This study compares the performance of two non-parametric classifiers and Gaussian Maximum Likelihood (GML) for the classification of LANDSAT TM 30-meter resolution six-band data. The mathematical assumptions made in developing GML are valid if the pixels that constitute the training classes are normally distributed. Since it requires a model of the data, GML is termed a "parametric" classifier. Of current interest are new classification methodologies that make no assumptions about the statistical distribution of the pixels in the training class; these approaches are termed "non-parametric" classifiers. This study will compare the n-Dimensional Probability Density Function (nPDF) essentially a projection technique that reduces data dimensionality, and an advanced neural network that utilizes fiizzy-set mathematics, the Fuzzy ARTMAP, to the traditional GML approach to image classification. The different approaches will be compared for statistical classification accuracy and computational efficiency.
Library of Congress Subject Headings
Remote sensing--Mathematics; Gaussian processes; Image processing--Statistical methods; Image processing--Digital techniques
Department, Program, or Center
Chester F. Carlson Center for Imaging Science (COS)
Nessmiller, Steven, "A Comparison of the performance of non-parametric classifies with Gaussian maximum likelihood for the classification of multispectral remotely sensed data" (1995). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus