Comparing parametric and non-parametric classifiers for remote sensing of tree species across a land use gradient in a Savanna landscape

Abstract

Several classification techniques have been used to map vegetation communities or land cover types using remote sensing data including maximum likelihood (ML), discriminant analysis and spectral angle mapper classifiers. ML classifier is a commonly used supervised classification method with conventional multispectral data that considers both first order variations (e.g. mean values) and second order variations (e.g. covariance matrices). However, there is a limitation with the application of the ML classifier in situations of high within-species variability. The objective of this study is to ascertain which classification techniques are suitable for classification of savanna tree species across a land-use gradient. Eight savanna tree species were classified for two sites in the vicinity of the Kruger National Park, South Africa using two parametric (ML and Mahalanobis distance classifiers) and three non-parametric classifiers (spectral angle mapper (SAM), artificial neural networks (ANN) and Random Forest (RF)). The spectral data used consisted of 8 WorldView-2 multispectral bands simulated from 72 VNIR bands image acquired over the study areas using the Carnegie Airborne Observatory (CAO) system. With the exception of SAM, the nonparametric classifiers provided higher classification accuracies (RF = 78%, ANN = 75%, SAM = 36%) when compared to the parametric classifiers (ML = 65%, Mahalanobis distance = 68). This study moves remote sensing closer towards classification of savanna tree species over large areas.