Rough set-derived measures in image classification accuracy assessment

Abstract

Currently, there are two types of measure in image classification accuracy assessment: pixel-level measures and category-level/map-level measures. These have their own limitations for representing the uncertainty at pixel and category/map levels. In addition, some of these measures derived from the error matrix are obtained by collecting reference data and then they may be affected by factors related to the sampling. This paper uses rough set theory to obtain the rough degree, rough entropy, quality of approximation and accuracy of approximation. Incorporating traditional measures, they compose one kind of three-level architecture for the classified image, which contains pixel-level measures, object/category-level measures and map-level measures. Unlike some conventional measures, these new measures can be derived directly from the supervised classification result without collecting reference data. A case study on the Landsat TM image is used to substantiate the conceptual arguments. The results demonstrate that the proposed measures are valid for measuring the accuracy of classified remotely sensed imagery and can provide additional information to conventional measures.

Acknowledgments

This research was supported in part by the National Natural Science Foundation of China (Grant No. 40671136) and the Open Research Fund from State Key Laboratory of Remote Sensing Science in China (LRSS0610), and the National High Technology Research and Development Program of China (Grant No 2006AA120106). Acknowledgement is given to Mr Jianghao Wang for his help on processing the SPOT 2 imagery. The authors are grateful to three anonymous referees for their constructive comments, which helped to improve the quality of the paper. Special thanks to Prof. Giles Foody who has given much encouragement and useful comments, and especially for his patience while this paper was produced.