrum image-forming technology, high spectrum image analyzing and processing technology is developed rapidly.Therefore, hyperspectral classification problem receives much concern.The existing defects of high spectrum image own, such as dimension is too high, and data redudancy is large, and mixed pixel etc., seriously hinder the development of hyperspectral technique.Wherein, mixed pixel is due to real space region corresponding to pixel single during high light spectrum image-forming, and not merely there is a kind of material in this region, and therefore, the information of this pixel record is the superposition of target complete information in one's respective area.The existence of mixed pixel has a strong impact on the nicety of grading of remote sensing.On the other hand, the reference sample of high spectrum image Supervised classification is selected very limited, is generally choose from known high spectrum image region, and the training data of process like this exists higher spectral mixing.Classic algorithm in hyperspectral classification, as spectral modeling coupling (SAM), k-arest neighbors (KNN) etc., all need to carry out matching to training set, due to the existence of mixed pixel, sample average center in single classification there will be poor fitting phenomenon, show as single center low to entirety training set fitting degree, directly cause classification accuracy low.

Summary of the invention

For solving the deficiency that prior art exists, the invention discloses multicenter approximating method in a kind of classification hyperspectral imagery, for small-sample learning classification and mixed pixel problem, divided by constraint, obtain the multiple matching center of same classification, while realizing the matching of the multicenter overall situation, the effective nicety of grading improving whole categorizing system.

For achieving the above object, concrete scheme of the present invention is as follows:

Multicenter approximating method in a kind of classification hyperspectral imagery, comprises the following steps:

Step (1): the sample of the label of Stochastic choice known class, builds training sample set X and label matrix y;

Step (2): initialization controling parameters: the maximum division number K that every class sample is set, deviation division threshold value σ _t, and smallest sample number N in the rear class of division _min;

Step (3): calculate all kinds of matching center c _iwith all kinds of mean deviation

Step (4): division decision-making, if mean deviation in classification be greater than the deviation division threshold value σ of setting _t, then need to divide this classification, adopt k means clustering method to obtain the multiple subclasses after dividing, be labeled as different labels respectively, otherwise division terminate;

Step (5): division efficiency assessment, sample number and smallest sample number N in the rear class of division in the subclass that comparison step (4) divides _minsize, if sample number to be all greater than after division smallest sample number N in class in two subclasses _min, then division effectively, repeats step (3) ~ (5) to subclass, until classification convergence; If sample number has one to be less than smallest sample number N in the rear class of division in subclass _min, then illustrate that this classification restrains, division terminate, division terminate after, obtain a component split after subclass;

Step (6): the matching center c calculating the rear each subclass of division _ij, i=1,2 ..., C, C are classification numbers, and j is the subclass number of the i-th classification division.

The concrete grammar that described step (1) builds training sample set X and label matrix y is:

Adopt the bootstrap method of sampling to randomly draw training sample, namely to specific class i, i ∈ [1,2 ..., C], C represents classification number, has the sampling of putting back at random, obtains training sample set and label matrix $y_i=[y_{i1}{y}_{i2},...,y_{{\mathrm{iN}}_i}];$

Wherein, x _ijrepresent a jth sample in classification i, l is characteristic number, y _ij=i, i=1,2 ..., C, j=1,2 ... N _i, N _irepresent the total sample number in classification i, it is set of real numbers.

Described step (3) calculates all kinds of matching center c _iconcrete grammar be:

The computing method at all kinds of matching center are the average of all samples in compute classes, are shown below:

$c_i=\frac{1}{N}\underset{j=1}{\overset{N_i}{\mathrm{\Σ}}}{x}_{\mathrm{ij}}---\left(1\right)$

Wherein, c _ifor the class matching center of classification i, x _ijrepresent a jth sample in classification i, N _irepresent the total sample number in classification i, i=1,2 ..., C, C are classification numbers.

All kinds of mean deviation account form as follows:

${\mathrm{\σ}}_j=\sqrt{{\left|\right|x_{\mathrm{ij}}-c_i\left|\right|}^2}$

${\stackrel{\‾}{\mathrm{\σ}}}_i=\frac{1}{N_i}\underset{j=1}{\overset{N_i}{\mathrm{\Σ}}}{\mathrm{\σ}}_j$

Wherein, σ _jrepresent a jth sample x in classification i _ijwith class center c _istandard deviation, j=1,2 ..., N _i, N _irepresent the total sample number in classification i, represent the mean deviation of classification i, i=1,2 ..., C, C are classification numbers.

The method of described step (4) division decision-making is:

Step (4-1): the mean deviation of classification i in comparison step (3) threshold value σ is divided with step (2) large deviations _tif, and when predivision number of times is less than maximum division number K, then need to carry out binary fission in classification to this; Otherwise, illustrate that this classification restrains, stop fission process;

Step (4-2): in classification, the method for binary fission is k mean cluster, k=2, carries out two taxonomic clusterings to all samples in classification i, classification i is split into two subclasses, and gives different class labels respectively, be designated as C _sub1, C _sub2.

Described step (4-2) k means clustering method is:

Step (4-2-1): to classification i, in random selecting classification i, any two samples are as initial cluster centre, are designated as μ ₁, μ ₂, μ ₁corresponding subclasses C _sub1, μ ₂corresponding subclasses C _sub2;

Step (4-2-2): to arbitrary sample x _ij, j=1,2 ..., N _i, calculate x respectively _ijwith two initial cluster center μ in step (4-2-1) ₁, μ ₂euclidean distance square:

$\left\{\begin{array}{c}{d}_1={\left|\right|x_{\mathrm{ij}}-{\mathrm{\μ}}_1\left|\right|}^2\\ d_2={\left|\right|x_{\mathrm{ij}}-{\mathrm{\μ}}_2\left|\right|}^2\end{array}\right.$

D ₁represent x _ijto cluster centre μ ₁euclidean distance square, d ₂represent x _ijto cluster centre μ ₂euclidean distance square;

If d ₁≤ d ₂, then sample x _ijbelong to subclasses C _sub1otherwise, sample x _ijbelong to subclasses C _sub2;

Step (4-2-3): upgrade cluster centre μ ₁, μ ₂, mode is as follows:

${\mathrm{\μ}}_1=\frac{1}{N_{\mathrm{sub}1}}\underset{x\∈C_{\mathrm{sub}1}}{\mathrm{\Σ}}x$

${\mathrm{\μ}}_2=\frac{1}{N_{\mathrm{sub}2}}\underset{x\∈C_{\mathrm{sub}2}}{\mathrm{\Σ}}x$

N _sub1represent subclasses C _sub1interior sample number, N _sub2represent subclasses C _sub2interior sample number, be a sample, l is characteristic number;

Step (4-2-4): repetition step (4-2-2) and step (4-2-3), until convergence, obtain two subclasses C _sub1, C _sub2.

The method of described step (5) division Efficient Evaluation is:

Subclasses C in comparison step (4) _sub1, C _sub2interior number of samples and smallest sample number N in the rear class of division _minsize, if sample number to be all greater than after division smallest sample number N in class in two subclasses _min, then division effectively, to two subclasses C in step (4) _sub1, C _sub2repeat step (3) ~ (5) until convergence.If sample number has one to be less than smallest sample number N in the rear class of division in subclass _minor reach maximum division number of times K, then illustrate that this classification restrains, division terminates.Division terminate after, obtain a component split after subclass.

Described step (6) calculates the matching center c of the rear each subclass of division _ij, i=1,2 ..., C, C are classification numbers, and j is the subclass number of the i-th classification division, and the computing method at the matching center of each subclass are the average calculating all samples in each subclass.

Beneficial effect of the present invention:

(1) core concept of the present invention is under pre-conditioned constraint, for each classification divides multiple matching center, minimizes class mean deviation, has good stability;

(2) the present invention is by arranging maximum division number of times and division threshold value, realizes the controllability of classification division;

(3) the present invention is directed to mixed pixel problem in hyperspectral classification, in multidimensional feature space, decision region can be divided more accurately.

(4) progressively divide convergence under category realization condition constraint of the present invention, obtain multiple matching center, the effective nicety of grading improving whole categorizing system.

Accompanying drawing explanation

Multicenter approximating method process flow diagram in Fig. 1 hyperspectral classification of the present invention;

Equalization fitting result in three class division classes in Fig. 2 a two-dimensional space;

Three class division multicenter fitting results in Fig. 2 b two-dimensional space.

Embodiment:

Below in conjunction with accompanying drawing, the present invention is described in detail:

As shown in Figure 1, in a kind of classification hyperspectral imagery, the process of multicenter approximating method is:

(1) .bootstrap sampling obtains training sample.The sample having the extraction of putting back to certain from the sample marked at random is as training sample.

(2). parameter initialization: parameter comprises the maximum division number K of every class, deviation division threshold value σ _tand smallest sample number N in the rear class of division _min.

(3). calculate all kinds of matching center c respectively _iand mean deviation

(4). the mean deviation of classification i in comparison step 3 threshold value σ is divided with initialized deviation in step (2) _tif, and when predivision number of times is less than maximum division number K, then carries out k mean cluster, k=2, current class is divided into two subclasses, and gives different class labels respectively; Otherwise, illustrate that this classification restrains, stop fission process.

(5). division efficiency assessment.For two subclasses that step (4) divides, if sample number is all greater than smallest sample number N in the rear class of division in two subclasses _min, then classification effectively, repeats step (3) ~ (5) until convergence to subclass.If sample number has one to be less than smallest sample number N in the rear class of division in subclass _minor reach maximum division number of times K, then illustrate that this classification restrains, division terminates.Division terminate after, obtain a component split after subclass.

(6). calculate the matching center c of the rear each subclass of division _ij, i=1,2 ..., C, C are classification numbers, and j is the subclass number of classification i division.

Detailed protocol is:

The concrete grammar that step (1) builds training sample set X and label matrix y is:

Adopt the bootstrap method of sampling to randomly draw training sample, namely to specific class i, i ∈ [1,2 ..., C], C represents classification number, has the sampling of putting back at random, obtains training sample set and label matrix $y_i=[y_{i1}{y}_{i2},...,y_{{\mathrm{iN}}_i}];$

Wherein, x _ijrepresent a jth sample in classification i, l is characteristic number, y _ij=i, i=1,2 ..., C, j=1,2 ... N _i, N _irepresent the total sample number in classification i, it is set of real numbers.

Step (3) calculates all kinds of matching center c _iconcrete grammar be:

The computing method at all kinds of matching center are the average of all samples in compute classes, are shown below:

$c_i=\frac{1}{N}\underset{j=1}{\overset{N_i}{\mathrm{\Σ}}}{x}_{\mathrm{ij}}---\left(1\right)$

Wherein, c _ifor the class center of classification i, x _ijrepresent a jth sample in classification i, N _irepresent the total sample number in classification i, i=1,2 ..., C, C are classification numbers.

The account form of all kinds of mean deviation is as follows:

${\mathrm{\σ}}_j=\sqrt{{\left|\right|x_{\mathrm{ij}}-c_i\left|\right|}^2}$

${\stackrel{\‾}{\mathrm{\σ}}}_i=\frac{1}{N_i}\underset{j=1}{\overset{N_i}{\mathrm{\Σ}}}{\mathrm{\σ}}_j$

Wherein, σ _jrepresent a jth sample x in classification i _ijwith class center c _istandard deviation, j=1,2 ..., N _i, N _irepresent the total sample number in classification i, represent the mean deviation of classification i, i=1,2 ..., C, C are classification numbers.

The method of step (4) division decision-making is:

Step (4-1): the mean deviation of classification i in comparison step (3) threshold value σ is divided with step (2) large deviations _tif, and when predivision number of times is less than maximum division number K, then need to carry out binary fission in classification to this; Otherwise, illustrate that this classification restrains, stop fission process;

Step (4-2): in classification, the method for binary fission is k mean cluster, k=2, carries out two taxonomic clusterings to all samples in classification i, classification i is split into two subclasses, and gives different class labels respectively, be designated as C _sub1, C _sub2.

Step (4-2) k means clustering method is:

Step (4-2-1): to classification i, in random selecting classification i, any two samples are as initial cluster centre, are designated as μ ₁, μ ₂, μ ₁corresponding subclasses C _sub1, μ ₂corresponding subclasses C _sub2;

Step (4-2-2): to arbitrary sample x _ij, j=1,2 ..., N _i, calculate x respectively _ijwith two initial cluster center μ in step (4-2-1) ₁, μ ₂euclidean distance square:

$\left\{\begin{array}{c}{d}_1={\left|\right|x_{\mathrm{ij}}-{\mathrm{\μ}}_1\left|\right|}^2\\ d_2={\left|\right|x_{\mathrm{ij}}-{\mathrm{\μ}}_2\left|\right|}^2\end{array}\right.$

D ₁represent x _ijto cluster centre μ ₁euclidean distance square, d ₂represent x _ijto cluster centre μ ₂euclidean distance square;

If d ₁≤ d ₂, then sample x _ijbelong to subclasses C _sub1otherwise, sample x _ijbelong to subclasses C _sub2;

Step (4-2-3): upgrade cluster centre μ ₁, μ ₂, mode is as follows:

${\mathrm{\μ}}_1=\frac{1}{N_{\mathrm{sub}1}}\underset{x\∈C_{\mathrm{sub}1}}{\mathrm{\Σ}}x$

${\mathrm{\μ}}_2=\frac{1}{N_{\mathrm{sub}2}}\underset{x\∈C_{\mathrm{sub}2}}{\mathrm{\Σ}}x$

N _sub1represent subclasses C _sub1interior sample number, N _sub2represent subclasses C _sub2interior sample number, be a sample, l is characteristic number;

Step (4-2-4): repetition step (4-2-2) and step (4-2-3), until convergence, obtain two subclasses C _sub1, C _sub2.

The method of step (5) division Efficient Evaluation is:

Subclasses C in comparison step (4) _sub1, C _sub2interior number of samples and smallest sample number N in the rear class of division _minsize, if sample number to be all greater than after division smallest sample number N in class in two subclasses _min, then division effectively, to two subclasses C in step (4) _sub1, C _sub2repeat step (3) ~ (5) until convergence.If sample number has one to be less than smallest sample number N in the rear class of division in subclass _minor reach maximum division number of times K, then illustrate that this classification restrains, division terminates.Division terminate after, obtain a component split after subclass.

Step (6) calculates the matching center c of the rear each subclass of division _ij, i=1,2 ..., C, C are classification numbers, and j is the subclass number of classification i division, and method is the average calculating all samples in subclass.

Three class division results in two-dimensional space: Fig. 2 a is equalization fitting result in class, and Fig. 2 b is multicenter fitting result.

By reference to the accompanying drawings the specific embodiment of the present invention is described although above-mentioned; but not limiting the scope of the invention; one of ordinary skill in the art should be understood that; on the basis of technical scheme of the present invention, those skilled in the art do not need to pay various amendment or distortion that creative work can make still within protection scope of the present invention.

Claims (6)

1. a multicenter approximating method in classification hyperspectral imagery, is characterized in that, comprise the following steps:

Step (1): the sample of the label of Stochastic choice known class, builds training sample set X and label matrix y;

Step (2): initialization controling parameters: the maximum division number K that every class sample is set, deviation division threshold value σ _t, and smallest sample number N in the rear class of division _min;

Step (3): calculate all kinds of matching center c _iwith all kinds of mean deviation

Step (4): division decision-making, if mean deviation in classification be greater than the deviation division threshold value σ of setting _t, then need to divide this classification, adopt k means clustering method to obtain the multiple subclasses after dividing, be labeled as different labels respectively, otherwise division terminate;

Step (5): division efficiency assessment, sample number and smallest sample number N in the rear class of division in the subclass that comparison step (4) divides _minsize, if sample number to be all greater than after division smallest sample number N in class in two subclasses _min, then division effectively, repeats step (3) ~ (5) to subclass, until classification convergence; If sample number has one to be less than smallest sample number N in the rear class of division in subclass _min, then illustrate that this classification restrains, division terminate, division terminate after, obtain a component split after subclass;

Step (6) calculates the matching center c of the rear each subclass of division _ij, i=1,2 ..., C, C are classification numbers, and j is the subclass number of classification i division;

The method of described step (4) division decision-making is:

Step (4-1): the mean deviation of classification i in comparison step (3) threshold value σ is divided with step (2) large deviations _tif, and when predivision number of times is less than maximum division number K, then need to carry out binary fission in classification to this; Otherwise, illustrate that this classification restrains, stop fission process;

Step (4-2): in classification, the method for binary fission is k mean cluster, k=2, carries out two taxonomic clusterings to all samples in classification i, classification i is split into two subclasses, and gives different class labels respectively, be designated as C _sub1, C _sub2;

Described step (4-2) k means clustering method is:

Step (4-2-1): to classification i, in random selecting classification i, any two samples are as initial cluster centre, are designated as μ ₁, μ ₂, μ ₁corresponding subclasses C _sub1, μ ₂corresponding subclasses C _sub2;

Step (4-2-2): to arbitrary sample x _ij, j=1,2 ..., N _i, calculate x respectively _ijwith two initial cluster center μ in step (4-2-1) ₁, μ ₂euclidean distance square:

$\left\{\begin{array}{c}{d}_1=\left|\right|x_{ij}-{\mathrm{\μ}}_1|{|}^2\\ d_2=\left|\right|x_{ij}-{\mathrm{\μ}}_2|{|}^2\end{array}\right.$

D ₁represent x _ijto cluster centre μ ₁euclidean distance square, d ₂represent x _ijto cluster centre μ ₂euclidean distance square;

If d ₁≤ d ₂, then sample x _ijbelong to subclasses C _sub1otherwise, sample x _ijbelong to subclasses C _sub2;

Step (4-2-3): upgrade cluster centre μ ₁, μ ₂, mode is as follows:

${\mathrm{\μ}}_1=\frac{1}{N_{sub1}}\underset{x\∈C_{sub1}}{\Σ}x$

${\mathrm{\μ}}_2=\frac{1}{N_{sub2}}\underset{x\∈C_{sub2}}{\Σ}x$

N _sub1represent subclasses C _sub1interior sample number, N _sub2represent subclasses C _sub2interior sample number, be a sample, l is characteristic number;

Step (4-2-4): repetition step (4-2-2) and step (4-2-3), until convergence, obtain two subclasses C _sub1, C _sub2.

2. multicenter approximating method in a kind of classification hyperspectral imagery as claimed in claim 1, is characterized in that, the concrete grammar that described step (1) builds training sample set X and label matrix y is:

Adopt the bootstrap method of sampling to randomly draw training sample, namely to specific class i, i ∈ [1,2 ..., C], C represents classification number, has the sampling of putting back at random, obtains training sample set and label matrix $y_i=\left[\begin{array}{cccc}{y}_{i1}& y_{i2},& ...,& y_{{\mathrm{iN}}_i}\end{array}\right];$

Wherein, x _ijrepresent a jth sample in classification i, l is characteristic number, y _ij=i, i=1,2 ..., C, j=1,2 ... N _i, N _irepresent the total sample number in classification i, it is set of real numbers.

3. multicenter approximating method in a kind of classification hyperspectral imagery as claimed in claim 1, is characterized in that, described step (3) calculates all kinds of matching center c _iconcrete grammar be:

The computing method at all kinds of matching center are the average of all samples in compute classes, are shown below:

$c_i=\frac{1}{N_i}\underset{j=1}{\overset{N_i}{\Σ}}{x}_{ij}---\left(1\right)$

Wherein, c _ifor the class center of classification i, x _ijrepresent a jth sample in classification i, N _irepresent the total sample number in classification i, i=1,2 ..., C, C are classification numbers.

4. multicenter approximating method in a kind of classification hyperspectral imagery as described in claim 1 or 3, is characterized in that, all kinds of mean deviation account form as follows:

${\mathrm{\σ}}_j=\sqrt{\left|\right|x_{ij}-c_i|{|}^2}$

${\stackrel{\‾}{\mathrm{\σ}}}_i=\frac{1}{N_i}\underset{j=1}{\overset{N_i}{\Σ}}{\mathrm{\σ}}_j$

Wherein, σ _jrepresent a jth sample x in classification i _ijwith class center c _istandard deviation, j=1,2 ..., N _i, N _irepresent the total sample number in classification i, represent the mean deviation of classification i, i=1,2 ..., C, C are classification numbers.

5. multicenter approximating method in a kind of classification hyperspectral imagery as claimed in claim 1, is characterized in that, the method for described step (5) division Efficient Evaluation is:

Subclasses C in comparison step (4) _sub1, C _sub2interior number of samples and smallest sample number N in the rear class of division _minsize, if sample number to be all greater than after division smallest sample number N in class in two subclasses _min, then division effectively, to two subclasses C in step (4) _sub1, C _sub2repeat step (3) ~ (5) until convergence, if sample number has one to be less than smallest sample number N in the rear class of division in subclass _minor reach maximum division number of times K, then illustrate that this classification restrains, division terminates.Division terminate after, obtain a component split after subclass.

6. multicenter approximating method in a kind of classification hyperspectral imagery as claimed in claim 1, is characterized in that, described step (6) calculates the matching center c of the rear each subclass of division _ij, i=1,2 ..., C, C are classification numbers, and j is the subclass number of the i-th classification division, and the computing method at the matching center of each subclass are the average calculating all samples in each subclass.