A Comparison of Mamdani and Sugeno Inference Systems for a Satellite Image Classification

Abstract

This research provides a comparison between the performances of Sugeno type versus Mamdani-type fuzzy inference systems. The main motivation behind this research was to assess which approach provides the best performance for satellite image classification. The performance of each approach has been evaluated for six bands (from Landsat-5) for West Iraq image classification and compared with traditional method (Maximum likelihood), based on pixel-by-pixel technique. Due to the importance of performance in online systems we compare the Mamdani model, used previously, with a Sugeno formulation using four types of membership function (MF) generation methods. The first method triangular membership function using the mean, minimum and maximum of the histogram attribute values. The second approach generates triangular membership function using the peak and the standard deviation of attributes values. The third procedure generates Gaussian membership function using the mean and the standard deviation of the histogram attributes values. The fourth approach generates Gaussian membership function using the peak and the standard deviation of the histogram attributes values. The results show that the Mamdani models perform better in most of the case under study