Takeaways

• Researchers have developed an unsupervised AI network that separates wood and leaves in 3D forest scans without requiring manual labels.
• The breakthrough reduces the cost and time of forest monitoring, improving estimates of biomass and carbon storage.
• The model showed strong accuracy across global datasets, offering reliable performance under varied conditions.

A new artificial intelligence breakthrough could transform how forests are monitored worldwide. Researchers at the Central South University of Forestry and Technology in China have unveiled an unsupervised AI network that can automatically separate wood and leaves in 3D LiDAR forest scans, a task that has long challenged scientists due to the complexity of tree structures and the expense of manual data labeling.

The study, published in Plant Phenomics on June 6, 2025, addresses one of the toughest hurdles in automated forestry: Distinguishing between trunks, branches, and leaves in dense 3D datasets. Such analysis is vital for assessing forest health, biodiversity, and carbon sequestration potential, yet conventional supervised machine learning requires millions of labeled data points, making large-scale monitoring impractical.

Read More: Understanding AI Pollution: Environmental Impact and Sustainable Solutions

Instead, the new system uses an unsupervised semantic segmentation network that learns directly from unlabeled LiDAR point clouds. Drawing on datasets from Norway, the Czech Republic, Australia, New Zealand, Austria, and China, the team’s model incorporates a sparse convolutional neural network with two custom modules, such as the dual point attention (DPA) and the point cloud feature convolutional integrator (PFCI), to enhance accuracy in differentiating wood from foliage.

Unlike conventional approaches, the network generates pseudolabels via super-point clustering, reducing reliance on costly human annotations. Testing showed the model achieved an overall accuracy of 67.6% at the plot level and up to 80.9% at the individual tree level, outperforming existing unsupervised baselines such as GrowSP and PointDC.

The AI also demonstrated resilience when parts of the data were removed, maintaining reliable performance even with up to 60% masking. Further tests confirmed that its capabilities extend beyond forestry applications, showing strong results in indoor 3D datasets.

Experts say this innovation has wide-reaching implications. By eliminating the need for labeled training data, it dramatically lowers the cost of analyzing forest structures while improving estimates of forest biomass, growth, and carbon storage, metrics critical for climate change modeling and verification in carbon markets.

Also Read: UN: AI Fuelled 150% Rise in Tech Giants’ Data Centre Emissions

The technology could also enhance applications like tree phenotyping, wood quality assessment, and monitoring growth dynamics. For policymakers and forest managers, this means faster, more accurate, and more affordable tools for managing natural resources in the face of climate change.

Follow more news and views via our ESG Tech and Featured Articles sections, and stay updated on the top ESG events to attend in 2025 for industry insights and networking.

Source: news wise