LAUSR

Planner surfaces play an important role in several mapping and planning applications in urban areas. Automated algorithms to extract such surfaces are essential since they reduce the time and cost when compared to editing these DEMs manually. These surfaces could be constructed from either correlation-based DEMs or LIDAR point clouds. The advantages of the first over the second make it more appealing in generating 3D planer surfaces. In this paper we outline an automated algorithm to delineate 3D surfaces in DEMs compiled from a set of high resolution digital aerial photographs. Each cell in the DEM is assigned three attributes representing: height from bare ground, local statistics of neighboring elevations, and homogeneity of pixel intensities in corresponding aerial photos. These attributes are used to classify DEM points to ground and non-ground points using a feedforward back-propagation neural network. Candidate non-ground points are further segmented into different surfaces based on their tendency to lie in a plane. The 3D parameters of roof planes are then estimated using a robust estimator and regression analyses. DEM points contiguous to the roof patches are augmented if they pass an intensity compatibility check. Results showed an average increase in the precision of the DEMs’ heights for roof posts from 48 cm to 15 cm.