Publications / 2014 Proceedings of the 31st ISARC, Sydney, Australia

Autocorrelation Statistics-Based Algorithms for Automatic Ground and Non-Ground Classification of Lidar Data

Sara Shirowzhan, Samsung Lim
Pages 897-902 (2014 Proceedings of the 31st ISARC, Sydney, Australia, ISBN 978-0-646-59711-9, ISSN 2413-5844)

Classification of lidar data to ground and nonground points is important for accurate topography mapping and reliable estimation of slope, volume and buildings’ geometry over urban areas. Manual or semi-automatic classification provides relatively good results, however, automatic classification in complex areas with diverse object sizes is still challenging. This research aims to propose two novel algorithms based on Getis-Ord Gi* (or Gi* for short) and Local Moran’s I (LMI) statistics to classify a lidar point cloud into a set of points representing ground and another set of points reflected from non-ground e.g. buildings and vegetation. The Two statistics, Gi* and LMI, have been widely used in cluster analysis to identify clustered features of high z-scores and low z-scores. Based on the two statistics, we proposed two classification algorithms that allow varying window sizes e.g. 100 m, 150 m and 200 m, and applied the algorithms to the lidar data in order to obtain optimal classification results. The results show that the Gi*-based algorithm decreases omission errors but increases commission errors when compared to the LMI-based approach. Overall the 100-m window size outperforms than the other window sizes in terms of feature extraction in slan areas, whereas the 150-m window size provides slightly better results in a complex scene of high-rise buildings and dense vegetation, and the 200-m window size is more efficient if large buildings are present in the study area. This feasibility study indicates that autocorrelation statistics such as Gi* and LMI can be effectively used to classify a lidar point cloud.

Keywords: Automatic classification, Bare earth extraction, Spatial association, DSM, DEM, NDSM