An Assessment of High-Density UAV Point Clouds for the Measurement of Young Forestry Trials

: The measurement of forestry trials is a costly and time-consuming process. Over the past few years, unmanned aerial vehicles (UAVs) have provided some signiﬁcant developments that could improve cost and time e ﬃ ciencies. However, little research has examined the accuracies of these technologies for measuring young trees. This study compared the data captured by a UAV laser scanning system (ULS), and UAV structure from motion photogrammetry (SfM), with traditional ﬁeld-measured heights in a series of forestry trials in the central North Island of New Zealand. Data were captured from UAVs, and then processed into point clouds, from which heights were derived and compared to ﬁeld measurements. The results show that predictions from both ULS and SfM were very strongly correlated to tree heights (R 2 = 0.99, RMSE = 5.91%, and R 2 = 0.94, RMSE = 18.5%, respectively) but that the height underprediction was markedly lower for ULS than SfM (Mean Bias Error = 0.05 vs. 0.38 m). Integration of a ULS DTM to the SfM made a minor improvement in precision (R 2 = 0.95, RMSE = 16.5%). Through plotting error against tree height, we identiﬁed a minimum threshold of 1 m, under which the accuracy of height measurements using ULS and SfM signiﬁcantly declines. Our results show that SfM and ULS data collected from UAV remote sensing can be used to accurately measure height in young forestry trials. It is hoped that this study will give foresters and tree breeders the conﬁdence to start to operationalise this technology for monitoring trials.


Introduction
Forestry trials are established to evaluate and monitor a variety of factors that affect growth. Typical examples include genetics trials, in which different provenances, families and individuals are assessed for their growth characteristics [1] or silvicultural trials, in which various silvicultural regimes are assessed to ascertain optimum management prescriptions to inform future establishment. As trials often require regular measurement of individual trees, this is a costly, time consuming, and frequently difficult task.
The use of remotely sensed data has considerable potential for overcoming difficulties associated with trial measurement. This type of data has been captured and utilised in forestry from manned aircraft since the 1920s, when aerial photography was first used in Canada [2], and satellites since limit at which height predictions can be robustly made using ULS, and the minimum tree height predicted to date has been 5.7 m [57].
Although costs have decreased, ULS is still cost-prohibitive for many operational uses. With low-cost, prosumer-grade UAVs becoming common, structure from motion photogrammetry (SfM) offers an appealing and cost-effective alternative to laser scanning for the assessment of key forestry metrics such as tree height. SfM photogrammetry utilises multiple overlapping images focused on an area of interest to generate a 3D representation of the area. SfM leverages computer vision and photogrammetric principles to simultaneously estimate camera position and scene geometry through common-feature mapping between overlapping images taken from different locations [58]. Through a combination of the SfM methodology and the prosumer-grade UAVs commonly used in forestry, 3D models of forest areas can be created in the form of a point cloud, which provides a low-cost alternative to lidar. SfM cannot penetrate through gaps in dense forest canopies to measure ground height to the same extent as ULS. Terrain under the canopy can only be reconstructed if there are sufficiently large gaps with adequate illumination to enable these gaps to be detected and reproduced through the SfM pipeline. With ULS, smaller gaps can be penetrated with multiple returns, meaning that partially occluded gaps can be resolved, owing to the active nature of the laser not requiring illumination. Pre-existing terrain information from other sources can be combined with SfM data if needed.
A number of studies have compared SfM to ALS [74,75,78,89], and of these, a few have compared the precision of height predictions between ULS and SfM [51,53,54,90]. However, studies comparing the two technologies have returned conflicting results. Cao et al. [51] and Wallace et al. [90] found that ULS outperformed SfM when used to predict Lorey's Mean Height and individual tree height. In contrast, SfM was found to outperform ULS for prediction of plot-level [53] and individual tree heights [54].
In this study, six large trials of the widely grown plantation species Pinus radiata D. Don, that covered a height range of 0.4-6.1 m, were measured and scanned using SfM and ULS. The P. radiata established in these trials included a common commercial seedlot, a selection of clonal stock and a P. radiata x P. attenuata hybrid; the hybrid was also sourced through clonal propagation. Using these data, our objectives were to (i) compare the precision and bias of height models derived from ULS and SfM data, (ii) determine if a model that combines SfM data with a ULS DTM significantly improves tree height estimation over the use of only SfM data and (iii) identity height thresholds above which these data sources can be used to accurately predict tree height.

Study Sites and Field Data
Data were collected from four field trial sites within the central North Island of New Zealand (Figure 1) that covered a broad range of environments to test the generality of the developed methods. Topography at these four sites ranged from flat (Scion), to gently rolling (Rangipo; Kaingaroa 127) and rolling (Kaingaroa 861) terrain. Varying levels of debris and weed cover were also represented through the sites, from regularly mowed research trials (Scion) and first-rotation sites with little harvest residue or weed cover (Rangipo) to multiple-rotation forest sites with moderate levels of harvest residue that are more typical of production forestry (Kaingaroa sites). The establishment techniques at each of the sites also differed. The Scion site was established directly into mowed, flat grass. The Rangipo site was established on a first rotation site with half of the site rip-mounded, with the other half planted into bare pasture. The two Kaingaroa sites were mechanically spot-mounded, a process by which harvest residue is cleared from the planting area and stacked in windrows, and the seedlings are established on a small mound of earth [91,92].
Remote Sens. 2020, 12, x FOR PEER REVIEW 4 of 21 mowed, flat grass. The Rangipo site was established on a first rotation site with half of the site ripmounded, with the other half planted into bare pasture. The two Kaingaroa sites were mechanically spot-mounded, a process by which harvest residue is cleared from the planting area and stacked in windrows, and the seedlings are established on a small mound of earth [91,92]. Within three of the sites (Rangipo, Kaingaroa 127 and Kaingaroa 861) a single trial was selected, while within the Scion site we selected three trials with different age classes (Scion: North, South and West). These six trials encompassed a variety of age classes and a wide range of tree heights. At the time of UAV data capture, the age of the stands at the Scion site ranged from 5 months at the west site to 3.5 years at the south site ( Table 1). The other three trials (Rangipo, Kaingaroa 127 and Kaingaroa 861) were around 3 years old at the time of capture (Table 1). There was a maximum difference of eight weeks between field measurements and UAV data capture and measurements were made in winter to mitigate the impact of this interval on height growth.  Within three of the sites (Rangipo, Kaingaroa 127 and Kaingaroa 861) a single trial was selected, while within the Scion site we selected three trials with different age classes (Scion: North, South and West). These six trials encompassed a variety of age classes and a wide range of tree heights. At the time of UAV data capture, the age of the stands at the Scion site ranged from 5 months at the west site to 3.5 years at the south site ( Table 1). The other three trials (Rangipo, Kaingaroa 127 and Kaingaroa 861) were around 3 years old at the time of capture (Table 1). There was a maximum difference of eight weeks between field measurements and UAV data capture and measurements were made in winter to mitigate the impact of this interval on height growth.
Tree heights were measured at the Kaingaroa and Rangipo sites using a height pole for trees up tõ 6m tall, and a Vertex 4 hypsometer (Haglöf, Langsele, Sweden) for taller trees. Within the Scion site, a height pole was used for the smaller trees at the West site, while a Vertex was used to measure trees at the North and South sites. A total of 6616 trees were measured across the six sites with heights that ranged from 0.12 to 6.1 m ( Table 2) and a mean height of 2.6 m. The distribution of measured heights was bimodal with peaks evident at both 0.4 m and 3.0 m ( Figure 2). Tree heights were measured at the Kaingaroa and Rangipo sites using a height pole for trees up to ~6m tall, and a Vertex 4 hypsometer (Haglöf, Langsele, Sweden) for taller trees. Within the Scion site, a height pole was used for the smaller trees at the West site, while a Vertex was used to measure trees at the North and South sites. A total of 6616 trees were measured across the six sites with heights that ranged from 0.12 to 6.1 m ( Table 2) and a mean height of 2.6 m. The distribution of measured heights was bimodal with peaks evident at both 0.4 m and 3.0 m ( Figure 2).

UAV Data Acquisition Methodology
Ground control was set up prior to data capture. A Trimble Geo7X handheld GPS unit (Trimble Inc., Sunnyvale, CA, USA) with a Trimble Zephyr Model 2 external aerial was used to survey in a minimum of five ground control points (GCPs) following best practice guidelines for Pix4D [93]. RMSE of GCPs collected per site ranged from 0.01 m to 0.11 m.

UAV Data Acquisition Methodology
Ground control was set up prior to data capture. A Trimble Geo7X handheld GPS unit (Trimble Inc., Sunnyvale, CA, USA) with a Trimble Zephyr Model 2 external aerial was used to survey in a minimum of five ground control points (GCPs) following best practice guidelines for Pix4D [93]. RMSE of GCPs collected per site ranged from 0.01 m to 0.11 m.
The UgCS software (SPH Engineering, Riga, Latvia) was used for ULS data capture as this software can incorporate adaptive banking turns into flight plans, reducing the potential error in the scanner's IMU position during turns. Map Pilot (Drones Made Easy, San Diego, CA, USA) was used for photogrammetry and was selected for its terrain-following functionality that ensured an even GSD across images. The altitude, line spacing and speed required to obtain a range in point density of 325-649 pt/m 2 between the sites are specified in Table 2.
Flight parameters for SfM broadly followed guidelines from Dandois et al. [75] who stipulate that flights should only be carried out under clear sky conditions with wind speeds of less than 20 km/h and high solar angle to reduce the effects of shadow. A flight altitude of approximately 80 m AGL (74 m was chosen to obtain a GSD of approximately 2 cm), and at least 80% forward overlap are also stipulated in the same study. We increased our overlap in line with guidance from Frey et al. [61], who recommended higher forward overlap as being optimal for the reconstruction of more complex forest environments. Flight parameters for the Scion North and South datasets were slightly different as these data were captured as part of a previous experiment [94]. The default camera settings were used with the Map Pilot application, which included infinity focus, auto exposure and a nadir camera angle. The only changes made to the default was to set white balance to "sunny". With these settings, the aperture, shutter speed and ISO were variable and controlled by the application. The point density of the SfM captures ranged from 410 to 939 pt/m 2 ( Table 2).

Data Processing and Analysis
The methodology for this project involved three broad phases: (1) processing of raw data into point clouds, (2) quality control and cleaning of the point cloud data, and (3) analysis of the point cloud data.

Processing Raw Data
UAV Laser Scanning ULS data generally have to be retrieved from the scanner in the manufacturer's native format and then processed into a more universal format, such as LAS or ASCII. The Snoopy V-Series can be operated as a PPK (post-processed kinematic) system, meaning that raw sensor data from the scanner can be post-processed with reference to GNSS base station data to increase the inherent accuracy of the point cloud. The system uses the Inertial Explorer Xpress (IE Xpress) (NovAtel Inc., Calgary, AB, Canada) software suite to apply a correction to the GNSS rover trajectory data from the GNSS log data captured by the CHCX900B base station (CHC Navigation, Shanghai, China) during the flight operations. ScanLook Point Cloud Export (Scanlook PC) (Fagerman Technologies INC., Somerville, AL, USA) was then used to combine the post-processed trajectory data with the raw sensor data. This data can then be exported in a universal point cloud format (in this case LAS). ULS data collected per site has a maximum RMSE of 0.01 m for horizontal values and 0.03 m for elevation values.
Rudimentary filtering was carried out on the raw lidar data within Scanlook PC. Further de-noising of the point cloud was carried out in LasTools (see Section 2.2.2). Turns and irrelevant flight lines, such as taxiing flights to and from the area of interest, were removed by a quality control process. This was carried out in the Graphics Mode feature of Scanlook PC, in which the trajectory of the flights can be filtered to remove sections of the point cloud that could provide potential sources of error. Crucially, ScanLook PC also applies boresight calibration angles and lever arm offsets. These two corrections are, respectively, the X, Y and Z offsets between the IMU and the scanner, and the X, Y and Z offsets between the IMU and the GNSS antenna and are critical measurements to ensure the accuracy of the point clouds. Failure to calibrate the data for either of these two measurements can build inherent errors into the point cloud data, such as the mismatching of opposing flight lines, or errors in point alignment.

Structure from Motion
The software Pix4Dmapper (Pix4D, Lausanne, Switzerland), hereafter referred to as Pix4D, was used to process the raw images from the UAV and generate the required spatial outputs. Pix4D follows three basic stages during the processing of SfM data that include: 1. Initial Processing; 2. Point Cloud and Mesh generation; and 3. DSM, Orthomosaic and Index generation. For this study, the standard 3D mapping processing settings were followed within Pix4D, except for the following adjustments. For the Initial Processing stage, Geometrically Verified Matching was disabled under the advanced and the Matching Table. This is designed to discard geometrically inconsistent matches and is promoted as being useful for projects that include rows of plants in agricultural fields [95]. Initial trials with the settings returned better results for datasets in this study with this option switched off. In the Point Cloud and Mesh generation stage, the Minimum Number of Matches was increased from the standard setting of three to five, to reduce noise and improve the quality of the point cloud. Trials with the minimum number of matches indicated three to be optimal for the data. The Classify Point Cloud option was also selected to aid in the generation of a DTM. Point clouds were exported as a single merged file in LAS format. No changes were made to the settings in stage three, other than to specify that all spatial outputs other than an orthomosaic should not be produced. Once the initial processing stage was completed, spatial reference data in the form of 3D GCPs were added to the projects to reprocess the models into a spatial coordinate system that matched the lidar (NZGD2000/New Zealand Transverse Mercator 2000 (EGM 96 Geoid)). A point cloud and orthomosaic were generated in subsequent steps and exported in LAS and TIF format, respectively.

Point Cloud Processing
The tree locations were manually identified to eliminate the error in data matching between the predicted and measured tree tips so that the precision of the methods at predicting height could be assessed. Using ArcGIS Pro 2.5.1 [96], individual trees were manually annotated on the RGB orthomosaic that was created for each site in Pix4D using the trial tree maps as a reference. These annotations were then buffered and used to confine the search area for tree tips within the point cloud. Figure 3 shows an example of these tree annotations overlaid on a raster of predicted height.

Structure from Motion
The software Pix4Dmapper (Pix4D, Lausanne, Switzerland), hereafter referred to as Pix4D, was used to process the raw images from the UAV and generate the required spatial outputs. Pix4D follows three basic stages during the processing of SfM data that include: 1. Initial Processing; 2. Point Cloud and Mesh generation; and 3. DSM, Orthomosaic and Index generation. For this study, the standard 3D mapping processing settings were followed within Pix4D, except for the following adjustments. For the Initial Processing stage, Geometrically Verified Matching was disabled under the advanced and the Matching Table. This is designed to discard geometrically inconsistent matches and is promoted as being useful for projects that include rows of plants in agricultural fields [95]. Initial trials with the settings returned better results for datasets in this study with this option switched off. In the Point Cloud and Mesh generation stage, the Minimum Number of Matches was increased from the standard setting of three to five, to reduce noise and improve the quality of the point cloud. Trials with the minimum number of matches indicated three to be optimal for the data. The Classify Point Cloud option was also selected to aid in the generation of a DTM. Point clouds were exported as a single merged file in LAS format. No changes were made to the settings in stage three, other than to specify that all spatial outputs other than an orthomosaic should not be produced. Once the initial processing stage was completed, spatial reference data in the form of 3D GCPs were added to the projects to reprocess the models into a spatial coordinate system that matched the lidar (NZGD2000/New Zealand Transverse Mercator 2000 (EGM 96 Geoid)). A point cloud and orthomosaic were generated in subsequent steps and exported in LAS and TIF format, respectively.

Point Cloud Processing
The tree locations were manually identified to eliminate the error in data matching between the predicted and measured tree tips so that the precision of the methods at predicting height could be assessed. Using ArcGIS Pro 2.5.1 [96], individual trees were manually annotated on the RGB orthomosaic that was created for each site in Pix4D using the trial tree maps as a reference. These annotations were then buffered and used to confine the search area for tree tips within the point cloud. Figure 3 shows an example of these tree annotations overlaid on a raster of predicted height. In order to extract the required metrics from the ULS and SfM point clouds, it was first necessary to process the data. Using LAStools software version 190,404 [97], the point cloud datasets were tiled, de-noised and thinned. The datasets were ground classified using the lasground_new algorithm with In order to extract the required metrics from the ULS and SfM point clouds, it was first necessary to process the data. Using LAStools software version 190,404 [97], the point cloud datasets were tiled, de-noised and thinned. The datasets were ground classified using the lasground_new algorithm with a terrain resolution of 10 m for the Scion data and 3 m for all other datasets. The Kaingaroa datasets were processed separately with care given to ensure the ground classification accounted for the spot mounds present at these sites. The point clouds were then height normalised to the terrain model generated from the ground classified points. For the datasets combining ULS ground data with SfM data, the ULS ground terrain models at 0.5 m resolution were merged with unclassified SfM point clouds before height normalisation.
Using these individual tree annotations, the predicted heights for each tree were extracted from the normalised point cloud heights using the LidR library [98] in R statistical software version 3.6.0 [99]. Specifically, the grid_canopy algorithm was used to identify the local maxima within the individual polygons defined around each tree canopy.

Integrating SfM DSM with ULS DTM
In order to integrate the ULS DTM with the SfM DSM, the models were first accurately aligned within the CloudCompare software version 2.1.1 [100]. This was undertaken by utilising the "Align (point pairs picking)" tool to align the point clouds by the ground control points laid out across each site. Aligned point clouds were then exported in LAS format and analysed following the methodology outlined in 2.2.2 and these point clouds will be referred to hereafter as SfM UTM .

Statistical Analysis
The precision and bias of height predictions using the three data types were assessed using the root mean square error (RMSE), mean bias error (MBE), and the coefficient of determination (R 2 ). These metrics were calculated using the following equations: where y i represents field measured heights,ŷ i represents predicted heights from UAV point clouds, y is the average of the observed values and n represents the sample size. The percentage RMSE (RMSE%) was also determined through expressing the RMSE as a percentage of the average observed values, y, as RMSE% = 100 (RMSE/y). The effect of height on error was examined using absolute error (AE) and percentage error (PE), which were both calculated at the tree level. The absolute error for each tree was determined from y i −ŷ i . The percentage error was determined for each tree from 100 × AE/y i . The effect of height on error was examined through plotting values of AE and PE against tree heights categorised into 0.5 m bins ranging from 0-0.5 to 6.5-7.0 m.

Results
Across all sites, the results from the ULS data show a high level of precision and a low level of bias (Table 3 and Figure 4). Overall, there was a very strong correlation between ULS predictions and field-measured heights with an R 2 of 0.99 (Figure 4), and RMSE of 0.15 (5.91%) ( Table 3). There was slight underprediction of height, as evidenced by the MBE of 0.05 m ( Table 3), but overall there was little apparent bias in predictions across the height range (Figure 4a).   Variation in precision between individual sites was relatively low for all sites, apart from Scion West, in which the trees were markedly smaller. For the five sites with the tallest trees, the R 2 ranged from 0.94 to 0.97, with RMSE and RMSE%, respectively, ranging from 0.11 to 0.19 m and from 4.01 to 6.61%. Bias was relatively low for these five sites, ranging from 0 to 0.06 m, which was also evidenced by the close correspondence of predictions to the 1:1 line for these sites (Figure 5a-e). Predictions of height at the Scion West site had far lower precision (R 2 = 0.27, RMSE = 32.1%) and were also more biased (MBE = 0.10 m) with ULS markedly underestimating tree height (Figure 5f). Variation in precision between individual sites was relatively low for all sites, apart from Scion West, in which the trees were markedly smaller. For the five sites with the tallest trees, the R 2 ranged from 0.94 to 0.97, with RMSE and RMSE%, respectively, ranging from 0.11 to 0.19 m and from 4.01 to 6.61%. Bias was relatively low for these five sites, ranging from 0 to 0.06 m, which was also evidenced by the close correspondence of predictions to the 1:1 line for these sites (Figure 5a-e). Predictions of height at the Scion West site had far lower precision (R 2 = 0.27, RMSE = 32.1%) and were also more biased (MBE = 0.10 m) with ULS markedly underestimating tree height (Figure 5f). Predictions of height from SfM had a lower precision and greater bias than predictions using ULS (Table 3 and Figure 4b). For the dataset that combined data from all sites, the precision was moderate to high with R 2 of 0.94 and RMSE of 0.48 m (18.5%) ( Table 3). Using this method, there was underprediction of height as shown by the MBE of 0.38 and this bias was also evident in Figure 4b.
There was little variation in model precision for the five sites with the tallest trees, with R 2 ranging from 0.76 to 0.87 (Table 3). However, the model precision diminished greatly for Scion West, which had an R 2 of 0.05 and RMSE of 75.9% (Table 3). At all sites, SfM underpredicted tree height ( Figure 6; Table 3) with the extent of this bias greatest for Scion West (Figure 6f). Predictions of height from SfM had a lower precision and greater bias than predictions using ULS (Table 3 and Figure 4b). For the dataset that combined data from all sites, the precision was moderate to high with R 2 of 0.94 and RMSE of 0.48 m (18.5%) ( Table 3). Using this method, there was underprediction of height as shown by the MBE of 0.38 and this bias was also evident in Figure 4b.
There was little variation in model precision for the five sites with the tallest trees, with R 2 ranging from 0.76 to 0.87 (Table 3). However, the model precision diminished greatly for Scion West, which had an R 2 of 0.05 and RMSE of 75.9% (Table 3). At all sites, SfM underpredicted tree height ( Figure 6; Table 3) with the extent of this bias greatest for Scion West (Figure 6f). Compared to the use of SfM data alone, there were slight gains in precision and bias when SfMUTM was used to predict height (Table 3). Using the SfMUTM dataset, predictions across all sites had an R 2 of 0.95 and RMSE of 0.43 m (16.5%) ( Table 3). Using this method, there was underprediction of height, as shown by the MBE of 0.32, and this bias was also evident in Figure 4c.
Height predictions made using the SfMUTM data showed little variation in model precision for the five sites with the tallest trees, with R 2 ranging from 0.75 to 0.88 (Table 3). Again, the model precision diminished significantly for Scion West, which had an R 2 of 0.02 and RMSE of 91.0% (Table  3). At all sites, SfMUTM underpredicted tree height with the extent of this bias greatest for Scion West (Table 3).
Using the ULS dataset, pooled across all six sites, the absolute height errors declined gradually from the smallest two categories of trees to the taller trees (Figure 7a). Values of the percentage error declined sharply from 87.2% and 63.7%, respectively, for the 0-0.5 and 0.5-1.0 m categories to 8.1% Compared to the use of SfM data alone, there were slight gains in precision and bias when SfM UTM was used to predict height (Table 3). Using the SfM UTM dataset, predictions across all sites had an R 2 of 0.95 and RMSE of 0.43 m (16.5%) ( Table 3). Using this method, there was underprediction of height, as shown by the MBE of 0.32, and this bias was also evident in Figure 4c.
Height predictions made using the SfM UTM data showed little variation in model precision for the five sites with the tallest trees, with R 2 ranging from 0.75 to 0.88 (Table 3). Again, the model precision diminished significantly for Scion West, which had an R 2 of 0.02 and RMSE of 91.0% (Table 3). At all sites, SfM UTM underpredicted tree height with the extent of this bias greatest for Scion West (Table 3).
Using the ULS dataset, pooled across all six sites, the absolute height errors declined gradually from the smallest two categories of trees to the taller trees (Figure 7a). Values of the percentage error declined sharply from 87.2% and 63.7%, respectively, for the 0-0.5 and 0.5-1.0 m categories to 8.1% for trees that had a height of 1-1.5 m (Figure 7c). The percentage error reached a minimum of 0.9% for the tallest group with heights of 6-6.5 m (Figure 7c).
for trees that had a height of 1-1.5 m (Figure 7c). The percentage error reached a minimum of 0.9% for the tallest group with heights of 6-6.5 m (Figure 7c). In contrast to ULS, there were increases in the mean absolute error of SfM predictions with increasing height (Figure 7b), with values reaching a maximum mean error of ca. 0.5 m at heights exceeding 3.5 m (Figure 7b). There were very steep reductions in percentage error from mean values of 71.7 and 64.1%, respectively, for the two smallest height categories to 25.1% for trees that were between 1 and 1.5 m (Figure 7d). More gradual reductions in percentage error occurred above this height, reaching values as low as 7.7% for trees within the 5-5.5 m and 5.5-6 m height classes.

Discussion
This study highlights the utility of point cloud data collected from UAVs for estimating tree height in young forestry trials. Although predictions made by ULS and SfM were both strongly related to measured height (R 2 = 0.99, and R 2 = 0.94, respectively), there was a three-fold difference in RMSE between the two methods which is a more sensitive indicator of precision (RMSE = 0.15 vs. 0.48 m). The substitution of the ULS DTM into the SfM dataset resulted in very little improvement in precision over the use of only SfM data. Tree heights that were less than 1 m could not be accurately estimated using SfM or ULS.
Our predictions of height using ULS had RMSE of 5.91%, which is consistent with previous literature. Similar levels of precision and accuracy were observed in mixed boreal forests by Jaakkola et al. [52] who reported a coefficient of determination of 0.95 and relative RMSE of 6.8% between ULS and field measurements. Wallace et al. [57] observed a relative RMSE of 6.8% when comparing individual ULS-derived tree heights to field-measured heights in stands of Eucalyptus globulus Labill. In contrast to ULS, there were increases in the mean absolute error of SfM predictions with increasing height (Figure 7b), with values reaching a maximum mean error of ca. 0.5 m at heights exceeding 3.5 m (Figure 7b). There were very steep reductions in percentage error from mean values of 71.7 and 64.1%, respectively, for the two smallest height categories to 25.1% for trees that were between 1 and 1.5 m (Figure 7d). More gradual reductions in percentage error occurred above this height, reaching values as low as 7.7% for trees within the 5-5.5 m and 5.5-6 m height classes.

Discussion
This study highlights the utility of point cloud data collected from UAVs for estimating tree height in young forestry trials. Although predictions made by ULS and SfM were both strongly related to measured height (R 2 = 0.99, and R 2 = 0.94, respectively), there was a three-fold difference in RMSE between the two methods which is a more sensitive indicator of precision (RMSE = 0.15 vs. 0.48 m). The substitution of the ULS DTM into the SfM dataset resulted in very little improvement in precision over the use of only SfM data. Tree heights that were less than 1 m could not be accurately estimated using SfM or ULS.
Our predictions of height using ULS had RMSE of 5.91%, which is consistent with previous literature. Similar levels of precision and accuracy were observed in mixed boreal forests by Jaakkola et al. [52] who reported a coefficient of determination of 0.95 and relative RMSE of 6.8% between ULS and field measurements. Wallace et al. [57] observed a relative RMSE of 6.8% when comparing individual ULS-derived tree heights to field-measured heights in stands of Eucalyptus globulus Labill. Dash et al. [101] also found an equivalent level of accuracy to this study in a mature stand of P. radiata with an RMSE of 6.8%.
Our results show that ULS slightly underestimates tree height and that this bias is highest for the smallest trees. The underestimation of tree height is a well-documented issue for single tree and area-based predictions from lidar [102][103][104][105]. This underestimation could be attributable to inadequate pulse density, use of a beam footprint that exceeds the width of the tree canopy [106], imperfections in the algorithms used in creating the DTM, DSM and CHM, pulse penetration through the canopy to the ground, and a lack of comprehensive pulse coverage [8]. Due to the very high pulse density, the open canopy of our study sites and efforts to minimise error in the DTM, errors were most likely to be attributable to imperfections in the algorithms used to create, DSM or CHM but could also result from a higher angle of incidence of the laser pulse or a larger beam footprint.
In our results, we observed a higher level of error in tree height accuracy (RMSE = 0.15 m at 45 m) than the reported accuracy of the MiniVUX-1UAV scanner (0.015 m at 50 m range, under Riegl test conditions [107]). Liang et al. [47], linked this to the influence of greater beam divergence under field conditions having a negative effect on positioning accuracy of points. A large footprint size, attributable to greater beam divergence, has been attributed to the creation of random error in the point cloud through differences in the incidence angle [52]. With an increased angle of incidence from the nadir, the beam footprint increases and, therefore, the strength of the return is averaged across a greater area introducing error into the point [108]. This same explanation would apply to a nadir pulse with a beam footprint wider than the target, such as a tree tip, or a small tree. The high overlap and minimal filtering of incidence angle in this study could explain some of the height error in our results.
Increasing the overlap has been demonstrated to introduce error into the point cloud and could remove the benefits of the additional point coverage over the target [47]. Further work to quantify the error introduced with various angles of incidence should be carried out to establish optimal overlap settings for forestry applications.
Although predictions of height using SfM data were relatively precise, these predictions had greater bias and a lower level of accuracy than the ULS, which is consistent with previous research [51,90]. Compared to ULS, the use of SfM point clouds resulted in a threefold increase in the RMSE (0.15 vs. 0.48 m) and RMSE% (5.91 vs. 18.5%). Previous research, that has used SfM point clouds for the measurement of tree heights, demonstrates a wide range of precision between studies, with RMSE% ranging from 1.89 to 19.4% and the coefficient of determination ranging from 0.21 to 0.99 [59,60,[62][63][64][65]69,71,[74][75][76][77][78]80,[83][84][85][86][87][88][109][110][111]. Although our overall results were at the less precise end of the RMSE% range, this was attributable to the inclusion of small trees within the study, which resulted in increases in the percentage error. When height error was expressed in absolute terms, our results are consistent with other studies that have found SfM to be capable of measuring tree heights to within 0.5 m of field-measured heights [59,62,63,65,85,88].
It is well known that SfM cannot describe the DTM as accurately as ULS [60] and the lower precision of the DTM could introduce error into tree height estimates from SfM. We tested the impacts of this factor on height predictions through integrating the actively sensed ULS DTM with the SfM data (SfM UTM ). The results showed a very modest increase in the correlation coefficient of 0.01 (R 2 = 0.95 vs. 0.94), a slight decrease in RMSE of 0.05 m (0.43 vs. 0.48 m), and a corresponding small reduction in RMSE% of 2% (16.5 vs. 18.5%). This supports previous findings [74] and is consistent with Hentz et al. [63] who hypothesised that it is possible to generate an accurate DTM from SfM when the canopy is sufficiently open. It was interesting to note that height predictions from ULS still had a significantly higher accuracy than those predicted from SfM UTM , which suggests these precision gains were attributable to other factors.
Further research should be undertaken to optimise settings and parameters for SfM acquisition. In this study, the flight parameters were carefully planned to reflect findings from Dandois et al. [75] and Frey et al. [61] in order to minimise the errors caused by image overlap, wind and weather.
The camera parameters that optimise image quality in aerial photography and photogrammetry are a large sensor, a focal length of between 24 and 35 mm, and an ISO setting optimised to ensure a shutter speed fast enough to minimise motion blur [112,113]. The focal length of the Phantom 4 Pro is 8.8 mm [114], which is considerably smaller than the suggested optimal focal length. A smaller focal length will introduce a greater level of lens distortion and could cause some error in the reproduction of the point cloud [112]. The focal length of the Phantom 4 Pro is fivefold smaller than the focal length of the camera that achieved the lowest RMSE from the literature [88]. The sensor size of the Phantom 4 Pro is almost three times smaller than the camera that achieved the lowest RMSE, which would result in reduced sensitivity [88]. For future study, a direct comparison between multiple cameras is suggested, so that the effect of camera parameters on tree height measurement can be explored in greater depth.
Within this study, Pix4D was used to process point cloud data as this is one of the most widely used software programmes for this purpose [65,68,74,79,83,86,87,115]. The SfM software could potentially introduce errors in the point cloud, notably in the creation of the DTM [62]. Studies that have trialled more than one SfM software to predict tree height have found significant differences in precision [68,116] and further research is required to compare the impact of this factor on the predictive precision of tree heights. As well as the SfM process itself, a number of studies have suggested that the algorithms used to create the CHM and individual tree crown delineation could have a possible smoothing effect on the SfM heights [62,63,117]. The methodology followed in this study to analyse tree heights did not create a CHM, and the negative bias was still observed. This is an interesting finding, as it means that the CHM creation phase can be ruled out as the sole cause of the bias, although it cannot be excluded as a contributing factor.
Similarly, there are many factors involved in ULS acquisition that could be refined to improve height predictions. Error from ULS acquisition is influenced by beam divergence (and consequently beam footprint), range accuracy, maximum range and the height that the scanner is flown above the target. There is now a range of UAV laser scanners on the market, and all have different attributes that could affect the accuracy of forestry measurements [118]. The interaction between these variables has not been well studied for forestry applications, and we echo the conclusion of Liang et al. [47] that further study in this area is required to optimise data collection and predictions of key metrics such as height.

Conclusions
This study has shown that point cloud data acquired from a UAV can be used to accurately predict tree heights. Our results extend previous research by demonstrating that tree height can be accurately quantified down to a threshold of 1 m. Below this height, error increases significantly. Although SfM was not as accurate as ULS for prediction of tree height, this method could provide a more cost-effective alternative to ULS. Further research should be undertaken to refine collection and data processing methods for both ULS and SfM.