Source-linked AI summary
Automated forest inventory: analysis of high-density airborne LiDAR point clouds with 3D deep learning
Binbin Xiang, Maciej Wielgosz, Theodora Kontogianni, Torben Peters, Stefano Puliti, Rasmus Astrup, Konrad Schindler
TL;DR
The paper addresses automated forest-point-cloud analysis by developing a 3D learning pipeline for segmentation and biophysical feature extraction. It reports 85.1% F-score for individual-tree segmentation, 73.5% mIoU for five-class semantic segmentation, and good predictive skill for several tree properties.
Problem
Automated point-cloud segmentation is a critical challenge in forest analysis, while comprehensive labeled datasets are an important prerequisite.
Method
The paper assembles a point-cloud processing pipeline using a shared feature-extraction backbone for multiple segmentation tasks and subsequent tree-parameter retrieval.
Results
85.1% F-score was achieved for individual-tree segmentation and 73.5% mIoU for five-class semantic segmentation, with good predictive skill for several biophysical properties.
Takeaways & Limitations
The pipeline sets a new state of the art for automated individual-tree analysis and supports extraction of a complete set of forest inventory properties.
Takeaways & Limitations
The 3D learning method is challenged by point densities below 100 pts/m2, and the proof-of-concept tends to reach lower performance in some conditions.
Abstract
from arXiv · showhide
Detailed forest inventories are critical for sustainable and flexible management of forest resources, to conserve various ecosystem services. Modern airborne laser scanners deliver high-density point clouds with great potential for fine-scale forest inventory and analysis, but automatically partitioning those point clouds into meaningful entities like individual trees or tree components remains a challenge. The present study aims to fill this gap and introduces a deep learning framework, termed ForAINet, that is able to perform such a segmentation across diverse forest types and geographic regions. From the segmented data, we then derive relevant biophysical parameters of individual trees as well as stands. The system has been tested on FOR-Instance, a dataset of point clouds that have been acquired in five different countries using surveying drones. The segmentation back-end achieves over 85% F-score for individual trees, respectively over 73% mean IoU across five semantic categories: ground, low vegetation, stems, live branches and dead branches. Building on the segmentation results our pipeline then densely calculates biophysical features of each individual tree (height, crown diameter, crown volume, DBH, and location) and properties per stand (digital terrain model and stand density). Especially crown-related features are in most cases retrieved with high accuracy, whereas the estimates for DBH and location are less reliable, due to the airborne scanning setup.
1. Introduction
High-density airborne LiDAR creates opportunities for fine-scale forest inventories, but reliable, transferable segmentation of individual trees and components remains a central challenge. The paper proposes an integrated deep-learning framework designed to support automated inventory across varied forests.
- Limitations of existing methods: Traditional and CHM-based individual-tree detection tends to miss intermediate and understory trees because it represents mainly the canopy top layer.This omission contributes to low detection rates and limits inventory uptake.
- Motivation: High-density airborne LiDAR provides unprecedented airborne detail, narrowing the gap between large-area ALS and high-resolution terrestrial scanning.ALS-HD data can contain 500–10,000 pts/m2 and reveal individual trees throughout the vertical canopy profile.
- Limitations of existing methods: Existing tree-instance methods often rely on ad-hoc sequences whose hyperparameters and step dependencies hinder transfer to new locations or sensor setups.The cited limitation concerns both repeated tuning and dependencies among steps.
- Motivation: The study addresses a data bottleneck by using comprehensive labeled datasets such as FOR-Instance to enable 3D deep learning for forest inventories.The introduction identifies labeled data as a prerequisite for supervised learning in this setting.
- Contribution: ForAINet integrates forest point-cloud segmentation with geometric retrieval to support automated individual-tree and stand inventories.The framework combines a segmentation engine with geometric algorithms for retrieving tree and stand properties.
- Contribution: The proposed deep-learning approach greatly outperforms a widely used point-cloud processing baseline across forests with different geographic and structural characteristics.The comparison is reported as part of the study’s contribution and evaluation motivation.
2.1. Dataset
FOR-Instance provides annotated airborne LiDAR point clouds spanning five geographic regions and diverse forest conditions, enabling evaluation across heterogeneous data.
- The dataset includes manually annotated semantic classes for low vegetation, ground, stems, live branches, and dead branches.
- FOR-Instance contains drone- and helicopter-acquired point clouds from 67 plots across diverse geographic regions and forest types.
- Incomplete or partially observed trees along plot borders are removed because they lack instance labels, while NIBIO is excluded from quantitative experiments because it lacks DBH references and omits understory annotations.
- Training and test splits contain all five geographic regions, with 42 training plots, 14 validation plots, and 11 test plots.
- Regions differ substantially in species composition, crown structure, point density, topography, and complex tree geometries such as bent or occluded stems.
2.2. Deep learning framework for multiple segmentation tasks
ForAINet jointly predicts semantic classes and individual-tree instances from 3D point clouds, then applies geometric processing to obtain forestry attributes.
- ForAINet labels each 3D point with both a semantic class and instance ID using a shared feature-extraction backbone.
- Data augmentation and balancing strategies: The network uses voxel-grid subsampling, class-balanced sampling, geometric augmentations, and optional point removal to address density imbalance and occlusion.
- Data augmentation and balancing strategies: TreeMix augments training data by replacing target-sample trees with non-overlapping trees from other samples, increasing diversity across locations and countries.
- Semantic and instance segmentation: Semantic prediction produces five-class scores and tree/non-tree probabilities, excluding non-tree points from subsequent individual-tree segmentation.
- Semantic and instance segmentation: Instance segmentation clusters complementary point embeddings and predicted center offsets to assign individual tree IDs.
2.3. Automated retrieval of tree parameters and stand structure
The pipeline derives stand-level terrain and density measures alongside per-tree dimensions from segmented point clouds using geometric computations.
- The system retrieves stand-level digital terrain models and stand density, plus per-tree height, crown diameter, crown volume, live crown volume, DBH, and location.
- Stand density is computed from the number of segmented trees divided by their estimated surface area from a two-dimensional convex hull.
- Tree height uses the highest point of each tree relative to interpolated ground elevation, after filtering isolated canopy outliers with HDBSCAN.
- Crown diameter is estimated from the smallest enclosing circle of projected live- and dead-branch points, while crown volume uses a three-dimensional convex hull.
- DBH and location are obtained by fitting a circle to projected stem points around breast height, with diameter defining DBH and center defining location.
2.4. Evaluation metrics
Evaluation combines segmentation metrics, instance matching and coverage measures, and geometric errors for terrain and tree attributes.
- Point cloud segmentation: Semantic segmentation is assessed with confusion matrices, overall and mean per-class accuracy, and mean intersection-over-union.
- Individual tree segmentation: Individual-tree predictions are matched to reference trees by point-set IoU, with predictions below 0.5 IoU counted as false positives.
- Individual tree segmentation: Coverage measures how closely predicted instance boundaries agree with ground truth by averaging the best IoU match for each reference tree.
- Biophysical attributes: Terrain and tree-attribute accuracy use regression diagnostics and RMSE, with RMSE% additionally normalizing crown-related errors by average reference values.
- Limitations: DBH estimates are relatively inaccurate because airborne viewing is vertical through the canopy, while understory evaluation is constrained by missing individual reference annotations.
2.5. Implementation details
The implementation used publicly available Python code, modified geometric-feature extraction, Torch-Point3D, and validation-set grid search for HDBSCAN parameters.
- The study's code is publicly available, and experiments used a machine with an 8-core Intel CPU, 8 GB memory per core, and one 24 GB Nvidia Titan RTX GPU.
- Geometric-feature extraction used a slightly modified version of code published by Guinard and Landrieu (2017).
- The point-cloud segmentation network was constructed with the Torch-Point3D library.
- HDBSCAN filtering hyperparameters were selected by grid search on the validation set.
- Table 4 reports instance- and semantic-segmentation results under different settings, with bold values marking the best result in each column.
3. Results
The experiments evaluate loss, sampling, input-feature, and augmentation choices for semantic and individual-tree segmentation, then analyze the selected TreeMix configuration across plots and forest-inventory attributes.
- 3. Results: Adding binary tree-versus-non-tree supervision increases tree-versus-non-tree IoU by 1.8 percentage points and individual-tree F-score by 1.6 percentage points.
- 3. Results: Return number provides a 5.1 pp gain in individual-tree F-score, while additional features generally improve semantic segmentation except intensity.
- 3. Results: Low-level augmentation increases instance-segmentation F-score by 4.4 pp but slightly reduces semantic-segmentation performance, whereas TreeMix reaches the best single-tree result.
- 3. Results: TreeMix achieves the best single-tree segmentation result: 85.1% F-score, a +6.1 pp gain, while also improving semantic segmentation.
- 3. Results: The selected model shows minimal tree/non-tree confusion, while errors mainly involve semantically similar categories and live-branch confusion.
- 3. Results: Most test plots achieve F-scores above 82%, but Plot 8 and Plot 11 perform relatively poorly, potentially because they have lower point densities.
- 3. Results: Estimated DTM coverage exceeds 98% with RMSE below 26 cm, and tree-height predictions have RMSE% below 0.06 across all plots.
4. Discussion
The proposed 3D learning pipeline generally outperforms conventional and unsupervised baselines for forest point-cloud segmentation, while producing useful tree- and stand-level inventory attributes. Performance is strongest in regular forests and at sufficient point density, but declines for complex structures, sparse scans, unfamiliar domains, and lower-stem measurements.
- The learned model handles plots with strongly varying tree sizes better than Treeiso, which can over-segment large trees or merge small trees.
- The deep learning segmentation approach achieved higher F-scores than Treeiso in all forest regions except CULS.
- The method more than doubled individual-tree F-score on the challenging TUWIEN plot, from 30% to over 60%.
- Accuracy is higher in relatively uniform coniferous forests but lower in structurally intricate forests, although limited training data prevents separating complexity effects from data scarcity.
- The pipeline estimates tree height, crown width, crown volume, stem density, and forest-floor terrain, supporting tree-level inventories and fine-scale management.
- Performance remains within 5 percentage points for point densities at least 75 pts/m2, whereas omission error rises markedly below that threshold and results are challenged below 100 pts/m2.
5. Conclusion
The pipeline uses ForAINet to segment high-density airborne LiDAR into semantic classes and individual trees, then derives tree- and plot-level inventory attributes. It performs strongly across forest types, while complex structure, low point density, and airborne scanning geometry constrain some results.
- ForAINet extracts semantic class labels and individual trees, enabling automatic estimation of tree- and plot-level attributes from ALS-HD data.The complete pipeline is designed for inventory of per-tree attributes and includes semantic segmentation, instance segmentation, and downstream attribute estimation.
- 85.1% F-score for individual tree segmentation and 73.5% mIoU for 5-class semantic segmentation were achieved on FOR-Instance.The semantic classes comprise five categories, while the pipeline also evaluates individual tree segmentation.
- Tree height, crown width, and crown volume show good predictive skill, whereas DBH and tree location are less accurate.The lower accuracy of DBH and location is attributed to unfavorable point density caused by airborne scanning geometry.
- The method performs well across varied forest types and detects many understory trees, supporting tree-level forest inventory from remotely sensed data.The reported conclusion links the segmentation outputs to a basis for tree-level inventory, including terrain-related outputs.
- Segmentation quality deteriorates in complex forest structures and when point density falls below approximately 100 pts/m2.The authors identify these conditions as remaining issues requiring further research.
Appendix A. List of abbreviations
The appendix defines abbreviations used for airborne and terrestrial laser scanning, point-cloud processing, forest inventory metrics, models, and evaluation measures.
- ALS denotes airborne laser scanning, ALS-HD denotes very high density ALS point clouds, and ULS denotes unmanned laser scanning.
- DEM, DSM, and DTM denote digital elevation, digital surface, and digital terrain models, respectively.
- FCN, MLP, NMS, and GUI denote fully convolutional network, multi-layer perception, non-maximum suppression, and graphical user interface.
- IoU and RMSE denote intersection over union and root mean square error, while UAV, MLS, TLS, and LiDAR identify related sensing technologies.
- ITC and ITD denote individual tree crown and individual tree detection, while DBH denotes diameter at breast height.