Synthetic Forest: A UAV Laser Scanning Benchmark Dataset for Individual Tree Segmentation, Classification, and Wood Volume Estimation

Accurate tree-level analysis from UAV-borne LiDAR is still limited by the lack of large, fully annotated point-cloud datasets, especially for Australian Eucalyptus forests. Existing public datasets are mostly real acquisitions with small areas and costly/manual labels, which makes it hard to develop and fairly compare deep-learning methods for tasks such as wood–leaf classification, individual-tree instance segmentation, and tree-level wood-volume estimation. To address this gap, we provide Synthetic Forest, a benchmark synthetic LiDAR dataset generated with a physics-based UAV laser scanning simulator and 3D Eucalyptus tree models. The dataset contains three 1-ha forest scenes (100 m × 100 m) representing Australia’s common canopy cover classes: woodland (38% canopy cover, 70 trees), open forest (63%, 121 trees), and closed forest (84%, 216 trees). Each scene was scanned virtually using HELIOS++, producing very dense point clouds (38–47 million points per scene, 3300–3860 pts/m², 2 cm average spacing) suitable for fine-grained 3D learning.