EM Algorithm And Its Application Of Remote Sensing Classification

Imaging process of remote sensing image subjects to random changes in various factors, resulting in image data obtained with a certain degree of randomness. remote sensing information has some degree of definite statistical characteristic, but as well as holds the high randomness and complexity, which generally behaves as mixture density distribution in feature space. Difficulty in mixed model using for classification is obtaining parameters.It is very difficult to solve the parameters by Newton-Raphson and scoring algorithms. EM algorithm by Dempster is a good solution to this problem, introducing a complete data model label likelihood problem. Two problems would be encountered when we directly use EM method to classify multi-spectral data, the first one is the singular variance-covariance matrix, which could generate incorrect results, and the other is the sensitivity of initial value, because different initial value would lead to different final results. This paper summaries the current status of Remote sensing classification and feature extraction, refers applications of EM in many fields,makes a more depth study of EM algorithm in image classification in remote sensing. Completed the major work and results achieved are as follows:1) This paper describes basic remote sensing image classification methods basing on statistical analysis, analysis of a variety of classification methods and characteristics, summarizes research results of remote sensing image classification.2) This paper describes mixed model and EM algorithm, derives the EM algorithm in the process of mixing model parameters in detail, analysis problems resulting by using directly EM algorithm in remote sensing image classification.3) For problems of EM algorithm, the algorithm proposed is based on Log-Principal Component Transform. multi-spectral data is logarithmically transformed firstly, and then implement principal component transformation. We analyze the histogram of the first principal component to determine the total number of initial classification. Through K-means method, we determine the tags of initial classification, and further adopt EM method to do iterative classification to avoid the singular variance-covariance matrix. Experiments show that the accuracy of the algorithm proposed is higher than ordinary EM method and traditional K-means.