The timing of silvicultural activities within a plantation forest is of critical importance for optimal management. Mistimed treatments, including pruning and thinning, can affect the growth of the trees and the quality and value of the produced timber (Pinkard et al. 2004; Muñoz et al. 2008). In order to ensure that these decisions are informed with accurate and timely information, forest inventories are collected at key stages of the
plantation growth cycle. These inventories involve the measurement of several properties that describe the geometry and spatial distribution of trees including stem count, tree height and diameter at breast height (DBH), and the visual and subjective assessment of canopy closure and stem form.
Forest inventory metrics have historically been acquired from sampling plots distributed within a stand or stands using ground-based field measurement techniques. Statisti- cal inference methods are then applied to these sample plot observations to achieve estimates of the stand-level conditions (Mandallaz 2007). Amongst many factors, the quality of these inferences relies on the variability in the stand being accurately captured within the set of observations (Lovell et al. 2005). This is a function of the number of observations in relation to the variance of the whole population. As a consequence, meeting precision requirements with traditional inventory methods can be both a costly and time consuming exercise (Hopkinson et al. 2004). Therefore, new approaches to obtaining these metrics have been continually developed with the goal of increasing measurement accuracy as well as reducing the cost of inventory collection.
The use of remote sensing data for the collection of forest inventory metrics has been widely researched and these data are now commonly used within the forestry industry. Air- and space-borne sensors allow spatially explicit data to be collected over large ar- eas in a timely and economic fashion (Wulder et al. 2008; Boudreau and Nelson 2008; Li et al. 2013). However, as spatial extent and resolution are inversely related, the precision and accuracy are often sub-optimal for many applications (Xie et al. 2008; Wulder et al. 2012). Airborne Laser Scanning (ALS) data captured at heights between 500 and 5000 m provide observations of 3D canopy structure and are often successfully used within model and design-based inference approaches to provide estimates of for- est properties such as biomass and Leaf Area Index (LAI) at the stand level (Nelson et al. 2003; Andersen et al. 2011; Ståhl et al. 2011). The detection of individual trees from this information has proved to be more difficult with reported detection rates vary- ing significantly (between 40 and 96%)(Reitberger et al. 2009; Holopainen et al. 2010; Kaartinen et al. 2012). Although, Reitberger et al. (2009), Ferraz et al. (2012) and Yao et al. (2012) demonstrated that 3D segmentation techniques improve the accuracy of the information derived from tree level analysis, the required accuracy of tree segmentation is still not sufïňĄcient to estimate a number of forest metrics. For example, Vastaranta et al. (2011) showed that, bias towards large dominant trees can cause significant over- estimation of final inventory values such as timber yield. As a consequence, extensive networks of ground plots are required to link statistical properties of the point cloud with properties of the forest to derive stand level metrics and allow ALS to be used as
an operational inventory tool (Wulder et al. 2012).
Terrestrial remote sensing techniques have also been deployed to provide estimates of inventory metrics (Macfarlane et al. 2007; Maas et al. 2008). In contrast to the data collected by air- and space-borne sensors, the woody components of the canopy are often visible within the data collected by terrestrial sensors, allowing for the objective and reproducible measurement of a number of key tree-level inventory metrics such as DBH, stem ovality and crown length that may improve estimation of timber quality at harvest (Maas et al. 2008). The precision of these data also allows direct observation of change to be made (Liang et al. 2012). However, terrestrial techniques can only be used to measure small areas, as the data collected with Terrestrial Laser Scanning (TLS) instruments are highly affected by occlusions and multiple viewing points are required within each plot to avoid downward bias in stem detection. Mobile Laser Scanning (MLS) systems overcome the small area restriction of TLS by deploying the laser scanner onboard a moving vehicle (Lin et al. 2012) or using hand-held instrumentation. These systems have shown potential in deriving individual tree level parameters (Lin et al. 2012), however, their use in forest inventories requires further investigation.
Recently, Unmanned Aerial Vehicle-Laser Scanning (UAVLS) systems have been pro- posed as a tool for mapping and measuring tree metrics (Jaakkola et al. 2010; Wallace et al. 2012b). These systems offer comparatively low-cost collection and generate data with point densities up to 1000 points per m2 (Jaakkola et al. 2010). UAVLS systems are a relatively new technology and to date their deployment within a forest inventory context has focussed on system development and potential, with limited observation of tree parameters. Jaakkola et al. (2010) presented a pilot study showing that the underestimation of tree height present in both ALS and TLS due to the tree tops not being observed was significantly reduced within UAVLS data. Jaakkola et al. (2010) suggested this was due to the increased point density of the collected point clouds in comparison to ALS. However, their study relied upon measurements from only 26 trees. Similar results for tree height observations were found by Wallace et al. (2012b) from repeat measurements of six isolated trees. Within this latter study it was also shown that measurements of the crown width of these trees are repeatable to within a standard deviation of 0.6 m.
In comparison to ALS systems, the scanner and the sensors used for direct georeferencing the point cloud on-board UAVLS platforms typically have higher errors and different sources of error. Although data capture from these platforms enables some desirable properties, the effects of the sensor characteristics on estimates of forest metrics are
unknown. The aim of this study is, therefore, to validate and verify the precision of forest metrics from data collected with a UAVLS system in a Eucalypt plantation forest. The paper describes a workflow for the automatic extraction of inventory metrics at the individual tree level. This workflow includes a determination of the effect of scan geometry on the output. The repeatability of all stages of this workflow and metrics derived at both the plot and tree level are assessed utilising multiple datasets collected over 6 plots.