Ensemble AI unlocks hidden tree crown structures in dense forests

The method not only overcomes limitations of traditional LiDAR scanning but also enables the extraction of other hard-to-measure crown parameters, promising to advance forest management, breeding, and ecological modeling.

Tree crown architecture reflects a tree’s competitive environment and influences growth, productivity, and photosynthesis. Among these traits, HMCW—marking the transition between the lower and upper crown—is a critical but often overlooked parameter. Its uneven distribution is shaped by directional competition, shading, and branch dieback, making it difficult to measure in dense canopies. While UAV and ground-based LiDAR have improved forest phenotyping, crown overlap still hampers accuracy. Traditional spatial structure analysis methods, such as the nearest four trees (NFT) or Voronoi diagrams, lack the directional precision needed to capture localized competition effects on crown shape. Based on these challenges, the research team sought a more realistic way to map crown interactions and couple them with advanced regression models.

A study (DOI: 10.1016/j.plaphe.2025.100018) published in Plant Phenomics on 28 February 2025 by Huaiqing Zhang’s team, Chinese Academy of Forestry, improves the accuracy of predicting hard-to-measure tree crown traits like HMCW, offering a scalable tool for forest structure analysis and management.

In this study, spatial structure units for 1,943 sample trees were first constructed using the proposed BSETC method, which accounts for crown distribution in four cardinal directions, and compared with the traditional nearest four trees (NFT) and Voronoi methods. The BSETC approach identified 2–8 neighboring trees per unit, versus a fixed four in NFT and 2–10 in Voronoi. Comparative analyses of neighbor selection showed that while all methods rely on inter-tree distances, BSETC achieved higher realism by incorporating crown width, distance, and shading effects, yielding neighbors with genuine spatial interaction. Overlap analysis indicated moderate-to-high similarity between BSETC and NFT (0.4–0.7 range) and lower overlap with Voronoi (0.1–0.4 range), while unique neighbor counts were lowest for BSETC (median=0), moderate for NFT (median=2), and highest for Voronoi (median=5). Similarity and dissimilarity metrics confirmed that BSETC aligns with forestry principles but captures directional competition more precisely. Using these spatial units, 11 machine learning algorithms were trained to couple HMCW with phenotype and competition parameters, with features including tree height, directional crown width, and vertical/horizontal competition indices. Hyperparameters were optimized via GridSearchCV, and evaluation across R², RMSE, MAE, MAPE, EVS, and MedAE identified Random Forest (RF) as the best-performing single model on test data (R² = 0.8186). To further enhance accuracy, five ensemble learning methods (Bagging, Boosting, Voting, Stacking, Blending) generated 10,180 model combinations; 398 exceeded RF’s performance. The top ensemble, a Bagging regressor integrating XGBoost, RF, SVR, GB, and Ridge, achieved R² = 0.8346, RMSE reduced by 6.66%, and EVS improved by 1.63% over RF. This confirmed that ensemble learning, combined with refined spatial structure mapping, provides a more accurate, generalizable solution for predicting HMCW from easily measured parameters.

The approach provides a scalable, non-destructive way to estimate HMCW and other challenging crown traits across species with similar architectures. By enabling more precise canopy morphology simulations, it supports studies on photosynthesis distribution, forest growth modeling, and selective breeding. In practical forestry, it can inform thinning strategies, optimize stand density, and improve timber yield predictions. The methodology also strengthens ecological research by allowing finer-scale coupling between environmental conditions and tree phenotypes, critical for climate change adaptation and biodiversity assessments.

##

References

DOI

10.1016/j.plaphe.2025.100018

Original Source URL

https://doi.org/10.1016/j.plaphe.2025.100018

Funding information

This work was funded by Fundamental Research Funds of CAF (CAFYBB2023PA003), Science and Technology Innovation 2030-Major Projects (2023ZD0406103) and National Natural Science Foundation of China (32271877).

About Plant Phenomics

Science Partner Journal Plant Phenomics is an online-only Open Access journal published in affiliation with the State Key Laboratory of Crop Genetics & Germplasm Enhancement, Nanjing Agricultural University (NAU) and distributed by the American Association for the Advancement of Science (AAAS). Like all partners participating in the Science Partner Journal program, Plant Phenomics is editorially independent from the Science family of journals. Editorial decisions and scientific activities pursued by the journal's Editorial Board are made independently, based on scientific merit and adhering to the highest standards for accurate and ethical promotion of science. These decisions and activities are in no way influenced by the financial support of NAU, NAU administration, or any other institutions and sponsors. The Editorial Board is solely responsible for all content published in the journal. To learn more about the Science Partner Journal program, visit the SPJ program homepage.