CN109543755A - Integrated study Remote Image Classification based on class weight vector - Google Patents

Abstract

The invention belongs to remote sensing technology fields, specifically disclose a kind of integrated study Remote Image Classification based on class weight vector.The sample of extraction is divided into training sample and test sample first by the present invention, then using different classifications algorithm training classifier, classified again with these classifiers to test sample later, obtain the class weight vector that classification can be made to be adaptive to different classifications device, to carry out integrating integrated classifier to different classifiers, realize that the classifier of different classifications algorithm realizes have complementary advantages.On this basis, using the method for AdaBoost iteration, promoted to obtain that a precision is higher, the better integrated classifier of integrated result to the integrated classifier of generation.The present invention realizes adaptive between classification and classifier, enhances the diversity of base classifier during integrated different classifications algorithm, ensure that effective promotion of classification of remote-sensing images precision.

Description

Integrated study Remote Image Classification based on class weight vector

Technical field

The invention belongs to remote sensing technology fields, are related to a kind of integrated study classification of remote-sensing images based on class weight vectorMethod.

Background technique

Remote sensing technology provides for the mankind most intuitively effectively to surface observation mode, but how to extract from remotely-sensed dataLULC (Land Use and Land Cover) information is still a long-term sciences problems.By means of computer science, passThe machine learning algorithm of system shows excellent performance (such as maximum likelihood method (Maximum in remote sensing image field of classifying automaticallyLikelihood, ML), decision tree method (Decision Tree, DT), minimum distance method (Minimum Distance, MD), branchHold vector machine (Support Vector Machine, SVM), naive Bayesian (Naive Bayes, NB), artificial neural network(Artificial Neural Network, ANN) etc.), realize the rapidly extracting LULC information from remote sensing image.But by groundThe influence of object complex characteristic still remains wrong point, leakage point, the disadvantages such as close atural object recognition capability is poor, and precision is low.

Currently, being improved by single sorting algorithm smaller and smaller to promote the space of classification of remote-sensing images precision.ResearchPerson has found that can be suitable for the algorithm of any classification problem is to be not present after comparing and analyzing to different sorting algorithms's.For same classification problem, the effect that different sorting algorithms obtains is different, and the same algorithm is for different classes of pointClass performance Ye You club difference.In classification of remote-sensing images, a certain sorting algorithm may distinguish certain LULC classifications well,But for the differentiation poor effect of other LULC classifications.Multi-classifier system (Multiple Classifier System, MCS)Classification can integrate the advantage of different classifications device, it is possible to prevente effectively from the deficiency of single classifier, further promotes nicety of grading,Extensive favor is received in terms of remote sensing image automatic interpretation, and forms a series of theory and technology and method system.

However, current MCS method still suffers from many problems.Such as, some MCS methods depend critically upon priori and knowKnow, only sample meets certain probability distribution, and integrated performance can be optimal.That there is stability is poor for these methods,Complexity is high, is easy over-fitting, the disadvantage that integrated precision is limited, therefore is difficult to be transplanted and promote.

Existing MCS method or the same sorting algorithm of use, start with from the quantity and feature space of sample and generate toolStandby multifarious base classifier；Has multifarious base classifier using different sorting algorithm buildings.Previous seed typeIt is easy to generate a large amount of and discrepant base classifier of tool, but is influenced by algorithm self-characteristic, the integrated performance meeting of multi-categorizerBy certain constraint；Latter type can integrate the advantage of different classifications algorithm, but since algorithm number is limited, and generate base pointClass device negligible amounts, and the integrated needs of multi-categorizer ensure that complicated operation with certain sample distribution.

Summary of the invention

It is an object of the invention to propose a kind of integrated study Remote Image Classification based on class weight vector, withOvercome the shortcomings of existing MCS method, achievees the purpose that efficient, high-precision classification of remote-sensing images.

The present invention to achieve the goals above, adopts the following technical scheme that

Integrated study Remote Image Classification based on class weight vector, includes the following steps:

S1. according to the basic step of AdaBoost, initialization sample concentrates the weight of sample；

S2. the sample of extraction is divided into training sample and survey using there is the method sample drawn for putting back to sampling from sample setSample sheet；

S3. the base classifier of a variety of different classifications algorithms is respectively trained using training sample；

S4. classified using the base classifier in step s3 to test sample, obtain the classification power of each base classifierWeight vector；

S5. each base classifier is integrated based on class weight vector, obtains integrated classifier；

S6. classified using the integrated classifier in step s5 to the sample in step s2 sample set, obtain classification and missDifference thens follow the steps s7 if error in classification is less than setting error threshold, no to then follow the steps s8；

S7. the sample weights in sample set are updated, updated sample is put back in the sample set in step s2；

S8. s1 is returned to step；

S9. AdaBoost the number of iterations is set, above-mentioned steps s2 to step s7 is repeated, is changed until terminating AdaBoostIn generation, generates a series of integrated classifiers with ballot weight；

S10. for remote sensing image to be sorted, a series of integrated classifiers with ballot weight is utilized respectively and are treated pointThe remote sensing image of class is classified, finally by all classification results of the method integration of Nearest Neighbor with Weighted Voting.

Preferably, the sorting algorithm include C4.5 decision tree, support vector machines, artificial neural network, naive Bayesian,K arest neighbors, Logistic recurrence, minimum distance method, expectation maximization, maximum likelihood method, mahalanobis distance method and random forest；

Multiple and different sorting algorithms is arbitrarily chosen in the step s3 from above-mentioned sorting algorithm for training base to classifyDevice.

Preferably, in step s4, the detailed process of class weight vector is obtained are as follows:

If each base classifier passes through class weight vector W for different classes of susceptibility_ijTo indicate；

If base classifier collection M={ M₁,M₂,M₃... Ms }, S is base classifier number；Sample set X={ X₁,X₂,X₃,…X_N, N is the number of sample；Classification collection Ω={ ω₁,ω₂,ω₃,…,ω_C, C is the number of classification；

Base classifier M is set again_iClass weight vector be W_i, 1≤i≤S, t_ijFor base classifier M_iTest sample is divided intoω_jThe number of class, e_ijFor base classifier M_iTest sample mistake is divided into classification ω_jNumber, 1≤j≤C；

Then base classifier M_iTo classification ω_jWeight expression are as follows:

W_ij=1-E_ij (1)

Therefore, base classifier M_iClass weight vector expression for identification test sample of voting are as follows:

W_i=(W_i1,W_i2,W_i3,...,W_ic) (3)。

Preferably, if M ' expression integrated classifier, x indicates sample；Then classification results of the integrated classifier M ' to sample xAre as follows:

Wherein, M ' (x) indicates that integrated classifier M ' acts on the classification results of sample x；

W_ijFor base classifier M_iFor the weight of j-th of classification；

M_i(x) base classifier M is indicated_iAct on the classification results of sample x.

Preferably, in step s7, detailed process is as follows for the sample weights in update sample set:

It improves by the weight of the sample of mistake classification, is chosen as training sample in AdaBoost iteration next time to increaseProbability.

The present invention has the advantage that

The sample of extraction is divided into training sample and test sample first by the present invention, then using the training of different classifications algorithmClassifier later again classifies to test sample with these classifiers, obtains that classification can be made to be adaptive to different classifications deviceClass weight vector realize point of different classifications algorithm integrate integrated classifier to different classifiersClass device, which is realized, to have complementary advantages.On this basis, using the method for AdaBoost iteration, the integrated classifier of generation is promotedObtain that a precision is higher, the better integrated classifier of integrated result.The method achieve adaptive between classification and classifierIt answers, enhances the diversity of base classifier during integrated different classifications algorithm, ensure that having for classification of remote-sensing images precisionEffect is promoted.

Detailed description of the invention

Fig. 1 is the principle frame of the integrated study Remote Image Classification based on class weight vector in the embodiment of the present inventionFigure；

Fig. 2 is that the process of the integrated study Remote Image Classification based on class weight vector in the embodiment of the present invention is shownIt is intended to.

Specific embodiment

With reference to the accompanying drawing and specific embodiment invention is further described in detail:

As shown in Figure 1, technology path of the invention is as follows:

First layer and the second layer be generate different classifications algorithm base classifier (base classifier is the base of Multi-classifers integratedPlinth classifier is the integrated main object of the present invention), and an integrated classifier is generated by the method for class weight vector.

Integrated classifier, i.e. Ensemble Learning based on Weight Vector, abbreviation EL_WV classificationDevice.

In integrated classifier, if each base classifier passes through class weight vector W for different classes of susceptibility_ijComeIt indicates.

If base classifier collection M={ M₁,M₂,M₃... Ms }, S is base classifier number；Sample set X={ X₁,X₂,X₃,…X_N, N is the number of sample；Classification collection Ω={ ω₁,ω₂,ω₃,…,ω_C, C is the number of classification.

Base classifier M is set again_iClass weight vector be W_i, 1≤i≤S, t_ijFor base classifier M_iTest sample is divided intoω_jThe number of class, e_ijFor base classifier M_iTest sample mistake is divided into classification ω_jNumber, 1≤j≤C.

Then base classifier M_iTo classification ω_jWeight expression are as follows:

W_ij=1-E_ij (1)

Therefore, base classifier M_iClass weight vector expression for identification test sample of voting are as follows:

W_i=(W_i1,W_i2,W_i3,...,W_ic) (3)。

Based on this, the solution procedure of available integrated classifier:

1. extracting training sample set D from sample set D_trainWith test sample collection D_test；

2. i is recycled to S from 1, step 2-1 and step 2-2 is executed；Wherein, i is base classifier M_iIndex；

2-1. uses training set D_trainTraining base classifier M_i；

2-2. test set D_testIn sample calculate base classifier M_iClass weight vector W_i。

3. setting M ' expression integrated classifier, x indicates sample；Then classification results of the integrated classifier M ' to sample x are as follows:

Wherein, M ' (x) indicates that integrated classifier M ' acts on the classification results of sample x；

W_ijFor base classifier M_iFor the weight of j-th of classification；

M_i(x) base classifier M is indicated_iAct on the classification results of sample x.

Third layer is collected by way of Nearest Neighbor with Weighted Voting using the multiple EL_WV classifiers of AdaBoost grey iterative generationAt obtaining ELM_CWV classifier (Ensemble Learning Method with the use of Class basedWeight Vector)。

Before introducing the detailed process step in the embodiment of the present invention, AdaBoost method is introduced first.

AdaBoost is that each sample assigns identical weight first, and the size of weight is for determining that the sample is selected to instructionPractice the probability concentrated and generate base classifier.Then, training sample is extracted from sample set in a manner of weighted sample and generates weak pointClass device classifies to all samples using Weak Classifier, calculates the error of current class device, calculates separately further according to errorThe weight of each sample keeps the sample weights by mistake classification higher.In this way, in next round iteration, the training sample of selectionOriginally it focuses more on those to be easy by the sample of mistake point, newly-generated classifier will give these samples higher concern.

And so on, by successive ignition, a series of Weak Classifier can be generated.For some unknown entity, ownWeak Classifier all classify to it, finally by all classification results of the method integration of Nearest Neighbor with Weighted Voting.

AdaBoost has derived many versions, and there are mainly two types of most typical algorithms: AdaBoost.M1 andAdaBoost.M2.AdaBoost.M1 is mainly used for multi-class classification, and AdaBoost.M2 is then mainly used for two-value classification.OneAs in the case of, classification of remote-sensing images belongs to more classification problems, therefore application of the AdaBoost in remote sensing automatic interpretation is mainIt is based on AdaBoost.M1.AdaBoost according to the present invention is primarily referred to as AdaBoost.M1, and specific algorithm is as follows:

Assuming that sample set S={ (x₁,y₁),(x₂,y₂),…,(x_n,y_n)}(x_p∈X,y_p∈Y)。

Wherein, n is the total number of sample, and X and Y respectively represent the feature space and class label of sample, and K is time of iterationNumber, the weight of W (p) representative sample p, p=1,2 .., n.The detailed process of AdaBoost algorithm are as follows:

Input: S；K；Weak typing algorithm WeakLearn；

Training:

1. by the weights initialisation of the sample p in S: W (p)=1/n；

2. working as q=1 ..., when K, following steps are executed:

3. having the method for putting back to sampling by weighting, training sample is obtained from S, calls one point of WeakLearn trainingClass predicts h_q:X→Y；H is calculated according to formula (5)_qError ε_q:

4. if ε_qThe weight of sample p, W (p)=1/n then go to step 3 in > 0.5, initialization sample collection S；

5. calculating β_q=ε_q/(1-ε_q), and W is updated according to following formula_q+1(p):

Wherein, Z_qFor the normalization factor that the sum of all sample weights are 1 can be made；

6. calculating λ_q=log (1/ β_q) it is used as h_qBallot weight；

7. terminating.

Classification: according to formula (7), the final prediction of the entity x of unknown classification is obtained by way of Nearest Neighbor with Weighted Voting.

As shown in Fig. 2, being carried out specifically to the integrated study Remote Image Classification based on class weight vector belowIt is bright:

S1. according to the basic step of AdaBoost, the weight of sample in initialization sample collection D.

S2. using from sample set D has the method sample drawn D for putting back to sampling_i, the sample of extraction is divided into training sample D_iaWith test sample D_ib；Wherein, D_iaFor the training of base classifier, D_ibFor determining class weight vector.

S3. the base classifier of a variety of different classifications algorithms is respectively trained using training sample.

Wherein, sorting algorithm includes C4.5 decision tree, support vector machines, artificial neural network, naive Bayesian, K nearestNeighbour, Logistic recurrence, minimum distance method, expectation maximization, maximum likelihood method, mahalanobis distance method and random forest etc..

Multiple and different sorting algorithms is arbitrarily chosen from above-mentioned sorting algorithm for training base classifier.

Sorting algorithm used in the present embodiment for example can be C4.5 (i.e. C4.5 decision tree), SVM (i.e. supporting vectorMachine), four kinds of ANN (i.e. artificial neural network) and NB (i.e. naive Bayesian).

Use D_iaC4.5, SVM, ANN, the base classifier C4.5 of NB is respectively trained_i, SVM_i, ANN_i, NB_i。

S4. base classifier C4.5 is utilized_i, SVM_i, ANN_i, NB_iTo test sample D_ibClassify, obtains each base classificationThe class weight vector of device, specific calculating process hereinbefore it is stated that.

S5. each base classifier is integrated based on class weight vector, obtains integrated classifier EL_WV_i。

S6. integrated classifier EL_WV is utilized_iClassify to the sample in step s2 sample set D, obtain error in classification,If error in classification is less than setting error threshold, s7 is thened follow the steps, it is no to then follow the steps s8.

It is 0.5 that error threshold is normally set up in the present embodiment, when error in classification meets condition less than 0.5.

S7. the sample weights in sample set are updated, that is, are improved by the weight of the sample of mistake classification, to increaseIt is chosen as the probability of training sample in AdaBoost iteration next time, updated sample is put back in the sample set in step s2.

S8. s1 is returned to step, the weight of sample in sample set D is reinitialized.

S9. AdaBoost the number of iterations is set, above-mentioned steps s2 to step s7 is repeated, is changed until terminating AdaBoostIn generation, generates a series of EL_WV integrated classifiers with ballot weight.

S10. for remote sensing image to be sorted, a series of integrated classifier EL_WV with ballot weight are utilized respectivelyClassify to remote sensing image to be sorted, finally by all classification results of the method integration of Nearest Neighbor with Weighted Voting.

Certainly, there are four types of the sorting algorithm in the present embodiment is not limited to, can also be certainly three kinds, five kinds even moreIt is more.In addition, being also not limited to the combination of above-mentioned four kinds of sorting algorithms, the combination of other sorting algorithms can also be.

For the present invention by integrating the base classifier of different classifications algorithm, the advantage realized between different classifications algorithm is mutualIt mends.

And the present invention then expands the scale of base classifier, enhances the multiplicity of base classifier by AdaBoost iterationProperty, realize it is adaptive between classification and base classifier, reached efficiently, high-precision classification of remote-sensing images purpose.

Certainly, described above is only that presently preferred embodiments of the present invention is answered the present invention is not limited to enumerate above-described embodimentWhen explanation, anyone skilled in the art is all equivalent substitutes for being made, bright under the introduction of this specificationAobvious variant, all falls within the essential scope of this specification, ought to be by protection of the invention.

Claims (5)

1. the integrated study Remote Image Classification based on class weight vector, which comprises the steps of:

S1. according to the basic step of AdaBoost, initialization sample concentrates the weight of sample；

S2. the sample of extraction is divided into training sample and test specimens using there is the method sample drawn for putting back to sampling from sample setThis；

S3. the base classifier of a variety of different classifications algorithms is respectively trained using training sample；

S4. classified using the base classifier in step s3 to test sample, obtain the class weight of each base classifier toAmount；

S5. each base classifier is integrated based on class weight vector, obtains integrated classifier；

S6. classified using the integrated classifier in step s5 to the sample in step s2 sample set, obtain error in classification, ifError in classification is less than setting error threshold, thens follow the steps s7, no to then follow the steps s8；

S7. the sample weights in sample set are updated, updated sample is put back in the sample set in step s2；

S8. s1 is returned to step；

S9. AdaBoost the number of iterations is set, above-mentioned steps s2 to step s7 is repeated, until terminating AdaBoost iteration,Generate a series of integrated classifiers with ballot weight；

S10. for remote sensing image to be sorted, a series of integrated classifiers with ballot weight are utilized respectively to be sortedRemote sensing image is classified, finally by all classification results of the method integration of Nearest Neighbor with Weighted Voting.

2. the integrated study Remote Image Classification according to claim 1 based on class weight vector, feature existIn,

The sorting algorithm include C4.5 decision tree, support vector machines, artificial neural network, naive Bayesian, K arest neighbors,Logistic recurrence, minimum distance method, expectation maximization, maximum likelihood method, mahalanobis distance method and random forest；

Multiple and different sorting algorithms is arbitrarily chosen in the step s3 from above-mentioned sorting algorithm for training base classifier.

3. the integrated study Remote Image Classification according to claim 1 based on class weight vector, feature existIn,

In the step s4, the detailed process of class weight vector is obtained are as follows:

If each base classifier passes through class weight vector W for different classes of susceptibility_ijTo indicate；

If base classifier collection M={ M₁,M₂,M₃... Ms }, S is base classifier number；Sample set X={ X₁,X₂,X₃,…X_N, N isThe number of sample；Classification collection Ω={ ω₁,ω₂,ω₃,…,ω_C, C is the number of classification；

Base classifier M is set again_iClass weight vector be W_i, 1≤i≤S, t_ijFor base classifier M_iTest sample is divided into ω_jClassNumber, e_ijFor base classifier M_iTest sample mistake is divided into classification ω_jNumber, 1≤j≤C；

Then base classifier M_iTo classification ω_jWeight expression are as follows:

W_ij=1-E_ij (1)

Therefore, base classifier M_iClass weight vector expression for identification test sample of voting are as follows:

W_i=(W_i1,W_i2,W_i3,...,W_ic) (3)。

4. the integrated study Remote Image Classification according to claim 3 based on class weight vector, feature existIn,

If M ' expression integrated classifier, x indicates sample；Then classification results of the integrated classifier M ' to sample x are as follows:

Wherein, M ' (x) indicates that integrated classifier M ' acts on the classification results of sample x；

W_ijFor base classifier M_iFor the weight of j-th of classification；

M_i(x) base classifier M is indicated_iAct on the classification results of sample x.

5. the integrated study Remote Image Classification according to claim 1 based on class weight vector, feature existIn,

In the step s7, detailed process is as follows for the sample weights in update sample set:

It improves by the weight of the sample of mistake classification, is chosen as the several of training sample in AdaBoost iteration next time to increaseRate.