Performance of different learning algorithms for object-based mapping of urban land cover using high-resolution satellite imagery

Recent advances in Earth observation technology have led to an increased availability of data products at high spatial resolutions and may open up new areas in the application of satellite imagery. Land-use/land-cover mapping in complex settings such as urban and suburban environments is one of the domains for which high resolution satellite imagery (Ikonos, Quickbird) shows great potential. Spatial resolutions of one meter or even less allow more accurate and detailed observation of the urban environment. Problem though is that, because pixels are smaller than the objects in which one is interested, one often observes a high spectral variation within the same object class. Therefore traditional per-pixel classifiers often produce unsatisfying results. The newly emerging object-oriented classification approach may offer some solutions. In this study several object-oriented classification strategies are applied to an Ikonos image covering part of the city of Ghent (Belgium), with the purpose of identifying the extent of sealed surfaces in the urban and semi-urban environment. The performance of the different classification scenarios is documented and compared.