The application of low-cost proximal remote sensing technologies for the biophysical measurement of forest structure

Structural measurements, that form part of forest inventories, are critical to the effective management of forest ecosystems at multiple levels. Proximal Three-Dimensional (3D) remote sensing approaches have been investigated, and more recently adopted for operational use, to meet growing inventory demands. This has largely been enabled by Light Detection and Ranging (LiDAR) technologies. Conventional LiDAR, however, is cost prohibitive to operational use outside established organisations overseeing the care of large extents of forested land. Due to this, there has been continued exploration into low-cost alternative hardware that provide 3D representations of forests. Modern Colour and Depth (RGB-D) sensors offer new opportunities due to advancements within sensor hardware, computational resources and spatial mapping algorithms. RGB-D sensors are now often integrated into consumer devices, increasing their accessibility for untrained operators. Presented across four Research Questions (RQ), this thesis aimed to explore the feasibility and application of RGB-D devices for the acquisition of biophysical forest measurements within both urban and native forest environments. RQ1 reviewed the current state of research surrounding the application of terrestrial low-cost 3D remote sensing technologies for forest inventory tasks and how this relates to current operational requirements. This was conducted through the examination of past literature and surveying forestry professionals regarding the importance and capture complexity of different structural forest measurements. Current research focus regarding inventory measurements captured by low-cost sensors was found to align with metrics identified on average as important, defined as ≤ 4 on a 5-point Likert scale, by survey respondents. Based on this investigation, a suite of research directions were proposed to promote the operational adoption of RGB-D devices for forest inventory tasks; (a) integration of RGB-D sensors into handheld or wearable devices for forestry professionals, (b) development of bespoke Simultaneous Localisation and Mapping (SLAM) algorithms for forestry environments, (c) development of a framework for RGB-D sensor operation and assessment in different forest environments, and (d) the exploration of plot-scale inventory capture that utilise low-cost devices from both terrestrial and airborne perspectives to overcome limitations associated with each approach. RQ2 and RQ3 were then designed to address point RQ1.c. The second research question assessed the accuracy and application of the Microsoft Azure Kinect, a Time of Flight (ToF) RGB-D device, for measuring individual stem Diameter at Breast Height (DBH) within urban parkland. This study also assessed the effect of ambient light and measurement distance when estimating DBH. Individual urban trees (n=51) were captured from one viewing angle at 1 m distance intervals, up to 5 m away, using the various capture settings available to the RGB-D sensor. DBH values were estimated and compared to measurements acquired with diameter tape. Optimal capture parameters with the Azure Kinect were observed to be at a distance of 2 m from the target stem and using the binned near field-of-view capture setting. Root Mean Square Error (RMSE) of DBH estimates when using this approach was 8.4 cm; however, after removing stems with obvious irregularities or non-circular deformation, RMSE was reduced to 3.5 cm (n=38). Ambient light was observed to have little effect on the accuracy of DBH estimates, however, strong ambient light was observed to reduce the effective range of the sensor. SLAM algorithms are commonly used by RGB-D devices to register depth images. However, SLAM may be influenced by spatial drift, resulting in point-cloud misalignment error. RQ3 aimed to evaluate the impact of environmental features on RGB-D SLAM when representing stem structure. Plots were established in urban parkland (S1) to assess the effect of stem proximity, and native woodland (S2), to assess the effect of surrounding vegetation (≤1.3 m), on the performance of three RGB-D devices. RGB-D measurements were then compared to those acquired using Terrestrial Laser Scanning (TLS). Depth-frame misalignment, as a result of accidental repeat stem observations, was visible in point clouds from all RGB-D devices. However, there was no significant difference in DBH error when comparing stem representations that were influenced by, and absent of, repeat observation errors at S1 (Kinect p = .16; iPad p = .27; Zed p = .79). When using a plot-scale capture approach, the iPad was the only RGB-D device to maintain SLAM position in all plots at S2. When assessing the effect of surrounding under-story vegetation on DBH measurement error, there was significant correlation with the Kinect RGB-D device (p = .04). Conversely, no significant relationship was observed for representations captured with the iPad (p = .55) and Zed (p = .86) devices. Of the assessed RGB-D devices, the iPad had the lowest DBH RMSE estimates across both individual-stem (DBH RMSE = 2.2 cm) and continuous-plot (DBH RMSE = 3.2 cm) capture approaches.   With evidence that RGB-D devices can provide representations of the lower area of stems (≤4 m) in complex native forest environments, RQ4 was formulated to address point RQ1.d. RQ4 aimed to determine the benefits acquired through the alignment and fusion of low-cost iPad RGB-D and drone Structure from Motion (SfM) point-clouds when representing forest structure. A 0.15 ha plot, established in native eucalypt forest, was captured using the low-cost devices. Registration marks, distributed beneath canopy gaps, were used to align the low-cost point clouds. Estimates of stem location, DBH, height and crown area were extracted from both the fused and standalone low-cost point clouds and compared with TLS measurements. The iPad was able to represent the majority of stems with DBH ≥ 5 cm (n=131/159), providing estimates of stem DBH (RMSE = 2.7 cm) and location (RMSE = 1.21 m). Conversely, drone SfM point clouds represented 21 stems at DBH height (DBH RMSE = 6.5 cm}, location RMSE = 0.33 m), however, did provide estimates of height to canopy (RMSE = 2.79 m) and crown area (RMSE = 12.09 m2). The fused Low-Cost 3D (LC3D) point cloud provided improvement when estimating stem location (RMSE = 0.37 m) and segmented crown area (RMSE = 7.6 m2). Whilst slight improvement was observed in estimates of stem DBH (RMSE = 2.5 cm), it was not significant (p = .98). Incomplete representation of structure between 2 m and 7 m meant that the fused LC3D approach struggled with estimates of stem height due to poor representation of forest mid-story (RMSE = 3.18 m). The research conducted as a part of this thesis has provided insight into the feasibility of contemporary RGB-D devices for the measurement of tree structures in both urban and native forests. The principal outcomes of the research presented within this thesis are: (i) modern Time of Flight (ToF) RGB-D sensors are more resilient to ambient light, making them appropriate for outdoor application; (ii) SLAM registration error does not significantly affect the accuracy of stem structural measurements; (iii) the first investigation and intercomparison of RGB-D devices in complex native forest environments; and (iv) the novel fusion of handheld RGB-D and consumer drone SfM point cloud products, presenting the benefits of aligning the two data sets. RGB-D sensors, integrated into consumer devices, offer the opportunity to improve access to 3D representations of forest structure. Future research should focus on how these sensors fit within the greater constellation of remote sensing approaches for representing forests. Recognising that with the decreasing price of hardware, cost-effective and sustainable forest management relies on efficient sampling and leveraging remotely sensed data captured from different platforms and across multiple scales.