Rough set approach is one of effective attribute reduction (also called a feature selection) methods that can preserve the meaning of the attributes(features). However, most of existing algorithms mainly aim at information systems or decision tables with discrete values. Therefore, in this paper, we introduce a novel rough set-based method followed by establishing a fuzzy discernibility matrix by using distance preserving strategy for attribute reduction, and only choose fisher discriminant analysis with kernels as discriminant criteria for testing the effectiveness of selected attribute subsets with relatively higher fitness values, since the proposed method is independent of post-analysis algorithms (predictors). Experimental results show that the classifiers developed using the selected attribute subsets have better or comparable performance on all eight UCI benchmark datasets than those obtained by all attributes. Thus, our newly developed method can, in most cases, get effective attribute subsets. In addition, this method can be directly incorporated into other learning algorithms, such as PCA, SVM and etc. and can also be more easily applied to many real applications, such as web categorization, image recognition and etc.