Research On The Uncertainties In The Classification Of Remote Sensing Image Based On Rough Set

Uncertainty with different types and varying degrees will be produced in some operations of remote sensing data, such as acquisition, processing, analysis, data conversion etc. The uncertainty will be spread in the subsequent processing and finally accumulated. It is necessary to investigate the uncertainty in the process of the classification of remote sensing image, since the classification data of remote sensing image is widely used as an important data source for the research of the land cover, land use and various models in GIS.Rough set is a mathematic approach to uncertainty. As a new inductive learning method, it is widely concerned by its advantages, such as "no priori assumptions on data are needed", "it can provide knowledge acquisition methods for incomplete and inconsistent data" and "easy understandability of knowledge obtained by it".Based on remote sensing image classification, we have investigated the uncertainty measurements in the process of the image classification based on rough set and the influence of discretization methods on the classification result. The following is the main content in this paper:1. In the classification of remote sensing image, we proposed uncertainty measurement indices based on rough set including the quality measurement of the sample data in supervised classification, the uncertainty measurement during the process of rule matching, and so on. The proposed measurements facilitate the research of the inherent uncertainty of the remote sensing image, and conduct all the stages in the classification of remote sensing image, such as, the selection of the discretization algorithms, sample area, classifier, and so on.2. In this paper, we proposed evaluation criteria of the discretization methods to investigate the influence of discretization methods on the result of image classification. These evaluation criteria include the capability of data compression; the proportion of inconsistent information; the compression ratio of the domain of attribute value, and so on. It provides a basis for the selection of suitable discretization in the classification of remote sensing images.3. The remote sensing image of Landsat 5 TM located at the Yellow River Delta is utilized as a case study for verifying the proposed theory and methods in the paper.