Comparison of standard maximum likelihood classification and polytomous logistic regression used in remote sensing

Abstract

Discriminant analysis, referred to as maximum likelihood classification within popular remote sensing software packages, is a common supervised technique used by analysts. Polytomous logistic regression (PLR), also referred to as multinomial logistic regression, is an alternative classification approach that is less restrictive, more flexible, and easy to interpret. To assess the utility of PLR in image classification, we compared the results of 15 classifications using independent validation datasets, estimates of kappa and error, and a non-parametric analysis of variance derived from visually interpreted observations, Landsat Enhanced Thematic Mapper plus imagery, PLR, and traditional maximum likelihood classifications algorithms.