Neural networks versus nonparametric neighbor-based classifiers for semisupervised classification of Landsat Thematic Mapper imagery

Abstract

Semisupervised classification is one approach to converting multiband optical and infrared imagery into landcover maps. First, a sample of image pixels is extracted and clustered into several classes. The analyst next combines the clusters by hand to create a smaller set of groups that correspond to a useful landcover classification. The remaining image pixels are then assigned to one of the aggregated cluster groups by use of a per-pixel classifier. Since the cluster aggregation process frequently creates groups with multivariate shapes ill suited for parametric classifiers, there has been renewed interest in nonparametric methods for the task. This research reports the results of an experiment conducted on six Landsat Thematic Mapper images to compare the accuracy of pixel assignment performed by four nearest neighbor classifiers and two neural network paradigms in a semisupervised context. In all the experiments, both the neighbor-based classifiers and the neural networks assigned pixels with higher accuracy than the maximum-likelihood approach. There was little substantive difference in accuracy among the neighborhood-based classifiers, but the feedforward network was significantly superior to the probabilistic neural network. The feedforward network classifier generally produced the highest accuracy on all six of the images, but it was not significantly better than the accuracy produced by the best neighbor-based classifier.