5.2 Study Area and Field Data
5.3.1 UAVLS data and processing
Discrete return small footprint UAVLS data were captured in transects with a multi- rotor UAV system (OktoKopter), equipped with an Ibeo LUX laser scanner (Wallace
Figure 5.1: Study area. Top Left: Location of the study area in Tasmania, Australia. Bottom left: The stand used for the study area showing the location of the six plots (GDA94 MGA zone 55). Upper right: image of theEucalyptus globulus within one of the chosen plots. Lower
right: Example of selected UAV flight lines over a single plot.
Table 5.1: The properties of the six measured plots within theEucalyptus globulusplantation. Showing mean and standard deviation (σ) tree height, DBH and crown radius.
stem Tree Height (m) DBH (m) Crown Radius (m)
Plot count mean σ mean σ mean σ
1 49 5.71 0.99 0.07 0.01 1.51 0.35 2 34 6.47 1.42 0.09 0.01 1.90 0.46 3 59 8.93 1.31 0.09 0.02 1.80 1.03 4 78 7.05 1.59 0.08 0.02 1.58 0.34 5 46 8.80 1.45 0.09 0.01 1.53 0.52 6 42 8.86 1.50 0.11 0.03 1.52 0.47
et al. 2012b). The laser scanner, which operates at a wavelength of 905 nm, employs four parallel scan lines in the along-track direction capable of recording up to 3 returns per pulse and has a beam divergence of 0.8◦ along-track and 0.08◦ across-track. Direct georeferencing of laser returns was achieved through the use of a dual frequency GPS receiver, a Micro-Electro-Mechanical (MEMS) based Inertial Measurement Unit and a HD Video camera. Observations of orientation based on frames taken from the video camera are used to constrain the error characteristics of the IMU and achieve a more reliable estimate of orientation. This allows the root mean square error (RMSE) of the directly georeferenced returns to be 0.30 m horizontally and 0.15 m vertically. See Wallace et al. (2012b) for a detailed description of the UAVLS system, including an assessment of the error budget.
Each transect was flown with a nominal velocity of 2.8 m/s and a flying height of 40 m Above Ground Level (AGL). Based on the analysis in Wallace et al. (2012a) the use of large scan angles causes significant occlusion within the final point cloud.The scan angle range was, therefore, restricted to±30◦from nadir. Under these conditions a point cloud with greater than 50 pulses per square meter (p/m2) is produced with an on ground swath width of 46 m and footprint ranges of 0.55 to 0.64 m and 0.05 to 0.06 m across and along track respectively. All of the following data processing steps were carried out using in-house MATLAB code.
Ten individual transects were flown over each of the 6 plots. Transects over plots 1, 2 and 3 were flown in May 2012 and transects over plots 4, 5 and 6 were flown in July 2012. As data were acquired in winter there were no significant seasonal differences in the vegetation between these periods. Two different merging strategies were trialled for this study. In both strategies data collected in two different transects were merged based on the georeferenced location of the laser returns without adjustment. In the first strategy (Fig. 5.2), two parallel transects approximately 10 m apart on either side of the plot centre were merged. This produced a merged point cloud in which the entire plot was observed with a minimal scan angle, but the point density in the middle of the plot was significantly greater than the point density towards the plot boundary. In the second strategy (Fig. 5.2), two perpendicular flight lines were merged. This created a merged point cloud in which the scan angles of individual points varied widely but the point density was more evenly distributed across the plot. Variations in flight parameters resulted in inconsistencies in the area captured within each transect. Only merged point clouds with complete coverage of the plot area plus a 2 m buffer were included in any further analysis. As a result, for each plot the final number of merged point clouds was as low as 10 and as high as 19.
Figure 5.2: Simulated distribution of scan angle and point density at the top of the canopy of a 12.62 m radius plot based on the a) parallel transect and b) crossed transect merging strategies applied in this study. Simulation assumes a flying height of 30 m above the canopy
height and a speed of 2.8 m/s. Flying height above canopy was used to ensure that the maximum scan angle used to observe the data is included in the simulated output.
For each merged point cloud, single and last returns were classified into ground and non- ground using the filtering algorithm outlined in Axelsson (1999). All first of many and intermediate returns were considered as non-ground, as these returns were unlikely to be from ground sources and found in initial investigations to be often erroneously classified by the filtering algorithm. Once the ground points were identified, natural neighbour interpolation was used to generate 0.25 m resolution Digital Elevation Models (DEMs). Subsequently, all non-ground points were normalised to vegetation height by subtracting the DEM elevation at the planimetric location using linear interpolation.
Due to the high spatial density of the data and an understory cover which varied sig- nificantly in height and coverage across each plot, a second filter was then applied to separate understory and tree returns. An understory filter was implemented in which all points below 2.5 m with no higher points (>2.5 m) within a 0.7 m radius were classified as understory. The 2.5 m height threshold was chosen as it is significantly higher than the expected maximum height of understory in this area. A 0.7 m radius threshold
allowed sufficient space to ensure low branches were not included as understory. This initial set of understory points was used to create an initial Triangular Irregular Network (TIN). All points above this TIN are consider as canopy returns and the lowest point below each facet of the TIN were added to the set of understory points from which a new TIN is created. This process was continued until there were no non-ground points below the TIN. Iterating in this manner ensured the lower parts of the crowns were included as tree vegetation and the majority of the understory removed by the filter. This approach was taken instead of a simple height threshold as the trees within the study were yet to be pruned and the field data indicated that the crown base height ranged from ground level to 1.2 m.