| Authors | Xiang et al. |
| Country | ETH Zürich, Switzerland |
| Paper (PDF) | View paper |
| Category | Planning & Inventory |
detailed per-tree attributes from ALS-HD data
0
Deep learning frameworks for processing LiDAR point clouds represent a convergence of advanced neural network architectures and three-dimensional spatial data, enabling computers to interpret the dense, unstructured outputs of Light Detection and Ranging sensors with high accuracy. LiDAR systems emit laser pulses and measure return times to generate point clouds — sets of millions of georeferenced 3D coordinates that capture the precise geometry of forests, terrain, and individual trees. Processing these point clouds with conventional algorithms is computationally expensive and often inadequate for complex canopy structures, which is where deep learning excels. Architectures such as PointNet, PointNet++, and more recent transformer-based models like Point Transformer are specifically designed to operate directly on unordered point sets without converting them to voxels or images, preserving spatial fidelity. In forestry applications, these frameworks are used to segment individual trees from dense stands, classify canopy layers, detect ground returns beneath dense vegetation, estimate tree height and crown volume, and differentiate species based on structural signatures. In Australian forestry contexts — including plantation eucalypt and pine estates managed by companies such as PF Olsen and Hancock Victorian Plantations — airborne LiDAR surveys combined with deep learning pipelines enable rapid, cost-effective inventories across vast and often remote areas. Key tasks include individual tree detection (ITD), above-ground biomass estimation, and the identification of structural defects such as forking or leaners before harvest. Frameworks such as PyTorch and TensorFlow provide the computational backbone, while forestry-specific libraries and pre-trained models are increasingly shared through platforms like GitHub and research institutions including the Australian National University and CSIRO. Integration with drone-mounted LiDAR sensors is extending these capabilities to operational forestry workflows at the coupe level, reducing reliance on costly manned aircraft surveys.
| Field | Value |
|---|---|
| Company | Xiang et al. |
| Country | ETH Zürich, Switzerland |
| Type | Paper |
| Year | 2024 |
| FWPA RD&E | 7.3 |