TreeLearn: A deep learning method for segmenting individual trees from ground-based LiDAR forest point clouds

TL;DR

TreeLearn introduces a fully automatic, deep-learning pipeline for segmenting individual trees from ground-based LiDAR forest point clouds by predicting semantic tree points and offsets toward the tree bases, followed by a single-step density-based clustering to yield complete tree instances. The method is trained initially with noisy, automatically generated labels and then fine-tuned on manually labeled datasets to improve generalization across forest types and sensor setups. A new manually segmented benchmark dataset (L1W) and extensions to LAUTx and Wytham Woods are provided to enable robust training and objective evaluation, with code and data publicly available. Across L1W and Wytham Woods, TreeLearn achieves competitive or superior instance-detection and instance-segmentation performance compared to ForAINet, SegmentAnyTree, TLS2Trees, and Lidar360, highlighting the value of end-to-end 3D-offset-based segmentation and tile-based processing for dense forest canopies.

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. 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. 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 as well as the algorithm used to generate its training data. Furthermore, the performance can be vastly improved by fine-tuning the model using manually annotated datasets. We evaluate TreeLearn on our benchmark dataset and the Wytham Woods dataset, outperforming the recent SegmentAnyTree, ForAINet and TLS2Trees methods. The TreeLearn code and all datasets that were created in the course of this work are made publicly available.

Figures (9)

• Figure 1: (a): predictions generated by TreeLearn on the benchmark forest dataset. (b): a top-view of the manually segmented benchmark forest dataset. For the green area at the edge there are no segmentation labels. It is still included in the dataset since it provides relevant context information for the labeled segment.
• Figure 2: The pipeline for segmenting a forest point cloud. The circled numbers correspond to the five steps of the pipeline that are described in detail in Section \ref{['sec:overview']}. Apart from the model input and output, all images depict a top-view of the forest point cloud. The blue square on the image of step 1 shows the full tile, which is used as input to the network, while predictions are generated only for the inner red area. For illustrative purposes, a side-view of only the inner part is displayed as the model input, although the network receives the entire tile. The blue dots on the image of step 1 indicate that several tiles are being processed one after the other to obtain predictions for the full input point cloud. The colored points on the image of step 4 represent the initial clustering results which only take into account points that have a sufficiently high verticality and are close to the tree base (see Section \ref{['sec:grouping']} for details). The colored points on the image of step 5 represent the final predicted trees after postprocessing.
• Figure 3: Visualizations regarding offset prediction. (a) depicts offsets for two example trees. The offset is the vector from a point towards the x- and y-coordinates of the tree base. The tree base is defined as the location of the trunk at a height of three meters. (b) visualizes the prediction area in relation to the whole tile. To make sure that the network has the necessary context information, predictions are only produced for the inner area of the tile. Offset prediction for points of the outer part of the tile is not possible in general since the corresponding tree bases might not be part of the tile.
• Figure 4: Comparison of instance segmentation results for L1W and Wytham Woods plots. TreeLearn results obtained from large (L1W) or mid setting (Wytham Woods).
• Figure 5: Instance segmentation performance. The top row depicts metrics for segments along the x- and y- axis, while the bottom row depicts metrics for segments along the z-axis. (a) shows results for L1W and (b) for Wytham Woods.