Application of evidential reasoning to improve the mapping of regenerating forest stands

Abstract

This study confirmed the ability of the Dempster–Shafer theory (DST) and the Dezert–Smarandache (Free DSm model) theory to significantly improve the quality of maps of regenerating forest stands in southern Quebec, Canada compared to a classical Maximum Likelihood Algorithm (MLA). The proposed approach uses data fusion methods that allow the integration of remotely sensed imagery with conventional maps of ecophysiographic features. While the MLA provided an overall accuracy of 82.75%, the DST and Free DSm models had overall accuracies of 90.14% and 91.13% respectively. In addition, this study showed that the data fusion methods can model the influence of biophysical parameters (e.g., surface deposits and drainage) on the growth potential of regenerating forest stands. This study illustrates the importance of the mass function allocation for each ancillary data source. We found that a Bayesian belief configuration provided results equivalent to those obtained when representing data uncertainty. This demonstrates the difficulty in modelling uncertainty associated with each ancillary source.