Tolerant Rough Set on the uncertainty of Satellite Remote Sensing Data classification

摘要

Using satellite remote sensing data to extract land cover information has a broad implication to various applications. However, several factors have been posing difficulties for accurate classification. One problem is that different targets in study sometimes share similar spectral identities, due to either the complexity of natural landscape or the limitation of remote sensing spectral resolution or the imperfect preparation of training sets. It causes an uncertainty in the classification process. So the need to seek a novel method to handle this uncertainty attracts great attention among researchers worldwide [1]. One methodology studying classification uncertainty is rough set [2, 3]. Traditional rough set method has its drawback in dealing with numerical data which prevails in the remote sensing. One solution is to employ methods of discretization [4]. However, the process of discretizing itself introduces another trouble. As a more advanced solution based on the similarity relation [5], tolerant rough set has been proposed and applied in the remote sensing data classification [6, 7]. However, the proposed method fails to take into consideration of the distribution information of the land cover spectral feature space. In this paper a spectral feature neighborhood based tolerant rough set classification method (SFNTRS)is proposed to handle the uncertainty in the process of satellite remote sensing data classification (See Figure 1 for flowchart)[8]. Experiment is carried out with Landsat-5 TM+ image of eastern Beijing. Classification result is compared with result from the current tolerant rough set method (See Figure 2 for results). Outcome indicates that our method is more interpretable and reliable, and can effectively handle the uncertainty in the process of satellite remote sensing classification. It is a promising tool at classifying areas with complex spectral feature.