Classification and extraction of trees and buildings from urban scenes using discrete return LiDAR and aerial color imagery

Abstract

Airborne Light Detection and Ranging (LiDAR) is used in many 3D applications, such as urban planning, city modeling, facility management, and environmental assessments. LiDAR systems generate dense 3D point clouds, which provide a distinct and comprehensive geometrical description of object surfaces. However, the challenge is that most of the applications require correct identification and extraction of objects from LiDAR point clouds to facilitate quantitative descriptions. This paper presents a feature-level fusion approach between LiDAR and aerial color (RGB) imagery to separate urban vegetation and buildings from other urban classes/cover types. The classification method used structural and spectral features derived from LiDAR and RGB imagery. Features such as flatness and distribution of normal vectors were estimated from LiDAR data, while the non-calibrated normalized difference vegetation index (NDVI) was calculated by combining LiDAR intensity at 1064 nm with the red channel from the RGB imagery. Building roof tops have regular surfaces with smaller variation in surface normal, whereas tree points generate irregular surfaces. Tree points, on the other hand, exhibit higher NDVI values when compared to returns from other classes. To identify vegetation points an NDVI map was used, while a vegetation mask was also derived from the RGB imagery. Accuracy was assessed by comparing the extraction result with manually digitized reference data generated from the high spatial resolution RGB image. Classification results indicated good separation between building and vegetation and exhibited overall accuracies greater than 85%.