Transferring instance and semantic segmentation across terrestrial laser scanning time series: towards automated forest monitoring plot measurements

Monitoring forest structural changes over time is important for sustainable management and ecological research. This requires repeated measurements on site, increasingly obtained using terrestrial laser scanning (TLS). However, conventional TLS-based methods are computationally intensive and not tailored for monitoring, as they do not exploit prior measurements. To improve the accuracy of forest monitoring and increase automation in data analyses, we developed a temporally consistent workflow, Tree Skeleton Transfer. This approach enables the transferring of both instance (tree ID labeling) and semantic segmentation (point class labeling) across TLS time series using skeletonized tree structures as stable spatial references. Using this method, instance and semantic labels are propagated to each point in a new, unclassified point cloud via a nearest neighbor assignment to the structural tree skeleton, which encodes label information from the initial segmentation of existing TLS data. The initial segmentation is performed in two stages: first, the sample plot point cloud is automatically segmented, and second, this segmentation is manually inspected and corrected if necessary. Subsequent point clouds are then segmented automatically, leveraging structural tree skeletons. Because changes in trees’ structures are often gradual, stem and leaf points are likely to be found in roughly the same locations over time. We applied this method to nine boreal forest plots in Evo, Southern Finland, using TLS data from 2014, 2023, and 2024. By leveraging semi-automated initial segmentations from 2023, tree ID and stem/crown labeling were transferred to unclassified 2014- and 2024-point clouds. The method achieved high instance segmentation accuracy, with point-level F1-scores of 0.889 when transferring labels backward to 2014, and 0.959 when transferring forward to 2024. For comparison, we adopted RayCloudTools as an alternative, automatic segmentation method applied in each data independently, achieving F1-scores of 0.791 for the 2014 data and 0.846 for the 2023 data. Semantic segmentation was robust across transfer directions, with F1-scores of 0.96–0.97 for stem points and 0.88–0.89 for crown points, exceeding the performance of an independent automated processing with F1-scores of 0.57–0.59. Compared to independent automated processing of subsequent datasets, the Tree Skeleton Transfer approach reduced computation times by 96–97% while enhancing segmentation consistency, offering clear benefits for monitoring forest structural changes over time. These findings suggest that skeleton-based segmentation transfer provides a scalable solution for operational forest inventories, with potential applications in other laser scanning contexts.