CN106355212A - Hyperspectral image classification method based on morphology contour characteristics and nonlinear multiple kernel learning - Google Patents

Abstract

The invention relates to a hyperspectral image classification method based on morphology contour characteristics and nonlinear multiple kernel learning and aims at overcoming the defects that the spatial information of a hyperspectral image can not be fully excavated and useful information produced by nonlinear interaction among base kernels is not considered in a hyperspectral image classification method. The hyperspectral image classification method provided by the invention comprises the following concrete steps: firstly, extracting a principal component of the hyperspectral image by utilizing a principal component analysis method, and obtaining multi-structure element morphology contour characteristics expanded by the hyperspectral image on the basis of the principal component; secondly, constructing linear base kernels; thirdly, obtaining a nonlinear combined kernel; fourthly, substituting the nonlinear combined kernel into a support vector machine, and obtaining optimal kernel weight by adopting a gradient descent method; and fifthly, classifying the hyperspectral image. The hyperspectral image classification method provided by the invention is applied to the field of image classification.

Description

Classification hyperspectral imagery based on morphological profiles feature and non-linear Multiple Kernel Learning Method

Technical field

The present invention relates to hyperspectral image classification method.

Background technology

Bloom spectrum sensor obtains the reflected radiation information of atural object by up to a hundred spectrum channels, and its wavelength band covers From visible ray to near-infrared or even LONG WAVE INFRARED region, high spectrum image contains spatial information, reflection or the spoke of atural object simultaneously Penetrate information and spectral information, its characteristic is commonly known as " collection of illustrative plates ".And, hyperspectral image data provides and almost connects Continuous spectrum sample information, can record the reflection differences of atural object spectrally very little.This characteristic is referred to as the diagnosis of atural object Characteristic, can be used as foundation atural object classified and detected.Research classification hyperspectral imagery new technique, has important reason By meaning and using value.In wisdom ocean, geology detecting, land use monitoring, urban planning, border monitoring, disaster management Etc. aspect had important application.

Existing typical hyperspectral image classification method has monokaryon support vector machine method (single kernel Support vector machine, single kernel svm), the Multiple Kernel Learning method (rule-based based on criterion Multiple kernel learning, rbmkl), the warp such as mixing kernel method (combined kernel learning, ckl) Allusion quotation method and proposed in recent years representative Multiple Kernel Learning (represented multiple kernel learning, Rmkl), multiple features study (multiple feature learning, mfl) etc..

Currently, a kind of high spectrum image space-optical spectrum feature extracting method of main flow is using a kind of shape, multiple dimensioned Structural element extract image extension morphological profiles, but the structural element of single shape can not be in high spectrum image Various geometry is detected；Additionally, current Multiple Kernel Learning method is all the linear combination mode using base core, ignore Useful information produced by non-linear interaction between base core, therefore can not obtain the nicety of grading of satisfaction.

Multiple Kernel Learning is a kind of effective means of feature learning in that context it may be convenient to embed feature various in a large number, therefore, such as What selects effective feature is a key issue.By the Multi-structure elements morphological profiles feature of extension and non-linear multinuclear Practise framework to combine, resolving ability produced by the reciprocal action between the resolving ability of feature and base core is obtained for fully Excavate, correspondingly, nicety of grading also can be improved.

Content of the invention

It is an object of the invention to overcome can not be to the spatial information of high spectrum image in hyperspectral image classification method Fully excavated and do not accounted for the deficiency of the useful information that non-linear interaction produces between base core, proposed a kind of base In the hyperspectral image classification method of non-linear Multiple Kernel Learning, extract the spatial information of high spectrum image using Multi-structure elements, Improve nicety of grading.

The invention aims to can not be to the space of high spectrum image in the existing hyperspectral image classification method of solution Information is fully excavated and is not accounted for the problem of the useful information that non-linear interaction produces between base core, and proposes Hyperspectral image classification method based on morphological profiles feature and non-linear Multiple Kernel Learning.

Based on the hyperspectral image classification method detailed process of morphological profiles feature and non-linear Multiple Kernel Learning it is:

Step one: extract the main constituent of high spectrum image using principal component analytical method, obtain high on the basis of main constituent The Multi-structure elements morphological profiles feature of spectrum picture extension, i.e. space-optical spectrum feature；

Step 2: (this is to carry inside high-spectral data by given EO-1 hyperion sample data.It is high-spectrum The one part of pixel of picture, these pixels are given class label, that is, have been formulated classification.), it is utilized respectively step one and obtain Take each of the Multi-structure elements morphological profiles feature of high spectrum image extension opening operation, closed operation and original main constituent structure Build linear base core；

Step 3: each linear base verification answers a core weight, by calculating the Hadamard of any two linear base core Amass and obtain Non-linear Kernel, Non-linear Kernel is weighted, obtains nonlinear combination core；

Step 4: the nonlinear combination nucleus band that step 3 is obtained enters in support vector machine, is obtained using gradient descent method Optimum core weight；

Step 5: high spectrum image is classified using the optimum core weight that step 4 obtains.

The invention has the benefit that

Because this method utilizes Multi-structure elements to extract the space characteristics of high-spectral data, polymorphic space can be believed Breath carries out depth excavation, and using non-linear Multiple Kernel Learning framework, feature is learnt, linear compared to currently employed base core The Multiple Kernel Learning method of combination, overcomes and can not carry out digging utilization by the useful information of internuclear non-linear interaction generation to base Shortcoming, for improve nicety of grading be very helpful.Based on the hyperspectral image classification method of non-linear Multiple Kernel Learning, adopt Extract the spatial information of high spectrum image with Multi-structure elements, improve nicety of grading.

In order to verify the performance of algorithm proposed by the invention, for one group of airborne visible ray/Infrared Imaging Spectrometer One group of farm data that (airborne visible infrared imaging spectrometer, aviris) gathers is carried out Experiment, the experiment show Multi-structure elements morphological profiles feature using extension proposed by the present invention and non-linear many The effectiveness to classification hyperspectral imagery algorithm for the core study.

Compare other typical hyperspectral image classification methods using single structure element morphology contour feature, this method exists 10% about is improve in nicety of grading.

Brief description

Fig. 1 be the present invention realize schematic flow sheet；

Fig. 2 is Indiana data；

Fig. 3 a is the single structure element morphology contour feature of Indiana data separate extension and non-linear Multiple Kernel Learning The classification hyperspectral imagery result figure of method；

Fig. 3 b is unijunction constitutive element (rhombus) the morphological profiles feature of Indiana data separate extension and is based on average The classification hyperspectral imagery result figure of the Multiple Kernel Learning method of criterion；

Fig. 3 c is unijunction constitutive element (rhombus) the morphological profiles feature of Indiana data separate extension and representative is many The classification hyperspectral imagery result figure of kernel learning method；

Fig. 3 d is unijunction constitutive element (rhombus) morphological profiles feature and the mixing multinuclear of Indiana data separate extension The classification hyperspectral imagery result figure of learning method；

Fig. 3 e is unijunction constitutive element (rhombus) morphological profiles feature and the multiple features of Indiana data separate extension The classification hyperspectral imagery result figure of learning method；

Fig. 3 f is that Multi-structure elements (rhombus, the square, circle) morphological profiles of Indiana data separate extension are special Levy and non-linear Multiple Kernel Learning method classification hyperspectral imagery result figure；

Fig. 3 g is that Multi-structure elements (rhombus, the square, circle) morphological profiles of Indiana data separate extension are special Levy and Multiple Kernel Learning method based on average criterion classification hyperspectral imagery result figure；

Fig. 3 h is that Multi-structure elements (rhombus, the square, circle) morphological profiles of Indiana data separate extension are special Levy and representative Multiple Kernel Learning method classification hyperspectral imagery result figure；

Fig. 3 i is that Multi-structure elements (rhombus, the square, circle) morphological profiles of Indiana data separate extension are special Levy and mix the classification hyperspectral imagery result figure of Multiple Kernel Learning method；

Fig. 3 j is that Multi-structure elements (rhombus, the square, circle) morphological profiles of Indiana data separate extension are special Levy and multiple features learning method classification hyperspectral imagery result figure；

Fig. 4 a is the structural element schematic diagram of the yardstick r=1 of rhombus and structural element, and the yardstick of r structural element, for setting Put structural element size；

Fig. 4 b is the structural element schematic diagram of the yardstick r=2 of rhombus and structural element；

Fig. 4 c is the structural element schematic diagram of the yardstick r=3 of rhombus and structural element；

Fig. 4 d is the structural element schematic diagram of the yardstick r=4 of rhombus and structural element；

Fig. 4 e is the structural element schematic diagram of the yardstick r=1 of square and structural element；

Fig. 4 f is the structural element schematic diagram of the yardstick r=2 of square and structural element；

Fig. 4 g is the structural element schematic diagram of the yardstick r=3 of square and structural element；

Fig. 4 h is the structural element schematic diagram of the yardstick r=4 of square and structural element；

Fig. 4 i is the structural element schematic diagram of the yardstick r=5 of square and structural element；

Fig. 4 j is circular and the structural element schematic diagram of the yardstick r=1 of structural element；

Fig. 4 k is circular and the structural element schematic diagram of the yardstick r=2 of structural element；

Fig. 4 l is circular and the structural element schematic diagram of the yardstick r=3 of structural element；

Fig. 4 m is circular and the structural element schematic diagram of the yardstick r=4 of structural element；

Fig. 4 n is circular and the structural element schematic diagram of the yardstick r=5 of structural element.

Specific embodiment

Specific embodiment one: with reference to Fig. 1 present embodiment is described, present embodiment based on morphological profiles feature and The hyperspectral image classification method detailed process of non-linear Multiple Kernel Learning is:

Step one: extract the main constituent of high spectrum image using principal component analytical method, obtain high on the basis of main constituent Multi-structure elements (as Fig. 4 a-4n) the morphological profiles feature of spectrum picture extension, i.e. space-optical spectrum feature；

Step 2: (this is to carry inside high-spectral data by given EO-1 hyperion sample data.It is high-spectrum The one part of pixel of picture, these pixels are given class label, that is, have been formulated classification.), it is utilized respectively step one and obtain Take each of the Multi-structure elements morphological profiles feature of high spectrum image extension opening operation, closed operation and original main constituent structure Build linear base core；

Step 3: each linear base verification answers a core weight, by calculating the Hadamard of any two linear base core Amass and obtain Non-linear Kernel, Non-linear Kernel is weighted, obtains nonlinear combination core；

Step 4: the nonlinear combination nucleus band that step 3 is obtained enters in support vector machine, is obtained using gradient descent method Optimum core weight；

Step 5: high spectrum image is classified using the optimum core weight that step 4 obtains.

Specific embodiment two: present embodiment from unlike specific embodiment one: obtain high in described step one The Multi-structure elements morphological profiles feature of spectrum picture extension；Detailed process is:

For any one pixel x in the high spectrum image having d wave band, many structures of the extension at this pixel Element morphology contour featureFormula be:

$\begin{array}{c}\stackrel{multises}{emp}\left(x\right)=\{\stackrel{1}{emp}\left(x\right),\stackrel{2}{emp}\left(x\right),\mathrm{...},\stackrel{s}{emp}\left(x\right)\}\\ =\{\stackrel{1}{{\mathrm{cp}}_c^{pc}}\left(x\right),\mathrm{...},\stackrel{2}{{\mathrm{cp}}_c^{pc}}\left(x\right),\mathrm{...},\stackrel{s}{{\mathrm{cp}}_c^{pc}}\left(x\right),i\left(x\right),\stackrel{1}{{\mathrm{op}}_c^{pc}}\left(x\right),\mathrm{...},\stackrel{2}{{\mathrm{op}}_c^{pc}}\left(x\right),\mathrm{...},\stackrel{s}{{\mathrm{op}}_c^{pc}}\left(x\right)\}\end{array},\∀c\&element;\[1,c\],$

Wherein, x represents any one pixel in the high spectrum image having d wave band, is a d dimensional vector, vector Each of value span be [0,1], d be positive integer；Represent and at pixel x in high spectrum image, utilize the Expanding morphology profile obtained by s kind structural element (processes high spectrum image with structural element, obtains morphological profiles special Levy),Represent and carry out the result obtained by closed operation using s kind structural element at pixel x in high spectrum image,Represent and carry out the result obtained by opening operation, c ∈ using s kind structural element at pixel x in high spectrum image [1, c] represent the yardstick of structural element, c is yardstick sum, the main one-tenth that i (x) obtains after principal component analysiss for high spectrum image Point；S and c is positive integer.

Other steps and parameter are identical with specific embodiment one.

Specific embodiment three: present embodiment from unlike specific embodiment one or two: logical in described step 2 Cross given EO-1 hyperion sample data, every in the Multi-structure elements morphological profiles feature of the extension being utilized respectively first step gained One opening operation, closed operation and original main constituent build linear base core；Detailed process is:

By given EO-1 hyperion sample data, it is utilized respectively the Multi-structure elements morphology wheel of the extension of first step gained Each of wide feature opening operation, closed operation and original main constituent build linear base core k_m(x_i,x_j)=< x_i,x_j>, x_iAnd x_jFor Any two pixel in the sample data of high spectrum image, i, j ∈ [1, n], n are given EO-1 hyperion sample data, and n is just Integer；If total s kind structural element, each of the configurations unit have c yardstick, then have m=2sc+1 linear base core.

Specific embodiment four: unlike one of present embodiment and specific embodiment one to three: described step 3 Each of linear base verification answer a core weight, amassed by the Hadamard calculating any two linear base core obtain non-linear Core, Non-linear Kernel is weighted, and obtains nonlinear combination core k_η(x_i,x_j), computing formula is as follows:

Wherein, k_m(x_i,x_j)=< x_i,x_j> it is linear base nuclear matrix, m=1,2 ..., m, m are linear base core sum, m= 2sc+1；η_mFor linear base core k_m(x_i,x_j) weight, η_hFor non-linear base core k_h(x_i,x_j) weight, ⊙ represents two linear bases The Hadamard of core is amassed, k_mFor linear base core, k_hFor non-linear base core, k_ηFor nonlinear combination kernel function.

One of other steps and parameter and specific embodiment one to three are identical.

Specific embodiment five: unlike one of present embodiment and specific embodiment one to four: described base core is total Number m value is the odd number more than 3.

One of other steps and parameter and specific embodiment one to three are identical.

Specific embodiment six: unlike one of present embodiment and specific embodiment one to five: described step 4 The middle nonlinear combination nucleus band obtaining step 3 enters in support vector machine, optimizes core weight using gradient descent method；Concrete mistake Cheng Wei:

1) first by nonlinear combination core k_η(x_i,x_j) bring in support vector machine,

$\underset{\mathrm{\&alpha;}\&element;r^n}{max}\underset{i=1}{\overset{n}{\mathrm{\&sigma;}}}{\mathrm{\&alpha;}}_i-\frac{1}{2}\underset{i=1}{\overset{n}{\mathrm{\&sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&sigma;}}}{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j{y}_i{y}_j{k}_{\mathrm{\&eta;}}(x_i,x_j),$

Wherein, α=[α₁,α₂,…,α_n] it is dual variable, α_i,α_j∈α；y_i,y_j∈ [- 1 ,+1] is sample label；R represents Real number field；N is given EO-1 hyperion sample data；rⁿRepresent n dimension real number space.

2) solve k_ηMin-max value optimization problem can be changed into,

Wherein ω is positive, bounded a 2- norm convex set；Positive number η can ensure that nonlinear combination kernel function k_ηPartly just it is Fixed, the regulationization on its border can control the norm of η；η is the set of core weight, η=[η₁,η₂,…,η_m], η₁,η₂,…,η_mIn Each element is positive number；

3) min-max value optimization problem is solved using the gradient descent algorithm based on mapping；First fix η, solveInMax problem, if The optimal solution of max problem is α^*, by α^*It is updated in former min-max value problem Obtain the minimum problems with regard to η:

For dual variable optimal solution, It is the minima with regard to η Problem, f (η) represents the function with regard to η.

One of other steps and parameter and specific embodiment one to five are identical.

Specific embodiment seven: unlike one of present embodiment and specific embodiment one to six: described step 5 Middle using the optimum core weight that step 4 obtains, high spectrum image is classified；Detailed process is:

Using equation below, each of high spectrum image pixel is classified:

$f\left(x\right)=\underset{i=1}{\overset{n}{\mathrm{\&sigma;}}}{\mathrm{\&alpha;}}_i^{*}{y}_i{k}_{\mathrm{\&eta;}}(x_i,x)+b$

Wherein b is side-play amount.

One of other steps and parameter and specific embodiment one to six are identical.

Using following examples checking beneficial effects of the present invention:

Embodiment one:

The hyperspectral image classification method based on morphological profiles feature and non-linear Multiple Kernel Learning for the present embodiment is specifically According to following steps preparation:

Data used by experiment is the state of Indiana high spectrum image of the U.S. central and north that aviris sensor obtains, data Comprise 224 spectral bands, wave-length coverage is 0.4～2.5 μm, ground resolution 20m, image size 144 × 144.Data is Through have passed through the pretreatment such as air, geometric correction, and eliminate 4 zero wave bands and 20 water body absorption bandses.Fig. 2 gives print The pseudo color composing figure of An Na data and atural object are truly schemed.Fig. 3 a-e using feature be extension single structure element morphology Learn profile, Fig. 3 a be using the non-linear Multiple Kernel Learning method in the present invention obtained by result figure, Fig. 3 b is using based on all Value criterion Multiple Kernel Learning method obtained by result figure, Fig. 3 c be using representative Multiple Kernel Learning method obtained by result Figure, Fig. 3 d be using mixing Multiple Kernel Learning method obtained by result figure, Fig. 3 e be using multiple features learning method obtained by Result figure；Fig. 3 f-j using feature be extension Multi-structure elements morphological profiles, Fig. 3 f be result figure of the present invention, Fig. 3 g It is that Fig. 3 h is using representative Multiple Kernel Learning method institute using based on the result figure obtained by the Multiple Kernel Learning method of average criterion The result figure obtaining, Fig. 3 i is the result figure obtained by employing mixing Multiple Kernel Learning method, and Fig. 3 j is using multiple features study side Result figure obtained by method.Table 1 is the corresponding nicety of grading of the above results figure, in conjunction with Fig. 3, truly schemes with reference to the atural object in Fig. 2, It can be seen that non-linear Multiple Kernel Learning method obtains best classifying quality under the conditions of identical input feature vector, adopting phase More preferable classifying quality can be obtained with conditions of feature learning method using the Multi-structure elements morphological profiles of extension, with When, the method for the present invention shows best classifying quality, reaches highest nicety of grading.Table 1 is that Indiana data is above-mentioned The nicety of grading of classification results.

Table 1

The present invention also can have other various embodiments, in the case of without departing substantially from present invention spirit and its essence, this area Technical staff when can according to the present invention make various corresponding change and deform, but these corresponding change and deformation all should belong to The protection domain of appended claims of the invention.

Claims (7)

1. the hyperspectral image classification method based on morphological profiles feature and non-linear Multiple Kernel Learning it is characterised in that: be based on The hyperspectral image classification method detailed process of morphological profiles feature and non-linear Multiple Kernel Learning is:

Step one: extract the main constituent of high spectrum image using principal component analytical method, obtain EO-1 hyperion on the basis of main constituent The Multi-structure elements morphological profiles feature of image spreading, i.e. space-optical spectrum feature；

Step 2: by given EO-1 hyperion sample data, be utilized respectively many structures that step one obtains high spectrum image extension The main constituent that each of element morphology contour feature opening operation, closed operation and high spectrum image obtain after principal component analysiss Build linear base core；

Step 3: a core weight is answered in each linear base verification, is amassed by the Hadamard calculating any two linear base core To Non-linear Kernel, Non-linear Kernel is weighted, obtains nonlinear combination core；

Step 4: the nonlinear combination nucleus band that step 3 is obtained enters in support vector machine, obtains optimum using gradient descent method Core weight；

Step 5: high spectrum image is classified using the optimum core weight that step 4 obtains.

2. the classification hyperspectral imagery side based on morphological profiles feature and non-linear Multiple Kernel Learning according to claim 1 Method it is characterised in that: in described step one obtain high spectrum image extension Multi-structure elements morphological profiles feature；Concrete mistake Cheng Wei:

For any one pixel x in the high spectrum image having d wave band, the Multi-structure elements of the extension at this pixel Morphological profiles featureFormula be:

$\begin{array}{c}\stackrel{multises}{emp}\left(x\right)=\{\stackrel{1}{emp}\left(x\right),\stackrel{2}{emp}\left(x\right),\mathrm{...}\stackrel{s}{emp}\left(x\right)\}\\ \begin{array}{cc}=\{\stackrel{1}{{\mathrm{cp}}_c^{pc}}\left(x\right),\mathrm{...},\stackrel{2}{{\mathrm{cp}}_c^{pc}}\left(x\right),\mathrm{...},\stackrel{s}{{\mathrm{cp}}_c^{pc}}\left(x\right),i\left(x\right),\stackrel{1}{{\mathrm{op}}_c^{pc}}\left(x\right),\mathrm{...},\stackrel{2}{{\mathrm{op}}_c^{pc}}\left(x\right),\mathrm{...},\stackrel{s}{{\mathrm{op}}_c^{pc}}\left(x\right)\}& \&forall;c\&element;\&lsqb;1,c\&rsqb;\end{array}\end{array},$

Wherein, x represents any one pixel in the high spectrum image having d wave band, is a d dimensional vector, in vector The span of each value is [0,1],Represent and at pixel x in high spectrum image, utilize s kind structural element institute The expanding morphology profile obtaining,Represent and carry out closing fortune using s kind structural element at pixel x in high spectrum image Result obtained by calculating,Represent and carry out opening operation institute using s kind structural element at pixel x in high spectrum image The result obtaining, c ∈ [1, c] represents the yardstick of structural element, and c is yardstick sum, and i (x) divides through main constituent for high spectrum image The main constituent obtaining after analysis；S and c is positive integer.

3. the classification hyperspectral imagery side based on morphological profiles feature and non-linear Multiple Kernel Learning according to claim 2 Method it is characterised in that: pass through given EO-1 hyperion sample data in described step 2, be utilized respectively step one and obtain high-spectrum Each of Multi-structure elements morphological profiles feature as extension opening operation, closed operation and high spectrum image are through principal component analysiss The main constituent obtaining afterwards builds linear base core；Detailed process is:

By given EO-1 hyperion sample data, the Multi-structure elements morphological profiles being utilized respectively the extension that step one obtains are special Each of levy the main constituent that opening operation, closed operation and high spectrum image obtain after principal component analysiss and build linear base core and build Linear base core k_m(x_i,x_j)=< x_i,x_j>, x_iAnd x_jFor any two pixel in the sample data of high spectrum image, i, j ∈ [1, n], n is given EO-1 hyperion sample data, and n is positive integer；If total s kind structural element, each of the configurations unit have c chi Degree, then have m=2sc+1 linear base core.

4. the classification hyperspectral imagery side based on morphological profiles feature and non-linear Multiple Kernel Learning according to claim 3 Method it is characterised in that: the linear base verification of each of described step 3 answers a core weight, by calculating the linear base of any two The Hadamard of core is amassed and is obtained Non-linear Kernel, Non-linear Kernel is weighted, obtains nonlinear combination core k_η(x_i,x_j), computing formula As follows:

Wherein, k_m(x_i,x_j)=< x_i,x_j> it is linear base nuclear matrix, m=1,2 ..., m, m are linear base core sum, m=2sc+1, Value is the odd number more than 3；η_mFor linear base core k_m(x_i,x_j) weight, η_hFor non-linear base core k_h(x_i, x_j) weight, ⊙ table Show that the Hadamard of two linear base cores is amassed, k_mFor linear base core, k_hFor non-linear base core, k_ηFor nonlinear combination kernel function.

5. the classification hyperspectral imagery side based on morphological profiles feature and non-linear Multiple Kernel Learning according to claim 4 Method it is characterised in that: described base core sum m value is the odd number more than 3.

6. the classification hyperspectral imagery side based on morphological profiles feature and non-linear Multiple Kernel Learning according to claim 5 Method it is characterised in that: in described step 4, the nonlinear combination nucleus band that step 3 obtains is entered in support vector machine, using gradient Descent method obtains optimum core weight；Detailed process is:

1) first by nonlinear combination core k_η(x_i,x_j) bring in support vector machine,

$\underset{\mathrm{\&alpha;}\&element;r^n}{max}\underset{i=1}{\overset{n}{\&sigma;}}{\mathrm{\&alpha;}}_i-\frac{1}{2}\underset{i=1}{\overset{n}{\&sigma;}}\underset{j=1}{\overset{n}{\&sigma;}}{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j{y}_i{y}_j{k}_{\mathrm{\&eta;}}(x_i,x_j),$

Wherein, α=[α₁,α₂,…,α_n] it is dual variable, α_i,α_j∈α；y_i,y_j∈ [- 1 ,+1] is sample label；R represents real number Domain；N is given EO-1 hyperion sample data；rⁿRepresent n dimension real number space；

2) k will be solved_ηChange into min-max value optimization problem,

Wherein ω is positive, bounded a 2- norm convex set；η is the set of core weight, η=[η₁,η₂,…,η_m], η₁, η₂,…,η_mIn each element be positive number；

3) min-max value optimization problem is solved using the gradient descent algorithm based on mapping；First fix η, solveInMax problem, if The optimal solution of big value problem is α^*, by α^*It is updated in former min-max value problem Obtain the minimum problems with regard to η:

$\underset{\mathrm{\&eta;}\&element;\mathrm{\&omega;}}{min}f\left(\mathrm{\&eta;}\right)=\underset{i=1}{\overset{n}{\&sigma;}}{\mathrm{\&alpha;}}_i^{*}-\frac{1}{2}\underset{i=1}{\overset{n}{\&sigma;}}\underset{j=1}{\overset{n}{\&sigma;}}{\mathrm{\&alpha;}}_i^{*}{\mathrm{\&alpha;}}_j^{*}{y}_i{y}_j{k}_{\mathrm{\&eta;}}(x_i,x_j);$

For dual variable optimal solution, It is the minimum problems with regard to η, F (η) represents the function with regard to η.

7. the classification hyperspectral imagery side based on morphological profiles feature and non-linear Multiple Kernel Learning according to claim 6 Method it is characterised in that: using the optimum core weight that step 4 obtains, high spectrum image is classified in described step 5；Specifically Process is:

Using equation below, each of high spectrum image pixel is classified:

$f\left(x\right)=\underset{i=1}{\overset{n}{\&sigma;}}{\mathrm{\&alpha;}}_i^{*}{y}_i{k}_{\mathrm{\&eta;}}(x_i,x)+b$

Wherein b is side-play amount.