*3.2. Verification of Height Accuracy of Surface Data Obtained by SfM Analysis by UAV*

The elevation values obtained at each point of the 3D model were compared with those measured by the TS. However, measurement by the TS takes time, which causes a significant time lag between drone photography and TS surveying. Therefore, verification of measurement accuracy was limited to four times: the morning of 22 August before the repair work, the morning of 23 August and 24 August during the repair work, and 13 September after the repair work. The results are shown in Table 1 and Figure 4.


**Table 1.** Height measurement error (m) in SfM analysis compared with the total station survey.

The overall measurement error averaged −0.146 m, leading to a general tendency toward underestimation. The RMSE was 0.098 m, indicating that elevation values could be obtained with an accuracy of about 10 cm. However, the average error varied greatly from one measurement session to the next. It was −0.102 m before the repair work, and −0.259 m and 0.198 m during the repair work, whereas it was −0.017 m after the repair work, a significant improvement. Lee and Lee [28] reported an RMSE error of 0.11 m in the vertical direction for the SfM analysis using vertical and 45-degree photographs from 50 m altitude at a construction earthwork site. The error in this study was similar, but the accuracy in this study was slightly less at closer distances to the ground surface and in forest areas.

The final survey was measured after the repair work had been completed and the area had been compacted. However, before and during the repair work surveys may not have measured the complete ground surface during TS surveying because of the softness of the soil on the ground.

Additionally, there was a considerable time lag between drone photography and TS surveying during the measurements before and during the repair work. This was because the survey was conducted safely so as not to interfere with the repair work as much as possible. In some locations, earthwork, or log construction may have been undertaken after drone flight and before the TS survey.

With this in mind, we discuss the accuracy of the measurements at each stage. The absolute value of the error at the TS measurement points before the repair work is shown in Figure 5a. It is shown in the figure that the error is small on slopes where the soil is visible, and slightly larger errors are observed in vegetated areas and on the shoulders of embankment slopes.

**Figure 4.** Frequency distributions of height measurement error in SfM analysis.

**Figure 5.** Absolute values of the error at each process of the repair work. (**a**) before the repair work, (**b**,**c**) during the repair work, (**d**) after the repair work.

Similarly, the situation during the repair work is shown in Figure 5b,c, and the situation after the repair work is shown in Figure 5d. It is shown in Figure 5b,c that errors were minor, even during the repair work, in areas where the soil was visible. However, substantial errors occurred in areas where work was being done, such as in areas where wood framing was being installed. A significant error at the shoulder of the embankment is shown in Figure 5c. Still, further verification was needed to determine whether the accuracy of edge detection in the SfM analysis was reduced by this problem or whether the soil tended to move easily at the shoulder of the embankment.

On the other hand, Figure 5d shows that the accuracy was generally good after the repair work, including the shoulders of the embankment. Even on the shoulders of the embankment, most of the errors were lower than 0.1 m. Slightly larger errors were observed in the upper right and lower left parts of the figure, which are close to the vegetation. Together with the edge mentioned above, a more precise analysis of the accuracy in areas easily shaded by vegetation is considered necessary.
