Answers are grounded in this paper and cite their sources.
Individual tree segmentation from terrestrial laser scanning (TLS) is fundamental to automated forest inventory, yet recent benchmarks have reached divergent conclusions about method performance. We evaluated six methods on TreeScanPL10K, a dataset of 10,417 manually segmented trees across 272 plots of diverse Central European forest. Two are algorithmic (treeX, RayExtract) and four are deep learning (TreeLearn, TreeAIBox, SegmentAnyTree, ForestFormer3D). Beyond aggregate accuracy, we characterized the per-method failure modes and the forest structural attributes that drive performance variation. A clear hierarchy emerged: ForestFormer3D achieved the highest mean IoU (0.76), followed by the other deep-learning methods (0.67–0.71) and the algorithmic approaches (0.60–0.62). Coniferous species were segmented more accurately than broadleaves (0.70 vs. 0.64), with the widest gap for the deep-learning methods. Higher accuracy came at substantial computational cost: 15 min/plot for ForestFormer3D versus under a minute for the fastest methods. To identify what makes a plot difficult to segment, we ran two complementary analyses: one on field-inventory attributes, the other on crown-geometry descriptors derived from the reference annotations. The dominant signal came from crown geometry: the plot-level Mean Crown Area Index (Mean CAI), which summarizes how many crown layers stack vertically above each ground cell, was the strongest single correlate and the top-ranked feature for every method individually. Multi-layered canopies are the hardest to segment. The field-inventory pool added a related but distinct dimension: structural diversity in species composition and spatial arrangement contributed beyond the vertical-layering signal. All six methods degraded in parallel along these gradients, indicating that the evaluated methods share common difficulties in structurally complex stands. We release the exact 2 cm benchmark point clouds, plot-level attributes and results, point-level predictions from all six methods for 18 selected plots, and the evaluation code.
No takes yet. Share an insight, caveat, or question.
Kulicki et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: