Comparison of 3D Point Cloud Acquisition Accuracy in a Large-Scale Japanese Pear Tree Orchard

Abstract

Practical three-dimensional (3D) phenotyping in large-scale orchards with repetitive row structures remains challenging, and systematic evidence comparing both accuracy and acquisition efficiency under outdoor conditions remains limited. This study presents a field-deployable evaluation framework and implements it in a 2-ha commercial Japanese pear orchard trained under a joint V-trellis system. Using a terrestrial laser scanner (TLS) as the reference, we evaluated two handheld LiDAR systems (a low-cost SLAM-based system and a high-performance system), structure from motion / multi-view stereo (SfM/MVS) reconstructions from three camera platforms (a digital camera, an action camera, and a 360° camera), and 3D Gaussian splatting (3DGS) constructed from action-camera video. Measurements were taken at two spatial scales to capture scale-dependent effects. In the span-scale survey (4 m), location error was derived from TLS-referenced target coordinate differences, and reconstruction error was quantified using cloud-to-mesh distances with cubic targets. In the row-scale survey (one tree row), positional stability during continuous mapping was evaluated as location error. Operational metrics (acquisition time, data volume, and processing effort) were also documented. The results demonstrate clear trade-offs among the methods: LiDAR enables rapid wide-area acquisition but is susceptible to cumulative drift in row-structured environments, whereas SfM/MVS provides superior geometric fidelity at the cost of increased time and data volume. Although 3DGS is less suitable for precise quantitative measurement, it demonstrates strong potential for intuitive visualization of orchard structure and fruit distribution. These findings highlight the need for staged, purpose-specific, and seasonally adaptive strategies for orchard-scale digital twin development.