Research On Ensemble Classification Method For Hyperspectral Remote Sensing Image Based On SVM

With the rapid development of remote sensing technology in the direction of "three comprehensive"(all-weather,all-day,global observation),"three high"(high spatial resolution,hyperspectral resolution,high time-phase resolution)and "three more"(multisensor,multiplatform,multiangles),people get more and more remote sensing information.However,the current used level of remote sensing information is far behind the development of remote sensing technology.Therefore,it is very important to study new theories and methods to analyze and process remote sensing information.In the field of hyperspectral remote sensing classification,data redundancy and dimensionality disaster are existed,and dimensionality reduction is an important preprocessing method.Because hyperspectral image has nonlinear characteristics,manifold learning may better excavate nonlinear structure of hyperspectral datas and improve the performance of data analysis.Support Vector Machine(SVM)is one of the representative algorithms based on kernel transformation technology,and has the advantages of high precision,fast operation speed and strong generalization ability.It is widely used in all aspects of remote sensing data processing.On this basis,in view of practical problems in classification of hyperspectral remote sensing images and substantive characteristics of image datas,manifold learning,twin support vector machine,semi-supervised learning and ensemble learning are introduced for classification of hyperspectral remote sensing image.The main research work is as follows:(1)The sample credibility model of hyperspectral image data is constructed and reduction method based on robust supervised feature isometric mapping is studied.In order to solve the problem of noise interference in hyperspectral remote sensing data set,the credibility of hyperspectral image samples is defined,and then a sample credibility model of hyperspectral image data is constructed.Weighted iterative least square method is used and robust estimation is introduced when weight function is constructed.A data sample credibility algorithm is proposed to compute the probability that the sample is singular value or noise.With aiming at the problems of high spectral resolution,large number of bands,high correlation between bands,data redundancy,dimensionality disaster and image noise in hyperspectral remote sensing images,local information,category information and credibility information of data samples are embedded to define the similarity measure based on triple geodesic distance.Fitting property of triple geodesic distance is proved.The new shortest distance matrix is calculated,and the embedded coordinates of training samples and test points are constructed by using multidimensional scale transformation and generalized regression neural network respectively.The method and flow chart of robust supervised feature isometric mapping are proposed.(2)Measure method of geometric mean membership and measure method of inter-class separation by class distribution are proposed.The separation characteristics of hyperspectral remote sensing image data after dimensionality reduction are described by the perspective of in-class relations and inter-class relations.By using the distance between the image data sample and the geometric center of the category,and the degree of tightness between samples,geometric mean fuzzy membership measurement method is defined.The performance of geometric mean fuzzy membership measurement method is analyzed.According to the distance between the image data class centers and the class distribution,inter-class separation metrics by class distribution is constructed.The effective basis for constructing the hyperspectral image classifier with excellent performance is provided by the separation characteristics.(3)The method of hyperspectral image classification based on fuzzy twin support vector machine is studied.With aiming at the problems about large amount of datas and data noise in hyperspectral remote sensing image,a linear fuzzy twin support vector machine model and a nonlinear fuzzy twin support vector machine model are proposed by different sample with different fuzzy membership.The linear multi-class classifier based on fuzzy twin support vector machine and the nonlinear multi-class classifiers based on fuzzy twin support vector machine are studied.(4)The method of hyperspectral image classification with decision tree based on inter-class separation metrics by class distribution is studied.With aiming at the structural defects of hyperspectral remote sensing decision tree S VM,an adaptive(or better)decision tree is built by using genetic algorithm with inter-class separability measure based on class distribution.Further studies are made in the active node classifiers of SVM and KNN(K-Nearest Neighbor).The basic framework and algorithm of decision tree are constructed.(5)The method of hyperspectral image classification with semi-supervised collaborative learning based on spectral spatial features is studied.With aiming at the problems about small amount of samples and "different spectra with the same object,different object with the same spectra" for hyperspectral images,posteriori probability output of binary classification SVM and multi-classification SVM are ratiocinated.An extended texture feature extraction method based on gray level co-occurrence matrix is proposed,and then the spectral feature and spatial texture feature of hyperspectral image are combined under the framework of cooperative learning.The cooperative learning algorithm based on spectral feature and spatial texture feature is constructed by using probabilistic output SVM as base classifier.(6)A multi-classifier ensemble method based on weighted voting by geometric mean accuracy is proposed.With aiming at complex and diverse characteristics of hyperspectral images,ensemble learning framework of hyperspectral remote sensing image classifier is studied.According to the advantages of Kappa coefficient and user accuracy in the hyperspectral image classification confusion matrix,the geometric mean accuracy is defined.On the basis of the classical voting method,with taking advantage of the differences and complementarities among hyperspectral image classifiers,a multi-classifier ensemble model based on weighted voting by geometric mean accuracy is constructed,and the process of the multi-classifier ensemble algorithm is designed.There are 58 figures,31 tables and 152 references.