The forest sample plots in this experiment represent a variety of stand conditions with regard to species, growth stages and management activities. The forests were measured using four approaches: ULS, MLS/PLS, TLS, and conventional in situ measurements. The performance of ULS was evaluated through the comparison among these observations: the conventional in situ measurements served as a reference, TLS represented the state-of-the-art of automatous terrestrial tree observation approach, and MLS and PLS exemplified emerging solutions with an enhanced mobility and data coverage.
Test area
The experimental site was established in 2014 and located in a boreal forest in Evo, Finland (61.19° N, 25.11° E). The main tree species were Scots pine (Pinus sylvestris L.), Norway spruce (Picea abies (H. Karst.) L.), Silver (Betula pendula Roth) and Downy (Betula pubescens Ehrh.) birches. On the site, 24 forest sample plots, 32 m by 32 m for each, were selected by foresters to test various in situ measurement approaches.
The sample plots were classified into three stand complexity categories, i.e., “easy”, “medium” and “difficult”, according to the amount of occlusion at the ground level, the spatial stem density and the distribution of the DBH. The category “easy” represented clear visibility with minimal understory vegetation and low stem density (~ 700 trees/ha); “medium” represented sample plots with moderate stem densities (~1000 trees/ha) and sparse understory vegetation; the “difficult” category represented plots with high stem densities (~2000 trees/ha) and dense understory vegetation.
Figure 1 illustrates tree maps of three example plots from the three different complexity categories. The mean DBH, mean tree height and mean basal areas are 20.5 cm, 18.3 m and 23.1 m2∙ha− 1, respectively, for the easy plot (Fig. 1a); 17.3 cm, 16.3 m and 30.2 m2∙ha− 1 for the medium plot (Fig. 1b), respectively; and 12.3 cm, 13.2 m and 29.3 m2∙ha− 1 for the difficult plot (Fig. 1c), respectively. In general, the mean DBH and tree height decrease and the basal area increases along with the growing complexity of stands.
Figure 2 shows the distribution of the DBH in each stand complexity category by the time of the UAV flight, which illustrates the variation of tree size within each complexity category. As shown by the DBH distribution, the higher the stand complexity, the higher is the population of small trees (e.g., DBH < 15 cm).
In situ observations using conventional field methods
Conventional forest field measurements were carried out between May and August of 2014 (Liang et al. 2018a). All trees with DBH larger than 5 cm were measured in the sample plots. The tree height and the DBH were measured using conventional field measurement methods, i.e., using calipers and inclinometer. Tree maps were produced for each plot by combining measurements in the field and in the TLS data. Preliminary tree positions were firstly mapped from TLS point clouds. The preliminary tree positions were then verified and updated during in situ investigations, and trees missed on the preliminary tree maps were added. The stem curves were manually digitized from the multi-scan TLS point clouds starting at the height of 0.65 m above the ground, continuing at the DBH height and then every meter above until the maximum measurable height in the point cloud. The stem volume was calculated directly using the tree height and stem curve measurements similarly as in Liang et al. (2018a). The total tree biomass was calculated using Finnish national allometric models (Repola, 2009). The DTM was generated through rasterization of the classified ground points using TerraScan software (TerraSolid Oy, Helsinki, Finland), where the ground points were automatically classified and manually edited when needed.
The forest plots were revisited in December 2017 to update the field reference. Possible changes since 2014 were first visually interpreted from the ULS data with respect to the existing reference tree maps. Among the 24 plots, 22 plots remained almost unchanged. One plot was completely cut. For another plot, the ULS data mis-matched the plot location. Therefore, 22 plots were employed in this study. In the 22 plots, altogether 72 trees had been felled in 18 plots during 2014–2017. The maximum number of felled trees in a sample plot was 10, which accounted for only a small proportion of the plot tree population. Thus, the classification of complexity categories did not change for the sample plots. The results of the ULS evaluations are therefore comparable with the results from TLS and MLS/PLS previously reported in other studies (i.e., Liang et al. 2018a, 2018b).
In situ measurements using TLS and MLS/PLS
TLS data were collected in 2014 using Leica HDS6100 (Leica Geosystems AG, Heerbrugg, Switzerland) with a multi-scan approach, that is, one scan at the plot center and four scans at the four quadrant directions. No pre-scan preparations were implemented in the field measurements. The data were registered using artificial spheres. The mutual scan registration accuracy was at a 2-mm level. The point spacing was 15.7 mm at a 25-m distance to the scanner in both horizontal and vertical directions.
The kinematic in situ measurements, namely, MLS from an all-terrain vehicle and PLS from a backpack, were also collected in 2014 (Liang et al. 2018b). The core measuring system for both platforms was identical, namely, AkhkaR2 (Finnish Geospatial Research Institute, FGI, Finland). Both platforms used the same scanning parameters: scanning frequency of 95 Hz, which resulted in an approximate 4-cm on ground point spacing along the profile at a range of 35 m and an on ground profile spacing of 1.0–1.4 cm at a typical platform moving speed of 1.0–1.45 m/s .
ULS measurements
The ULS data were collected in September 2017 using a Riegl RiCOPTER with VUX-1UAV (RIEGL, Horn, Lower Austria, Austria), as shown in Fig. 3. The UAV campaign lasted for 3 days and the average time spent on UAV flight per plot was 10–20 min. The flight altitude was approximately 50 m above the ground. Each plot was covered by 4–5 flight lines. The overlap of all flight lines at the plot area was typically high, e.g., 100%.
The GNSS-IMU system installed on the UAV was Applanix AP20 GNSS-Inertial System (Trimble Applanix, Ontario, Canada). The position measurement accuracy was better than 0.1 and 0.2 m in horizontal and vertical directions, respectively. The roll, pitch and heading were measured at an accuracy of better than 0.015, 0.015 and 0.035 degrees, respectively. These give a point location accuracy of better than 3.6 cm at the nadir and 7.1 cm at the far end of the field-of-view at 50 m above the ground, if the positioning and ranging errors are left out of consideration.
The applied laser sensor was a Riegl VUX-1UAV. The scanner was mounted to scan nadir profiles and it operated at the 1550-nm wavelength. The maximum measurement range of the scanner was 300 m. The beam divergence was 0.5 mrad, providing 2.5-cm and 5.0-cm footprints at 50-m and 100-m distances from the scanner, respectively. Since the targeted flight altitude was 50 m, the 50-m and 100-m distances corresponded to the scanning distances at nadir and at far end of the field of view. In order to gain better stem visibility, the data were collected with a 120-degrees field-of-view and a 550-kHz laser pulse rate, resulting in 106 scan lines per second and in a 0.07-degrees (1.2-mrad) measurement resolution along each scan line. Alone each scan line, the on ground point spacing was 6.1 cm at nadir and 24.4 cm at the far end of the scan line. The typical flight speed was 2.0–4.0 m∙s− 1, resulting in a 2.0–4.0 cm on ground spacing between scan lines. The point density was around 100–800 points∙m− 2 on the horizontal ground surface if only one echo per pulse was considered. In practice, due to the high overlaps between flight lines, the point density increased vastly, resulting in 4000–18,000 points∙m− 2 at the sample plot areas. Two Sony ILCE-6000 cameras were also mounted on the UAV for colorizing the point cloud. Figure 4 illustrates the examples of colored ULS point cloud data in the three stand complexity categories. Color information was not used in the following processing.
The visibility of trees varied significantly according to tree species and forest structure, as shown in Fig. 5. For example, pine trees in easy forest stands typically had an excellent visibility and the stem was visible for most of its length. For spruce trees, the visibility of the stem decreased significantly as the tree stem can be either partly visible or totally occluded by the tree’s own canopy and/or by canopies of the surrounding trees.
Figure 6 illustrates a pine tree under moderate occlusion in the ULS data. ULS points at five different stem heights are shown in the subfigures. The stem structure is visually identifiable in two lower slices, but is totally missing in other three upper slices, indicating the impact of tree’s own crown on its stem digitization.
Geometric inconsistency among different flight lines was visible in the delivered point clouds. Possible reasons for such errors include registration errors, influences of wind, and/or varying geometric accuracies from different flight lines at identical locations. Figure 7 shows four examples of mismatches between the flight lines, which indicate the challenge of tree parameter estimation brought by geometric inconsistencies.
Automated retrieval of tree and plot parameters
The point cloud from stationary TLS, MLS/PLS and ULS, were processed through the same processing chain as described in Liang et al. (2018b). The specialty here is that the ULS data were understood equal to the terrestrial point cloud and processed with a stem detection and modeling algorithm that was identical to the other terrestrial point clouds.
In a preprocessing stage, the original point cloud was first sampled through an equivalent sampling method. The point cloud was digitized into a voxel space, and in each voxel the point closest to the center of gravity was selected as the representative point for the point distribution within the voxel. The voxel size was 1 cm for both the TLS and MLS and 0.5 cm for the ULS considering the given density of raw point clouds. The DTM was reconstructed using a morphological filter and linear interpolation. The point cloud was firstly rasterized in 2D. The lowest point in each pixel (20 cm by 20 cm) of a 2D raster space was selected as a seed point and the largest connected group was interpreted to be part of the ground. Detached groups were accepted as ground if they were smoothly connected with the accepted ground. Due to the above canopy viewing position, ULS collects tremendous amount of canopy points that are less important for the stem analysis. Therefore, a canopy filtering was applied to remove the topmost canopy layer (i.e., 20% topmost) before the stem detection and modeling. After the stem detection, all canopy points were used in the tree height estimation.
Stem points were identified through a point-based approach. Points on stems were identified by analyzing the structure in their immediate neighborhood using principal component analysis. Tree stem models were built from the recognized stem points as a series of 3D cylinders representing the changes in the growth direction of stems. The DBH and location of a stem were then estimated from the cylinder element at breast height (1.3 m above the ground). The stem curve was estimated from the cylinder elements at predefined heights. The tree height was estimated differently for big and small trees, which were separated according to a DBH threshold of 15 cm. Big trees were assumed mostly to be dominant or co-dominant trees that are exposed directly to the sunlight and no trees are above them (Wang et al. 2016). The tree top was assumed to be the highest point around the stem. The small trees were mostly intermediate and suppressed trees, and the treetop was therefore found from the largest connected point group around the tree stem. For both big and small trees, the elevation difference between the tree top and the DTM beneath it was used as a height estimate for the tree.
The MLS/PLS data were processed with one additional step. Since a tree may be observed several times from different trajectories and spatial inconsistency from different observations prevails in the mobile data, the tree mapping followed the multi-single-scan type of processing (Liang et al. 2018a).
Retrieval of tree and plot parameters from ULS data by manual measurements
In addition to the automated processing, 7 plots, 3 from easy, 2 from medium and 2 from difficult categories, were randomly selected and manually measured in order to benchmark a data-level accuracy, thus, to enhance understandings of the automatically derived tree parameter estimates. The manually measured tree parameters include the tree position, the DBH and the tree height of individual trees that are visually identifiable from the ULS data. A circle at the breast height was fitted to each visually identified tree. If only few or none points were recorded for a stem at the DBH height, a circle at a higher position was searched for and fitted when possible. The tree position and the DBH were therefore the center of the fitted circle and the diameter of the fitted circle, respectively. The treetop and the stump were also visually identified, and the elevation difference between the tree stump and top was recorded as the tree height.
Evaluating the accuracy of individual tree parameters
The performance of TLS, MLS/PLS, ULS was evaluated by comparing the accuracies of the estimated tree-level parameters with respect to the field measurements.
The evaluation followed the procedure in Liang et al. (2018a). The detected trees were matched with the reference trees based on the horizontal stem locations and the DBHs of the trees. The search distance was 50 cm for TLS and ULS, and 150 cm for the MLS/PLS. The mapping accuracy was evaluated using the completeness, which indicates the proportion of reference trees that were automatically detected. The accuracy of the tree position, tree height, DBH, stem curve, stem volume and AGB were all evaluated using the relative Root Mean Squared Error (RMSE) and relative bias, both in percentages, with an exception of the tree location where only the absolute RMSE was calculated. The evaluation was carried out at an individual-tree-level, i.e., the estimated and reference parameters were compared, and reported at a plot-level, i.e., relative RMSE and bias were calculated at a plot level and further averaged at the plot-complexity-category-level. The results are reported separately for each stand complexity category.