Segmentation of living and dead tree crowns using terrestrial laser scanning and deep learning

Detailed data on individual tree crown characteristics improve the accuracy of tree biomass, carbon sequestration, and wood quality estimation. While machine-learning methods exist to reconstruct detailed tree models, most lack partitioning of live and dead crown parts (i.e., semantic segmentation) and are insufficient for real-time applications. On the other hand, deep learning may enable more flexible and robust semantic segmentation of trees with potential for real-time processing. In this study, we trained deep-learning models to semantically segment stem, living crown, and dead crown from point clouds of individual Norway spruce ( Picea abies (L.) H. Karst.) and Scots pine ( Pinus sylvestris L.) trees. Data were collected from southern Finland. For training, we manually annotated 100 spruce and 100 pine point clouds and generated two additional augmented versions of each tree. The data were subsequently divided into training and validation sets, and the segmentation models for each species were trained. We used RandLA-Net as an example of a deep-learning architecture optimized for semantic segmentation of large-scale point clouds containing hundreds of thousands of points. The performance of the trained models in semantic segmentation was evaluated using overall accuracy and overlap ratio, expressed as Intersection over Union (IoU). For Norway spruce, the overall accuracy and IoU were 0.90 and 0.82, respectively; the corresponding values for Scots pine were 0.84 and 0.76. For testing the robustness of the trained models, we used an external dataset of 50 Norway spruce and 50 Scots pine point clouds. Using the semantic groups, we delineated the boundaries (i.e., base heights) of living crown (LB) and dead crown (DB) of these trees. The resulting crown boundaries were then compared with field measurements and manual point cloud measurements taken from these trees. The assessed values showed moderate agreement with the manually measured references (RMSE: 2.30 m for LB and 0.74 m for DB in Norway spruce; 0.93 m for LB and 1.78 m for DB in Scots pine). However, when compared with the field references, the predictions exhibited larger errors (RMSE: 3.10 m for LB and 1.69 m for DB in Norway spruce; 1.91 m for LB and 2.21 m for DB in Scots pine). Delineating the living crown was found particularly challenging in trees with complex and asymmetric crown structures. The uncertainties in assessing the DB were due to misclassification of points from small branches and branch stubs or incomplete scanning of the tree.