Comparing techniques for vegetation classification using multi- and hyperspectral images and ancillary environmental data

Abstract

This paper evaluates the predictive power of innovative and more conventional statistical classification techniques. We use Landsat 7 Enhanced Thematic Mapper Plus (ETM+), Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) and airborne imaging spectrometer (HyMap) images to classify Mediterranean vegetation types, with and without inclusion of ancillary data (geology, soil classes and digital elevation model derivatives). When the number of classes is low, both conventional and innovative techniques perform well. For larger numbers of classes the innovative techniques of random forests and support vector machines outperform the other techniques. Compared to conventional techniques, classification trees, random forests and support vector machines proved to be better suited for the incorporation of continuous and categorical ancillary data: overall accuracies and accuracies for individual classes improve significantly when many, difficult to separate, classes are taken into account. Therefore, these techniques are definitely worth including in common image analysis software packages.