CN104408478A - Hyperspectral image classification method based on hierarchical sparse discriminant feature learning - Google Patents

Abstract

The invention particularly discloses a hyperspectral image classification method based on hierarchical sparse discriminant feature learning. The hyperspectral image classification method based on the hierarchical sparse discriminant feature learning disclosed by the invention is mainly used for solving the problem of being incapable of well learning feature representation of a hyperspectral data neighbourhood block in the prior art. The hyperspectral image classification method disclosed by the invention comprises the following realization steps of: inputting hyperspectral image data sample sets, and selecting a training set and a testing set from the hyperspectral image data sample sets; based on the selected training set and sampling set, obtaining a first-layer discriminant feature and a second-layer discriminant feature by utilizing a hierarchical discriminant feature learning method based on sparse coding; combining the first-layer discriminant feature with the second-layer discriminant feature to obtain a hierarchical discriminant feature; and, based on the hierarchical discriminant feature, classifying by utilizing a support vector machine, and outputting a classification result. On the basis of a spatial pyramid sparse coding model, discriminant dictionary leaning of classification identifier supervisory information is added; furthermore, the spatial pyramid sparse coding model is subjected to two-layer discriminant feature leaning; therefore, the feature discriminant is increased; the classification precision is increased; and thus, classification of hyperspectral data is more accurate.

Description

A kind of hyperspectral image classification method based on the sparse differentiation feature learning of layering

Technical field

The invention belongs to technical field of image processing, relate to machine learning and Hyperspectral imagery processing, specifically a kind of hyperspectral image classification method based on the sparse differentiation feature learning of layering, the present invention can by carrying out differentiation feature learning to high-spectral data, the appropriate feature symbolizing the different atural object of high spectrum image, thus realize on this basis computing machine autonomous Classification and Identification is carried out for the different atural object of high spectrum image.

Background technology

The terrain classification of high spectrum image is the study hotspot in current Hyperspectral imagery processing field, and its research is mainly devoted to find learn and identify different images target technical method with making computer intelligence.High spectrum image has higher spectral resolution, usually can reach 10 ^{-2 λ}the order of magnitude, wave band is many simultaneously, spectrum channel number nearly tens of more than even hundreds of, and each interchannel continuous print often.The terrain classification of high spectrum image all has applications well prospect in fields such as geologic examination, crops disaster monitoring, atmospheric pollution and military target strikes.High spectrum image terrain classification method the most general is normally: (1) inputs a panel height spectrum picture; (2) training sample and test sample book is therefrom chosen; (3) by the method for feature learning respectively to training sample and test sample book learning characteristic; (4) learned feature is classified by sorter; (5) classification results is obtained.One of them key issue is exactly how to extract useful information from a large amount of with the high-spectral data of redundancy, uses suitable feature learning method to symbolize the expression of different atural object, because the whether reasonable UPS upper performance score determining subsequent classification represented.In addition, because EO-1 hyperion has the unfavorable factors such as data volume is large, redundant information is many, wave band is many, therefore require efficient to the technical method used during high-spectral data feature learning, simple and have certain anti-noise jamming ability.

The people such as Jianchao Yang are at paper " Linear Spatial Pyramid Matching Using Sparse Coding for Image Classification " (CVPR, 2009) utilize in and carry out spatial pyramid maximum pond feature coding based on the method for Sparse Coding to original hyperspectral image data, last combining classification device is classified.The concrete steps of the method are the 1st step: extract sample SIFT feature; 2nd step: training dictionary; 3rd step: carry out coding according to dictionary to SIFT feature and obtain sparse coding vector, does to sparse coding vector the final feature that maximum pond algorithm obtains each sample; 4th step: the linear support vector machine method of final feature is classified.Although this method is relatively accurate to feature coding, the weak point still existed is, the method compares the quality depending on sparse coding.

Summary of the invention

The present invention is directed to above-mentioned the deficiencies in the prior art, propose a kind of layering based on sparse coding newly and differentiate feature learning method, class mark information is added when carrying out sparse coding to hyperspectral image data, structural information is added in the process of layered characteristic study, make terrain classification feature have more identification, thus improve the Intelligent Recognition ability to the different atural object of high-spectral data image further.

Technical scheme of the present invention is: a kind of hyperspectral image classification method based on the sparse differentiation feature learning of layering, comprises the following steps:

(1) input comprises the high-spectrum remote sensing data of C class atural object, and each pixel is sample, sample is used spectral signature vector representation, and the intrinsic dimensionality of sample is h, all composition of sample sample sets wherein y _ibe i-th sample, N is the total number of sample, and R represents real number field;

(2) from every class sample set, the sample of 10% is selected as training set at random n ₁represent training set number of samples, remaining 90% sample is as test set n ₂represent test set number of samples;

(3) based on training set Y _trainwith sample set Y, utilize the layering based on sparse coding to differentiate feature learning method, obtain ground floor and differentiate feature set and the second layer differentiates feature set wherein, for the ground floor corresponding to sample set Y i-th sample differentiates feature, for the second layer corresponding to sample set Y i-th sample differentiates feature:

3a) random selecting K from training set ₁individual training sample differentiates the initialization dictionary of dictionary as ground floor utilize and differentiate K-SVD dictionary learning method, obtain ground floor and differentiate dictionary D;

3b) differentiate dictionary D based on ground floor, utilize orthogonal matching pursuit algorithm to obtain the ground floor sparse coding feature of all samples $z=[z_1,z_2,...,z_N]\∈R^{K_1\×N};$

3c) according to the ground floor sparse coding feature of all samples, utilize ground floor to differentiate feature learning method, obtain ground floor and differentiate feature set $\hat{F}={\left\{{\hat{F}}_i\right\}}_{i=1}^N\∈R^{20K_1\×N}$ And second layer input feature vector collection ${\hat{D}}_i\∈R^{5K_1\×4},i=\mathrm{1,2},...,N;$

3d) concentrate random selecting K from the second layer input feature vector that training set is corresponding ₂the individual initialization dictionary D ' differentiating dictionary as the second layer ₂, in conjunction with corresponding class mark matrix and discrimination matrix, be similar to ground floor and differentiate that the optimization of dictionary learning method differentiates dictionary objective function, obtain the second layer and differentiate dictionary

3e) based on second layer input feature vector collection and the second layer differentiation dictionary of sample set Y, orthogonal matching pursuit algorithm is utilized to obtain the second layer sparse coding feature of each sample i=1,2 ..., N, to the second layer sparse coding characteristic use maximum pond algorithm of all samples, obtains the second layer and differentiates feature set

(4) merge ground floor and differentiate feature set feature set is differentiated with the second layer the layering obtaining sample set Y differentiates feature set F, $F=[\hat{F};\hat{\hat{F}}];$

(5) training set and layering corresponding to test set are differentiated that feature set is input to supporting vector machine, obtain the tag along sort vector of test set, such label vector is the classification results of this high spectrum image.

Above-mentioned steps 3a) in differentiate that the concrete steps of K-SVD dictionary learning method are:

1st step, based on training set Y _train, differentiate that the objective function of K-SVD dictionary learning method is as follows:

$\begin{array}{c}\arg \underset{D,W,A,X}{\min }{\left|\right|Y_{\mathrm{train}}-\mathrm{DX}\left|\right|}_2^2+\mathrm{\α}{\left|\right|Q-\mathrm{AX}\left|\right|}_2^2+\mathrm{\β}{\left|\right|H-\mathrm{WX}\left|\right|}_2^2\\ s.t.\∀i,{\left|\right|x_i\left|\right|}_1\≤\mathrm{\ϵ}\end{array}$

Wherein, above-mentioned formula Section 1 is reconstruct error term, and Section 2 is for differentiating sparse coding bound term, and Section 3 is error in classification item, and D represents that ground floor differentiates dictionary, comprises K ₁individual dictionary atom, each atom dimension is d, W presentation class transformation matrix, and A represents the matrix of a linear transformation, and X represents sparse coding matrix of coefficients, represent l ₂the quadratic sum of norm, α and β represents that balanced class mark differentiates the regular parameter of item and error in classification item, and span is 1 ~ 5, represent differentiation sparse coding matrix of coefficients ideally, if a kth dictionary atom and training sample set Y in D _trainin i-th sample when belonging to same class, then Qki value is 1, is 0 during inhomogeneity, represent the class mark matrix of training sample, if Y _trainin i-th sample belong to c (c=1,2 ..., C) and class, H _cibe 1, otherwise be 0, x _irepresent i-th column vector of sparse coding matrix of coefficients X, || || ₁represent l ₁norm, ε is 10 of definition ^-6;

2nd step, in order to solve the objective function differentiating K-SVD dictionary learning method, is rewritten as:

$\begin{array}{c}\arg \underset{D_{\mathrm{nwe}},X}{\min }\left\{{\left|\right|Y_{\mathrm{new}}-D_{\mathrm{new}}X\left|\right|}_2^2\right\}\\ s.t.\∀i,{\left|\right|x_i\left|\right|}_1\≤\mathrm{\ϵ}\end{array}$

Wherein, $Y_{\mathrm{new}}={(Y_{\mathrm{train}}^T,\sqrt{\mathrm{\α}}{Q}^T,\sqrt{\mathrm{\β}}{H}^T)}^T,D_{\mathrm{new}}={(D^T,\sqrt{\mathrm{\α}}{A}^T,\sqrt{\mathrm{\β}}{W}^T)}^T,$ () ^tthe transposition of representing matrix, utilizes K-SVD dictionary learning method to solve to this objective function, thus obtains ground floor differentiation dictionary D.

Above-mentioned steps 3b) in the concrete steps of orthogonal matching pursuit algorithm be:

1st step, differentiate dictionary D based on ground floor, the objective optimization function of orthogonal matching pursuit algorithm is as follows:

$\begin{array}{cc}\underset{z_i}{\min }{\left|\right|y_i-D{z}_i\left|\right|}_2^2& s.t.{\left|\right|z_i\left|\right|}_1\≤\mathrm{\δ},i=\mathrm{1,2},...,N\end{array}$

Wherein, y _irepresent i-th sample of sample set Y, z _irepresent y _isparse coding coefficient, δ be definition 10 ^-6;

2nd step, structure residual error item, residual error item is configured to r ⁽⁰⁾=y _i, i=1,2 ... N, indexed set Λ ⁰for K ties up null vector, initializing variable J=1;

3rd step, finds out residual error r ^(J-1)with the jth row d in dictionary D _jthe maximum corresponding subscript λ of inner product, namely $\mathrm{\λ}=\arg {\max }_{j=\mathrm{1,2},...,T_0}|\⟨r^{(J-1)},d_j\⟩|;$

4th step, upgrades indexed set Λ ^(J), Λ ^(J)(J)=λ; The set D that dictionary atom row selected by renewal are formed ^(J)=D (:, Λ ^(J)(1:J)), obtain by least square method that J rank approach new residual error r ^(J)=y _i-D ^(J)z _i, J=J+1;

5th step, judges whether that iteration terminates: if J≤K and still have y _inot as residual error item, then return the 2nd step, otherwise, if J≤K and y _i, i=1,2 ... N, as residual error item then EOP (end of program), if J > is K, then turns back to the 3rd step and continues to perform.

Above-mentioned steps 3c) in ground floor differentiate that the concrete steps of feature learning method are:

1st step, with the sparse coding feature z of each sample _i, i=1,2 ..., centered by N, get the sparse coding structural feature sparse coding block Z that neighborhood window size is all samples in (2m+1) × (2m+1) _i, i=1,2 ..., N, Z _ifor (2m+1) × (2m+1) × K ₁a three-dimensional matrice;

2nd step, to the sparse coding block Z of each sample _icarry out piecemeal, utilize the moving window of (m+1) × (m+1), drawing window step-length is m, from top to bottom, from left to right travels through Z _i, extract sparse coding successively and represent sub-block Z _i ⁽¹⁾, Z _i ⁽²⁾, Z _i ⁽³⁾and Z _i ⁽⁴⁾, 4 sub-blocks altogether, the scale of each sub-block is (m+1) × (m+1) × K ₁;

3rd step, carries out spatial pyramid maximum pond algorithm to 4 sub-blocks obtained successively

$\mathrm{SM}\left(Z_i^{\left(j\right)}\right)=[M\left({\left(Z_i^{\left(j\right)}\right)}_1^1\right),...,M\left({\left(Z_i^{\left(j\right)}\right)}_1^{V_1}\right),...,M\left({\left(Z_i^{\left(j\right)}\right)}_u^{V_u}\right),...,M\left({\left(Z_i^{\left(j\right)}\right)}_U^{V_U}\right)],j=\mathrm{1,2,3,4}$

Wherein, SM () represents that carrying out the maximum pondization of spatial pyramid operates, u represents spatial pyramid Decomposition order, V _ube the total number of all pieces being positioned at spatial pyramid u layer, M () represents maximum pond algorithm, $M\left({\left(Z_i^{\left(j\right)}\right)}_u^{V_u}\right)=\left[\underset{t\∈Z_i^{\left(j\right)}}{\max }\right|z_{1t}|,...,\underset{t\∈Z_i^{\left(j\right)}}{\max }|z_{K_{1}t}\left|\right];$

4th step, by the mode of row matrix combination ${\hat{F}}_i=[\mathrm{SM}\left(Z_i^{\left(1\right)}\right);\mathrm{SM}\left(Z_i^{\left(2\right)}\right);\mathrm{SM}\left(Z_i^{\left(3\right)}\right);\mathrm{SM}\left(Z_i^{\left(4\right)}\right)]$ The ground floor obtaining i-th sample differentiates feature by the mode of rectangular array combination ${\hat{D}}_i=[\mathrm{SM}\left(Z_i^{\left(1\right)}\right),\mathrm{SM}\left(Z_i^{\left(2\right)}\right),\mathrm{SM}\left(Z_i^{\left(3\right)}\right),\mathrm{SM}\left(Z_i^{\left(4\right)}\right)]$ Obtain the second layer input feature vector of i-th sample.

Beneficial effect of the present invention: the present invention inputs hyperspectral image data, the training sample of a part for random selecting is utilized initially to differentiate dictionary as one deck, through differentiating that dictionary learning obtains one deck and differentiates dictionary, the sparse coding solving the neighborhood block of each high-spectral data according to one deck differentiation dictionary obtained represents coefficient, through pyramid maximum pond method, obtain two layers and initially differentiate dictionary and one deck coding characteristic, recycle two layers and initially differentiate that dictionary is through differentiating that dictionary learning algorithm obtains two layers and differentiates dictionary, coefficient is represented according to the sparse coding that the two layers of differentiation dictionary obtained solve second layer feature coding corresponding region block, through pyramid maximum pond method, obtain two layers of coding characteristic, one deck coding characteristic is combined with two layers of coding characteristic, as the feature finally learning to obtain, this characteristic use sorter is classified, thus reach the object of EO-1 hyperion terrain classification, and achieve higher terrain classification precision.The present invention compared with prior art, has the following advantages:

First, the present invention utilizes the method differentiating dictionary learning, when ground floor dictionary learning and second layer dictionary learning, consider class mark information, overcome the deficiency that traditional K-SVD dictionary learning does not make full use of class mark information, the dictionary making the present invention learn to obtain and had more the advantage of identification by the sparse coding coefficient that this dictionary learning obtains.

The second, the present invention utilizes the method for multilayer sparse coding feature learning, overcomes tradition and uses the shortcoming that individual layer sparse coding coefficient directly carries out classifying and nicety of grading is lower, make the present invention have the high advantage of nicety of grading.

3rd, the feature learning method that the present invention utilizes empty spectral domain to combine, overcomes the deficiency not considering surrounding neighbors information of carrying out the algorithm of feature learning with a pixel, makes the present invention have the better advantage of feature robustness obtained study.

Below with reference to accompanying drawing, the present invention is described in further details.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the inventive method;

Fig. 2 is the image of Indianan Pine in emulation experiment of the present invention.

Concrete implementing measure

Below in conjunction with accompanying drawing, invention is described further.

1 concrete steps of the present invention are described below by reference to the accompanying drawings:

Step 1, input comprises the high-spectrum remote sensing data of C class atural object, and each pixel and the vector representation of sample spectral signature, the intrinsic dimensionality of sample is h, all composition of sample sample sets wherein y _ibe i-th sample, N is the total number of sample, and R represents real number field;

Step 2, in this N number of sample, gets rid of background sample point, selects the sample of 10% as training set at random from every class sample set n1 represents training set number of samples, and remaining 90% sample is as test set n ₂represent test set number of samples;

Step 3, based on training set and sample set, utilizes the layering based on sparse coding to differentiate feature learning method, obtains ground floor and differentiate feature set and the second layer differentiates feature set wherein, for the ground floor corresponding to sample set i-th sample differentiates feature, for i-th second layer corresponding to sample set i-th sample differentiates feature:

The first step, from all kinds of training set, a Stochastic choice part, chooses K altogether ₁individual training sample differentiates the initialization dictionary of dictionary as ground floor utilize and differentiate K-SVD dictionary learning method, obtain ground floor and differentiate dictionary D, differentiate that the objective function of K-SVD dictionary learning method is as follows:

$\begin{array}{c}\arg \underset{D,W,A,X}{\min }{\left|\right|Y_{\mathrm{train}}-\mathrm{DX}\left|\right|}_2^2+\mathrm{\α}{\left|\right|Q-\mathrm{AX}\left|\right|}_2^2+\mathrm{\β}{\left|\right|H-\mathrm{WX}\left|\right|}_2^2\\ s.t.\∀i,{\left|\right|x_i\left|\right|}_1\≤\mathrm{\ϵ}\end{array}$

Wherein, Section 1 is reconstruct error term, and Section 2 is for differentiating sparse coding bound term, and Section 3 is error in classification item, and D represents that ground floor differentiates dictionary, comprises K ₁individual dictionary atom, each atom dimension is d, W presentation class transformation matrix, and A represents the matrix of a linear transformation, and X represents sparse coding matrix of coefficients, represent l ₂the quadratic sum of norm, α and β represents that balanced class mark differentiates the regular parameter of item and error in classification item, and span is 1 ~ 5, represent differentiation sparse coding matrix of coefficients ideally, if a kth dictionary atom and training sample set Y in D _trainin i-th sample when belonging to same class, then Q _kivalue is 1, is 0 during inhomogeneity, represent the class mark matrix of training sample, if Y _trainin i-th sample belong to c (c=1,2 ..., C) and class, H _cibe 1, otherwise be the i-th column vector that 0, xi represents sparse coding matrix of coefficients X, || || ₁represent l ₁norm, ε is 10 of definition ^-6;

In order to solve the objective function differentiating K-SVD dictionary learning method, be rewritten as:

$\begin{array}{c}\arg \underset{D_{\mathrm{nwe}},X}{\min }\left\{{\left|\right|Y_{\mathrm{new}}-D_{\mathrm{new}}X\left|\right|}_2^2\right\}\\ s.t.\∀i,{\left|\right|x_i\left|\right|}_1\≤\mathrm{\ϵ}\end{array}$

Wherein, $Y_{\mathrm{new}}={(Y_{\mathrm{train}}^T,\sqrt{\mathrm{\α}}{Q}^T,\sqrt{\mathrm{\β}}{H}^T)}^T,D_{\mathrm{new}}={(D^T,\sqrt{\mathrm{\α}}{A}^T,\sqrt{\mathrm{\β}}{W}^T)}^T,$ () ^tthe transposition of representing matrix, utilizes K-SVD dictionary learning method to solve to this objective function, thus obtains ground floor differentiation dictionary D;

${\left|\right|Y_{\mathrm{new}}-D_{\mathrm{new}}X\left|\right|}_2^2=\left|\right|Y_{\mathrm{new}}-\underset{j=1}{\overset{L}{\mathrm{\Σ}}}{d}_j{x}_j^T\left|\right|={\left|\right|(Y_{\mathrm{new}}-\underset{j\≠k}{\overset{L}{\mathrm{\Σ}}}{d}_j{x}_j^T)-d_k{x}_k^T\left|\right|}_2^2={\left|\right|E_k-d_k{x}_k^T\left|\right|}_2^2$

Wherein, d _jrepresent D _newjth row atom, represent the jth row of X, L represents D _newtotal columns, d _krepresent D _newkth row atom, represent the row k of X, E _krepresent and do not use D _newkth row atom d _kcarry out the error matrix that Its Sparse Decomposition produces;

Wherein K-SVD dictionary learning method is as follows:

1. pair differentiation dictionary objective function be out of shape, by Y _newuse vector form E _krepresent, by D _newuse vector form d _krepresent, X is used vector form represent, to the formula of gained after distortion be multiplied by Ω _k, obtain goal decomposition formula:

${\left|\right|E_k{\mathrm{\Ω}}_k-d_x{x}_k^T{\mathrm{\Ω}}_k\left|\right|}_2^2={\left|\right|E_k^R-d_k{x}_k^R\left|\right|}_2^2$

Wherein distortion inaccuracy matrix represent error matrix E _kdistortion, Ω _ksize be P × | ω _k|, P represents training sample set Y _newcolumns, | ω _k| represent ω _kmodulus value, and Ω _kat (ω _k(j), j) place is 1, other places are 0, wherein 1≤j entirely≤| ω _k|, ω _kj () represents ω _kjth number;

2. pair gained goal decomposition formula in distortion inaccuracy matrix carry out SVD decomposition to obtain wherein U represents left singular matrix, V ^Τrepresent right singular matrix, Δ represents singular value matrix;

3. remove more fresh target train word allusion quotation D with the first row of gained left singular matrix U _newkth row atom d _k;

4. repeat step 1 to step 3 couple D _newin all atoms carry out update process, obtain the dictionary D that K is new ₁', D ₂' ... D _k'.

Second step, differentiates dictionary D based on ground floor, utilizes orthogonal matching pursuit algorithm to solve following objective function, obtain the ground floor coding characteristic of all samples

$\begin{array}{cc}\underset{z_i}{\min }{\left|\right|y_i-D{z}_i\left|\right|}_2^2& s.t.{\left|\right|z_i\left|\right|}_1\≤\mathrm{\δ},i=\mathrm{1,2},...,N\end{array}$

Wherein, y _irepresent i-th sample of sample set Y, z _irepresent y _isparse coding coefficient, δ be definition 10 ^-6, orthogonal matching pursuit algorithm is as follows:

First construct residual error item, residual error item is configured to r ⁽⁰⁾=y _i, indexed set Λ ⁰for K ties up null vector, initializing variable J=1;

Then circulation performs following steps 1-5

1. find out residual error r ^(J-1)with the jth row d in dictionary D _jthe maximum corresponding subscript λ of inner product, namely $\mathrm{\λ}=\arg {\max }_{j=\mathrm{1,2},...,T_0}|\⟨r^{(J-1)},d_j\⟩|;$

2. upgrade indexed set Λ ^(J), Λ ^(J)(J)=λ.The set D that dictionary atom row selected by renewal are formed ^(J)=D (:, Λ ^(J)(1:J));

3. utilizing least square method to obtain, J rank approach

4. upgrade residual error r ^(J)=y _i-D ^(J)z _i, J=J+1;

5. judge whether that iteration terminates.If J > is K, then terminate, otherwise continue 1.

3rd step, according to the ground floor sparse coding feature of all samples, utilizes ground floor to differentiate feature learning method, obtains ground floor and differentiate feature set $\hat{F}={\left\{{\hat{F}}_i\right\}}_{i=1}^N\∈R^{20K_1\×N}$ And second layer input feature vector collection ${\hat{D}}_i\∈R^{5K_1\×4},i=\mathrm{1,2},...,N;$ Wherein ground floor differentiates that feature learning method is as follows:

1. with the sparse coding feature z of each sample _icentered by, getting neighborhood window size is (2m+1) × (2m+1), m=1,2 ..., the sparse coding feature construction of each sample is become sparse coding and represents block Z _i, i=1,2 ..., N, namely a scale is (2m+1) × (2m+1) × K ₁three-dimensional matrice;

2. the sparse coding of pair each sample represents block Z _icarry out piecemeal, utilize the moving window of (m+1) × (m+1), drawing window step-length is m, and from top to bottom, the sparse coding from left to right traveling through each sample represents block, extracts sparse coding successively and represents sub-block Z _i ⁽¹⁾, Z _i ⁽²⁾, Z _i ⁽³⁾and Z _i ⁽⁴⁾, 4 sub-blocks altogether, the scale of each sub-block is (m+1) × (m+1) × K ₁;

3. successively spatial pyramid maximum pond algorithm is carried out to 4 sub-blocks obtained

$\mathrm{SM}\left(Z_i^{\left(j\right)}\right)=[M\left({\left(Z_i^{\left(j\right)}\right)}_1^1\right),...,M\left({\left(Z_i^{\left(j\right)}\right)}_1^{V_1}\right),...,M\left({\left(Z_i^{\left(j\right)}\right)}_u^{V_u}\right),...,M\left({\left(Z_i^{\left(j\right)}\right)}_U^{V_U}\right)],j=\mathrm{1,2,3,4}$

Wherein, SM () represents that carrying out the maximum pondization of spatial pyramid operates, u represents spatial pyramid Decomposition order, V _ube the total number of all pieces being positioned at spatial pyramid u layer, M () represents maximum pond algorithm, $M\left({\left(Z_i^{\left(j\right)}\right)}_u^{V_u}\right)=\left[\underset{t\∈Z_i^{\left(j\right)}}{\max }\right|z_{1t}|,...,\underset{t\∈Z_i^{\left(j\right)}}{\max }|z_{K_{1}t}\left|\right];$

4. press the mode of row matrix combination ${\hat{F}}_i=[\mathrm{SM}\left(Z_i^{\left(1\right)}\right);\mathrm{SM}\left(Z_i^{\left(2\right)}\right);\mathrm{SM}\left(Z_i^{\left(3\right)}\right);\mathrm{SM}\left(Z_i^{\left(4\right)}\right)]$ The ground floor obtaining i-th sample differentiates feature by the mode of rectangular array combination ${\hat{D}}_i=[\mathrm{SM}\left(Z_i^{\left(1\right)}\right),\mathrm{SM}\left(Z_i^{\left(2\right)}\right),\mathrm{SM}\left(Z_i^{\left(3\right)}\right),\mathrm{SM}\left(Z_i^{\left(4\right)}\right)]$ Obtain the second layer input feature vector of i-th sample.

4th step, from training set obtain the second layer input feature vector concentrate select a part at random, altogether choose K ₂the individual initialization dictionary D ' differentiating dictionary as the second layer ₂, in conjunction with corresponding class mark matrix and discrimination matrix, be similar to ground floor and differentiate that dictionary construction method is by differentiating that dictionary objective function can obtain two layers and differentiate dictionary

5th step, layer 2-based input feature vector and the second layer differentiate dictionary, utilize orthogonal matching pursuit algorithm to obtain the second layer sparse coding feature of each sample wherein j=1,2,3,4 correspond to the input feature vector of i-th second layer the jth row second layer sparse coding feature obtained, to the second layer sparse coding characteristic use maximum pond algorithm of all samples, obtains the second layer and differentiates feature set

Step 4, differentiates feature set by the ground floor of all samples and the second layer differentiates feature set in conjunction with, obtain layering and differentiate feature set F

$F=[\hat{F};\hat{\hat{F}}]$

By training set and layering corresponding to test set, step 5, differentiates that feature set is input to supporting vector machine, obtain the tag along sort vector of test set, such label vector is the classification results of this high spectrum image.

Below in conjunction with accompanying drawing 2, effect of the present invention is described further.

Emulation of the present invention is that the high spectrum image Indiana Pine that obtains in June, 1992 in the northwestward, Indiana at the AVIRIS of representative NASA NASA carries out, Indiana Pine image size is 145 × 145 pixels, 220 wave bands are comprised in image, 20 wave bands removed by waters absorbs remain 200 wave bands, and this image comprises 16 class atural objects as shown in table 1 altogether.

Emulation experiment of the present invention is at AMDA4-3400APU, dominant frequency 2.69GHz, and the MATLAB2011a on internal memory 4G, Windows732 bit platform realizes.

16 class data in table 1 Indiana Pine image

Classification	Item name	Number of samples	Training sample number
				1	Clover	46	4
2	Corn-do not plough plough	1428	142
				3	Corn-irrigation	830	83
4	Corn	237	23
				5	Herbage	483	48
6	Trees	730	73
				7	The herbage of cutting	28	2
8	Hay stockpile	478	47
				9	Buckwheat	20	2
10	Soybean-do not plough plough	972	97
				11	Soybean-irrigation	2455	245
12	Soya bean	593	59
				13	Wheat	205	20
14	The woods	1265	126
				15	Buildings-grass-tree	386	38
16	Stone-reinforcing bar	93	9

2. emulate content and analysis

The present invention and existing three kinds of methods are used to classify to high spectrum image, existing three kinds of methods are respectively: supporting vector machine SVM, based on the sorting technique SRC of rarefaction representation, based on spatial pyramid coupling sorting technique SCSPM, the wherein penalty factor of SVM method of rarefaction representation nuclear parameter determined by 5 times of cross validations, the regular terms parameter lambda of SRC method is set to 0.1, the Sparse parameter of SRC method and SCSPM method is set to 20, SCSPM method and spatial domain of the present invention scale parameter are set to 7 × 7, from 16 class data, every class gets the pixel of 10% at random as training sample, remaining 90% conduct test, carry out 5 experiments and be averaged, then the experimental precision of three kinds of methods experiment precision and this method is as shown in the table:

The existing three kinds of methods of table 2 and experimental precision result of the present invention

Method	Nicety of grading
		SVM	89.23％
SRC	83.70％
		SCSPM	92.34％
Method of the present invention	96.54％

As can be seen from Table 2, method of the present invention shows optimum in nicety of grading, methodology acquistion of the present invention to the nicety of grading that obtains through SVM classifier of feature more direct than SVM high to the classify precision that obtains of raw data, illustrate that the feature that the present invention learns to obtain is more suitable for SVM classifier, reflect that the feature learning to obtain is effective from the side; It is more effective that method of the present invention learns through the aspect ratio SCSPM that two-layer dictionary learning and sparse coding obtain the feature that obtains, is more applicable to SVM classifier, thus describe the present invention and have obvious advantage compared with the existing methods.

To sum up, the layering that the present invention is based on sparse coding differentiates that feature learning method carries out classification hyperspectral imagery, make full use of sparse characteristic and the spatial domain contextual information of high spectrum image, can classify more accurately to original high spectrum image, with the contrast of existing three kinds of image classification methods after, describe accuracy of the present invention and validity.Compared with prior art, have the following advantages:

First, the present invention utilizes the method differentiating dictionary learning, when ground floor dictionary learning and second layer dictionary learning, consider class mark information, overcome the deficiency that traditional KSVD dictionary learning does not make full use of class mark information, the dictionary making the present invention learn to obtain and had more the advantage of identification by the sparse coding coefficient that this dictionary learning obtains.

The second, the present invention utilizes the method for multilayer sparse coding feature learning, overcomes tradition and uses the shortcoming that individual layer sparse coding coefficient directly carries out classifying and nicety of grading is lower, make the present invention have the high advantage of nicety of grading.

3rd, the feature learning method that the present invention utilizes empty spectral domain to combine, overcomes the deficiency not considering surrounding neighbors information of carrying out the algorithm of feature learning with a pixel, makes the present invention have the better advantage of feature robustness obtained study.

The part do not described in detail in present embodiment belongs to the known conventional means of the industry, does not describe one by one here.More than exemplifying is only illustrate of the present invention, does not form the restriction to protection scope of the present invention, everyly all belongs within protection scope of the present invention with the same or analogous design of the present invention.

Claims (4)

1., based on a hyperspectral image classification method for the sparse differentiation feature learning of layering, it is characterized in that, comprise the following steps:

(1) input comprises the high-spectrum remote sensing data of C class atural object, and each pixel is sample, sample is used spectral signature vector representation, and the intrinsic dimensionality of sample is h, all composition of sample sample sets wherein y _ibe i-th sample, N is the total number of sample, and R represents real number field;

(2) from every class sample set, the sample of 10% is selected as training set at random n ₁represent training set number of samples, remaining 90% sample is as test set n ₂represent test set number of samples;

(3) based on training set Y _trainwith sample set Y, utilize the layering based on sparse coding to differentiate feature learning method, obtain ground floor and differentiate feature set and the second layer differentiates feature set wherein, for the ground floor corresponding to sample set Y i-th sample differentiates feature, for the second layer corresponding to sample set Y i-th sample differentiates feature:

3a) random selecting K from training set ₁individual training sample differentiates the initialization dictionary of dictionary as ground floor utilize and differentiate K-SVD dictionary learning method, obtain ground floor and differentiate dictionary D;

3b) differentiate dictionary D based on ground floor, utilize orthogonal matching pursuit algorithm to obtain the ground floor sparse coding feature of all samples $z=[z_1,z_2,\·\·\·,z_N]\∈R^{K_1\×N};$

3c) according to the ground floor sparse coding feature of all samples, utilize ground floor to differentiate feature learning method, obtain ground floor and differentiate feature set and second layer input feature vector collection ${\hat{D}}_i\∈R^{5K_1\×4},i=\mathrm{1,2},\·\·\·,N;$

3d) concentrate random selecting K from the second layer input feature vector that training set is corresponding ₂the individual initialization dictionary D ' differentiating dictionary as the second layer ₂, in conjunction with corresponding class mark matrix and discrimination matrix, be similar to ground floor and differentiate that the optimization of dictionary learning method differentiates dictionary objective function, obtain the second layer and differentiate dictionary

3e) based on second layer input feature vector collection and the second layer differentiation dictionary of sample set Y, orthogonal matching pursuit algorithm is utilized to obtain the second layer sparse coding feature of each sample i=1,2 ..., N, to the second layer sparse coding characteristic use maximum pond algorithm of all samples, obtains the second layer and differentiates feature set

(4) merge ground floor and differentiate feature set feature set is differentiated with the second layer the layering obtaining sample set Y differentiates feature set F,

(5) training set and layering corresponding to test set are differentiated that feature set is input to supporting vector machine, obtain the tag along sort vector of test set, such label vector is the classification results of this high spectrum image.

2. a kind of hyperspectral image classification method based on the sparse differentiation feature learning of layering according to claim 1, is characterized in that, described step 3a) in differentiate that the concrete steps of K-SVD dictionary learning method are:

1st step, based on training set Y _train, differentiate that the objective function of K-SVD dictionary learning method is as follows:

$\arg \underset{D,W,A,X}{\min }{\left|\right|Y_{\mathrm{train}}-\mathrm{DX}\left|\right|}_2^2+\mathrm{\α}{\left|\right|Q-\mathrm{AX}\left|\right|}_2^2+\mathrm{\β}{\left|\right|H-\mathrm{WX}\left|\right|}_2^2$

$s.t.\∀i,{\left|\right|x_i\left|\right|}_1\≤\mathrm{\ϵ}$

Wherein, above-mentioned formula Section 1 is reconstruct error term, and Section 2 is for differentiating sparse coding bound term, and Section 3 is error in classification item, and D represents that ground floor differentiates dictionary, comprises K ₁individual dictionary atom, each atom dimension is d, W presentation class transformation matrix, and A represents the matrix of a linear transformation, and X represents sparse coding matrix of coefficients, represent l ₂the quadratic sum of norm, α and β represents that balanced class mark differentiates the regular parameter of item and error in classification item, and span is 1 ~ 5, represent differentiation sparse coding matrix of coefficients ideally, if a kth dictionary atom and training sample set Y in D _trainin i-th sample when belonging to same class, then Q _kivalue is 1, is 0 during inhomogeneity, represent the class mark matrix of training sample, if Y _trainin i-th sample belong to c (c=1,2 ..., C) and class, H _cibe 1, otherwise be 0, x _irepresent i-th column vector of sparse coding matrix of coefficients X, || || ₁represent l ₁norm, ε is 10 of definition ^-6;

2nd step, in order to solve the objective function differentiating K-SVD dictionary learning method, is rewritten as:

$\arg \underset{D_{\mathrm{new}},X}{\min }\left\{{\left|\right|Y_{\mathrm{new}}-D_{\mathrm{new}}X\left|\right|}_2^2\right\}$

$s.t.\∀i,{\left|\right|x_i\left|\right|}_1\≤\mathrm{\ϵ}$

Wherein, $Y_{\mathrm{new}}={(Y_{\mathrm{train}}^T,\sqrt{\mathrm{\α}}{Q}^T,\sqrt{\mathrm{\β}}{H}^T)}^T,$ $D_{\mathrm{new}}={(D^T,\sqrt{\mathrm{\α}}{A}^T,\sqrt{\mathrm{\β}}{W}^T)}^T,$ () ^tthe transposition of representing matrix, utilizes K-SVD dictionary learning method to solve to this objective function, thus obtains ground floor differentiation dictionary D.

3. a kind of hyperspectral image classification method based on the sparse differentiation feature learning of layering according to claim 1, is characterized in that, described step 3b) in the concrete steps of orthogonal matching pursuit algorithm be:

1st step, differentiate dictionary D based on ground floor, the objective optimization function of orthogonal matching pursuit algorithm is as follows:

$\underset{z_i}{\min }{\left|\right|y_i-{\mathrm{Dz}}_i\left|\right|}_2^2,s.t.{\left|\right|z_i\left|\right|}_1\≤\mathrm{\δ},i=\mathrm{1,2},...,N$

Wherein, y _irepresent i-th sample of sample set Y, z _irepresent y _isparse coding coefficient, δ be definition 10 ^-6;

2nd step, structure residual error item, residual error item is configured to r ⁽⁰⁾=y _i, i=1,2 ... N, indexed set Λ ⁰for K ties up null vector, initializing variable J=1;

3rd step, finds out residual error r ^(J-1)with the jth row d in dictionary D _jthe maximum corresponding subscript λ of inner product, namely $\mathrm{\&lambda;}={\arg \max }_{j=\mathrm{1,2},...,T_0}|<r^{(J-1)},d_j>|$

4th step, upgrades indexed set Λ ^(J), Λ ^(J)(J)=λ; The set D that dictionary atom row selected by renewal are formed ^(J)=D (:, Λ ^(J)(1:J)), obtain by least square method that J rank approach new residual error r ^(J)=y _i-D ^(J)z _i, J=J+1;

5th step, judges whether that iteration terminates: if J≤K and still have y _inot as residual error item, then return the 2nd step, otherwise, if J≤K and y _i, i=1,2 ... N, as residual error item then EOP (end of program), if J > is K, then turns back to the 3rd step and continues to perform.

4. a kind of hyperspectral image classification method based on the sparse differentiation feature learning of layering according to claim 1, is characterized in that, described step 3c) in ground floor differentiate that the concrete steps of feature learning method are:

1st step, with the sparse coding feature z of each sample _i, i=1,2 ..., centered by N, get the sparse coding structural feature sparse coding block Z that neighborhood window size is all samples in (2m+1) × (2m+1) _i, i=1,2 ..., N, Z _ifor (2m+1) × (2m+1) × K ₁a three-dimensional matrice;

2nd step, to the sparse coding block Z of each sample _icarry out piecemeal, utilize the moving window of (m+1) × (m+1), drawing window step-length is m, from top to bottom, from left to right travels through Z _i, extract sparse coding successively and represent sub-block with 4 sub-blocks altogether, the scale of each sub-block is (m+1) × (m+1) × K ₁;

3rd step, carries out spatial pyramid maximum pond algorithm to 4 sub-blocks obtained successively $\mathrm{SM}\left(Z_i^{\left(j\right)}\right)=[M\left({\left(Z_i^{\left(j\right)}\right)}_1^1\right),\&CenterDot;\&CenterDot;\&CenterDot;,M\left({\left(Z_i^{\left(j\right)}\right)}_1^{V_1}\right),\&CenterDot;\&CenterDot;\&CenterDot;,M\left({\left(Z_i^{\left(j\right)}\right)}_u^{V_u}\right),\&CenterDot;\&CenterDot;\&CenterDot;,M\left({\left(Z_i^{\left(j\right)}\right)}_U^{V_U}\right)],$ j＝1,2,3,4

Wherein, SM () represents that carrying out the maximum pondization of spatial pyramid operates, j=1,2,3,4, U represent spatial pyramid Decomposition order, V _ube the total number of all pieces being positioned at spatial pyramid u layer, M () represents maximum pond algorithm, $M\left({\left(Z_i^{\left(j\right)}\right)}_u^{V_u}\right)=\left[\underset{t\&Element;Z_i^{\left(j\right)}}{\max }\right|z_{1t}|,\&CenterDot;\&CenterDot;\&CenterDot;,\underset{t\&Element;Z_i^{\left(j\right)}}{\max }|z_{K_{1}t}\left|\right];$

4th step, by the mode of row matrix combination ${\hat{F}}_i=[\mathrm{SM}\left(Z_i^{\left(1\right)}\right);\mathrm{SM}\left(Z_i^{\left(2\right)}\right);\mathrm{SM}\left(Z_i^{\left(3\right)}\right);\mathrm{SM}\left(Z_i^{\left(4\right)}\right)]$ The ground floor obtaining i-th sample differentiates feature by the mode of rectangular array combination ${\hat{D}}_i=[\mathrm{SM}\left(Z_i^{\left(1\right)}\right);\mathrm{SM}\left(Z_i^{\left(2\right)}\right);\mathrm{SM}\left(Z_i^{\left(3\right)}\right);\mathrm{SM}\left(Z_i^{\left(4\right)}\right)]$ Obtain the second layer input feature vector of i-th sample.