The use of artificial intelligence (AI) in understanding forestry lidar datasets can revolutionise monitoring and forest managment. Key application of AI is the ability to analyze point clouds to identify individual trees' location, extent and shape.

In this study, we present a validation of automatic tree detection on aerial lidar scanning (ALS) against ground-truth data that were obtained by a combination of manual and terrestrial lidar measurements. The first step of the ALS data analysis was classification using our AI semantic segmentation approach into five relevant classes: ground, trunk, canopy, understory and other points. Classifications were used to create digital elevation and canopy height models requiered for determining tree top centres, their height and canopy shape. Additionally, trunk detections were clustered into individual stems and analysed by line fitting to describe them. When a sufficient number of points describes a trunk, we can also extract its radius.

The study area has a point density of 95 points per m2 and spans over 4600 hectares of mainly coniferous forest. The algorithm detected 1.7 million tree trunks with determined tree height, location at breast height and trunk inclination. Validation was performed on a manually generated single tree inventory at 129 circular sample plots with a radius of 25 m. The recall of our trunk detection method in the validation plots is 0.46 for trees of any height. It struggles to analyze small and obscured trees where no trunks are visible in the data. Therefore the recall value for trees above 12 m is higher and reaches score of 0.85.

Our findings provide insights into the strengths and limitations of the algorithm and can guide future research aimed at AI tree instance segmentation.

This abstract is based upon work from COST Action 3DForEcoTech, CA20118, supported by COST (European Cooperation in Science and Technology).

Terrestrial laser scanning (TLS) is a widely used technology in forestry to derive a broad range of parameters for both single trees and whole stands. Quantitative structural models (QSM) can be used to simplify point clouds using geometric primitives and to derive e.g. topological and volumetric information about trees. Although already broadly applied for various purposes, the uncertainties of this technology are still largely unexplored. In our study, we investigated the effect of scanning distance on length and volume estimates of branches when deriving QSM from TLS point clouds. We scanned ten European beech (Fagus sylvatica L.) branches with an average length of 2.6 m from different distances under controlled conditions. The branches were scanned from distances ranging from 5 m to 45 m at intervals of 5 m from three scan positions each. For the ground truthing, twelve close range scans were performed. For each distance and branch, QSMs were derived. With increasing distance, the point cloud density and the cumulative length of the branches, including all sub-branches, decreased (Fig. 1). The volume, however, increased with scanning distance. Depending on the hyperparameters of the QSMs, at a scanning distance of 45 m, the cumulative branch length was on average underestimated by −75%, while branch volume was overestimated by up to +539%. We assume the strong misestimation of the length and volume of small branches is related to point cloud quality. As the scanning distance increases, the size of the individual laser footprints and the distances between them increase, making it impossible to fully capture small branches. Previous studies have shown high agreement between TLS and QSM volume estimates and field measured data. Among other reasons, this could be due to occlusion effects offsetting the overestimation caused by low point cloud density.

Tree structure (3D size and arrangement of material) plays a critical role in various biological and physical processes that control tree growth, development, reproduction, and adaptability to environmental conditions. Accurate quantification of tree structure provides valuable insights into the growth, health, and overall function of individual trees and entire forests. The advent of terrestrial LiDAR technologies delivers 3D point cloud data with millimetre-level accuracy. However, accurately extracting topological structure from point clouds remains challenging. Current tree reconstruction methods have mostly focused on retrieving tree volume estimations, which often involve assumptions that can significantly impact and even constrain the resulting structural properties.

Here, we present Treegraph, an open-source Python software package that automatically extracts structural parameters from point clouds of individual trees. Crucially Treegraph is designed to retrieve topology, and makes minimal assumptions about tree form. Treegraph also offers a batch processing solution on high-performance computing (HPC) facilities. The approach firstly generate a skeleton graph, then estimate branch radii from point clouds, and finally generating a quantitative structure model (QSM). The QSM enables the retrieval of a tree’s morphological and topological traits at the node-level, branch-level, and whole-tree level, providing valuable information for studying tree structure and functions. This approach is fully automatic and entirely data-driven, eliminating the need for user input of empirical parameters. Another key benefit arising from the minimal assumptions around structure is the ability to provide an objective assessment of the quality of the input point clouds.

We demonstrate Treegraph on a diverse set of tree species from tropical, temperate and urban forest plots, including a total of 595 trees, showcasing its applicability for a wide range of tree shape and size. This approach will facilitate the analysis of fine-level branching network patterns and promote new insights into the understanding of tree structure and functions.

Forests are the largest terrestrial ecosystem that provide wide range of ecosystem services. Better understanding of forest functioning processes, such as tree-growth allocation, is essential for developing smart forest management methods for meeting the requirements of both carbon neutrality and wood productivity. This is especially important in an extensively forested country like Finland, where the forests have a significant economic and ecological value.

Terrestrial laser scanning (TLS) time-series have a clear, yet not fully explored potential to study the allocation of tree growth over time at branch-level of detail. In this study, we tested the feasibility of estimating branch-level growth using quantitative structure modelling (QSM) on multitemporal TLS point clouds. The possibilities in branch identification between different time points are going to be examined.

We used dense spatio-temporal point cloud time series collected with the permanent FGI Lidar Phenology station (LiPhe) that scans the surrounding forest hourly. Despite the unique potential of the data, the geometry of the data acquisition is sensitive to occlusions that leave gaps in data that have significant effects on the temporal comparability of the QSM metrics. We will present our first results and discuss the possibilities and challenges found in our analysis.

We verified that the QSM metrics extracted from the LiPhe data were stable for a single time point as two consecutive runs yielded the same results for the branch metrics. This is the fundamental requirement for conducting temporal studies using TreeQSM. Also, statistical distributions of all tree-level metrics extracted with TreeQSM stayed similar across multiple time points, indicating that the overall structural information remains broadly consistent.

Quantification of the aboveground sapwood and heartwood volume of trees is challenging because they typically must be harvested to obtain the data, and most studies only focus on the main stem, leaving the branch components uncertain. Terrestrial laser scanning (TLS) technology provides novel and detailed three-dimensional data of trees that can be used to quantify their aboveground woody volume in a non-destructive manner. Here, TLS data were used to generate quantitative structure models (QSMs) of twenty-four urban Gleditsia triacanthos trees that were scanned in leaf-off condition on the Michigan State University campus. QSMs provided detailed quantification of the woody volume of main stem and branches of the study trees. Woody disks of stems and branches were collected from a subset of the study trees that were destructively sampled after scanning. Based on these disks, models of heartwood-sapwood proportions were developed, and they were combined with QSM-based volumes to estimate sapwood and heartwood volume of stems and branches. The results showed opposite patterns of vertical accumulation of heartwood and sapwood volume for stems versus branches. A positive relationship was found between the sapwood volume and the woody surface area of branches and stems. These results have important implications for studying the growth and the architecture of trees.