Unsupervised image classification for remotely sensed imagery is very challenging due to the fact that the unknown image background generally varies with a wide range of spectral deviations. Additionally, spectral similarity among subtle small calsses also causes tremendous difficulty in classification. This paper investigates three major issues, (1) image background removal, (2) generation of training sample data, (3) determination of the number of classes to be classified, p which are encountered in unsupervised image classification. The study on these three issues is conducted via a well-known Airborne Visible InfraRed Imaging Spectrometer (AVIRIS) image scene, Indiana Pine test site available online at Purdue University's website. Since image background varies with different applications, it is generally difficult to perform background removal without prior knowledge. In order for unsupervised classification to be effective, a good set of training data is also necessary. These training samples must be generated directly from the image data in an unsupervised manner. This paper develops an unsupervised training sample generation algorithm (UTSGA) that can generate a good sample pool of training data for supervised classification. In determining p, a newly developed concept, called virtual dimensionality (VD) is used to estimate the p where a Neyman-Pearson-based eigen-analysis approach developed by Harsanyi, Farrand and Chang, called noise-whitened HFC (NWHFC)'s method, is implemented to find the VD to be used for the p. Finally, an unsupervised image classification algorithm can be derived by implementing a supervised classifier in conjunction with teh UTSGA algorithm and NWHFC's method.