ENTROPIC DISTANCE BASED K-STAR ALGORITHM FOR REMOTE SENSING IMAGE CLASSIFICATION

Thematic maps produced through the classification of satellite images are main resources for many applications about the Earth's surface. Many methods exist in the literature for remotely sensed image classification, but none is regarded as the standard, mainly due to the underlying assumptions on the sample distribution and requirement of user interaction for their design and parameter selection. In this study, K-star classifier, an instance based classifier using entropic distance measure, is introduced for the classification of remotely sensed images. The classifier has a simple mathematical description with a single parameter (blending parameter) taking values between 0 and 100. In order to validate its use, classification problems are constructed using Landsat TM and Terra ASTER images, of Gebze district of Kocaeli in Turkey. The performance of K-star algorithm was compared with Mahalanobis distance and maximum likelihood classifiers. Statistical significance of classifier performances were thoroughly analyzed using McNemar's test on three data sets. Results confirm the potential of the K-star algorithm in the use of remote sensing image classification.