A STUDY ON THE FEASIBILITY OF SELECTION OF TRAINING DATA FOR IMAGE ANALYSIS

Abstract

Training data refer to the data obtained from training samples,which are areas of known ground identity. The training data so obtained are used to generate various statistical parameters which help in the classification of the whole image of the area for which the training data were collected. The methodology of classification involves the characterization of the classes of interest through the analysis of training data which are representative of the desired classes. All the remaining data are classified by means of numerical rules - which utilize the class characteristics. The classification of an image may be influenced significantly depending upon the purity of the training samples from which the training data is extracted. The selection of training samples are influenced by many factors. The present work is focussed on these factors to find the feasible method of selection of training samples from which the training data is extracted. The study shows that time of sampling is a very important factor. It also shows that 4 x 4 sample size results in higher accuracies than 5 x 5 sample size. Stratified sampling and systematic sampling have been found to give representative training samples over random sampling.