Object-based image analysis of QuickBird imagery for the classification of savanna vegetation in the monsoonal tropics of northern Australia

The tropical north of Australia provides a number of challenges to the remote sensing practitioner. The spectral dynamics of vegetation associated with a monsoonal climate and seasonal bushfires can make land cover classification difficult. A landcover type might be presented that is both spatially and spectrally heterogeneous making it difficult to distinguish using traditional pixel-based methods of classification. This is particularly so when using high spatial resolution imagery such as QuickBird. The recent emergence of robust object-based methods involving segmentation of an image into objects and the subsequent classification of these objects has opened new possibilities for image interpretation. The ability to incorporate ancillary data, such as a vegetation index, in developing a knowledge base for image objects can enhance image classification. Analysis was undertaken on a QuickBird image captured in August 2004. Classification involved an initial segmentation of the image using three weighted parameters. Class rules for the objects were then developed using the spectral information from all bands within the image. Further rules were then developed based on information provided from the vegetation index. These rules are then applied during a fuzzy classification of the image. Initial results show the object-based approach for classification can be used to map savanna vegetation. The classification process was able to distinguish between spectrally similar but compositionally different vegetation types.