SelectAnyTree: One-click tree segmentation from 3D LiDAR forest scans

Automated instance segmentation of forest LiDAR point clouds is crucial for scalable 3D forest monitoring, but manual annotation of individual trees is extremely laborious—a single hectare can contain millions of points and hundreds of overlapping crowns. Existing methods rely on costly pre-segmentations with no interactive refinement. To address this, researchers propose SelectAnyTree, a promptable instance segmentation model that can delineate any individual tree from just a few user clicks.

SelectAnyTree introduces two key innovations: a Click-to-Query prompt encoder that turns each click into a content query encoding 3D position and polarity, and a Canopy Height Model (CHM)-guided first prompt that automatically proposes treetops without user input. A state-space query decoder then efficiently captures long-range context in large-scale forest scenes. Evaluated across seven diverse forest regions and a held-out test set, SelectAnyTree achieves 78.2 Intersection over Union (IoU) from a single click—24.8 points above the strongest baseline—while using far fewer parameters and less inference time. The source code is publicly available.