Abstract
Laser-scanned point clouds of forests make it possible to extract valuable information for forest management. To consider single trees, a forest point cloud needs to be segmented into individual tree point clouds. Existing segmentation methods are usually based on hand-crafted algorithms, such as identifying trunks and growing trees from them, and face difficulties in dense forests with overlapping tree crowns. In this study, we propose TreeLearn, a deep learning-based approach for tree instance segmentation of forest point clouds. Unlike previous methods, TreeLearn is trained on already segmented point clouds in a data-driven manner, making it less reliant on predefined features and algorithms. Furthermore, TreeLearn is implemented as a fully automatic pipeline and does not rely on extensive hyperparameter tuning, which makes it easy to use. Additionally, we introduce a new manually segmented benchmark forest dataset containing 156 full trees, and 79 partial trees, that have been cleanly segmented by hand. The data is generated by mobile laser scanning and contributes to create a larger and more diverse data basis for model development and fine-grained instance segmentation evaluation. We trained TreeLearn on forest point clouds of 6665 trees, labeled using the Lidar360 software. An evaluation on the benchmark dataset shows that TreeLearn performs equally well or better than the algorithm used to generate its training data. Furthermore, the method's performance can be vastly improved by fine-tuning on the cleanly labeled benchmark dataset. The TreeLearn code is available from https://github.com/ecker-lab/TreeLearn. The data as well as trained models can be found at https://doi.org/10.25625/VPMPID.
Introduction
Extracting detailed information about individual trees—such as height, species, and biomass—from forest point clouds is essential for modern forest management and climate change research. However, "unmixing" a forest into its constituent trees, a process known as instance segmentation, is notoriously difficult.
Traditional methods rely on hand-crafted heuristics, such as identifying trunks and "growing" trees upward. These often fail in dense forests where tree crowns overlap and intertwine. While deep learning has revolutionized 2D image segmentation, its application to 3D forest data has been limited by a lack of high-quality labeled datasets and the computational complexity of processing millions of points. TreeLearn addresses these challenges with a fully automatic, data-driven pipeline.
Methods
TreeLearn moves away from rigid rules in favor of a data-driven approach. The process follows five key steps:
- Tiling: Large forest plots are broken into overlapping 35m x 35m tiles to manage memory.
- Neural Network Prediction: A 3D U-Net processes the voxels to predict a semantic score (tree vs. ground) and an offset vector for every point.
- Projection: Points are "shifted" by their predicted offsets. Ideally, all points belonging to the same tree collapse toward a single coordinate representing the tree's base.
- Single-Step Clustering: Unlike other methods that segment small chunks and try to stitch them together (which causes fragmentation), TreeLearn clusters the entire projected point cloud at once using density-based clustering.
- Post-processing: Any remaining points are assigned to the nearest identified tree instance.
Key Methodological Innovations
A critical innovation in TreeLearn is the 3-meter offset strategy. The model predicts offsets to the tree trunk at a 3-meter height rather than at ground level. At this height, trunks are typically more distinct and further apart than at the very base, significantly simplifying the task for the clustering algorithm.
Furthermore, TreeLearn avoids the "Frankenstein effect" common in block-merging approaches. By performing global clustering on the projected points, the model maintains the cohesion of tree instances, preventing a single tree from being split into multiple fragments.
Datasets and Evaluation
The study utilizes a mix of existing and brand-new data:
- Lidar360 Labels: 6,665 trees from 18 plots in Germany, labeled automatically for "noisy" pretraining.
- L1W (New Benchmark): A high-quality, manually segmented dataset of 156 trees serving as the "gold standard."
- External Datasets: Evaluation on the Wytham Woods (UK) and LAUTx (Austria) datasets demonstrated the model's generalizability.
The authors show that while pretraining on "noisy" labels provides a strong foundation, fine-tuning on a small amount of high-quality manual data (like the L1W set) provides a massive performance boost, particularly in identifying smaller trees.
Results
TreeLearn sets a new state-of-the-art across multiple benchmarks:
- L1W Dataset: TreeLearn achieved a Coverage (Cov) of 91.8%, significantly outperforming ForAINet (73.8%) and SegmentAnyTree (77.5%).
- Wytham Woods: On this difficult "leaf-on" dataset, TreeLearn reached 49.2% Coverage, surpassing all other tested deep learning and classical methods.
- Efficiency: The pipeline processed an 85-million-point forest plot in approximately 60 minutes on a standard workstation, demonstrating its practical utility for large-scale forestry.
Conclusion
TreeLearn represents a significant step forward in forest point cloud analysis. While it still faces challenges with very small trees (under 10m) and occasional branch misassignments in extremely dense areas, its data-driven nature and global clustering approach offer superior performance over traditional methods. Future work may incorporate additional parameters like tree height into the offset prediction to further refine segmentation. The TreeLearn code, data, and trained models are publicly available to support further research in the community.
Summary updated