| With the development of spectral imaging technique,the technology of remote sensing based on hyper-spectral data has developed rapidly.From the literal meaning,remote sensing means the perception from far away.Broadly speaking,it is the technology to detect the target within a certain distance without touching.From the narrow sense,it means a kind of method to explore the surface of earth by the electromagnetic wave from sensors on the platform of ground,space craft and aviation.Remote sensing acquires the information by the mutual effect between the signal and ground,and then processing the acquired information for further applying and analyzing.Compared to the normal image,the data provided by hyper-spectral image has more specific feature:(1)abundant approximate continuous spectral information of ground objects;(2)the higher ability in recognizing.(3)the agility in the method of recognizing(4)the possibility of classifying of certain element.At the same time,The hyper-spectral image contains more information,has higher dimension of data and correlation between bands,these are the challenges in the classification of hyper spectral image.Traditional method of classification can be divided into supervised and non-supervised.Before the supervised classification,we need to measure the objects that we are interested in by field research,or to select the classical samples to train,so this kind of classification needs a big amount of prior knowledge,the consequence depends greatly on the quality and the number of the training samples and the process to acquire mass prior knowledge cost too much manpower and material resources.Non-supervised classification doesn't need prior knowledge,it only depends on the spectral similarity and diversity among samples,thus the consequence is not satisfied usually.The usual supervised and non-supervised classification utilizes the spectral information in the hyper-spectral image only but ignores the information of spatial feature,which led to the waste of information and the drop of the accuracy of classification.Semi-supervised classification is a kind of method belongs to the field of machine learning,which combines the advantage of supervised classification with non-supervised classification,utilizes a small quantity of labeled samples and a big amount of unlabeled samples to reduce the complexity in classifying.It could reduce the time in classification andimprove the sufficiency of classifying to have a better result.As a consequence,this thesis pays more attention to the research of semi-supervised classification.K-means clustering is a kind of target clustering method based on distance,which calculates the distance between samples and the centroid?K-means has the fast speed in iteration and convergence,it could adapt to the large-scale data set with a big amount of samples.However,K-means clustering classifies all the samples without the reduction of the number of samples.Accordingly,this thesis came up with a method to improve the traditional k-means algorithm,using the samples that are near to the center of the whole samples instead of the whole samples to achieve the aim of samples reduction.This thesis adopts the Least Square Twin Support Vector Machine(LSTSVM)based on the one-against-rest classifying structure to classify the hyper-spectral image.During the research of clustering,this thesis puts forward the further improvement based on this theory by using the samples that are near to the center to replace the whole data set,so that we could control the rate of reduction by controlling the number of clustering center.To achieve the aim of reduction,this method is used in the negative samples to reduce the time consumed in classification and improve the efficiency.And then,the paper focuses on the spatial information includes in the hyper-spectral image to improve the accuracy in classifying.The results of experiment show that the method in this paper can greatly improve the performance of classification. |
PDF Full Text Request