Abstract:
The significance of urban land-cover information has increased greatly in the past decade following the rapid growth of geo-location services. Efforts to automate existing labour-intensive and time-consuming thematic data acquisition and updating procedures gained considerable momentum once very high-resolution images were provided by a fleet of satellites. However, the procedures for extracting land-cover information from such images failed to produce acceptable results for spectrally and spatially-complex urban land features. Attempts have been made to integrate information that is complementary to the spectral and spatial data in order to successfully overcome the typical problems faced by reflectance-based classification methods. Most such data fusion endeavours are specially customised and feature specific procedures, but fail to adapt to different urban scenarios. This study tackles the problems commonly associated with urban land-cover classification through fusing multi-sensory data. A multistage segmentation classification workflow using LiDAR and WorldView-2 data based on existing data fusion methods is developed to identify the land-covers that are normally found in core urban areas. The results show that integration of the height products derived from the 3D LiDAR data brought many benefits. Elevation (nDSM), used as a main data source during image segmentation, produced sharper image object boundaries, thus reducing the impact of shadows. A novel elevation-based feature differentiation step was implemented prior to thematic classification effectively separated a range of spectrally-identical impervious land-covers into two intermediate sub-classes. As a result, it was possible to derive and integrate two separate classification processes targeting a particular group of land-covers. Collectively, the integration of multi-stage procedures and multi-sensory data products significantly improved the thematic classification accuracy, from the 75% obtained using the normal method of spectral classification, to 93%. As a secondary outcome of this study, a spectral index called ZABUD is successfully modified for WorldView-2 multispectral channels. This index is found to be beneficial in identifying impervious asphalt features. The standard nearest neighbour classifier was later replaced with other methods such as the K-nearest neighbour, the support vector machine and the DC-CART, with the obtained results being comparatively assessed at the end of the study. A comparison of the classification accuracies obtained further confirmed the flexibility of the proposed workflow in adapting various procedures to the different scenarios of land-covers.