Comparative performance of regression tree and parametric classification of savannah woody cover on SPOT 6 NAOMI imagery

Abstract

Earth observation data potentially can provide ecosystem state indicative data for use in nature conservation. Full realisation of this potential requires robust algorithms that can extract the required ecosystem attributes from imagery. Woody cover is an indicative ecosystem state variable in savannah wilderness areas, whose extraction poses a challenge in reproducing the characteristically discontinuous cover. In this work the performance of the more modern, nonparametric regression tree and Random Forest classifiers in mapping savannah woody cover is compared with the parametric, potentially highly accurate sub-pixel classification approach. Detailed field quantification of % woody cover in 1 ha plots yielded data for the training stage of the two decision tree-based nonparametric classifiers. SPOT 6 NAOMI imagery whose timing excluded grass from the spectral signature of non-senescent vegetation was used. In order to improve the capability to extract woody cover the 6 m resolution of the multispectral bands was enhanced to 1.5 m, using high pass filter pan-sharpening. For the decision tree-based classifiers multiresolution image segmentation was then employed as preliminary step in extracting the woody cover. Objects approximating the 1 ha field plots were delineated at the coordinates of the respective field sampling sites. The univariate decision tree training stage used 60% of the sampling sites. The woody cover classes 0–19%, 20–39%, 40–59% and 60–79% were the labels of recursive partitioning terminal tree leaves. For Random Forest, up to 500 decision trees (Ntree) were specified. For regression tree modelling the 10-fold cross validation approach using the 1-SE rule was used for optimal tree selection, and pruning using 20% of the field sampling sites was accomplished using the cost complexity pruning technique. The remaining 20% of the field sampling sites were reserved for accuracy assessment, for which an additional stratified random sample of coordinates verified using a concurrent high resolution QuickBird image was used. Spectral signatures approximating Gaussian distributions were generated using woody end member sites for sub-pixel classification. Random Forest produced the highest overall accuracy (95.7%), followed by regression tree modelling (90.6%), and sub-pixel classification (73.3%). The two nonparametric classifiers correctly modelled the landscape distribution, occasional continuous formations and discontinuities in the woody cover. Based on the results, nonparametric classifiers appeared more appropriate for mapping the discontinuous savannah woody cover.