Unsupervised classification

In unsupervised classification, statistical approaches are applied to image pixels to automatically identify distinct spectral classes in the image data. These classes are usually referred to as clusters because two or more of these may represent a single land cover class that may display high spectral heterogeneity. On the other hand, one cluster may represent two or more land cover classes. For this reason, results of unsupervised classification routines must sometimes be further processed to merge or split clusters. This type of techniques does not require prior knowledge of the exact number of classes in the area of interest. What is required is a specification of the number of clusters (or a range) to be identified, and then the classifier automatically aggregates the image pixels into the required clusters by minimizing some predefined error function (Tso and Mather, 2009). K-means and the Iterative Self Organizing Data Analysis (ISODATA) technique are some of the more widely used algorithms (Jensen, 1996; Mather, 2004). Further details on these techniques can be found here.