This paper analyzes the classification of meteorological images provided by geostationary satellites. Classified datasets are used often to gather spatial statistics efficiently from large areas and to simplify the analysis of complex images. The algorithm uses a modified version of k-means standard method to classify the images and statistical distance between classes in order to reduce and optimize the final partition to be obtained. The method is checked for several meteorological situations, both in summer and in winter. The program was written in C and was tested both in UNIX work stations and in personal computers. In all situations the computer time required to make the classification can be considered acceptable. The methodology presented can be applied directly to the analysis of meteorological images, providing a tool for interpreting complex scenes. Additionally, minor changes introduced in the algorithm allow classification of any image provided that two (or more) channels of the same scene are available.