Sometimes the problem of data classification arises without any knowledge of the underlying classes. Furthermore, not enough sample data from the classes may be available to develop reliable parametric estimates and, hence, data classification rules. However, when the size of the mixed data to be classified is large, computer aided techniques for displaying data with histograms and frequency tables can be exploited to compensate partially for the lack of class knowledge and sample data. Based on this concept, a numerical classification procedure is developed and is described in this paper. The classification rule is completely data directed and utilizes the data marginal distributions in a sequential manner. The technique was employed for partitioning certain remotely sensed multispectral scanner data for a number of sites in the U.S. and was found to be fairly competitive when compared to some commonly used parametric approaches utilizing training sample data.