CN105389597A - Hyperspectral data multi-classification method based on Chernoff distance and SVM (support vector machines) - Google Patents

Abstract

The invention discloses a hyperspectral data multi-classification method based on Chernoff distance and SVM (support vector machines). The method comprises the following steps: I, preprocessing input data into normalized data; II, calculating the Chernoff distance between any two types to obtain a Chernoff distance matrix; III, determining an execution sequence table under a multi-classification task OAA (one-against-all) strategy to obtain a separability criterion of each wave band and that of all-wave bands; IV, establishing a sub-classifier guiding coefficient based on the Chernoff distance; and V, executing the classification task determined by the whole OAA strategy via a weighed SVM classifier based on the Chernoff distance till the final categorical attribute of each tested sample is single. The method not only improves the accuracy of the conventional SVM method, but also improves the classification accuracy of small sample classification to a relatively great extent, and is suitable for application to hyperspectral image pattern recognition based on the OAA classification strategy.

Description

A kind of many sorting techniques of high-spectral data based on Chernoff Distance geometry SVM

Technical field

The invention belongs to area of pattern recognition, relate to a kind of svm classifier method based on Chernoff distance optimizing training data.

Background technology

Hyperspectral image data is made up of the wave band of hundreds of continuous distribution usually.Whole data are regarded as three dimension data rectangular parallelepipeds, wherein the position of bidimensional determination object in plane of vision, the position of third dimension determination object in spectral wavelength.For AVIRIS (AirborneVisible/InfraredImagingSpectrometer) high-spectral data, about 10nm of being generally only separated by between adjacent two wave bands.Due to image space adjacent band wave band between correlativity very strong, make conventional sorting methods must carry out dimensionality reduction after could continue process data, conventional method is band selection.Because kernel method (kernelmethod) is very little by the impact of input space height dimension, so increasing researcher selects kernel method.No matter whether select band selection, kernel method all has excellent classification performance, such as our support vector machine (SupportVectorMachines, SVM) sorter of being familiar with very much.But, seldom have research to be devoted to expansion SVM method and make it be more suitable for many classification application of high-spectral data.

Classification error probability is the best quantitive measure of characteristic validity in pattern-recognition, and it is minimum that the dreamboat of feature selecting makes to reach classification error probability.But this point is often difficult to accomplish.Therefore Upper bound of error probability is minimum is usually a kind of reasonably selection.The Upper bound of error probability that Chernoff proposes is minimum, claims the Chernoff upper bound.The nicety of grading of single core sorter effectively can be improved in the Chernoff upper bound, and has certain directive function for the classification policy of core sorter.

Because SVM can only complete two classification task in itself, the many classification application of typical case of high-spectral data often need by multiple SVM and certain strategy to build multi-categorizer.High-spectral data third dimension information is utilized mainly to concentrate on following two aspects to the work that SVM improves at present: one is carry out filtering process to EO-1 hyperion input data, and two is kernel functions of generating custom.But these methods are all only for unified SVM.For each SVM setting up multi-categorizer, research in the past all takes the simple scheme unanimously treated.More accurate method takes one more favorably many classification schemes, and the feature of two class objects namely handled by each sub-classifier carries out the independent customization of kernel function to it.On many classification policys, widespread use be OAA (One-Against-All, one-to-many) and OAO (One-Against-One, two kinds of strategies one to one), they respectively have relative merits, the former sub-classifier negligible amounts, but the training time of each sub-classifier is longer, the sub-classifier quantity of the latter is more, but the training time of each sub-classifier is shorter.And nicety of grading aspect, the two is also little in difference after parameter optimization.

Summary of the invention

The object of the invention is to a kind of svm classifier method based on Chernoff distance proposing improvement, by introducing the distance weighted matrix of Chernoff, information between the class making sorter make full use of sample in the process of training, and utilize Chernoff distance to instruct the select progressively of OAA policy class, not only increase the degree of accuracy of traditional SVM method, and the nicety of grading of small sample classification is greatly improved, be applicable to the high spectrum image application of pattern recognition based on OAA classification policy.

The object of the invention is to be achieved through the following technical solutions:

Based on the many sorting techniques of high-spectral data of Chernoff Distance geometry SVM, comprise following five steps:

One, pre-service is carried out to input data, obtain normalization data;

Two, calculate the Chernoff distance between any two classifications, obtain Chernoff distance matrix;

Three, determine the execution sequence table under many classification task OAA strategy, obtain each wave band and full wave dissociable basis;

Four, the sub-classifier built based on Chernoff distance instructs coefficient;

Five, the weighed SVM sorter based on Chernoff distance is adopted to perform the determined classification task of whole OAA strategy, until obtain the final single category attribute of each test sample book.

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

1, in traditional svm classifier method, introduce Chernoff distance weighted matrix, information between the class making sorter make full use of sample in the process of training, thus improve classification effectiveness.Have more way targetedly because this method takes in the optimization of each sub-classifier, thus effectively can improve the nicety of grading of each sub-classifier, thus improve the final nicety of grading of multi-categorizer.

2, compared with traditional OAA classification policy, utilize in this method and each sub-classifier execution sequence under OAA strategy is instructed, the separability of the collectivity tolerance calculated particular by utilizing Chernoff determines this execution sequence table, this just makes the good classification of separability first divide away from test sample book, and the bad classification of separability is then successively placed on finally to be distinguished.And, when each sub-classifier is classified, the weighting coefficient adopted, i.e. guidance to different types of areas coefficient, all carry out COMPREHENSIVE CALCULATING according to remaining classification, the classification separated is not considered, this also enhances the specific aim of weighting classification, thus the error in classification of each sub-classifier is reduced further.

Accompanying drawing explanation

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

Fig. 2 is hyperspectral image data original graph;

Fig. 3 is hyperspectral image data label figure;

Fig. 4 is standard SVM kernel function classifying quality figure;

Fig. 5 is the kernel function classifying quality figure based on Chernoff distance.

Embodiment

Below in conjunction with accompanying drawing, technical scheme of the present invention is further described; but be not limited thereto; everyly technical solution of the present invention modified or equivalent to replace, and not departing from the spirit and scope of technical solution of the present invention, all should be encompassed in protection scope of the present invention.

Embodiment one: present embodiments provide for a kind of many sorting techniques of high-spectral data based on Chernoff Distance geometry SVM, data normalization will be inputted by pre-service, calculate the Chernoff distance between any two classifications and utilize dissociable basis to instruct OAA (One-Against-All, one-to-many) policy class order to improve nicety of grading, utilize Chernoff distance vector to be weighted single sub-classifier, then repeatedly use weighed SVM sorter to sample classification.

As shown in Figure 1, be divided into five steps, concrete steps are as follows:

Step one: training sample and test sample book are set.

1) for the multi-spectral remote sensing image of shooting wherein Row, Column represent the wide of multi-spectral remote sensing image and length, and B represents the wave band number of multi-spectral remote sensing image, if x _z∈ R ^b, z=1,2 ..., N is that B ties up sample, y _z∈ 1,2 ..., L} is and x _zrelevant class label, wherein N is number of samples, and L is class number.

2) other pixel of same class is all brought together, the pixel relating to class categories in all pixels is processed all like this.

3) grouping coefficient Group assignment is given, simultaneously the broken number of " a few folding cross validation " namely often said of this coefficient.

4) original sample is divided into training sample and test sample book, represent with two-dimensional matrix TrainSamples and TestSamples respectively, the spectral information data of each wave band of the corresponding single pixel of column vector, wherein training sample accounts for the 1/Group of total sample, and it is test sample book that all the other (1-Group)/Group remain sample.

5) training sample and test sample book are normalized.

Step 2: calculate two different classes of between Chernoff distance.

1) select the classification that two different, be set to p class and q class.

If n _pand n _qbe respectively the number of p class and q class sample in training sample.For kth-1, k and k+1 tri-wave bands of sample, first the data of these three wave bands are normalized, then the data of the p class on these three wave bands and q class sample are deposited to two temporary variable matrix D according to the form of column vector _pand D _qin, wherein D _pfor n _pthe matrix of × 3 dimensions, D _qfor n _qthe matrix of × 3 dimensions.

2) D is calculated _pwith D _qthe average of each row, obtains the Mean Matrix Mean of two 1 × 3 _pwith Mean _q.

3) D is calculated _pwith D _qcovariance, obtain the covariance matrix Cov of two 3 × 3 _pwith Cov _q.

4) carry out Chernoff distance to all classifications different between two to calculate:

$\begin{array}{c}{{\mathrm{\Ω}}^k}_{p,q}=\frac{\mathrm{\β}\left(1-\mathrm{\β}\right)}{2}{\left({\mathrm{Mean}}_p-{\mathrm{Mean}}_q\right)}^T{\left({\mathrm{\βCov}}_p+\left(1-\mathrm{\β}\right){\mathrm{Cov}}_q\right)}^{-1}\left({\mathrm{Mean}}_p-{\mathrm{Mean}}_q\right)\\ +\frac{1}{2}\ln \left(\frac{\left|{\mathrm{\βCov}}_p+\left(1-\mathrm{\β}\right){\mathrm{Cov}}_q\right|}{{\left|{\mathrm{Cov}}_p\right|}^{\mathrm{\β}}{\left|{\mathrm{Cov}}_q\right|}^{1-\mathrm{\β}}}\right)\end{array}.$

Wherein, Ω ^k _{q, q}be the Chernoff distance between p class and q class, k is current band number, Mean _pwith Mean _qbe respectively p class and the Mean Matrix of q class on kth-1, k and k+1 tri-wave bands, Cov _pwith Cov _qbe respectively p class and the covariance matrix of q class on kth-1, k and k+1 tri-wave bands, β is Chernoff distance adjustment parameter, and 0 < β < 1.Repeatedly change β and can find suitable upper error for sorter.

5) by 4) the middle Chernoff distance structure Chernoff distance matrix calculated:

Due to Chernoff distance be two different classes of between parameter, so the element on diagonal line is 0, C _{p, q}for the vector of B × 1, p={1,2 ..., L}, q={1,2 ..., L} and p ≠ q.Distance between p class and q class and the distance between q class and p class are merged into a distance.

6) the 1st is repeated) to the 5th) step, until all carried out the computation process of Chernoff distance to all dimensions of any two classifications of training sample.

Step 3: determine the execution sequence table under many classification task OAA strategy.

Many classification task consider that classification sum is more than or equal to the situation of 3.Be different from common OAA strategy, this method utilizes Chernoff distance to instruct OAA classified order.

1) the Chernoff distance vector of p class training sample and other classifications is sued for peace, obtain the Chernoff distance vector to other all categories on each wave band coefficient is instructed as sub-classifier.On each wave band, the Chernoff distance of p class training sample and other classifications is ${{\mathrm{\&Omega;}}^l}_p=\underset{q=1,q\&NotEqual;p}{\overset{L}{\&Sigma;}}{\mathrm{\&Omega;}}_{p,q}^l.$

2) will in each wave band be added, obtain p class and on all wave bands, the separability of the collectivity of other all categories measured

3) by the order arrangement that the separability of the collectivity of all categories tolerance is successively decreased by numerical values recited, then the execution sequence table under OAA strategy is this and puts in order, and uses ordered set <Class ₁, Class ₂... Class _l..., Class _l> represents, wherein Class _l∈ 1,2 ..., and L}, l=1,2 ..., L.

Step 4: the sub-classifier built based on Chernoff distance instructs coefficient.

After the execution sequence under OAA strategy is determined, namely we carried out the classification task of all L classes by L sub-classifier (each subclassification is two sorters), concrete execution sequence is: first pass through Class ₁determine that in test sample book, label sequence number equals Class with other two classification of all residue classes ₁sample, then pass through Class ₂(note now not containing Class with all residue classifications ₁classification) two classification determine that in test sample book, label sequence number equals Class ₂sample, successively perform down, last two sorter will judge Class _l-1and Class _lclassification ownership.

Each sub-classifier concrete instructs coefficient by containing a sub-classifier based on Chernoff distance, and for improving the classification performance of each sub-classifier, obtain the classification results that confidence rate is higher, this sub-classifier instructs the computation process of coefficient specific as follows:

1) for the classification Class classified at first ₁if, Class ₁=i, wherein i ∈ 1,2 ..., L}, i.e. Class ₁corresponding classification is the i-th class, is arranged in the i-th row of Chernoff matrix, then sub-classifier instructs coefficient to be:

${\tilde{C}}_i=\underset{q=1,q\&NotEqual;i}{\overset{L}{\&Sigma;}}{C}_{i,q}.$

Wherein, C _{i, q}for the i-th row in Chernoff matrix, the element of q row.

2) according to the execution sequence table <Class under OAA strategy ₁, Class ₂..., Class _l..., Class _l>, proceeds to Class _ltime, if Class _l=j, wherein j ∈ 1,2 ..., L}, the Chernoff distance vector C of class categories during calculating _{p, q}foreclose, then sub-classifier instructs coefficient to be ${\tilde{C}}_j=\underset{q=1,q\&NotEqual;j,...,q\&NotEqual;i}{\overset{L}{\&Sigma;}}{C}_{j,q}.$

3) finally to Class _l-1and Class _lclassification, only need carry out a subseries.

Step 5: adopt the weighed SVM sorter based on Chernoff distance to perform the determined classification task of whole OAA strategy, until obtain the final single category attribute of each test sample book.

1) suitable kernel function is first selected.The adaptability checked in numerous types of data due to RBF is stronger, and to easily causing the larger classification problem of the quantity variance of a kind and plurality of classes to have good result of calculation under OAA strategy, so the present invention selects RBF core as the kernel function of SVM classifier; Meanwhile, our guidance to different types of areas coefficient for p class in order to Chernoff distance be adopted before to obtain be incorporated in the middle of each sub-classifier, we need the RBF kernel function adopting weighting to improve:

$K_{RBF}({\tilde{C}}_{p}x,{\tilde{C}}_p{x}^{\&prime;})=\exp (-\frac{\left|\right|{\tilde{C}}_p(x-x^{\&prime;})|{|}^2}{2{\mathrm{\&sigma;}}^2}).$

Wherein σ is width constant, and σ is less, and function more has selectivity.

In order to ensure the correct execution of SVM sub-classifier, the sample data also tackling input after weighting is normalized.

2) according to the determined classification policy of step 3, each step all adopts step 1) in the constructed weighed SVM sorter based on Chernoff distance removed training and testing.

3) repeatedly step 2 is performed), until all OAA strategy is after determined classification task all completes, namely each test sample book is all single category attribute by final decision, and at this moment whole many classification task for test sample book can stop.

Embodiment two: the SVM approach application based on Chernoff distance of improvement in hyperspectral image data assorting process, is improved nicety of grading in conjunction with OAA strategy by present embodiment further.

First the description of hyperspectral image data is provided:

Experimental subjects is EO-1 hyperion image data.This packet is containing 520 continuous wave bands, and experimental subjects is EO-1 hyperion image data.These data comprise 75x75 pixel, and each pixel comprises 520 continuous wave bands, and wavelength coverage is 400nm-1000nm, and spectral resolution is 8nm.

Present embodiment have chosen 4 maximum classes of pixel count in pictures taken as experiment sample, refers to table 1.

Table 1 corresponding experiment sample number of all categories

Mark	Classification	Sample number
			A	Sky background	4369
B	Tree	1027
			C	Cloud	222
D	Aircraft	7

Fig. 2 is the original image of high-spectral data, wherein sky, and tree and cloud occupy the overwhelming majority; Aircraft only account for several pixel in view picture figure.Fig. 3 is the label figure of sample.In the process labelled, due to the relation of resolution, a part for treetop has been denoted as sky.But because the spectral differences of treetop and sky is apart from comparatively large and the sample point of two classes is enough, so the effect of this part on classification does not almost affect, and can as the checking of classification results.

Present embodiment takes two kinds respectively and checks the classification experiments that 4 classes are carried out in high-spectral data set, and they are standard RBF core and the weighting RBF core based on Chernoff distance.These two kinds of kernel functions all construct multi-categorizer by OAA strategy, and the classification experiments for 4 class atural objects needs 3 sub-classifiers.

Perform step one: input hyperspectral image data with the label of correspondence, if need the wave band B calculating Chernoff distance to be 520.Experiment grouping coefficient selects Group=2, i.e. 2 folding cross validations.Class number L=4.Input data are normalized, give the label that training sample is corresponding with its classification.

Perform β numerical value in the calculating of step 2: Chernoff distance generally to select for less or larger β value, because the complexity calculated and the Singular Value problem that may exist are not considered.This method is selected calculate any two different classes of Chernoff distances of each wave band, obtain Matrix C hernoff distance matrix C.

Perform step 3: the separability of the collectivity tolerance calculating each classification the separability of the collectivity of all categories is measured the order arrangement of successively decreasing by numerical values recited is as the execution sequence under OAA strategy.Obtain the execution sequence table <Class under OAA strategy ₁, Class ₂, Class ₃, Class ₄>.

Perform step 4: each sub-classifier calculated based on Chernoff distance instructs coefficient i ∈ { 1,2,3,4}.

Perform step 5: utilize guidance to different types of areas coefficient to each sub-classifier weighting, then to training sample normalization carry out svm classifier to it again.Because RBF has good parameter adaptation in hyperspectral data processing field, therefore we get penalty factor be 100, parameter σ is 0.4.

Add standard SVM kernel function to test as a comparison.Keep input amendment constant, single sub-classifier is not weighted, and OAA strategy is classified successively for category.

Conclusion: contrast test the results are shown in Table 2.Many sorting techniques based on Chernoff distance all have the lifting of about 10% compared with the average nicety of grading of standard SVM method, overall accuracy also has the lifting of 2%, but the latter is not fairly obvious relative to the former lifting.From support vector sum, the support vector number of the kernel method of improvement is slightly more than kernel function originally, and the many sorting techniques as seen based on Chernoff distance can make the process of training slightly complicated, but training time and test duration reduce.Although improve kernel method both increases weighting coefficient product calculation to each element in kernel function, no matter be training or classification, the time that the kernel method of improvement consumes all can be more less than standard SVM method.But calculate the time of Chernoff weighting matrix much larger than the time of training and classify, and do not count the time of this part.So comprehensive these two aspects factor, the time of this method actual consumption is greater than standard SVM method.

The comparison of nicety of grading when table 2 adopts RBF core, support vector sum and time loss

Parameter index	Standard SVM	Chernoff distance kernel method
			Mean accuracy	55.8688	65.0868
Overall accuracy	88.0697	90.1323
			Support vector number	4046	4634
Training time	16.8302	9.5793
			Test duration	26.4759	23.0401

Fig. 4 is the classification results figure of standard SVM kernel function, Fig. 5 is the classification results figure of CHernoff apart from kernel function.By in figure, we can find, for sky, tree and cloud three classifications, because the quantity of training sample is enough, so the effect difference of classification is little, are only the differences of indivedual point.But for aircraft, because sample point itself is considerably less, so the difficulty of training is very large.And relative to the SVM method of standard, the classifying quality of the many sorting techniques based on Chernoff distance that this method proposes wants better.

Claims (9)

1., based on the many sorting techniques of high-spectral data of Chernoff Distance geometry SVM, it is characterized in that described method step is as follows:

One, pre-service is carried out to input data, obtain normalization data;

Two, calculate the Chernoff distance between any two classifications, obtain Chernoff distance matrix;

Three, determine the execution sequence table under many classification task OAA strategy, obtain each wave band and full wave dissociable basis;

Four, the sub-classifier built based on Chernoff distance instructs coefficient;

Five, the weighed SVM sorter based on Chernoff distance is adopted to perform the determined classification task of whole OAA strategy, until obtain the final single category attribute of each test sample book.

2. the many sorting techniques of high-spectral data based on Chernoff Distance geometry SVM according to claim 1, is characterized in that the concrete steps of described step one are as follows:

1) for the multi-spectral remote sensing image of shooting wherein Row, Column represent the wide of multi-spectral remote sensing image and length, and B represents the wave band number of multi-spectral remote sensing image, if x _z∈ R ^b, z=1,2 ..., N is that B ties up sample, y _z∈ 1,2 ..., L} is and x _zrelevant class label, wherein N is number of samples, and L is class number;

2) other pixel of same class is all brought together, the pixel relating to class categories in all pixels is processed all equally;

3) grouping coefficient Group assignment is given;

4) original sample is divided into training sample and test sample book, represents with two-dimensional matrix TrainSamples and TestSamples respectively, the spectral information data of each wave band of the corresponding single pixel of column vector;

5) training sample and test sample book are normalized.

3. the many sorting techniques of high-spectral data based on Chernoff Distance geometry SVM according to claim 2, is characterized in that described training sample accounts for the 1/Group of total sample, and it is test sample book that all the other (1-Group)/Group remain sample.

4. the many sorting techniques of high-spectral data based on Chernoff Distance geometry SVM according to claim 1, is characterized in that, in described step 2, Chernoff distance calculating method is as follows:

$\begin{array}{c}{{\mathrm{\&Omega;}}^k}_{p,q}=\frac{\mathrm{\&beta;}(1-\mathrm{\&beta;})}{2}{({\mathrm{Mean}}_p-{\mathrm{Mean}}_q)}^T{({\mathrm{\&beta;Cov}}_p+(1-\mathrm{\&beta;}\left){\mathrm{Cov}}_q\right)}^{-1}({\mathrm{Mean}}_p-{\mathrm{Mean}}_q)\\ +\frac{1}{2}\ln \left(\frac{\left|{\mathrm{\&beta;Cov}}_p+\left(1-\mathrm{\&beta;}\right){\mathrm{Cov}}_q\right|}{{\left|{\mathrm{Cov}}_p\right|}^{\mathrm{\&beta;}}{\left|{\mathrm{Cov}}_q\right|}^{1-\mathrm{\&beta;}}}\right)\end{array};$

Wherein, Ω ^k _{p, q}be the Chernoff distance between p class and q class, k is current band number, Mean _pwith Mean _qbe respectively p class and the Mean Matrix of q class on kth-1, k and k+1 tri-wave bands, Cov _pwith Cov _qbe respectively p class and the covariance matrix of q class on kth-1, k and k+1 tri-wave bands, β is Chernoff distance adjustment parameter, and 0 < β < 1;

Chernoff distance matrix computing method are as follows:

$C_{p,q}=({\mathrm{\&Omega;}}_{p,q}^1,{\mathrm{\&Omega;}}_{p,q}^2,...,{\mathrm{\&Omega;}}_{p,q}^d),p=\{1,2,...,L\},$ Q={1,2 ..., L} and p ≠ q.

5. the many sorting techniques of high-spectral data based on Chernoff Distance geometry SVM according to claim 4, is characterized in that described β numerical value is generally selected

6. the many sorting techniques of high-spectral data based on Chernoff Distance geometry SVM according to claim 1, is characterized in that the concrete steps of described step 3 are as follows:

1) the Chernoff distance vector of p class training sample and other classifications is sued for peace, obtain the Chernoff distance vector to other all categories on each wave band instruct coefficient as sub-classifier, on each wave band, the Chernoff distance of p class training sample and other classifications is ${{\mathrm{\&Omega;}}^l}_p=\underset{q=1,q\&NotEqual;p}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{\&Omega;}}_{p,q}^l;$

2) will in each wave band be added, obtain p class and on all wave bands, the separability of the collectivity of other all categories measured

3) by the order arrangement that the separability of the collectivity of all categories tolerance is successively decreased by numerical values recited, then the execution sequence table under OAA strategy is this and puts in order, and uses ordered set <Class ₁, Class ₂... Class _l..., Class _l> represents, wherein Class _l∈ 1,2 ..., and L}, l=1,2 ..., L.

7. the many sorting techniques of high-spectral data based on Chernoff Distance geometry SVM according to claim 1, is characterized in that the concrete steps of described step 4 are as follows:

1) for the classification Class classified at first ₁if, Class ₁=i, wherein i ∈ 1,2 ..., L}, i.e. Class ₁corresponding classification is the i-th class, is arranged in the i-th row of Chernoff matrix, then sub-classifier instructs coefficient to be

2) according to the execution sequence table <Class under OAA strategy ₁, Class ₂..., Class _l..., Class _l>, proceeds to Class _ltime, if Class _l=j, wherein j ∈ 1,2 ..., L}, the Chernoff distance vector C of class categories during calculating _{p, q}foreclose, then sub-classifier instructs coefficient to be ${\tilde{C}}_j=\underset{q=1,q\&NotEqual;j,...,q\&NotEqual;i}{\overset{L}{\&Sigma;}}{C}_{j,q};$

3) finally to Class _l-1and Class _lclassification, only need carry out a subseries.

8. the many sorting techniques of high-spectral data based on Chernoff Distance geometry SVM according to claim 1, is characterized in that the concrete steps of described step 5 are as follows:

1) the RBF kernel function selecting weighting to improve as the kernel function of the weighed SVM sorter based on Chernoff distance, wherein for the guidance to different types of areas coefficient for p class, and be normalized after weighting;

2) according to the determined classification policy of step 3, adopt step 1) in the constructed weighed SVM sorter based on Chernoff distance removed training and testing;

3) repeatedly step 2 is performed), until all OAA strategy is after determined classification task all completes, at this moment whole many classification task for test sample book can stop.

9. the many sorting techniques of high-spectral data based on Chernoff Distance geometry SVM according to claim 8, is characterized in that the RBF kernel function that described weighting improves computing method as follows:

$K_{\mathrm{RBF}}({\tilde{C}}_{p}x,{\tilde{C}}_p{x}^{\&prime;})=\exp (-\frac{{\left|\right|{\tilde{C}}_p(x-x^{\&prime;})\left|\right|}^2}{{2\mathrm{\&sigma;}}^2});$

Wherein, σ is width constant.