A COMPARATIVE STUDY OF SOME SUPERVISED NEURAL NETWORK ALGORITHMS FOR IMAGE CLASSIFICATION

Abstract

Over the years, Artificial Neural Network (ANN) classifiers have gained popularity in classifying remote sensing images. Compared to conventional statistical classifiers, these classifiers are non-parametric and distribution-free. Thus, they are less restrictive in approximation, especially when distributions of various classes are non-Gaussian. Among the various available neural network algorithms, most attention has focused on Multi Layer Perceptron (MLP) with Back Propogation (BP) learning algorithm. However, this type of network takes relatively longer training times and is prone to convergence to local minimum. Other neural networks such as Radial Basis Function (RBF) and Probabilistic Neural Network (PNN) train much faster and are more stable than MLP. In this study, the performance of MLP, RBF and PNN for classification of remote sensing data has been evaluated. Several classifications have been performed using these classifiers by varying training sample sizes and number of hidden nodes. A Landsat ETM+ image of Syracuse area, New York, U.S.A. has been used as experimental data set. All the remaining parameters in each algorithm have been optimally fixed. The results show that PNN has produced a maximum accuracy of 86.76% which is the best among the three algorithms. Its training time was also significantly less than the other two. It has taken 56.424 seconds which was just 1% of MLP and 4% of RBF training times to produce maximum accuracies. RBF has also shown its potential in classifying larger training sample sizes with less number of hidden nodes than MLP. It achieved a higher accuracy than MLP with half the hidden nodes and one-third training time as taken by MLP.