Unsupervised AI network separates wood and leaves in 3D forest scans

This innovation overcomes one of the biggest challenges in automated forest inventory and management, offering a faster, more cost-effective way to assess forest structure and carbon storage potential.

Forests are vital for global climate regulation, biodiversity, and sustainable resource management. Analyzing tree architecture—including branch, trunk, and leaf structures—is critical for monitoring growth and estimating carbon sequestration. Light detection and ranging (LiDAR) has transformed forestry surveys by providing detailed 3D point clouds that capture tree geometry. However, reliably distinguishing between wood and leaves within these datasets has remained difficult due to data occlusion, irregular tree shapes, and the high cost of manually annotated training data. Traditional supervised machine learning methods, while effective, require millions of labeled points—making them impractical for large-scale applications. Based on these challenges, researchers set out to develop an unsupervised approach capable of performing wood-leaf separation automatically.

A study (DOI: 10.1016/j.plaphe.2025.100064) published in Plant Phenomics on 6 June 2025 by Shuai Liu’s team, Central South University of Forestry and Technology, holds major potential for sustainable forestry and climate research.

In this study, the team introduced an end-to-end unsupervised semantic segmentation network for wood–leaf separation and evaluated it across diverse ecological conditions. Methodologically, they processed UAV-based LiDAR point clouds drawn from the international FOR-instance dataset (Norway, Czech Republic, Australia, New Zealand, Austria) plus a self-constructed Masson pine dataset from Guilin, China. The model uses a sparse convolutional neural network backbone augmented by two custom modules—dual point attention (DPA) and point cloud feature convolutional integrator (PFCI)—to strengthen multi-scale feature extraction and fusion. Instead of human labels, the system generates pseudolabels via super-point clustering, enabling learning directly from unlabeled data. Training ran on high-performance hardware with a voxel size of 0.05, batch size 10, SGD optimization, and phase-specific super-point schedules reaching up to 30,000 iterations. Corresponding results show strong performance across six datasets: overall accuracy (oAcc) 67.6%, mean accuracy (mAcc) 50.2%, and mean IoU (mIoU) 38.5% at the plot level; at the individual-tree level, accuracy increased to 80.9% oAcc, 64.0% mAcc, and 49.7% mIoU. Against mainstream unsupervised baselines GrowSP and PointDC, the proposed network consistently performed better, especially in overlapping wood–leaf regions. Ablation studies verified each module’s contribution: compared to a sparse-CNN baseline (oAcc 56.6%), adding DPA and PFCI lifted accuracy to 67.6% while improving discrimination at branch–leaf junctions. Robustness tests with random masking demonstrated resilience to incompleteness: performance declined only slightly at 30% masking and remained useful even at 60–90% masking, reflecting compensation from geometric features when neighborhood information is lost. Finally, cross-domain tests on the indoor S3DIS dataset confirmed strong generalizability beyond forestry. Collectively, the method reduces reliance on costly manual annotation while delivering stable, high-quality wood–leaf separation across varied forest types and challenging data conditions.

By eliminating the dependency on labeled data, the network dramatically reduces the time and cost of analyzing LiDAR datasets, accelerating the adoption of automated forest monitoring. Accurate wood-leaf separation enables improved estimates of forest biomass, growth rates, and carbon sequestration—critical parameters for climate change modeling and carbon market verification. It also enhances applications such as tree phenotyping, wood quality assessment, and growth monitoring, providing powerful tools for researchers, policymakers, and forest managers worldwide.

###

References

DOI

10.1016/j.plaphe.2025.100064

Original URL

https://doi.org/10.1016/j.plaphe.2025.100064

Funding information

This work was supported by the Scientific Research Program of Hunan Provincial Department of Education (No. 22B0258) and Natural Science Foundation of Hunan Province (No. 2024JJ5649).

About Plant Phenomics

Plant Phenomics is dedicated to publishing novel research that will advance all aspects of plant phenotyping from the cell to the plant population levels using innovative combinations of sensor systems and data analytics. Plant Phenomics aims also to connect phenomics to other science domains, such as genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics, and computer sciences. Plant Phenomics should thus contribute to advance plant sciences and agriculture/forestry/horticulture by addressing key scientific challenges in the area of plant phenomics.