**2. Materials and Methods**

The measurement test was conducted at a 2.5 m wide road repair site in Kizugawa City, located in the southern part of Kyoto Prefecture, Japan. The length of the road is 706.3 m, and it was constructed in 2020 as a branch line from a larger forest road. This road was constructed using the Ohashi-type road construction method, which assumed that 3–5-ton class excavators and 2–3-ton trucks will be used in the construction [32]. The logs used to make the retaining wall with log structure were used for the frame to distribute the loads [33].

Repair work was conducted from 22–25 August 2022 to repair a defective embankment near the middle of this line caused by heavy rainfall in May 2021. According to official rainfall data in the neighboring city of Kyotanabe, the total rainfall from 16–22 May was 204.5 mm, 160 mm of which was concentrated over two days. The road was carefully constructed with logs on the upper part of the embankment and the roadbed, but the bedrock was exposed in the affected area, making it challenging to form the embankment.

The repair consisted of first creating a simple road to access the lower part of the missing embankment and then reshaping the embankment with logs from the valley flowing under the lower part of the road embankment to the upper part of the road. A rectangular plot with a length of about 30 m and a width of about 16 m was set along the work road in the repair work area, and this was used as the analysis range.

UAV photography was conducted six times: once before the repair work (morning of 22 August), four times during the repair work (afternoon of 22 August, morning of 23 August, afternoon of 23 August, and morning of 24 August), and once after the repair work (13 September). A DJI Mavic3 drone was used for photography. We used PiX4Dmapper (Pix4D S.A., Prilly, Switzerland) for the SfM analysis. ArcGIS 10.8.1 (ESRI Inc., Redlands, CA, USA) and QGIS 3.16.7 (QGIS Development Team, 2021) were used for earthwork volume estimation.

Four aerial markers were set up as ground control points (GCPs) to measure absolute coordinates and eliminate distortion of the obtained model. In addition, surveying was conducted at two other open locations in the sky using a survey GNSS receiver (Spectra Precision SP80). Absolute coordinates of the GNSS points were precisely measured using electronic reference points of the Geospatial Information Authority of Japan, and absolute coordinates of the GCPs were calculated using a total station (TS) TAJIMA TT-N45 (TJM Design Co., Tokyo, Japan) from GNSS points. According to the TS specifications, the ranging accuracy, in this case, is "(2 + 2 ppm × D) mm (for measuring distances of 4 m or more)," with an error of 0.01 mm at a distance of 10 m. The angular accuracy is 5 and the error at a distance of 10 m is 0.24 mm.

The ground height was measured a total of four times using TS during the lunch break on each day when repair work was suspended. The measurement results were compared with the 3D model obtained by SfM analysis from the images taken by the UAV every morning. As the UAV imaging on the 2nd and 3rd days was conducted during the repair work, there was a slight time lag with the TS survey and the topography may have slightly changed. We used 41, 76, 91, and 88 TS survey points for comparison on days 1 (before the repair work), 2, 3, and 4 (after the repair work), respectively. The accuracy of the ground height obtained by the UAV was calculated by the TS measurement points as the validation points, and the results of the TS measurements were true values.

The UAV was operated at two altitudes: close up (3–5 m) and low altitude (7–10 m). In the SfM analysis, 3D models were created using two altitude images at the same time. It captured the area where the repair work was being done, with a flight duration of about 30 min. The UAV was operated manually along the area, and the image was automatically

captured at 2 s intervals to ensure a high overlap between images. We also captured images from various directions in the flyable area.

DEMs of the ground surface were created from the 3D model obtained by SfM analysis using a UAV, and the amount of cut and fill at each stage was calculated using the Cut Fill tool in ArcGIS.

## **3. Results and Discussion**

#### *3.1. Generation of 3D Models for Each Phase of the Small-Scale Forest Road Repair Project*

The 3D models of each phase were successfully constructed using SfM analysis. The resolution of the generated models ranged from 1.1 cm to 1.4 cm. For example, the 3D models generated from the data taken before, during, and after the work was completed and viewed from various angles are shown in Figures 1–3.

In this test, the flight route of the UAV was not fixed, and the appropriate course for photography was determined manually during the operation. Because the site was located in a valley, which made GNSS reception difficult, and the flight altitude was lower than the height of the surrounding trees, there was a risk of colliding with branches or tree trunks. Under these conditions, GNSS control is insufficient, and UAV flight control technology for real-time obstacle detection using various sensors and SLAM technology is needed.

**Figure 1.** A 3D model generated by SfM analysis before the repair work started.

**Figure 2.** A 3D model generated by SfM analysis during the repair work.

**Figure 3.** A 3D model generated by SfM analysis after the repair work completed.

Because the shooting area differed depending on the shooting session, the area where the 3D model was obtained also differed in each shooting session. There were some differences between the modeled trees and other land objects. However, no inconsistencies in their positions or shapes were observed. Even for the complex and finely patterned subsoil vegetation and soil and stone surfaces, no significant distortions were observed, and it was found that 3D models with clean shapes could be obtained.
