TL;DR
This paper introduces a novel "bottom-up" algorithm for precisely segmenting individual rubber trees from dense plantations using UAV-LiDAR data. By first identifying tree trunks and then fusing multiple features (density, geometry, direction) to delineate crowns, the method achieves high accuracy even in challenging, overlapping canopy scenarios, paving the way for better forest management and biomass estimation.
What Problem Does This Paper Address?
Accurately identifying and separating individual trees in dense forest environments, particularly tropical plantations like rubber trees, is a significant challenge. These plantations often feature overlapping canopies, indistinct tree tops, and complex branch structures, making traditional segmentation methods struggle. Existing approaches, whether based on projected images or point clouds, often falter in these "high-closure" conditions, leading to over- or under-segmentation. This inaccuracy hinders precise measurements of tree structural parameters, 3D modeling, and biomass estimation, which are crucial for effective forest management and ecological assessments.
Methods
The proposed method employs a "bottom-up" approach, starting from the base of the trees and working upwards. The core innovation lies in a multi-feature fusion strategy that moves beyond single-threshold constraints.
- Initial Skeleton Point Generation: The point cloud is first processed to remove noise and ground points. Then, it's vertically stratified, and a density-based clustering algorithm (DBSCAN) is used to generate initial "skeleton points" that represent the tree's basic structure.
- Trunk Extraction: This stage focuses on identifying the main tree trunk. It uses density variations and directional features of the skeleton points, along with neighborhood density and upward search angles, to distinguish trunk points from canopy points. This is crucial for separating the trunk from the complex canopy structure.
- Multi-Feature Fusion for Canopy Segmentation: This is the most novel part. Instead of relying on a single feature, it integrates geometric, directional, and density attributes to classify points into:
- Core Canopy Points: The central, most dense parts of the crown.
- Boundary Points: The outer edges of the crown.
- Disputed Points: Points in overlapping branch regions. These disputed points are then assigned to the correct tree based on the consistency of growth direction angles within their neighborhood, ensuring robust segmentation even with significant overlap.
Datasets
The study utilized UAV-LiDAR point cloud data collected from three sample plots (A, B, and C) within a rubber tree experimental base in Danzhou City, Hainan Province, China. These plots were specifically chosen to represent varying levels of canopy closure: low (Plot A), medium (Plot B), and high (Plot C). Ground truth data for crown width and canopy projection area was obtained through manual segmentation.
Key Insights and Ideas
- Bottom-up is Better for Dense Canopies: By starting with the trunk and moving upwards, the algorithm avoids the common pitfall of needing to accurately identify indistinct tree tops, which is a major limitation of top-down methods in dense forests.
- Multi-Feature Fusion is Key: Moving beyond single-threshold segmentation and integrating density, geometry, and direction provides a more robust way to handle the complex and overlapping nature of rubber tree crowns.
- Handling "Disputed" Points: The method's ability to intelligently assign points in overlapping areas based on growth direction consistency is a critical innovation for accurate individual tree separation.
- Adaptability to Canopy Closure: The algorithm demonstrates consistent high performance across plots with low, medium, and high canopy closure, highlighting its versatility.
Key Results
The proposed algorithm achieved high segmentation accuracies:
- Plot A (Low Canopy Closure): 0.97
- Plot B (Medium Canopy Closure): 0.98
- Plot C (High Canopy Closure): 0.95
Furthermore, the extracted crown width and canopy projection area showed excellent agreement with ground truth data, with R² values exceeding 0.98 and 0.97, respectively. The method also outperformed several existing algorithms (hierarchical DBSCAN, ForAINet, PCS, layer stacking, K-means, and watershed) in these challenging conditions.
Limitations and Future Work
The authors acknowledge that understory vegetation can influence segmentation accuracy. They suggest future work could focus on:
- Mitigating the influence of understory vegetation.
- Developing adaptive thresholding techniques.
- Exploring soft segmentation methods.
- Integrating multi-sensor data fusion.
- Enhancing automation for greater efficiency.
Summary updated