A comparison of three neural network classifiers for remote sensing classification

This paper evaluates the use of neural network classifiers for the pattern classification problem in remote sensing. The performance of multi-layer perceptron (MLP), radial basis function, and fuzzy ARTMAP networks is evaluated using a Landsat-5 TM scene of the northern section of the city of Vienna, Austria. Classification accuracies obtained from the neural network classifiers are compared with a benchmark, the maximum likelihood classifier. In addition to the evaluation of classification accuracy, the neural networks are analyzed for their generalization capability and stability of results. Best overall results (in terms of accuracy and convergence time) are obtained using fuzzy ARTMAP followed by MLP (with weight elimination). Their classification error on the training data set are zero and 7.87% respectively; classification error on the testing data set are 10.24% and less than 2 percent. Simulation results serve to illustrate the properties of the various classifiers in general, as well as the stability of the result with respect to various critical control parameters, initial parameter conditions, training time, and different training and testing data sets.