A drone lidar study on canopy complexity shows how far three labeled plots can go, and where it fits among the open models now used for forest lidar classification.

Canopy structural complexity has become one of the more useful numbers to pull from a forest point cloud. It describes how leaves, branches, and stems fill three-dimensional space and has been linked to productivity, habitat quality, and disturbance recovery. Calculating it first requires isolating canopy returns from shrubs, understory, and ground, usually with a height threshold that falls apart on steep slopes and in layered stands.
A Plant Phenomics study led by Nanjing University of Information Science and Technology, with the Chinese Academy of Forestry and the University of Waterloo, replaces that threshold with a deep learning classifier trained on just three hand-labeled plots. It is a good entry point into a growing set of open models for forest lidar classification, all wrestling with the cost of labeled training data.
How the Canopy Extraction Model Works
The team flew a DJI Matrice 300 with a Zenmuse L1 over the Fengshushan Forest Farm in Jiangxi Province, China, at 100 meters, producing point clouds averaging more than 412 points per square meter. Of 135 plots measuring 20 by 30 meters, four were labeled point by point as canopy or non-canopy: three for training, one for testing.
The classifier works in three stages. A RandLA-Net network first learns from the three labeled plots. That model then classifies every unlabeled plot, and the 40 percent it is most confident about are recycled as pseudo-labels, meaning the model’s own best guesses are fed back in as extra training data, weighted by confidence.
Finally, the network retrains on real labels, pseudo-labels, and the remaining unlabeled plots, and is penalized whenever its answer changes for a copy of the same plot that has been rotated or normalized for terrain differently. That terrain trick keeps the model stable on steep ground, which matters because the team skipped conventional ground filtering and height normalization to avoid burning in their errors.
On the test plot, the model reached 85.37 percent mean intersection over union (a measure of how closely predicted classes overlap the true ones) and 92.64 percent overall accuracy. That narrowly beats six fully supervised networks trained on the same three plots; Point Transformer finished 0.15 points behind. The clearer gap shows up on unlabeled terrain, where competing models mislabeled shrubs and ground as canopy.
The authors then calculate a joint canopy entropy index that combines three projected views of the canopy with local measures of how evenly points are spread. Against canopy cover, top rugosity, fractal dimension, and single-dimension canopy entropy, it posted the best combined score (0.76) for consistency, stability under point thinning, and sensitivity to stand density.
One caveat on reuse: the press release says the code and pseudo-labeling pipeline are available, but the paper lists no repository, and the raw lidar is restricted under third-party licensing.
Why Open-Source Forest Lidar Classifiers Matter
Classification is the step that turns raw returns into something usable, and in forestry it has largely relied on commercial software or whatever classes the data vendor delivered. Open-source classifiers change that in practical ways. Agencies and research groups can run a model across archived acquisitions without per-seat license costs, see exactly how it was trained, and retrain it on their own forest types instead of accepting someone else’s defaults.
Public benchmark datasets add a second benefit: results can be compared head to head, which is a big part of why the field has moved so quickly from height thresholds to deep learning.
Most open forest models target a handful of class schemes, each tied to a downstream product. Ground versus non-ground underpins every DTM and canopy height model. Canopy versus non-canopy feeds complexity, cover, and gap metrics. Instance segmentation separates individual trees for inventory, and species classification then labels each one.
Wood versus leaf separation matters because the two components feed different estimates. Woody points are what stem and branch reconstruction models use to calculate wood volume and biomass, and stray foliage returns inflate those numbers. Going the other way, lidar-based leaf area is really plant area unless the wood is removed, since branches and stems intercept pulses too.
For fire behavior, foliage and fine branches make up the canopy fuel that carries a crown fire, so isolating them gives a cleaner path to canopy bulk density than treating every canopy return the same. Lidar News recently covered a related finding that low ladder fuels, not total biomass, best predict wildfire severity.
Open-Source Models for Forest Lidar Classification
Most available open-source tools run on a short list of point-based networks. PointNet++ groups nearby points and learns patterns from each neighborhood. RandLA-Net randomly thins points as it works, giving up a little precision for the speed needed on large tiles. Transformer models such as Point Transformer let each point weigh which of its neighbors matter most, and they now lead most forest benchmarks.
France’s national mapping agency, IGN, has published FRACTAL, a RandLA-Net model trained on 80,000 patches of Lidar HD data that sorts points into seven classes, including ground and vegetation, with 93.7 percent IoU on vegetation. It runs in IGN’s open Myria3D library.
ForestFormer3D, presented at ICCV 2025 in Honolulu, labels ground, wood, and leaf while separating individual trees, and arrived with the open FOR-instanceV2 benchmark. SegmentAnyTreeV2, a June 2026 preprint from three of its co-authors, moves to a Point Transformer v3 backbone and expands the benchmark to 427 scenes and 26,496 annotated trees from multiple lidar platforms.
For wood and foliage, the University of Florida’s points2SBL is a pip-installable package aimed at wildfire fuel mapping, with a pretrained Point Transformer for TLS, MLS, and drone data.
The FOR-species20K dataset covers more than 20,000 trees from 33 species. In its benchmark, models that first convert each tree’s point cloud into a set of flat images viewed from different angles, then classify those images, were more accurate on average (77 percent) than models that work on the 3D points directly (72 percent).
For cloud-native workflows, PyForestScan streams COPC and EPT data through PDAL to produce canopy height, cover, and foliage height diversity rasters at landscape scale.
Labels Remain the Bottleneck
The common thread is doing more with fewer labels. Sensor-agnostic models train on mixed TLS, MLS, and drone data so one network serves many instruments, while pseudo-labeling stretches every hand-labeled plot further. For programs sitting on archives of unclassified drone and airborne lidar, that matters more than a fraction of a point of accuracy.
Sources
- Newswise: Forest complexity decoded: how AI and drone lasers read the canopy’s hidden language
- Chen et al. (2026), Weakly supervised forest canopy extraction and multi-dimensional joint canopy entropy, Plant Phenomics
- IGN FRACTAL-LidarHD RandLA-Net model card
- Xiang et al. (2025), ForestFormer3D, ICCV 2025
- Wielgosz, Puliti, and Astrup (2026), SegmentAnyTreeV2 preprint
- points2SBL on PyPI
- Puliti et al. (2025), FOR-species20K benchmark
- PyForestScan, FOSS4G 2025 talk














