Comparative Analysis of Tree Segmentation Techniques on High- and Low-Density LiDAR data

LiDAR systems are powerful tools for sustainable forest management and ecological research, offering the capability to extract detailed information about canopy structure, tree height, biomass, carbon storage, and biodiversity. Unmanned Aerial Systems (ULSs) mounted on drones provide high spatial resolution, density, and flexibility for capturing detailed forest metrics. Their ability to fly at low altitudes enables the collection of fine-scale details. Conversely, Airborne Laser Scanning (ALS) systems, mounted on aircraft, are ideal for large-scale assessments, despite providing a lower point-cloud density compared to ULSs. This study evaluates individual tree segmentation algorithms in a coniferous forest ecosystem using two data sources: (1) high-density ULS LiDAR data collected with the Zenmuse L1 sensor on a DJI Matrice 300RTK drone, and (2) lower-density ALS LiDAR data from the third coverage of Spain’s PNOA-LiDAR project (Plan Nacional de Ortofotografía Aérea). The general methodology for processing LiDAR data involves preliminary steps to generate Digital Elevation Models (DEMs), Digital Surface Models (DSMs), and Canopy Height Models (CHMs). Subsequently, segmentation techniques are applied to assist tree-level forest analysis. Segmentation is critical for understanding forest structure; however, selecting the most suitable segmentation technique remains an active area of research. To address this issue, a comparative assessment of four commonly used segmentation algorithms (Watershed, Dalponte2016, Silva2016 and Li2012) was conducted using CHMs and normalized 3D point clouds derived from both high- and low-density LiDAR data. Ground-truth reference data were generated by the manual segmentation of individual trees in a representative plot. The results revealed that the Li2012 algorithm demonstrated the best performance in properly segmenting trees in both types of datasets.