TreeSegAgent : Agentic tree instance segmentation via a multi‐anchor pipeline

Individual‐tree segmentation from three‐dimensional forest point clouds underpins forest inventory, biomass estimation and ecosystem monitoring, but remains challenging in structurally complex stands, where vertical heterogeneity and point‐density degradation weaken the robustness of centre‐offset‐based instance grouping. We present TreeSegAgent, an extensible framework for forest point‐cloud analysis, whose core segmentation component is a multi‐anchor framework for tree segmentation (MA‐TreeSeg). MA‐TreeSeg stabilizes instance formation through two modules. First, a multi‐anchor (three‐vertical‐anchor) structural grouping module within an enhanced PointGroup‐style instance segmentation network, explicitly encoding tree vertical structure to improve grouping stability under sparse canopy conditions and to better recover low‐stature and understory trees. Second, we incorporate segment anything model (SAM)‐driven trunk constraint module. SAM is applied on two‐dimensional projections to identify trunk regions and generate trunk candidates, which are then employed as spatial hard constraints to regularize centre‐offset predictions and suppress clustering drift under canopy overlap. To support reproducible ecological workflow, TreeSegAgent further provides a task‐driven agent for automated, specification‐driven orchestration of common point‐cloud operations (e.g. instance segmentation, DEM generation and height‐based filtering) across datasets. Across multiple forest point‐cloud benchmarks, our method yields more spatially coherent individual‐tree instances and improved robustness to density variation, offering an extensible solution for large‐scale forest point‐cloud analysis.