Mitch Bryson

Deep learning, neural networks and other data-driven processing techniques are increasingly used in the analysis of LiDAR point cloud data in forest environments due to the benefits offered in accuracy and adaptability to new environments. One of the downsides of these techniques in practical applications is the requirement for manually annotated data necessary for training neural networks, which can be time consuming and costly to attain. We develop an approach to training neural networks for forest tree stem segmentation from point clouds that uses synthetic data from a custom tree simulator, which can generate large quantities of training examples without manual human effort. Our tree simulator captures the geometric characteristics of tree stems and foliage, from which automatically-labelled synthetic point clouds can be generated for training a semantic segmentation algorithm based on the PointNet++ architecture. Using evaluations on real aerial and terrestrial LiDAR point clouds from a range of different forest sites, we demonstrate our synthetic data-trained models can out-perform, or provide comparable performance with models trained on real data from other sites or when available real training data is limited (increases in IoU from 1–7%). Our simulation code is open-source and made available to the research community.

Accurate measurement of the geometric parameters of trees is a vital part of forest inventory in forestry management. Aerial and terrestrial Light Detection and Ranging (LiDAR) sensors are currently used in forest inventory as an effective and efficient means of forest data collection. Many recent approaches to processing and interpreting this data make use of supervised machine learning algorithms such as Deep Neural Networks (DNNs) due to their advantages in accuracy, robustness and the ability to adapt to new data and environments. In this paper, we develop new approaches to deep-learning-based forest point cloud analysis that address key issues in real applications in forests. Firstly, we develop a point cloud segmentation framework that identifies tree stem points in individual trees and is designed to improve performance when labelled training data are limited. To improve point cloud representation learning, we propose a handcrafted point cloud feature for semantic segmentation which plays a complementary role with DNNs in semantics extraction. Our handcrafted feature can be integrated with DNNs to improve segmentation performance. Additionally, we combine this feature with a semi-supervised and cross-dataset training process to effectively leverage unlabelled point cloud data during training. Secondly, we develop a supervised machine learning framework based on Recurrent Neural Networks (RNNs) that directly estimates the geometric parameters of individual tree stems (via a stacked cylinder model) from point clouds in a data-driven process, without the need for a separate procedure for model-fitting on points. The use of a one-stage deep learning algorithm for this task makes the process easily adaptable to new environments and datasets. To evaluate our methods for both the segmentation and parameter estimation tasks, we use four real-world datasets of different tree species collected using aerial and terrestrial LiDAR. For the segmentation task, we extensively evaluate our method on the three different settings of supervised, semi-supervised, and cross-dataset learning, and the experimental results indicate that both our handcrafted point cloud feature and our semi-supervised and cross-dataset learning framework can significantly improve tree segmentation performance under all three settings. For the tree parameter estimation task, our DNN-based method performs comparably to well-established traditional methods and opens up new avenues for DNN-based tree parameter estimation.