A variable precision rough set approach to the remote sensing land use/cover classification

  • Authors:

  • Xin Pan;Shuqing Zhang;Huaiqing Zhang;Xiaodong Na;Xiaofeng Li

  • Affiliations:

  • Northeast Institute of Geography and Agricultural Ecology, Chinese Academy of Sciences, Changchun 130012, China and Graduate University of Chinese Academy of Sciences, Beijing 100039, China and Sc ...;Northeast Institute of Geography and Agricultural Ecology, Chinese Academy of Sciences, Changchun 130012, China;Institute of Forest Resources Information, Chinese Academy of Forestry, Beijing 100091, China;Northeast Institute of Geography and Agricultural Ecology, Chinese Academy of Sciences, Changchun 130012, China and Graduate University of Chinese Academy of Sciences, Beijing 100039, China;Northeast Institute of Geography and Agricultural Ecology, Chinese Academy of Sciences, Changchun 130012, China

  • Venue:
  • Computers & Geosciences
  • Year:
  • 2010

Quantified Score

Hi-index 0.00

Visualization

Abstract

Nowadays the rough set method is receiving increasing attention in remote sensing classification although one of the major drawbacks of the method is that it is too sensitive to the spectral confusion between-class and spectral variation within-class. In this paper, a novel remote sensing classification approach based on variable precision rough sets (VPRS) is proposed by relaxing subset operators through the inclusion error @b. The remote sensing classification algorithm based on VPRS includes three steps: (1) spectral and textural information (or other input data) discretization, (2) feature selection, and (3) classification rule extraction. The new method proposed here is tested with Landsat-5 TM data. The experiment shows that admitting various inclusion errors @b, can improve classification performance including feature selection and generalization ability. The inclusion of @b also prevents the overfitting to the training data. With the inclusion of @b, higher classification accuracy is obtained. When @b=0 (i.e., the original rough set based classifier), overfitting to the training data occurs, with the overall accuracy=0.6778 and unrecognizable percentage=12%. When @b=0.07, the highest classification performance is reached with overall accuracy and unrecognizable percentage up to 0.8873% and 2.6%, respectively.