Clustering Algorithm Application In Hyperspectral Imaging Anomaly Detection

Hyperspectral imagery contains rich spectrum information, it can reflect the slightdifferences between the goals of the remote sensing data, and has the ability of divising tinyspectral differences of background. Anomaly detection algorithm can test goals which isspectral different from the surrounding environment without priori information, thus hasstrong practicability, and has now become a hot research topic. In this paper, we use thehyperspectral imagery anomaly detection technology as the research object, by clusteringanalysis method, in order to improve and enhance the performance of hyperspectral imageryanomaly detection, focus on the problems of using the EM clustering algorithm inhyperspectral anomaly target detection. In the following, three aspects of the study has beendiscussed:First of all, with the characteristics of the correlation in hyperspectral data, clusteringalgorithm is used in hyperspectral imagery anomaly detection field. The EM clusteringalgorithm is used to deal with the hyperspectral imagery data, considering the spatialcorrelation between of it; initialization of EM algorithm can influence the final result, so weuse the principal component to initialize EM algorithm, which use two methods, the vectorquantization and estimating the density of principal component. The algorithm make full useof the spectrum characteristics while better give dual attention to the hyperspectral imagespace characteristics at the same time, and obtained a good test results.Secondly, in view of the influence of estimates by the background in the traditionalclassical RX algorithm, This paper use EM algorithm to smooth background by clustering theadjacent area of the pixel under test (PUT). Considering the nature of the clustering, in thehigh dimension space, the pixels with same features has a strong correlation, while them withdifferent features is far away from each other, the paper introduces the algorithm using thecenter of the clustering to replace the vectors of the pixel to reduce the effects of abnormaldata in the background matrix, resulting in better description state of actual backgrounddistribution, in order to remove the noise interference, to smooth detection effect and toimprove the detection effect.Finally, on the basis of analyses the orthogonal subsapce projrction(OSP) algorithm, thepaper and hyperspectral data will be projected on the background orthogonal subsapce, to test the anomaly goals in the new space which is less influence by the background. This paperpropose an extraction algorithm based on EM algorithm and we can select the backgroundendmember according to the number of all the endmembers, then the data projected on thebackground orthogonal subsapce using OSP, and in the new space use RX algorithm and EMalgorithm testing abnormalities, to get a better test results. With AVIRIS hyperspectral data,the simulation experiment has good detection effect.