TLDR: The research paper introduces a revised treeX algorithm for unsupervised tree instance segmentation in dense forest 3D point clouds. It combines density-based stem detection with region growing for crown delineation, offering improved accuracy and reduced runtime over its predecessor. The algorithm includes parameter presets for ground-based (TLS/PLS) and UAV-borne (ULS) laser scanning data, making it a resource-efficient and interpretable alternative to deep learning, and a valuable tool for generating training data for deep learning models.
Forests are vital ecosystems, and understanding their structure is crucial for effective management and environmental monitoring. Traditional methods for measuring individual trees can be labor-intensive and time-consuming. Modern laser scanning technologies, which capture detailed 3D point clouds of forest stands, offer a powerful alternative. However, processing this vast amount of 3D data to identify and segment individual trees remains a significant challenge.
Recent advancements have seen the rise of deep learning methods for tree instance segmentation, which can be highly accurate but often demand extensive annotated datasets and substantial computational resources. As a more resource-efficient and interpretable alternative, a revised version of the treeX algorithm has been introduced, offering an unsupervised approach to segment individual trees in dense forest point clouds.
Understanding the treeX Algorithm
The treeX algorithm is designed to take a 3D point cloud (containing xyz-coordinates and optionally reflectance intensity values) and output individual tree instance IDs for each point, along with tree locations and stem diameters at breast height (DBH). The process unfolds in three main stages:
First, a **Digital Terrain Model (DTM) is constructed**. This involves classifying the input point cloud into terrain and non-terrain points using an unsupervised method called cloth simulation filtering (CSF). Once terrain points are identified, a rasterized DTM is created, allowing the algorithm to calculate the height of points relative to the ground.
The second stage focuses on **detecting tree stems**. This is a multi-step process. Initially, a horizontal ‘stem layer’ is extracted from the point cloud, containing points within a specific height range above the DTM. These points are then clustered using the DBSCAN algorithm, first in 2D (xy-coordinates) to identify dense stem-like groupings, and then refined in 3D (xyz-coordinates) to further separate closely spaced stems. To ensure accuracy, these candidate clusters undergo a series of filtering steps. Filters are applied based on the number of points in a cluster, its vertical extent, reflection intensity (if available), and critically, the consistency of circle fitting to horizontal layers within the cluster. This last step helps to confirm that the cluster represents a cylindrical stem rather than irregular crown elements. Finally, for valid stem detections, the algorithm estimates stem positions and diameters at breast height using a generalized additive model (GAM).
The third and final stage is **tree crown delineation**. Building on the detected stem positions and diameters, a region growing method is employed. This process starts by selecting initial ‘seed points’ around each detected stem. Iteratively, neighboring points are added to a tree segment if they haven’t been assigned to another tree, effectively expanding the tree’s crown. The algorithm dynamically adjusts its search radius to balance computational efficiency with segmentation accuracy, promoting upward growth to capture the full crown structure.
Key Innovations and Performance
This revised treeX algorithm significantly improves upon its original version, offering enhanced accuracy and reduced runtime. A major innovation is the introduction of two parameter presets: one for ground-based laser scanning data (Terrestrial Laser Scanning – TLS and Personal Laser Scanning – PLS), which typically have high point densities in the stem layer, and another for UAV-borne laser scanning (ULS) data, which tends to be sparser. These presets optimize the algorithm’s performance for different data characteristics.
Evaluated on six public datasets, the revised treeX algorithm demonstrates competitive accuracy, particularly for TLS and PLS data, often matching or outperforming recent deep learning methods. For ULS data, while deep learning approaches generally perform better due to their ability to handle sparser data, the ULS preset of treeX still achieves a respectable F1-score, whereas the original algorithm failed to segment any correct instances for ULS data. A notable advantage of treeX is its computational efficiency, often requiring less processing time and memory compared to many deep learning models, and importantly, it does not require a GPU.
Also Read:
- Pinpointing Earth Changes: A Network That Connects ‘When’ to ‘Where’
- Autonomous Rovers Get Smarter: Prioritizing Scientific Targets with Depth and Curiosity
Applications and Future Outlook
The treeX algorithm, implemented as an open-source Python package called pointtree, offers two main applications. Firstly, it serves as a resource-efficient and interpretable alternative to deep learning, especially in scenarios where data characteristics (sufficient stem visibility and point density) align with its design. Secondly, it can be invaluable for the semi-automatic generation of labels for deep learning models. By providing pseudo-labels, it can help mitigate pre-training bias and support the creation of more diverse and balanced training datasets, thereby improving the robustness and generalizability of deep learning approaches.
Future work may involve integrating treeX with deep learning methods, for example, by incorporating deep-learning-based semantic segmentation into its stem detection stage or using its detected stem positions as prompts for transformer-based instance segmentation models. This hybrid approach could combine the strengths of both methodologies to address more challenging forest conditions.
