PERFORMANCE EVALUATION OF BACK- PROPAGATION NEURAL NETWORK, FUZZY C- MEANS AND STATISTICAL CLASSIFIERS FOR REMOTELY SENSED DATA

Abstract

Artificial neural network and fuzzy approaches are considered as attractive alternatives for image classification. These approaches seem to possess many advantages over conventional classifiers such as maximum likelihood, minimum distance to mean etc. With the increase in the availability of uigital spatial data, data from different sources are often used in the classification process to improve the quality of the end product. Conventional statistical classifiers have limitations to handle such a variety of multi-source data and thus have paved the way to advanced techniques. In the present study, the performance of an Artificial Neural Network and fuzzy-c-means clustering has been evaluated in terms of their accuracy and efficiency. Maximum likelihobd classifier has been used as the bench mark to evaluate their performances. Three data sets have been utilised to assess the quality of these classifiers. The effect of some important factors on the classification accuracy has also been investigated. Computer programs have been developed in MATLAB environment for various kinds of formulations used. Several classifications were performed using various combinations of factors considered. The findings of the present work showed that in general, neural network classifiers produced the highest accuracy while the fuzzy-c-means (iii) clustering produced the lowest. A significant increase in accuracy was also observed with all the classifiers when the number of bands and the training sample size was increased. From efficiency point of view, ANN took longer time in training but once trained these were able to classify the unknown data sets more quickly than others. It can he concluded that the performance of ANN is significantly better than the other two classifiers on the data sets considered.