Light Detection and Ranging (LiDAR) has become an important data source in urban modelling. Traditional methods of LiDAR data processing for building detection require high spatial resolution data and sophisticated algorithms. The aerial photos, on the other hand, provide continuous spectral information on buildings. However, the accuracy of classified building boundaries from aerial photos is constrained when building roofs and their surroundings share analogous spectral characteristics. This paper develops a statistical approach that can integrate characteristic variables derived from sparse LiDAR points and air photos to detect buildings by estimating object heights and identifying clusters of similar heights. Within this study, the approach chooses a local regression method, namely geographically-weighted regression (GWR), to account for local variations of building surface height. In the GWR model, LiDAR data provide the height information of spatial objects, which is the dependent variable, while the brightness values from visible bands of the aerial photo serve as the independent variables. The established GWR model estimates the height at each pixel based on height values of its surrounding pixels with consideration of the distances between the pixels as well as similarities between their brightness values in visible bands. Clusters of contiguous pixels with higher estimated height val ues distinguish themselves from surrounding roads or other surfaces. A case study is conducted to evaluate the performance of the proposed method. It is found that the accuracy of the proposed statistical method is better than those by image classification of aerial photos alone or by building extraction of LiDAR data alone. The results demonstrate that this simple and effective method can be very useful for automatic detection of buildings in urban areas. The approach can be most helpful for studies of urban areas where more suitable but expensive high resolution data are not available.
