A Remote Sensing Feature Discretization Method Accommodating Uncertainty in Classification Systems
Guifeng Zhang, Zhaocong Wu and Lina Yi
School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079, China
Abstract. Most of the classification methods in remote sensing can only process the discrete feature data, such as rough set. Thus the discretization of feature plays a very important role in the remote sensing imagery classification system. In general, the remote sensing features are discretized currently by means of the methods from fields other than remote sensing. Because there is a lack of consideration of uncertainty of the classification system in these methods, it is not predicted whether the discretization influences the classification accuracy or not. This paper introduces a discretization method considering the uncertainty of the classification system. It comprises three components: the building of the initial candidate cut points set, the selection of cut points based on information entropy and the deletion of redundant cut points. All the three parts are executed iteratively. The first two are iterative processes from top to bottom, while the last is an iterative process from bottom to top. The stopping criterion of iterative process is a threshold which represents the max possible change of the uncertainty of the classification. Therefore, the change of the uncertainty of the classification system resulted from discretization is limited to a certain range, and its influence on classification can be predicted and controlled. The experiment shows that the proposed method produces comparative results with those of Ent-MDLC and can lessen the influence on the classification accuracy from the discretization.
Keywords: uncertainty, discretization, segment range, rough entropy, accuracy
In: Wan, Y. et al. (eds) Proceeding of the 8th international symposium on spatial accuracy assessment in natural resources and environmental sciences, World Academic Union (Press).