Evaluating the accuracy of extracting urban land cover/use from remotely sensed imagery and its potential application in urban planning

Remote sensing exhibits a great potential for enhancing current urban planning processes. Literature review, however, indicates that there is a significant gap between its technology development and applications in urban planning because of discipline barriers. This study is, therefore, intent to address such a gap by studying the effectiveness of two image classification software (ENVI 4.5 and Definiens Professional 5) on extracting urban subzonal land covers, and investigating the usefulness/accuracy of the extracted information in urban planning process. Several satellite images, including Landsat ETM+, SPOT4, IKONOS, and QuickBird, are used in our testing. It is found that medium-resolution images, such as Landsat ETM+ and SPOT4, are only good at extracting large-size homogeneous objects, such as vegetation and water bodies, but less powerful in extracting small urban features, including buildings, streets and parking lots, due to their low spatial resolutions. Later experiments focus on extracting these small-size urban features with VHR imagery from IKONOS and QuickBird. Classification results show that both software packages have more or less problems in distinguish parking lots, streets and building roofs because of similar materials used and therefore, very close spectral signatures. It is found that the object-oriented hierarchical algorithm applied to QuickBird images offers the highest extraction accuracy for building and street classes with a producer's accuracy (PA) of 78% and 63% respectively, when compared to other algorithms, such as maximum likelihood and nearest neighbour, and images (QuickBird and IKONOS) combinations. Recommendations and discussions about how to improve the current results by using off-leave images and supplementary data are also presented.