Canopy structural complexity (CSC) describes how leaves, branches, and trunks are arranged in space—a property that influences everything from light penetration and microclimate to wildlife habitat and carbon cycling. Traditional field surveys rely on plot‑level metrics like tree height variation or canopy cover, but these two‑dimensional or discrete measures miss the true three‑dimensional internal arrangement of canopy elements. LiDAR has revolutionized forest monitoring by capturing millions of points that map canopy volume in 3D, yet extracting pure canopy from these dense point clouds remains a major obstacle. Most methods depend on fully labeled training data—point‑by‑point annotations that are prohibitively expensive and time‑consuming for large‑scale surveys across diverse terrain. Given these challenges, there is a pressing need to develop efficient and generalizable methods for canopy extraction and structural quantification from large‑scale, unlabeled forest point clouds.

A team led by researchers from Nanjing University of Information Science and Technology, in collaboration with the Chinese Academy of Forestry, the University of Waterloo, and Nanjing Forestry University, reports a new weakly supervised framework for canopy extraction and structural complexity assessment. The work, published (DOI: 10.1016/j.plaphe.2026.100169) on January 16, 2026, in Plant Phenomics, combines a three-stage deep learning strategy with a novel multidimensional joint canopy entropy index. Using UAV LiDAR data from a subtropical forest in Jiangxi Province, China, the method achieves segmentation accuracy that surpasses fully supervised models while relying on just three labeled plots for training.

The team’s approach unfolds in three stages. First, a RandLA-Net backbone is trained on a small set of manually labeled plots to establish a baseline model. Second, this model predicts labels across all unlabeled plots, and high-confidence predictions are selected as pseudo-labels based on plot-level average confidence scores. Third, the model is retrained jointly on three data sources: the original labeled plots, the newly generated pseudo-labeled plots with confidence-based weighting, and the remaining unlabeled plots through a consistency-based self-supervised loss. Crucially, the team introduced elevation normalization perturbations during data augmentation—randomly varying how terrain height is normalized—which forces the model to learn terrain-invariant features, dramatically improving performance on steep slopes and complex topography.

For quantification, the new multidimensional joint canopy entropy index fuses five features: three directional projection entropies (horizontal, and two vertical planes) that capture global spatial distribution, plus the mean and standard deviation of local entropy derived from graph-based neighborhood analysis. This combination preserves both the broad canopy silhouette and fine-scale structural heterogeneity. When tested across varying point cloud densities, the index remained robust, achieving a normalized composite score of 0.76 across monotonicity, robustness, and sensitivity—outperforming conventional metrics like canopy cover, top rugosity, fractal dimension, and single-dimension canopy entropy.

“The real breakthrough here is that we can now map forest structure at scale without drowning in annotation costs,” the authors said. “With just three labeled plots, our model learns to distinguish canopy from understory and ground—even on steep slopes and in fragmented stands—by leveraging the structural patterns hidden in unlabeled data. And the new entropy index doesn’t just give a single number; it reads the forest’s three-dimensional signature from both the skyline and the understory, which we believe will change how ecologists monitor forest health and carbon dynamics across landscapes.”

This work has immediate implications for large-scale forest monitoring and sustainable management. The weakly supervised strategy dramatically reduces the manual labor required for LiDAR data processing, making high-resolution canopy mapping feasible across vast, remote, or topographically challenging regions. The multidimensional joint canopy entropy index offers a standardized, ecologically interpretable metric that could support carbon stock estimation, biodiversity assessment, and disturbance monitoring—from selective logging to storm damage. As UAV LiDAR data become increasingly abundant, this framework provides a practical pathway to turn raw point clouds into actionable ecological intelligence. The method is open access under a Creative Commons license, and the authors have made their code and pseudo-labeling pipeline available to facilitate adoption by the broader remote sensing and forestry community.

###

References

DOI

10.1016/j.plaphe.2026.100169

Original Source URL

https://doi.org/10.1016/j.plaphe.2026.100169

Funding information

This work was supported by the Key Laboratory of Land Satellite Remote Sensing Application, Ministry of Natural Resources of the People's Republic of China under Grant KLSMNR-G202506 and the National Natural Science Foundation of China Grants (42371447), the National Key Research and Development Program of China Grants (2024YFD2201202) and the Postgraduate Research and Practice Innovation Program of Jiangsu Province Grants (KYCX25_1621).

About Plant Phenomics

Plant Phenomics publishes breakthrough research that integrates genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics and computer science to advance plant phenotyping—from the cellular level to entire populations. The journal tackles key scientific challenges in phenomics by showcasing cuttingedge technologies for data acquisition, management, interpretation and modelling. Its scope spans practical applications across all plant sciences, aiming to translate novel phenotyping tools into realworld solutions for breeding, crop management and fundamental plant biology. By fostering interdisciplinary innovation, Plant Phenomics serves as a hub for researchers developing highthroughput, accurate and scalable methods to decipher complex plant traits, ultimately accelerating progress in agriculture and plant science.