Optimal Ground Control Point Utilization for Aligning 3D Surface Models of Forest Areas with Steep Slopes

for the precise alignment of 3D surface models while utilizing the same GCPs for aligning pre-and post-3D surface models.


Introduction
Forest operations can cause various disturbances to the forest environment.(3)(4) Sustainable forest management requires continuous monitoring, especially after implementing forestry operations.For continuous monitoring, it is essential to conduct direct surveys (for detailed investigations and measurements).However, surveying forests with human effort is difficult because of the challenging terrain and limited available facilities in such areas.
To reduce the labor intensity in surveying forests, a 2D-sensor-based unmanned aerial vehicle (UAV) system was implemented in various forestry fields owing to their cost efficiency when compared with 3D scanning sensors, such as LiDAR.Several attempts have also been made to introduce novel technologies for monitoring post-timber harvesting effects such as the elevation changes caused by the surface deformation of forest soils (e.g., soil erosion and deposition). (5,6)To detect the changes in soil surfaces, a few recent studies used UAV-derived 3D surface models to identify the wheel tracks of forestry machines (left behind after timber harvesting). (7,8)In this study, we collected approximately 200 2D images for monthly UAV surveys of the slope using UAV photogrammetry technology; a 3D surface model of the posttimber harvesting site was created using the structure from motion (SfM) photogrammetry algorithm, and the soil changes in the target site were also observed. (8)Moreover, ground control points (GCPs) were used to align the monthly point cloud data (PCD) acquired through georeferencing and calculate the elevation changes (e.g., erosion and deposition) in the soil surface.
(11) A recent study carried out multiple simulations based on UAV photogrammetry to determine the optimization of GCP utilization; the study configured 24 GCP networks to confirm the optimal method of GCP use. (11)Although extensive studies have been conducted in various related fields, there is a lack of studies on the optimal installation and utilization of GCP methods in mountainous and forested terrains.
To monitor the topographical changes at the target site, enabling the continuous acquisition of 3D data is important; notably, in such cases, the technology used for data acquisition and the alignment of the acquired 3D data are the key factors.Therefore, in this study, we monitored the impact of post-timber harvesting operations on the forest environment using UAVs to determine the optimal number of GCPs required to effectively monitor an area.First, we used GCPs to implement a nonrestorative extraction method and obtained the spatial root mean square error (RMSE) for each number of GCPs utilized to confirm the optimal number.Then, after carrying out geometric correction by applying the same GCP used for the verification of the time-lapse data, the overlapping state of both PCD points was examined as the distance of the GCP identified in the orthomosaics to confirm the overlap accuracy according to the spatial RMSE.

Site description
We selected the forest area of Kangwon National University, Gangwon Province, Republic of Korea (37°46′34.4″N, 127°49′41.1″E) as our study area (Fig. 1).Approximately 3 ha of timber was harvested at the site in March 2022.The site has an altitude of 508-628 m above sea level, with an average slope of 47%.After clearcutting, stumps and logging debris remained on the slopes of the area, with the debris being arranged horizontally downslope.A real-time kinematic (RTK) system was not available for this region because of the mountainous site location, as the signal from the base station could not be received at the site.However, the GPS signals from the site could be recorded using satellites; therefore, we surveyed the area using a UAV.

UAV survey and field data collection
In this study, surveys were conducted using UAVs equipped with a global navigation satellite system (GNSS).The UAV survey was performed with DJI Matrice 300 (Da-Jiang Innovation, Shenzhen, China) as the platform and DJI Zenmuse H20T (Da-Jiang Innovation, Shenzhen, China) as the sensor.The DJI Matrice 300 weighed approximately 7 kg and provided with both GPS and RTK modes.The maximum flight duration of the DJI Matrice 300 attached with the DJI Zenmuse H20T was about 43 min.The DJI Zenmuse H20T weighed 0.82 kg and provided an optical camera with a resolution of 12 megapixels.
To georeference the 3D surface models derived from UAV surveys, a 40 × 40 cm 2 Fomex texture plate of 29 GCP markers was prepared for the preprocessing of the study site, where the RTK mode is unavailable.We selected stumps, which are undeforming objects for monitoring the forest field, to install GCP markers.During the GCP installation, we considered 1 GCP per 30 × 30 m 2 to cover approximately 3 ha of the total study site.
Vertically parallel flights were performed at heights of 100 and 140 m [Fig.2(a)] for every monthly UAV survey with the GPS mode because of the unavailability of the RTK mode.The flight speed was set at 5 m/s, optimized for the battery duration of the DJI Matrice 300 RTK system.The overlaps were set to 90 and 80% on the front and back sides, respectively.The margins were set at 20 m to ensure the high quality of the PCD during processing.The coordinates of the GCP points were collected to validate the georeferencing results.On the other hand, field survey was conducted with Trimble R12i GNSSs (Trimble, Westmister, USA).The spatial coordinates of GCPs (with the least coordinate error of 3 cm) were collected using GNSS from the center point of each GCP marker [Fig.2(b)].
from 235 2D images captured in July 2022 and 230 2D images captured in September 2022.By calculating the coordinates and correlations of the 2D images, generally referred to as PCD points, we simulated dense clouds from the data of sparse clouds.The PCD points were subsequently generated into digital surface models (DSMs) and orthomosaics by calculating the height of each PCD point and the color of each pixel point from the 2D images.

Georeferencing by sampling without replacing GCPs
The number of GCPs required for optimum 3D surface model alignment was validated using two different methods: (1) random GCP selection for georeferencing and (2) the calculation of alignment error.The optimization results of the GCP configuration were compared on the basis of the number of GCPs utilized for data collection.In related studies, the number of validated GCPs was not unified. (11,12)However, the checkpoints and validated GCPs were assessed on the basis of spatial errors, which were calculated using the RMSEs of the XYZ coordinates. (11)Thus, for analyzing the probabilistic equity, we randomly sampled 4-28 GCPs in two multiples and calculated the GCP and checkpoint RMSEs in three different trials.To calculate the spatial RMSEs, the total coordinates of the 3D surface model GCPs were compared with those of the GNSS-GCPs (i.e., ground truth data) using the XYZ coordinate data.The equation for calculating the RMSEs of the 3D surface model GCPs from the ground truth is ( ) where The purpose of georeferencing was to align the 3D surface model as precisely as possible to detect the surface deformations caused by the events that occurred in the region.Notably, the georeferencing process in multiple 3D surface models can automatically align multiple DSMs from different periods.However, determining the GCP center points in PCD is impossible with the automatic method.Therefore, the precision of the alignment was assessed by manually calculating the spatial distance between the GCP center points of the aligned orthomosaics in ArcGIS Pro.

Results and Discussion
The GCPs used for georeferencing were selected in two multiples (from 4 to 28); thus, the GCP networks were classified into 13 validation sets.The GCPs used for georeferencing each validation set were randomly selected by confirming the center of the GCP markers in the PCD acquired in July 2022 (Fig. 3).Because of vegetation growth, a few GCPs (e.g., GCP-1 and GCP-4) could not be identified clearly in the September 2022 orthomosaic.Thus, in the georeferencing of the September data, we excluded GCP-1 and GCP-4 for georeferencing the 3D surface models.
The georeferencing RMSE of each validation set was used to calculate the average RMSE (Fig. 4).Despite the increasing number of GCPs, the GCP RMSEs among all the GCP network designs did not vary significantly.In contrast, the checkpoint RMSE decreased by 2.03 cm until 26 GCPs were used.Moreover, the two validation sets confirmed precise georeferenced GCP networks by subtracting the GCP RMSE from the checkpoint RMSE.The lowest GCP and checkpoint RMSEs of all the GCP network designs were 6 GCP (10.69 and 14.52 cm, respectively), 14 GCP (11.21 and 15.24 cm, respectively), 16 GCP (11.62 and 14.83 cm, respectively), 20 GCP (12.09 and 16.53 cm, respectively), and 26 GCP (12.62 and 14.83 cm, respectively).(13) Although the results were shown as RMS in Hastaoglu et al., (12) a comparison of checkpoint RMS data obtained for all-network design (46 GCPs) and sparse network design (32 GCPs) showed differences by an average of 2.87 cm when additional 14 GCPs were used from the sparse network design.In the comparison between the 4-GCP group and the 10-GCP or more group, which was presented by Zhang et al., (13) the total horizontal error was reduced by an average of 30.63 cm.In particular, Villanueva and Blanco (11) investigated the optimization of GCP networks with various patterns (e.g., clumped -center, distributed -edge, clumped -edge, and distributed -edge and center), which was different from our study.The GCP RMSE of the study decreased significantly (by 1.6 m) with increasing number of GCPs.Moreover, Villanueva and Blanco (11) revealed that the GCP RMSE results showed a significant reduction (by approximately 20 cm) from GCP 4 to GCP 24 of the XY coordinates, which is comparably similar to our study.However, the GCP RMSEs of XY and Z may not be perfectly comparable to those in our study because we considered the total RMSE of the XYZ coordinates.Thus, we compared the total GCP RMSE magnitude and noted a reduction of approximately 4 m.This result indicated that the DEM resolution of the comparison study was 10 cm, which was higher than that considered in our study (which is approximately 2.7 cm), and that the center points in our study may have been confirmed more precisely than the comparison site.Furthermore, the DSM resolution acquired for our study was significantly lower than that in the compared study by approximately 6.3 cm, which may have also affected the insignificant differences in GCP RMSEs in the 13 validation sets (Fig. 4).However, the checkpoint RMSE in our study showed a decreasing trend, similar to that in the compared studies.
Furthermore, we generated orthomosaics from all GCP validation sets for June and September 2022 by comparing the lowest GCP-RMSE of each GCP used for georeferencing (Table 2).The precision of the alignment was assessed using the orthomosaics for June and September 2022 (Table 2).The average alignment distances of the validation sets using 6 GCPs (VS-6) and 26 GCPs (VS-26) were calculated as 3.78 and 2.96 cm, respectively; the GCP-26 set was confirmed to be the most precisely aligned PCD (based on the total validation process).
Georeferencing is a key process that allows the alignment of different 3D surface models for effective monitoring; (14) the total alignment error in the study was 1.28-3.33cm in the construction field.The lowest alignment errors in the XY and Z coordinates in an artificial mound, considered an idealized field condition, were calculated to be in the ranges of 1.31-3.39cm for the XY error and 0.94-1.8for the Z error (when using GCPs on the alignment). (15)A recent study that monitored timber-harvesting sites utilized 29 GCPs for georeferencing to align four different PCD points; (8) the results indicated that the alignment error for models developed for June and September 2022 was 3 cm.This error may be compared with that calculated as 2.96 cm for VS-26 in our study [Fig.5(a)].Our alignment error calculated for 3 rd VS-26 was the most accurate validation set, with the error being less than 4 cm, according to related studies. (8,15)oreover, the alignment error of the VS-6 was also acceptable compared with that of the VS-26 (Fig. 5).This result indicates that reducing the number of GCPs used for georeferencing may not negatively affect the alignment results when the post-3D surface model is generated from the same GCPs that were utilized in the pre-3D surface model, which was the approach adopted in this study.Furthermore, we determined the optimal number of GCPs required for aligning the 3D surface models.A comparison of the average GCP-checkpoint RMSEs of the VS-6 and VS-26 showed a strong relationship with respect to the alignment error; the GCP-checkpoint RMSE of the VS-26 was approximately 3.48 cm lower than that of the VS-6 (Table 2).The decreasing gap between the GCP and checkpoint RMSEs also decreased the alignment error by 0.82 cm with respect to the average GCP-checkpoint RMSE.However, the standard deviation of the VS-6 and VS-26 validation sets was approximately 1 cm, indicating an insignificant difference.Furthermore, we confirmed that the VS-14, VS-16, and VS-20 sets had sufficiently small (<4 cm) alignment errors in the validation results (Table 2).Moreover, no significant difference was confirmed in the alignment performance (e.g., average GCP-checkpoint RMSE, standard deviation, and alignment error) of five cases (e.g., VS-6, VS-14, VS-16, VS-20, and VS-26) (Table 2).These results indicated no significant correlation between georeferencing and the  number of GCPs.Another GCP-based 3D data alignment study indicated that using 6 GCPs for the alignment of multi-photogrammetry data at construction sites was possible; this approach is considered to be effective even in larger areas. (16)Thus, we conclude that the georeferencing method used in this study, wherein we considered the same GCPs used for collecting the PCD in June 2022 to collect the PCD in September 2022, could yield low alignment errors.We also conclude that the least alignment error can be acquired using 26 GCPs; however, similar results can be obtained using 6 GCPs.

Conclusions
The installation of GCPs in actual forest fields is not easy, as finding undeformed but sturdy objects is difficult.This can affect the distribution of GCPs in the field and the precision of the georeferencing results for the target site.To address this problem, we generated a total of 13 validation sets in three trials through random selection and assessed the GCP RMSE for each of the 13 validation sets; the results did not show significant differences.The alignment error confirmed from the generated visual 3D surface models, which were orthomosaics, showed no correlation with the GCP RMSE when the same GCPs were used for surveys during two different periods.Notably, we conclude that the GCP georeferencing methods and analysis results of this study can be utilized as base data for conducting or comparing UAV surveys in forest fields for which RTK data were not available.We suggest the use of more than 6 GCPs for georeferencing and aligning 3D surface models.Furthermore, we suggest using the same GCPs for developing the pre-and post-3D surface models.Although the proposed method can yield precise alignment results, validating this method by applying it to other forest areas in future studies is important.

Fig. 1 .
Fig. 1. (Color online) Location and panoramic view of the timber-harvesting area of Kangwon National University considered in Kim et al.(8) and this study.

Fig. 2 .
Fig.2.(Color online) Illustration of the surveying method carried out in this study using UAVs and field references, which was the same as the method used by Kim et al.:(8) (a) shows a vertically parallel flight for monthly UAV surveys and (b) shows the GCP-GNSS survey carried out using GCPs installed on tree stumps in the study area.

Fig. 3 .
Fig. 3. (Color online) GCP network designs from random GCP selection.The GCPs (4-28) were selected on each GCP network by multiples of 2 in three different trials.

Fig. 4 .
Fig. 4. (Color online) GCP and checkpoint RMSEs in the three validation sets considered in this study.

Fig. 5 .
Fig. 5. (Color online) Comparison of GCPs and checkpoints in average RMSE and alignment error: (a) total RMSE of GCPs and checkpoints; error bars represent standard deviation and (b) average GCP-checkpoint RMSE.

Table 1
Parameters used in each process to generate 3D data.

Table 2
RMSE and alignment error estimated from the confirmed validation set of the lowest average GCP-checkpoint RMSE.