Analysis of UAS-LiDAR Ground Points Classification in Agricultural Fields Using Traditional Algorithms and PointCNN

Abstract

Classifying bare earth (ground) points from Light Detection and Ranging (LiDAR) point clouds is well-established research in the forestry, topography, and urban domains using point clouds acquired by Airborne LiDAR System (ALS) at average point densities (≈2 points per meter-square (pts/m2)). The paradigm of point cloud collection has shifted with the advent of unmanned aerial systems (UAS) onboard affordable laser scanners with commercial utility (e.g., DJI Zenmuse L1 sensor) and unprecedented repeatability of UAS-LiDAR surveys. Therefore, there is an immediate need to investigate the existing methods, and to develop new ground classification methods, using UAS-LiDAR. In this paper, for the first time, traditional ground classification algorithms and modern machine learning methods were investigated to filter ground from point clouds of high-density UAS-LiDAR data (≈900 pts/m2) over five agricultural fields in North Dakota, USA. To this end, we tested frequently used ground classification algorithms: Cloth Simulation Function (CSF), Progressive Morphological Filter (PMF), Multiscale Curvature Classification (MCC), and ArcGIS ground classification algorithms along with the PointCNN deep learning model were trained. We investigated two aspects of ground classification algorithms and PointCNN: (a) Classification accuracy of optimized ground classification algorithms (i.e., fine adjustment is user-defined parameters) and PointCNN over training site, and (b) transferability potential over four yet diverse test agricultural fields. The well-established evaluation metrics of omission error, commission error, and total error, along with kappa coefficients showed that deep learning outperforms the traditional ground classification algorithms in both aspects: (a) overall classification accuracy, and (b) transferability over diverse agricultural fields.

UAS Quality Control and Crop Three-Dimensional Characterization Framework Using Multi-Temporal LiDAR Data

“…In the pre-processing step, acquired VHR images were processed using the Pix4D mapping application to obtain orthophotos [66]. The ULS point clouds were obtained by processing DJI L1 raw data using the DJI Terra software application, and then strip adjustments between consecutive flight paths were performed to obtain seamless consistent ULS point clouds [25,54,67]. As shown in Figure 3b, TLS point clouds obtained using the Livox sensor were first registered with ULS point clouds obtained on the same date, i.e., 27 September 2022, using control points found in both datasets (see reference scale in Figure 2d) using CloudCompare (https://www.cloudcompare.org/, accessed on 30 November 2022)-an open-source software application [68,69].…”

“…UAS-LiDAR equipped with Real Time Kinematic (RTK) provided the reference and TLS point clouds as registration targets. The outlier-noise points then were removed from the ULS and TLS point clouds using the statistical outlier removal (SOR) filter implemented in CloudCompare [25,70]. The quality of CHM is directly influenced by the quality of corresponding derivative products, i.e., DEM and DSM [71].…”

“…The quality of CHM is directly influenced by the quality of corresponding derivative products, i.e., DEM and DSM [71]. Multi-temporal ULS point clouds were classified into ground and non-ground points using the previously published author's trained PointCNN model to achieve high classification accuracy [25]. In addition, a thorough ground point QC was performed to improve the classification results due to some limitations of the PointCNN model previously mentioned by authors [25].…”

“…Moreover, a substantial portion of the chosen study site consists of bare ground, allowing for the LiDAR bare-ground point cloud evaluation in an agricultural context (Figure 1c). The selected sites contain weeds and shrubs among crops appropriate to ascertain the ULS mapping capabilities over fully vegetated and semi-vegetated landscapes, i.e., ground partially covered with weeds/grasses [25].…”

“…Nonetheless, the plant's vertical structural complexity cannot be fully understood using VHR optical imagery due to the limited penetration of light through the foliage [23,24]. Furthermore, optical sensors are constrained by absolute, geometric, and turbid occlusions caused by plants' leaves, branches, and stems [25,26].…”