Remote sensing images, in particular those having coarse spatial resolution are fraught with mixed pixels (a pixel containing more than one class). Conventional per pixel based classification approach results in erroneous classification since the noise in the form of mixed pixels is ignored. This necessitates the use of fuzzy set based methods, which assign class membership values to the pixels to produce a sub-pixel classification. Fuzzy c-means (FCM) clustering has generally been adopted for the classification of remote sensing data. However, the major limitation of this method is the probabilistic sum to one constraint, which may distort the actual class membership values. In possibilistic c-means (PCM) clustering, this constraint is relaxed, which may provide meaningful class membership values. In this paper, we report a comparative study of these two fuzzy set based methods to produce accurate sub-pixel classification of a remote sensing image acquired from IRS LISS II sensor. The results show that PCM has produced significantly higher classification accuracy than FCM in the presence of mixed pixels (i.e., noise) in the dataset.