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
\nstatistical classification techniques. We use Landsat 7 Enhanced Thematic Mapper
\nPlus (ETMþ), Advanced Spaceborne Thermal Emission and Reflection
\nRadiometer (ASTER) and airborne imaging spectrometer (HyMap) images to
\nclassify Mediterranean vegetation types, with and without inclusion of ancillary
\ndata (geology, soil classes and digital elevation model derivatives). When the number
\nof classes is low, both conventional and innovative techniques perform well. For
\nlarger numbers of classes the innovative techniques of random forests and support
\nvector machines outperform the other techniques. Compared to conventional techniques,
\nclassification trees, random forests and support vector machines proved to
\nbe better suited for the incorporation of continuous and categorical ancillary data:
\noverall accuracies and accuracies for individual classes improve significantly when
\nmany, difficult to separate, classes are taken into account. Therefore, these techniques
\nare definitely worth including in common image analysis software packages.