Benchmarking tree instance segmentation of terrestrial laser scanning point clouds

Terrestrial laser scanning (TLS) has revolutionized forest measurement techniques by providing detailed three-dimensional (3D) point cloud data that captures the structure of forests and individual trees. Instance segmentation of point clouds, i.e. separating the forest into individual tree point clouds, remains a key challenge in automated processing due to complex, diverse tree structure and interactions. Furthermore, comparing segmentation performance is difficult, as new methods are often tested on new data with varying evaluation practices. Establishing a standardized benchmark and evaluation pipeline is key to consistent comparison and development of new algorithms and models. To this end, we manually segmented point clouds of four different forest types into almost 3000 individual trees spanning over 2.7 ha. We then evaluated five open-source segmentation methods, three theory-driven and two deep learning-based, using an evaluation pipeline with both plot and tree-scale metrics, independent of downstream application. Our results showed that a graph-based approach currently outperforms data-driven models for metrics such as plot-level F1-score and tree-level mean F1 score. Segmentation performance varied greatly across forest types, underscoring that instance segmentation remains difficult to automate and highlighting the need for diverse training and evaluation data. The benchmark dataset and evaluation code are publicly available to facilitate development and evaluation of generalized automated segmentation methods.