Developing novel methods for the capture and analysis of digital and physical samples in complex forest environments

Forest measurements and samples are of great importance when modelling, mitigating, and adapting to the impacts of climate change, habitat degradation and bushfires. However, fine scale measurements and samples of trees are largely still collected through manual methods, limiting the scope and scale of the information that can be practically collected, and in some situations, putting workers at risk of injury. Ongoing advancements in Unoccupied Aircraft Systems (UAS) present opportunities for safer, more efficient, and more scalable alternatives to the manual collection of detailed forest information. In this thesis, there are three major sets of contributions described across four major chapters. First, the application of UAS photogrammetry to the mapping of complex forests under the canopy is demonstrated. The next two chapters describe a novel deep-learning based approach to semantically segmenting complex and diverse forest point clouds and how this segmentation method can simplify the fully automated extraction of detailed forest measurements from such point clouds. The last major chapter describes and demonstrates a novel UAS design for the collection of physical canopy samples, as a step towards combining autonomous forest mapping, automated data processing, and autonomous canopy sample collection based upon the point cloud derived information. Conventionally, UAS are flown above the forest canopy for remote sensing of forests, yet dense canopy cover can impede the completeness and quality of data when point clouds of complex forest structures and individual stem characteristics are desired. Under-canopy UAS are rapidly emerging as a valuable addition to both ground based and aerial measurement techniques in areas of dense canopy cover and dense undergrowth. This thesis begins with an investigation into the application of UAS based photogrammetry beneath the forest canopy, the purpose of which, is to map areas of dense canopy and dense undergrowth using a safer and more easily automated alternative to terrestrial LiDAR or manual plot measurements. An off-the-shelf UAS was used to photogrammetrically map two structurally complex plots, with the resulting 3D map being compared against 378 benchmark diameter measurements. The root-mean-squared errors (RMSE) of these point-cloud based diameter measurements relative to the manually measured benchmark were between 0.011 m and 0.021 m, demonstrating the approach can achieve sufficient accuracy for plot measurements. The absence of an existing suitable point cloud analysis tool for these datasets necessitated the use of time-consuming, semi-manual methods of analysis for these point clouds. This unmet need for an automated point cloud analysis tool resulted in the following two chapters. Extracting useful measurements from forest point clouds has remained a tedious and difficult to automate process in all but the most structurally simple forests, inhibiting the uptake of these technologically-advanced remote sensing approaches to forest measurement. For this reason, this thesis seeks to address the challenge of fully automatically measuring forest structural metrics from diverse and complex forest point clouds. The first part of the approach taken to this problem is to semantically segment these complex point clouds into four categories of points: terrain, vegetation, coarse woody debris (CWD) and stem, and this is described in Chapter 3. The segmentation model achieved an overall label accuracy of 95.4% on the manually labelled benchmark point clouds, greatly simplifying the task of automatically extracting measurements from these complex point clouds. Chapter 4 describes a series of algorithms which are made publicly available as an open-source Python package called the Forest Structural Complexity Tool (FSCT). FSCT, which exploits the semantic segmentation model described in Chapter 3, is capable of fully automated and robust measurement of complex and diverse forest point clouds. The accuracy of this tool for extracting stem diameters and tree heights was evaluated against a dataset containing 49 Terrestrial Laser Scanned (TLS) point clouds with 7022 corresponding benchmark diameter measurements. Diameters were automatically measured with an RMSE of 0.092 m and a mean error of 0.014 m and heights were automatically measured with an RMSE of 3.807 m and mean error of -1.677 m. With rapid and fully automated digital measurements of complex forests successfully demonstrated with FSCT, structural information can now be relatively easily captured, however, structural information alone, while useful, does not provide a complete picture of the state of a forest; there remains a need for physical forest samples, which are largely still obtained in manual, expensive, and in some cases dangerous ways. Physical canopy samples are commonly required in forest management and are currently acquired via either arborists climbing trees, cranes, or even by shooting branches down with shotguns. All of these methods are expensive, time consuming and come with significant workplace health and safety issues. The final chapter describes the development of a novel forest canopy sampling UAS, as a step towards the integration of digital and physical samples into a holistic forest information capture system. This system was successfully demonstrated collecting canopy samples from 30 trees and an associated video of this process has been provided for demonstration purposes. This system also provides a novel solution to take-off and landing on steep terrain, enabling safe operation in areas lacking suitably flat take-off and landing sites. The collection of research works presented in this thesis are important steps towards the fully autonomous capture and processing of high-resolution forest information, even in areas of considerable structural complexity. Ongoing efforts will seek to combine the digital sampling techniques in the form of point clouds and automated processing, with the physical samples that can be captured using the canopy sampling UAS presented. Future work may use the physical sampling UAS developed in this work to aid the calibration of spectral remote sensing techniques, with the longer term goal of fusing multi/hyper-spectral, spatial, and physical forest information into a holistic forest information capture system.