Cluster Analysis Based On Geometry Probability

Cluster analysis is one kind of multiple statistical analysis, and an important branch of non-supervising classifying either, it collects a sample without classification mark according to a certain criterion to divide into several subclasses, what is more, it makes the similar sample belong to the same class as much as possible, and the dissimilar sample tries one's best to divide to the different kind. At present, this method has already been widely applied to many fields such as biology, climatology, economics, remote sensing etc. Its purpose lie in distinguishing different things and knowing similative among things. So, the research of cluster's analysis has important meanings. Cluster analysis to pass through several decades development, has formed the extremely huge content system, and has been realized many algorithms. But there are at least two disadvantages in the current situation of studying cluster analysis: First, it must assign class numbers in advance before the classification, and it is possible to occur that assigning class numbers in advance does not tally with actual existence class of the research object, which affects cluster result; Second, A lot of incorrect cases of classification exist in the classification outcome directly derived from the currently available cluster analysis methods. Centering on these disadvantages of cluster analysis, this paper combining with the Bayes theory proposes the new cluster methods—— cluster analysis based on geometry probability. Remotely sensed data belong to complicated large samples. A lot of incorrect cases of classification exist in the classification outcome directly derived from the currently available cluster analysis methods, which are usually inferior to supervised classification methods in terms of accuracy. This paper proposes a cluster analysis approach based on geometric probability, which determines the classification number directly according to the distributive characteristics of pixels in the spectral space by utilizing geometric probability and gradually generates the hierarchical classification system beginning at the top level. The experiment indicates this approach can obviously improve the accuracy of classification of remotely sensed images.