Automated ground filtering of LiDAR and UAS point clouds with metaheuristics

Abstract

The ground filtering is essential for the extraction of the topography of the bare Earth surface. Various ground filtering methods have been developed, especially in the last three decades. The main disadvantage of the ground filtering methods is that their performances are highly dependent on some user-defined parameter values. Hence, the analysts usually have to try a large number of parameter values until the optimum ground filtering result is achieved, which is neither practical nor time-efficient, especially for topographies with abrupt elevation changes. In addition, inappropriate parameter values may lead to the misclassification of the points that belong to the ground surface and to the above-ground objects. In cases where the analyst is not experienced in ground filtering, classification errors are expected to increase significantly. This reveals the necessity of an automated ground filtering strategy to avoid the user intervention for minimum classification errors. Hence, this study proposed to automate one of the most successful ground filtering methods, cloth simulation filtering (CSF), through algorithm-specific parameter-free metaheuristic optimization algorithms Grey Wolf Optimizer (GWO) and Jaya. The performances of the proposed GWO-based CSF (GWO-CSF) and Jaya-based CSF (Jaya-CSF) methods were tested on three LiDAR and two UAS point clouds. The results of the GWO-CSF and Jaya-CSF methods were qualitatively and quantitatively compared against those of the widely-used ground filtering methods progressive morphological 2D (PM2D), maximum local slope (MLS), elevation threshold with expand window (ETEW), multi-scale curvature classification (MCC), Boise Centre Aerospace Laboratory LiDAR (BCAL), gLiDAR, pro-gressive triangulated irregular network densification (PTD) and standard CSF in five test sites. The performance evaluations revealed that the proposed GWO-CSF and Jaya-CSF methods did not only outperform the standard CSF, but also the other filtering methods used. The GWO-CSF and Jaya-CSF methods were also found to achieve the best balance between the omission and commission errors. It was also concluded that the GWO-CSF and Jaya- CSF methods did not only perform well on gentle slopes, but also on sloping terrains with various large complex- shaped above ground objects. Another important conclusion is that the GWO-CSF and Jaya-CSF methods pre-sented a very high filtering performance on both LiDAR and UAS point clouds. The proposed methods managed to automate the filtering process, minimizing the filtering errors.