This paper presents a region based approach for doing image classification. An attempt is made to use both spectral and spatial information of an input image. The procedure first divides the image into several groups (in the spectral space) based on a selected sample set using the Mahalanobis distance as a measure of similarity. The result is then segmented into regions using region growing segmentation. The region growing process also incorporates edge information to avoid growing over vital inter-region boundaries. Each region is defined by a number of measurements or attributes such as the unique ID, the list of all its member pixels, the mean intensity and covariance matrix of the spectral values in that region.

The extracted regions are classified using a minimum distance decision rule. Several regions are selected as training samples for region classification. Each region is compared to the training samples and is assigned to its closest class.

The procedure has been implemented using MathWorks MATLAB software. It has been tested on a number of images including Landsat TM, ASTER and ADS40 imagery. The procedure significantly reduces the mixed pixel problem suffered by most pixel based methods.