COMPARISON OF CLASSIFIERS IN VARIOUS IMAGE PROCESSING SOFTWARES

Abstract

Thematic map is derived from remote sensing data through a digital image classification procedure. The procedure may supervised or unsupervised. Supervised classification has three distinct stages of analysis: training, allocating and testing. It is the allocating stage where classifier is selected to perform image classification. Various classifiers are available in image processing software. These include minimum distance to means, maximum likelihood, mahalanobis distance, and parallelepiped classifier. Some image processing software uses only minimum distance to means classifier in unsupervised classification. In supervised mode all classifiers stated above are available in software. Selecting a proper classifier is crux of classification process. An attempt has been made to evaluate performance of classifiers in three image processing software namely ERDAS IMAGINE-8.5, ER Mapper-6.3, and MGE Advance Imager. Supervised classification mode is used for study. Accuracy assessment for each classified image is carried out to assess the performance of classifier. The performance of classifiers has been evaluated using IRS 1D LISS III data. A thorough comparison between selected classifiers has been made for each class. It has been observed that for the data set considered the minimum distance classifier has performed well on shallow water. The Mahalanobis classifier has performed well for deep water. For other classes maximum likelihood classifier has performed well. The maximum overall accuracy is achieved by maximum likelihood in ERDAS IMAGINE-8.5, MGE Advance Imager, and maximum likelihood enhanced neighbor in ER Mapper-6.3.