An unsupervised classification method provides the interpretation, feature extraction and endmember estimation for the remote sensing image data without any prior knowledge about the ground quality. We explore such method and construct an algorithm based on the non-negative matrix factorization (NMF). The use of the NMF is to match the non-negative property in sensing spectrum data.. The data dimensionality is estimated by using the partitioned noise-adjusted principlal component analysis (PNAPCA). The initial matrix used to start the NMF is obtained by using the fuzzy c-mean (FCM). This algorithm is capable to produce a region- or part-based representation of objects in images. Both simulated and real sensing data are used to test the algorithm.