Combined impact of semantic segmentation and quantitative structure modelling of Southern pine trees using terrestrial laser scanning.

Southern pine forests play a key role in the ecological function and economic vitality of the southeastern United States. High-resolution terrestrial laser scanning (TLS) has become an indispensable tool for advancing tree structural research and monitoring. A critical challenge in this field is the accurate segmentation of leaf and wood components, which directly impacts the reliability of Quantitative Structure Models (QSMs). Segmentation techniques have progressed, but most existing methods are tailored for broadleaf species, with limited exploration for coniferous species such as southern pines. Addressing this gap, our study evaluates the performance of multiple segmentation algorithms on TLS data from southern pines, providing valuable insights into improving structural analysis and supporting more precise and efficient forest research and monitoring methodologies. We collected TLS data from longleaf pine (Pinus palustris Mill.) and loblolly pine (Pinus taeda L.) trees in Florida, USA, and compared the performance of four segmentation algorithms: TLSep, Graph, DBSCAN, and KPConv to separate leaf and wood. We found that KPConv was the most accurate method of segmenting wood and leaf points, with an overall accuracy (OA) of 98% and F1 score of 98% for loblolly pine and 95% and 94%, respectively, for longleaf pine. Although KPConv requires a substantial initial investment for training, its inference time is fast, making it a strong candidate for high-accuracy large-scale applications. These results led to highly reliable QSMs across trunk, branch, and total volume estimates. In contrast, DBSCAN, while slightly less accurate (OA of 92% for loblolly pine and 90% for longleaf pine), does not require training data and offers a favorable trade-off between performance and efficiency. These findings highlight the importance of selecting segmentation algorithms based on specific research goals, balancing accuracy and computational feasibility in forest structural modeling.