Abstract
The management of vegetated areas by urban planners relies on detailed and updated knowledge of their nature and distribution. Manual photo-interpretation of aerial photographs is efficient, but is time consuming. Image segmentation and object-oriented classifications provide a tool to automatically delineate and label vegetation units. Object-oriented techniques were tested with a very high-resolution multispectral Ikonos image to produce fine scale maps of vegetation communities in Dunedin (New Zealand). The Ikonos image was orthorectified and a first classification produced a map with 4 strata: industrial/commercial (with amenity pastures and tree groups), residential (with amenity pastures and private gardens), vegetation (with other vegetation classes), and water. A hierarchical network of image objects was built to extract vegetation patches of various sizes such as small private gardens and larger exotic plantations. The classification of the image objects was performed using the nearest neighbour (NN) method. Thirteen variables were considered to build the NN feature space, including mean object spectral value and standard deviation for each spectral band, and object compactness. The vegetation map was validated using an independent dataset collected in the field. The original classification scheme included 17 vegetation categories, of which ten were successfully discriminated: forests, exotic plantations, tree groups, exotic scrubs, mixed scrubs, native scrubs, pastures, amenity grasslands, rough grasslands, private gardens. Classes of ecological interest characterized by various canopy densities could not be discriminated (e.g. low and high density gardens, shrublands and scrubs). Vegetation patches smaller than 0.05 ha were efficiently extracted within the city. The overall classification accuracy was 92% and the kappa coefficient was 0.89 (i.e. 89% more accurate than a random classification). Object-oriented techniques and Ikonos imagery proved to be a promising technique to produce GISready vegetation map.