3D PHOTOGRAMMETRY MODELING FOR SAFETY MONITORING OF FORTRESS WALL S

: In the monitoring to prevent collapse or tipping in the structure of linear-shaped fortress walls, the record of shapes like slope or protrusion is very important(Jo and Hong, 2019). As the way to record the shapes, laser scan or photogrammetric technology is employed. A laser scan that has to install several points of measurement has a limitation in recording them if the terrain is undulating, or the structure becomes huge. Photogrammetry has a shortening effect in terms of cost and process(Kim, 2020) compared to that. This study examined a measure for producing a 3D mesh model with aerial and ground photography for fortress walls. In addition, this study tested if it would be possible to monitor the slope and protrusion, etc. of fortress walls. This study tested the accuracy of the model produced with photogrammetry to use in the diagnosis and monitoring of the structural safety of fortress walls. The research subject was Korea’s cultural heritage, Seosan Haemieupseong Fortress, and a model was produced for the section about 100m of the fortress walls. The positional accuracy, elevation/section accuracy, LOD, and texture joint precision of two models (Model A & Model B) produced with photogrammetry and laser scan were compared. This study revealed that the model produced at LOD Level 4 would be effective for monitoring the displacement since it could measure the shape, curvature, and slope of stones of the fortress walls.


INTRODUCTION
It is necessary to develop measures to repair and preserve the cultural heritage by frequently monitoring it. In particular, the displacement and damage including collapse and overturn are likely to occur over time in the structures such as fortresses or Eupseong, due to the earth pressure or the inflow of running water, so it is essential to frequently check the safety of them. The Korean building cultural properties are regularly examined every three or five years, for the safety monitoring of them. They are designated as the intensive control lists, if more systematic examination is required because serious faults are found, and then, the precise measurement of them is conducted by using the visual examination and equipment. It is very important to record several figures including slops and projections, in monitoring long fortress structures (Jo and Hong, 2019). A 3D model established allows to do the efficient and dimensional monitoring on them. Although the laser scanning commonly used for producing the 3D model for the cultural heritage can acquire high-precise data, it needs to takes long time to acquire the data and undergo the post-processing process. It is necessary to repeat the periodical measurements, to discover the displacement, but there is a limit in comparing and discovering the displacement by using big point cloud data. In addition, it is also difficult to apply scanners to the fortress walls located at steep terrain or cliffs and acquire the data, as the blockade occurs, and therefore, the production costs increase. It becomes possible produce an efficient 3D model using the digital photogrammetry, as the drone shooting can be recently possible. The accuracy is also actively examined. The digital photogrammetry is effective in reducing expenses and working hours and its errors is lower as the GSD less than 3cm(Kim, 2020), compared to the laser scanning. O Junyeong et al.(2017) proved that the digital photogrammetry is more useful than the laser scanning by applying both techniques to a rock cliff Buddha. Kimg Seonghan et al.(2019) noted the potential to acquire the 3D model of fortress walls which persons cannot easily access, through the digital photogrammetry using drones and the laser scanning as a short monitoring method, it did not examine the usability of it as the long-term technology. This study analyzed the accuracy and the precision of the 3D model of complex and long fortress walls among the Korean cultural heritage, which was produced with the digital photogrammetry using drones. It examined whether it is possible to regularly and accurately diagnose the safety conditions of fortress walls, such as the slopes and projections, by using the model established through the digital photogrammetry.

STUDY AREA AND METHOD
This study examined the Seosan Haemi Eupseong, among the Korean cultural heritage, which was constructed in 1491 ( Figure  1). Eupseong is located at Seosan-myeon, Chungnam and its whole circumstance, internal area are 1,800m and 196,381㎡, respectively. The average height of the fortress walls is 4.9m, and the top width is about 2.1m and the soil is slantingly accumulated at the inside of the fortress walls (Kim et al., 2016). About 23-26 facing stones are accumulated on the fortress walls and the size of them gradually increase from the bottom to the top. It selected the target area by considering the presence of the neighboring obstacles and hurdles, the topography for the drone shooting and the safety in shooting it( Figure 1). It produced a 3D model by selecting the interval of about 100m, from the west gate to Poru(artillery bastions) along the whole fortress walls, and analyzed the accuracy and the precision for about 2m, the part of the interval. It collected the data from Jun., 08 to Jun., This study was conducted by surveying the ground control point and the control point, followed by collecting the 3D data based on the laser scanning and the digital photogrammetry, and then analyses on the unification of two models, and the accuracy and the precision of the digital photogrammetry 3D model. The laser scanning commonly used for producing 3D models, which is a measurement for determining the relative position of an observation point, by using the time difference and the rotation angle, calculated while infrared laser reflects off the observation point and returns to its original point, can collect high precise data.
There are smallest errors between The laser scanning outcomes and real objects, so the analysis on the accuracy used the 3D models produced from the laser scanning and the digital photogrammetry as the control and the comparison, respectively.

Ground Control Point and Control Point Surveying
This study surveyed a total of 11 ground control points (one ground control point per 0.01㎢) with the position accuracy less than 3cm, by using Leica GS08+, based on VRS( Figure 2). It selected the areas that can be clearly distinguished by using topographical features such as manholes, paving blocks, etc., and those with the visual field more than 45° from the ceiling, to prevent the distortion during the drone shooting.  The control point surveying was conducted in the same way in which the ground control point surveying was conducted. Then, the control points were used in analyzing the accuracy. The amount of the control points should be at least 1/3 of the amount of the ground control points, according to the UAVbased public surveying guideline. This study surveyed 9 control points against 11 ground control points, to verify the accuracy of precise positions.

Collection of 3D Data
It used the DJI Mavic Air and DJI Mavic 2 Pro, equipped with GPS, for the digital photogrammetry. The data were collected by initially making the shooting plan, followed by checking the drones and the camera performance, and the, taking vertical and oblique shots in the air, taking horizontal and oblique shots at the lower altitude, taking occlusion areas and examining and arranging the shooting data.
It acquired 1,227 photographs by taking vertical and oblique shots in the air, at the altitude of 25m, given the height of the fortress walls. It compensated the curved and occlusion areas of the fortress walls, by taking horizontal and oblique shots at the low altitude. It acquired 297 photographs by manually operating a drone to taking vertical shots of the point 10 m away from the fortress walls at the low altitude (3, 5, and 7m height), with the cameral angle of about 45°. It acquired a total of 110 photographs about the occlusion area by adjusting the set points such as the camera exposure, ISO, etc., so a total of 2,658 photographs were used in the analysis (Figure 3). It spent about 7 hours to taking shots.
It planned the longitudinal overlap of the shooting more than 80% and the crossing overlap of it more than 70%, to minimize the occlusion area and matching errors, given the area of Eupseong, and took shots with the resulting overlap of 80% or 85-90% (Table 1) It used FARO Focus 3D X330 HDR for the razer scanning. It scanned the top, facade and the rear of the fortress walls, given the height of them and the occlusion area (Figure 4). The it collected a total of 51 scanned data sets consisting of 18 data sets on the top, 17 data sets at the facade and 16 data sets at the rear, and there was the interval from 5 to 7 cm between two scan points (

Production of 3D Model
The digital photogrammetry-based 3D model production was conducted by inputting shot images, followed by matching the ground control points, and then, the aerotraingulation and the post-processing. It used the Agisoft Metashape for the 3D model production. It inputted aerial and ground images shot by drones and the exterior orientation parameters acquired from GPS applied to the drones. The exterior orientation parameters, which produce the posture values of drones by using the geometric conditions of the drone shooting, and the X, Y and Z position and posture information of drones as Omega, Phi and Keppa, can be used the initial ones for establishing the model. It tried to match with the images into which the exterior orientation parameters are inputted, by using the WGS84 ellipsoid coordinate system set in surveying the ground control points. The aerotriangulation was conducted by the cameral calibration, image matching and he bundle block adjustment. It automatically matched with overlapped images and removed the images with high automatic tie points and RMSE between images( Figure 5).
The production of a 3D model using the laser scanning was conducted by the matching of the scan data, followed by the merging of them., and the, the post-processing of them. It used the Cyclone and Geomagic Design X, for the production of it. It matched with the whole scan data based on the common control points of more than three points and merged them into a model ( Figure 5). For the efficiency, it removed the data outside of this study scope, in the post-processing of two models.

Unification of Two Models
It unified the forms and coordinates, to compare two models. It abstracted Model A and Model B as the same point cloud data, through the Reality Capture and the Cyclone, respectively, to minimize the errors occurring in comparing them. Then, it makes two model have the same coordinates, by applying the absolute coordinates of the model A to the model B, based on the cloud comparison ( Figure 6). a. Model A b. Model B c. Unification of Two Models Figure 6. Unification of Two Models

Accuracy of Positions:
It compared a total of nine control points acquired with VAR, for the model A. It calculated the differences in the coordinates between the control points of the model A and those surveyed. The findings show that the minimal, the maximal and the average error of the coordinates were 0.008m, 0.037m and 0.023m, respectively, with the standard deviation of 0.009m and the RMSE of 0.025m, indicating the very high accuracy (Table 3) The façade accuracy was conducted by initially producing the 3D mesh model based on the digital photogrammetry, followed by selecting fortress stones, and then, comparing the size of them. For a precise comparison of two data sets, it conducted a test by using the 3D mesh model, not in the form of point. It compared the width of fortress stones, by randomly selecting 10 fortress stones from the façade of the fortress walls. The results show that each of the average, the maximal and the minimal error was 1.40mm, 2.5mm and 0.2mm, respectively, indicating the high accuracy (Figure 7). It compares the increases of Y with the differences in X, for 20 sections (Figure 9). It estimated a total of 50 difference in X, by equally dividing the sections at 0.1mcm gaps for the model A and B as follows: y=0, y=0.1m, y=0.2m, ⋯ , y=4.9m. It reviewed a total of 50 differences in X for 20 sections each in the model A and B.

No
a. Model A b. Model B

Figure 9. Comparison of Positional Coordinate by Points of Sections
The results show that each of the average, the maximal and the minimal error was 0.013m, 0.062m and 0.000m, respectively, with the average RMSE of 0.263m, indicating that there is little difference between two models ( The slop and projecting and depressed parts of the fortress walls were expressed, and the real image texture was also embodied for the model A, indicating that it satisfies the criteria of the LOD Level 4 specified in the rule.

Joining Precision:
A Test of the texture joining precision for the model A shows that each resolution of the vicinity of fortress walls and the fortress walls was 6mm/pixel and 0.12mm/pixel, respectively, indicating that the resolution of the fortress walls was higher than that of their vicinity( Figure  10). The number of overlapped images was up to 810 for the fortress walls, while aerial shots of their vicinity were often taken, so the number of their overlapped images was less than 810. In addition, multiangle shots of the vicinity of the fortress walls were not taken, so errors occurred or the resolution was low.
Not only the aerial shooting using drones but also the shooting from various angles, including vertical and oblique shooting and the occlusion area shooting are required for the large-area cultural heritage, such as Haemieupseong, and it is effective to use both automatic and manual shooting modes in producing the 3D model with the high joining precision.

CONCLUSIONS
This study established the 3D model for the fortress walls, the large-area cultural heritage, through the laser scanning and the digital photogrammetry using drones. It was found that it is effective to use the aerial shooting, the shooting from various angles, including vertical and oblique shooting and the occlusion area shooting, and both automatic and manual shooting modes, for increasing the precision of the models during the digital photogrammetry.
The accuracy and precision of the digital photogrammetry were as good as those of the laser scanning. The digital photogrammetry was expected to be effectively used for the displacement survey monitoring which helps quantitatively understand the form, curve, slop, etc. of the fortress walls, as it could save more than five hours in collecting the data. It proved that the laser scanning commonly used for currently producing the 3D model can be replaced with the digital photogrammetry which is more efficient in terms of both cost and time, in monitoring the large-area cultural heritage such as the fortress walls for a long time. It would be possible to do the fast and efficient structural safety diagnosis monitoring on the large-area cultural heritage including the fortress walls, if the digital photogrammetry is used for it.