Zhong et al. (2025) A 3D point cloud instance segmentation network for extracting individual trees from complex forest scenes

Identification

• Journal: Computers and Electronics in Agriculture
• Year: 2025
• Date: 2025-12-23
• Authors: Yijun Zhong, Shuai Liu, Hua Sun
• DOI: 10.1016/j.compag.2025.111333

Research Groups

• Central South University of Forestry and Technology, Changsha 410004, China

Short Summary

This study proposes a novel 3D point cloud instance segmentation network to accurately extract individual trees from complex forest scenes, addressing challenges like under-segmentation and over-segmentation. The network achieves superior performance compared to existing methods, providing reliable technical support for automated forest resource management.

Objective

• To develop a novel 3D point cloud instance segmentation network capable of high-precision individual-tree instance segmentation in complex forest scenes, specifically aiming to improve accuracy and mitigate under-segmentation and over-segmentation issues.

Study Configuration

• Spatial Scale: Individual trees within complex forest scenes.
• Temporal Scale: Static analysis of 3D point cloud data.

Methodology and Data

• Models used: A novel 3D point cloud instance segmentation network incorporating a sliding window mechanism, contextual feature enhancement module, breast-height centroid prediction, and adaptive clustering strategy. Compared against: SoftGroup, ForAINet, TreeLearn, and traditional individual-tree segmentation algorithms.
• Data sources: Public datasets and self-constructed datasets.

Main Results

• The proposed network achieved an average precision (AP) of 63.28 %, AP50 of 71.69 %, and AP25 of 80.54 %.
• It outperformed traditional individual-tree segmentation algorithms and current mainstream deep learning-based methods (SoftGroup, ForAINet, TreeLearn) in terms of correct detection, omission, and commission.
• The method excelled in tree canopy boundary delineation and effectively suppressed under-segmentation and over-segmentation.
• Ablation studies confirmed the critical contributions of the contextual feature enhancement module and adaptive clustering to segmentation accuracy.

Contributions

• Introduction of a novel 3D point cloud instance segmentation network specifically designed to overcome challenges in complex forest scenes.
• Integration of a contextual feature enhancement module, breast-height centroid prediction, and an adaptive clustering strategy to significantly improve segmentation accuracy.
• Demonstrated superior performance against existing state-of-the-art deep learning and traditional methods in individual-tree instance segmentation.
• Provides a robust technical solution for automated forest resource management and ecological monitoring by enabling high-precision individual-tree extraction.

Funding

• Not specified in the provided text.

Citation

@article{Zhong20253D,
  author = {Zhong, Yijun and Liu, Shuai and Sun, Hua},
  title = {A 3D point cloud instance segmentation network for extracting individual trees from complex forest scenes},
  journal = {Computers and Electronics in Agriculture},
  year = {2025},
  doi = {10.1016/j.compag.2025.111333},
  url = {https://doi.org/10.1016/j.compag.2025.111333}
}