EFFECTS OF TRAINING DATA ON THE CLASSIFICATION OF REMOTELY SENSED IMAGES

Abstract

The Maximum Likelihood Classification (MLC) is presently the most widely known and utilized. The MLC is often used as a standard classification due to the fact that MLC is the optimal classifier in the sense of minimizing Bayesian error. However, the MLC belongs to a parametric classification method where the underlying probability density function must be assumed a priori. We may obtain a poor MLC performance if the true probability density function is different from that assumed by the model. In recent years, the Layered Neural Networks (LNN) have been proposed as a method suitable for the efficient classification of remotely sensed images to overcome this disadvantage of the MLC. The relationship between MLC and LNN classifier has been already discussed and the conclusion is that the output of the LNN, when trained with a sufficient number of sample data by the least squares, approximates the Bayesian posterior probability. This paper introduces the experimental results in the LNN and MLC classifiers and shows that the potential of the LNN approach to land cover mapping in comparison with the MLC on the same training data.