Unsupervised Classification Using Gravity Centers from Scatter Plot

Abstract

Unsupervised classification creates clusters by grouping pixels based on the reflectance properties of pixels. This paper presents a new approach to classify multi-spectral remotely sensed image using pixels' density in N-dimensional scatterplot. It first finds the densely populated clusters in N-dimensional scatter plot and then finds gravity centers of these densely populated clusters. Later, multispectral image is classified using minimum distance to gravity classifier. At the beginning of classification, this approach neither makes extreme assumption of considering each pixel as a different cluster nor goes to the other extreme by considering all the pixels in a single cluster. It follows the middle path by making assumption of some pixels as gravity centers of different clusters before classifying the image. It creates clusters of equal size in N-dimensional scatter plot and picks up the densely populated clusters. All these clusters are recursively iterated for self-adjustment of gravity centers using mathematical algorithm. The approach uses gravitational force for merging two nearby clusters. Here, gravity center of a cluster is calculated by summing up the spectral bands' values and dividing it by number of pixels within that cluster. When two nearby clusters are merged, their gravity centers are also adjusted accordingly. This provides the most densely populated clusters and their gravity centers. Now, these gravity centers can be used to classify the remotely sensed image using minimum distance to gravity center classifier.