Abstract
The main purpose of this study is to derive a method suitable for producing a land cover map of southern hemisphere Africa at a spatial resolution of 1 km. Daily SPOT‐Vegetation images from the year 2000 were used to build a dataset of monthly composite images. The composites were used in the development of two different classifiers obtained through the induction of classification trees. The selection of image data for training the classifiers and for accuracy assessment was supported by maps at several different scales, expert knowledge, Landsat Thematic Mapper (TM) imagery, and a preliminary unsupervised classification of the monthly image composites. One classification is based on the construction and application of a single tree classifier, and a second classification relies on the construction and application of an ensemble of tree classifiers using bootstrap aggregation (bagging). Classification accuracy was assessed using a validation dataset. The ensemble of trees produced better results than the single tree classifier. The advantages and limitations of the methods used are discussed, and suggestions for future work presented.
Acknowledgments
SPOT‐Vegetation Images were provided by the VEGA 2000 initiative of the Vegetation Programme to the Global Land Cover 2000 project of the Joint Research Centre of the European Commission.