A COMPARATIVE ASSESSMENT OF SOME UNSUPERVISED CLASSIFICATION ALGORITHMS

Abstract

With the advent of advances in technology, Data Acquisition has increased many manifolds. The modern systems used today for acquisition and storage are capable of collecting the data in Petabytes rather than in Gigabytes. This continuous collection of data adds on to the huge volumes of data stored in libraries/databases, the size of which has grown with each passing day. In such a scenario data, analysis becomes an impossible task for human to analyze data of each and every moment. Thus there arises a need of automated data analysis tool, which reduces burden of human beings, of deriving meaningful information of.all the data collected. One such big repository of data is being formed by the remote sensing data, sent by the satellites; which provides rich and valuable information of the earth surface.. Such repository forms spatial databases which gives a pictorial representation to data analysis. Therefore, automated knowledge discovery has become important in spatial databases. One of the vital means of dealing with these is to classify or group them into a set of categories or clusters. Cluster analysis is a primitive exploration technique which with little or no prior knowledge, are able to perform the classification process with minimum human intervention. Clustering algorithms provides important methods of unsupervised class identification in spatial databases. However, not many algorithms have yet been explored and incorporated in the image processing software(s) currently being used by the industry. In this dissertation, remote sensing data is assessed by applying DENCLUE (density-based) and STING (grid-based) clustering algorithms. A relative comparison is made between the outcomes the K-Means/ISODATA, DENCLUE and STING algorithms