Sam‐Stars: A Multimodal, Prompt‐Guided Approach for Individual Tree Crown Delineation and Species Mapping

ABSTRACT Tree‐level information is increasingly required to reduce uncertainty in biodiversity assessment and species‐aware carbon accounting, yet individual tree crown (ITC) delineation and species mapping remain challenging in compositionally heterogeneous forests at landscape scales. We present a scalable multimodal workflow for crown‐resolved, species‐level mapping that fuses very high‐resolution (VHR) WorldView‐3 imagery with airborne lidar. The approach uses lidar‐guided prompting to drive a minimally fine‐tuned Segment Anything Model (SAM) for high‐recall canopy detection, followed by shape‐aware instance separation with StarDist to delineate individual crowns. Across benchmarks on Hawaiʻi Island, the proposed sam‐stars framework achieves strong delineation performance with limited annotation, outperforming an end‐to‐end Detectron2 (Mask R‐CNN) baseline most strongly at small training sizes. In an object‐based species‐mapping workflow, crown‐level multispectral and structural features are classified with gradient‐boosted trees, yielding the highest species‐aware semantic panoptic quality (PQ = 0.47) among evaluated methods. Applied at landscape scale using a delineation model trained on only 254 crowns and a larger independent species inventory, the workflow produced a wall‐to‐wall map of 333,204 crowns. When aggregated to 30 m resolution, crown‐derived canopy cover was 19.0% compared with 14.9% in a publicly available land‐cover product, indicating substantial under‐detection in existing maps. By enabling label‐efficient, crown‐resolved, species‐aware mapping from widely available lidar and VHR imagery, this framework supports applications ranging from invasive species detection to restoration planning and species‐informed carbon accounting in structurally heterogeneous forest landscapes.