Application of a Bayesian network for land-cover classification from a Landsat 7 ETM+ image

Abstract

This article describes the use of a Bayesian network (BN) for the classification of land cover from satellite imagery in northern Swaziland. The main objective of this work was to apply and evaluate the efficacy of a BN for land-cover classification using gap-filled and terrain-corrected Landsat 7 Enhanced Thematic Mapper Plus (ETM+) imagery acquired on 15 May 2007. The posterior probabilities (parameters) were estimated using the expectation-maximization (EM) and conjugate gradient descent (CGD) algorithms. A comparison of the results obtained from the algorithms indicates similar and excellent overall classification accuracies of 93.01%, and kappa coefficient values of 0.9143. The main result obtained in this study is that both algorithms considered here provide relatively similar and accurate solutions for the classification of the multispectral image although the EM algorithm is marginally competitive relative to CGD algorithm when measured in terms of the Brier score and the logarithmic loss.

Acknowledgement

The author acknowledges Brent Boerlage of Norsys Software Corporation for the complementary license of the Netica^TM software used in this study.