Roof Surface Classification with Hyperspectral and Laser Scanning Data – An Assessment of Spectral Angle Mapper and Support Vector Machines

The urban environment is characterised by a variety of different surface materials. For the discrimination of urban materials, hyperspectral imaging proved a valuable tool. In this study, two methods for classification, Spectral Angle Mapper and Support Vector Machines, are compared on a hyperspectral dataset to derive a detailed map of roof materials. Spectral similarity of different materials, especially with low reflectance and no distinct absorption features can complicate the classification process. Therefore, hyperspectral data were supplemented with laser scanning data to not only discriminate roof from ground data but also to use roof inclination to distinguish roof spectra. A binary roof mask from a laser scanning dataset was used to restrict the classification to roofs only. After testing the two classifiers on this reduced dataset, the approach was extended by incorporating inclination information in the classification process. Comparison between the classified images is done visually and quantitatively using confusion matrices. It can be shown that both classifiers are suitable for the classification of roof materials with the Spectral Angle Mapper results yielding higher classification accuracies than Support Vector Machines. For both classification approaches, the confusion between several materials was reduced by the incorporation of roof inclination, thus improving overall accuracy.