Classification Of Remote Sensing Image By Spectral Graph Theory

With the continuous development of technology,the quality of the remote sensing image data is more and more high.Remote sensing information technology has got unprecedented development and concern,fully processing analysis of remote sensing image data to obtain useful information for human is the hot topic of research,and the important step of the research is the classification of image.The main explore content of this article is in view of the image classification by spectral graph theory.Supervised classification and unsupervised classification are the most famous classification.This paper respectively discusses the maximum likelihood,BP neural network and support vector machine(SVM)that several kinds of common remote sensing image classification methods.And a comparative experimental study is carried out on the BP neural network,support vector machine(SVM)maximum likelihood and spectral clustering algorithm.The spectral theory of graph is not only an important field in graph theory but also an active topic.The spectral theory of graph mainly include the adjacency spectrum,Laplace-spectrum,Q-spectrum,C-spectrum,S-spectrum,which adjacency spectrum and Laplace-spectrum are most universal.In many applications,good upper bounds for the largest Laplacian eigenvalues of a graph G are needed.Spectral clustering algorithm in the spectral graph theory is based on eigenvectors.The first step of the process is to construct the similarity matrix of the graph and calculate the Laplace matrix.Then get the eigenvalues and eigenvectors of the Laplacian matrix.At last,complete the classification of the corresponding data for the selected eigenvectors which is needed.Through comparative tests conclude that the method based on spectral clustering is not only the highest precision,and calculate the smallest time,more suitable for large amount of data,high dimension,strong correlation between adjacent bands of remote sensing image classification.But the spectral clustering method is sensitive to outliers more,so you need to determine the K value in advance,and prior to do more accurately determine eigenvalues eigenvector,division and comparison of spectrum accurately.