Case Study on the Use of an Unmanned Aerial System and Terrestrial Laser Scanner Combination Analysis Based on Slope Anchor Damage Factors

Abstract

This study utilized unmanned aerial systems (UAS) and terrestrial laser scanners (TLS) to develop a 3D numerical model of slope anchors and conduct a comprehensive analysis. Initial data were collected using a UAS with 4 K resolution, followed by a second dataset captured 6 months later with 8 K resolution after artificially damaging the anchor. The model analyzed damage factors such as cracks, destruction, movement, and settlement. Cracks smaller than 0.3 mm were detected with an error margin of ±0.05 mm. The maximum damaged area on the anchor head was within 3% of the designed value, and the volume of damaged regions was quantified. A combination analysis examined elevation differences on the anchor’s irregular bottom surface, resulting in an average difference at 20 points, reflecting ground adhesion. The rotation angle (<1°) and displacement of the anchor head were also measured. The study successfully extracted quantitative damage data, demonstrating the potential for an accurate assessment of anchor performance. The findings highlight the value of integrating UAS and TLS technologies for slope maintenance. By organizing these quantitative metrics into a database, this approach offers a robust alternative to traditional visual inspections, especially for inaccessible facilities, providing a foundation for enhanced safety evaluations.

1. Introduction

The frequency and magnitude of climate change-induced disasters are increasing globally, manifested by (among others) increases in annual temperatures and precipitation. The aging and performance degradation of SOC facilities are accelerating in tandem with climate change. In South Korea, 2020 recorded the longest rainy season in the past 40 years, with rainfalls of ≥5 mm for 53 consecutive days [1]. Additionally, the number of slope collapses influenced by typhoons more than doubled compared with those in the previous year, totaling 326 incidents. This surge in collapses is attributed to intensified external environmental changes, such as concentrated heavy rainfall and typhoons due to climate change, ongoing developments in mountainous areas, and the aging of slopes [2]. In South Korea, slope maintenance is assessed based on numerous factors, such as usability, safety, and durability, with maintenance goals set accordingly. Specifically, slope reinforcement is evaluated qualitatively by examining the adhesion between the anchor head and the ground, as well as the presence of cracks and failures in the anchor head. Moreover, only approximately 10% of the anchors installed on slopes are selectively inspected, leading to potential inadequacies in visual inspections, particularly in areas that are difficult for inspectors to access. Consequently, it remains challenging to utilize quantitative data on cracks and failures in concrete hydraulic structures and historical data on occurrence sections owing to reliance on text-based inspection reports. Addressing slope safety management in the context of climate change is becoming increasingly critical, prompting efforts to incorporate smart technology or 4th industrial revolution technologies as potential solutions [3].

Spatial information is gaining prominence in the realm of smart facility management. Digital twin technology creates a digital replica of a real-world object, integrating various physical attributes to analyze past and present operating conditions and predict future responses. Notably, the use of unmanned aerial systems (UAS), light detection and ranging (LiDAR), and global navigation satellite systems (GNSS) to generate 3D numerical models is central to digital twins and is increasingly applied in the construction industry [4]. Advances in rapidly acquiring high-density terrain data and analyzing displacement using big data have progressed significantly. Techniques for detecting changes in terrain and infrastructure have evolved and range from visual comparison methods [5,6] to measurement-based approaches [7], and currently, they integrate image and laser sensor technologies. Combining GNSS positioning sensors, high-resolution image sensors, and LiDAR sensors in ground-based laser scanners like terrestrial laser scanners (TLS) and UAS-based aerial photogrammetry allows for smart fusion technology in precise 3D displacement detection. This technology provides detailed positioning information, pixel data, and point clouds for various infrastructure facilities, including terrain, bridges, tunnels, dams, and slope anchors. Specifically, the filtering of multiple return information and the classification of surface vegetation removal (based on reflection intensity as demonstrated by UAS [8] and TLS [9,10]) offer valuable insights for identifying and detecting terrain features and changes.

2. Research Trends

Xiao et al. [11] developed a 3D numerical model from images of an excavated slope and quantitatively assessed stability to explore the feasibility of using UAS for inspections. Cho et al. [12] compared the accuracy of 3D modeling of road slopes using photogrammetry in conjunction with actual measurements. The inverse analysis of discrepancies between onsite displacement measurements and the 3D model’s position showed that horizontal displacements along the x- and y-axes had a root-mean-square error of 3 mm. In contrast, vertical displacements along the z-axis exhibited a larger error of 15 mm compared with horizontal displacements. Typically, photogrammetric measurements yield horizontal displacement errors in the range of 20–30 mm and vertical displacement errors in the range of 40–60 mm. Given these ranges, the numerical accuracy of the 3D modeling was assessed to be very high. Kang and Kim [13] evaluated the quality of a 3D slope model using drone photography with ground control points (GCP). For vertical facilities like slopes, where vertical displacement errors are critical, the model was found to be underutilized. Thus, it is essential to address data gaps in blind spots caused by environmental factors that affect photographing and scanning distances, as well as the unevenness of vertical installations such as slopes.

Before utilizing UAS for collecting data to create 3D digital models, aerial photogrammetry and TLS were employed for generating 3D structural models [11,14]. Hong and Park. [15] and Hellmy et al. [16] demonstrated the feasibility of assessing the external stability of rock slopes by measuring discontinuities with TLS. Additionally, Park [17] and Lee and Park. [18] suggested that slope displacements could be monitored for maintenance purposes. Despite these advances, limitations persist owing to blind spots caused by vegetation or data acquisition failures by slope, which introduce errors and obstruct optimal monitoring conditions.

The creation of 3D numerical models, including various environmental factors and vertical structures such as dams, retaining walls, excavation faces, and slopes, often results in blind spots in both TLS and UAS images due to uneven vertical structures. To build more accurate 3D numerical models, it is essential to address these data gaps based on the photographing angle and position. UAS data acquisition often results in occlusion areas from the ground or the underside of roofs, while TLS generates occlusion areas in the higher areas of facilities. Combining these two data sources can resolve these occlusion areas [19]. Furthermore, integrating TLS and UAV data using scan-to-building information modeling (BIM) enables the mutual enhancement and construction of BIM data for geometric representation with a level of detail 3 [20]. Kim et al. [21] proposed a hazard risk assessment method for steep slopes using UAS images and terrestrial LiDAR. Kang and Lee [22] constructed 3D spatial information of an artificial climbing wall by combining UAS images and terrestrial LiDAR, demonstrating the potential for precise 3D reproduction by addressing blind spots with supplementary data and suggesting its applicability for maintenance. However, research applying this method to slope maintenance remains limited, highlighting the need for further investigation.

To evaluate the quantitative metrics for the reinforcement performance of the anchor method installed on slopes, this study implemented a 3D numerical model twice on a slope with an artificially constructed anchor (10 places) using UAS image filming, structure-from-motion (SfM)-based image analysis, and a combination analysis of TLS’s point cloud data. In the first implementation, images were recorded at 4 K resolution. In the second implementation, after artificially inducing damage, 3D modeling was performed with 8 K resolution to derive damage factors such as adhesion, displacement, and failure of the slope anchor head. In particular, to examine the crack reproducibility of the 3D model by resolution, the data were configured as 4 K and 8 K, respectively. Additionally, laser scanning using TLS was performed. The study first validated the accuracy and reproduction of the constructed 3D numerical model. It then analyzed the damage factor of the slope anchor, focusing on cracks, by comparing the measured crack width. Failure and adhesion with the ground were assessed by detecting quantitative numerical data in the 3D model, and error differences in crack width were analyzed by distinguishing between 4 K and 8 K UAS image resolutions. For anchor head failure, the percentage of failures was analyzed by detecting the area and volume of the failure. Rotation angle and displacement were measured when the head caused rotational displacement. The ground settlement was assessed by measuring the adhesion of the anchor head, and a comparative analysis was conducted to determine the maximum depth of settlement in the numerical model based on a combination of UAS and terrestrial LiDAR analysis. Finally, the study examined the potential for utilizing the extracted damage factors from the 3D numerical model for maintenance purposes. Figure 1 shows the process of this study.

3. Trends in Point Cloud Coordinate Displacement Analysis

3.1. Point Cloud Combination

This study investigates these methods for 3D point cloud-based combination analysis. The first step in detecting changes based on point cloud data involves merging point clouds acquired at different times, using either fixed or free points as the basis for combination. The fixed point-based combination employs fixed points as artificial or natural conjugate points at the same location between point clouds. In contrast, the free point-based combination relies on geometric characteristics or the iterative closest point (ICP) algorithm proposed by Besl and McKay [23], as well as the extended ICP algorithm suggested by Chen and Medioni [24]. Combining point clouds from actual facilities is highly complex, and the convergence rates and accuracy vary depending on the composition of the objective functions. Thus, this process is executed through rigid body transformation. ICP is a technique that aligns a comparison with a reference point cloud sequentially without requiring a precise fixed point between the two clouds. This alignment uses the least-squares adjustment method to determine the translation matrix (T) and rotation matrix (R), among other transformation matrices, at a tolerance level of “minimum distance” or “minimum search distance”. The goal is to optimize the correspondence between the two point clouds. However, since the introduction of the ICP theory, various derived algorithms or combinations of these algorithms have been proposed to enhance efficiency due to the low performance caused by the characteristics of point cloud data or issues with divergence during the combination adjustment stage [25,26,27].

3.2. Displacement Analysis and Calculation

The typical combination process involves integrating 3D point clouds (obtained from sensors) embedded in TLS, UAS, and similar technologies, into software that can manage them. A single reference point cloud, selected for its high accuracy and precision, is chosen from overlapping areas between multiple-point clouds. Feature points within standardized categories, such as voxels, are identified in the reference point clouds. Conjugate candidate points on the comparison point clouds are then located based on iterative calculations using least-squares adjustment, with noncorresponding points removed. The weighting factor is determined based on elements such as distance, normal deviation, and surface gradient from each point cloud combination that has established a correspondence relationship with the feature points. Conversion factors, including rotation and translation, are estimated to minimize the distance deviation between the point clouds. These estimated conversion factors are applied to the point clouds for comparison; after conversion, the distance deviation or convergence of the voxel search radius—an acceptable criterion for combination decision between each point cloud—is assessed (ε: acceptable range; n = nmax, maximum number of cycles). If divergence occurs, iterative operations are performed on different feature points within a standardized category with duplicate point clouds. If convergence is achieved, displacement analysis and displacement calculations are conducted [28].

3.3. Review of Displacement Calculation Algorithm

The method used to calculate displacement can be classified based on selected criteria [29,30,31,32]. Algorithms such as cloud-to-cloud, cloud-to-mesh [33], and mesh-to-mesh [34] compute displacements as Euclidean distances by using the closest points between two-point clouds. Although these methods tend to underestimate displacement values, they are commonly used in point cloud-based displacement calculations.

Other algorithms, including F2S3 [35], CD-PB M3C2 [36], image-based correlation [37], and image-based feature points [38], use Euclidean distances between extracted or corresponding feature elements from two-point clouds or converted 2D images. However, these approaches have limitations, such as decreased spatial resolution and accuracy when dealing with repetitive patterns [28].

Algorithms based on parameterization methods, such as geometric primitives [39] and B-spline surfaces [32], parameterize the surface of point clouds to calculate variations in parameters and estimate the range. To utilize point clouds that include geometric primitive elements or those capable of detecting parameter variations, prior information on global or local parameters is required.

Algorithms such as DoD [40], M3C2 [41], and M3C2-EP [42] are direction-based methods that identify corresponding points between two point clouds along a predefined direction, such as the gravity direction or the surface normal direction of the point clouds. These algorithms calculate displacement using Euclidean distances, making them suitable for point clouds with directional displacement information. However, if the actual displacement direction differs from the predefined direction, these methods may exhibit biases, leading to either displacement overestimation or underestimation.

A representative partial combination-based method is local ICP [43]. This approach involves selecting overlapping subsets from two point clouds and performing a partial combination procedure, akin to the nearest neighbor-based method, to derive transformation matrices and displacement vectors. Local ICP is effective for estimating displacements in point cloud data with a strong correspondence relationship, as it identifies a subset group with corresponding points between the two point clouds without requiring prior information.

Recently, various efforts have been reported across different fields to utilize point cloud data for combination and displacement analysis using artificial-intelligence-based learning methods, such as SiamPointNet++, SiamGCN, and Siamese KPConv [44,45].

Additionally, several commercial analysis tools for point cloud data combinations and displacement calculations have been introduced. For example, Dewez et al. [46] developed a dedicated plugin called FACETS, which was integrated into the CloudCompare analysis tool. This plugin uses the Kd-tree and fast-marching algorithms to partition large 3D point cloud datasets into various planar forms. It also supports data storage in .shp or .csv formats, calculates dominant directions and slopes, and offers interactive visualization.

In this study, the ICP algorithm was applied to combine UAS-based point clouds and TLS point clouds, and the conversion formulas to calculate the optimal translation matrix ($\mathrm{T}$) and rotation matrix (R) are as follows:

$$p_i^{\prime }=R{p}_i+T$$

(1)

Here, $p_i$ is the original point and $p_i^{\prime }$ is the point after transformation.

$$T_{final}=\left[\begin{array}{cc}R& T\\ 0& 1\end{array}\right]$$

(2)

Afterwards, the formula for deriving the final transformation matrix $T_{final}$ is derived as follows, and data storage and post-processing are performed through this.

4. Date Processing

4.1. Experimental Area

The slope examined in this study was reinforced with 10 anchor holes, as illustrated in Figure 2, to assess the extent of damage to the anchors as a reinforcement method. This slope is situated in the Gyeonggi Massif region, known for the Gyeonggi Metamorphic Complex, as well as the Chuncheon Formation and the Yeoncheon Group, which are categorized as metamorphic sedimentary rocks and Jurassic igneous rocks, respectively. Additionally, the small-scale Daedong Group, also known as the Gimpo Coalfield or Gyeonggi Coalfield from the Mesozoic Jurassic period, is present alongside Precambrian gneisses and schists. The slope comprises sedimentary rocks from the Yeoncheon Group, dating to the Precambrian period. It is highly weathered, with a thick weathering layer on the bedrock, resulting in the transformation of residual soil through weathering processes. The slope specifications include a height of 4.5 m and an angle of 33.6° from the ground to the first anchor point. Throughout the 84 m section, 10 anchor holes were installed at 2 m intervals along a 27 m segment.

To apply the SfM algorithm to a 3D numerical model, positional attribute information and geotagging of digital image data are essential. Additionally, a setup point for the TLS is required.

First, two temporary bench marks (TBMs) were placed on the road adjacent to the study area as shown in Figure 3, and a survey was conducted using a GNSS receiver with the specifications listed in

Table 1. Global navigation satellite systems (GNSS) specifications.

Trimble R8	Parameters
	weight	1.52 kg	channel	440 channels
	stop positioning vertical	3.5 mm + 0.4 parts per million (ppm) root-mean-square (RMS)	input	CMR+, CMRx, RTCM2.1–3.1
	stop positioning horizontal	3 mm + 0.1 ppm RMS	output	16 NMEA
	RTK vertical	20 mm + 1 ppm RMS	radio modem	450 MHz
	RTK horizontal	10 mm + 1 ppm RMS	signal update cycle	1–20 Hz

. Based on the acquired coordinates, a total station with the specifications listed in

Table 2. Total station specifications.

HTS-420R	Parameters
	weight	5.5 kg
	accuracy	0.00001 rad
	minimum focusing distance	1.5 m
	prism mode	±(2 mm + 2 ppm)
	reflectorless	±(3 mm + 2 ppm)

was installed at TBM-01, and TBM-02 was used as a backsight to provide the machine with distance and angle information.

Subsequently, coordinates for five GCPs for UAS image-based image analysis and three scan control points (SCPs) for TLS setup were obtained. Moreover, for accurately reconstructing the 3D numerical model of vertical structures based on UAS image analysis, vertical control points (VCPs) are required in addition to GCPs. Therefore, in this study, a total of 169 VCPs were measured across the vertical sections of all anchors.

4.2. UAS Data

In the first case, the equipment used was the DJI Phantom 4 Pro with specifications including a 4 K (5472 × 3648) resolution, a 15.7 mm complementary metal-oxide semiconductor (CMOS) image sensor, an 84° field of view (FOV), and a 24 mm focal length lens, The point cloud was created as shown in Figure 4. For the second case, 3D numerical modeling was performed by inducing artificial displacements at the anchor head. The equipment used was the Autel EVO 2 Dual 640T, featuring an 8 K (8000 × 6000) resolution, a 12.7 CMOS image sensor, and a lens with a 79° FOV. The lens supports a focal length range of 4.3–17.2 mm, with lossless 4× optical zoom. The point cloud was created as shown in Figure 5. Data acquisition was performed using a dual-grid system with a maximum photographing height of 60 m and a photographing angle of 90°. Owing to the slope’s nature, multiple blind spots occurred during automatic flight at a 90° photographing angle. To minimize these blind spots, the camera angle was adjusted according to the slope angle during data collection. Close-range data were obtained by photographing at heights between 5 m and 10 m and angles between 35° and 90°. The average photographing height was set at 50 m with consistent photographing conditions maintained for each case. The UAS aerial image data for the 3D numerical model included 269 images acquired at heights from 5 m to 10 m and 763 images acquired at heights from 10 m to 60 m.

4.3. TLS Data

The TLS device utilized the Trimble SX10 device. SX10 has an effective range of 600 m, a scanning angle accuracy of 1”, and a distance measurement accuracy of 1 mm + 1.5 ppm. Approximately 18 million data points were collected from the entire slope area, with approximately 4.9 million data points from the anchor area. Data were measured once at the TBM-01 point with the high point density setting among the SX10 options, and three times at the SCP 3 points with the normal point density setting. Additionally, the TLS-based point cloud in Figure 6 was generated before artificial damage, and the TLS-based point cloud in Figure 7 was generated after artificial damage was applied to the anchor.

4.4. Create a Combination Model

For the image data, a point cloud was generated using Sfm by matching feature points between images and reverse-calculating the camera pose and position of each image. Although acquiring image data over a large area is a relatively fast process, the accuracy depends on factors such as image overlap, resolution, and errors in camera orientation. TLS data provides precise 3D point cloud data collection and direct processing using a laser as the light source. This method ensures fast data processing and yields high-accuracy outcomes, but area analysis based on visuals can be limited. To address this issue, the study combined image-based 3D data point cloud and TLS point cloud data to reconstruct the point clouds as shown Figure 8 and Figure 9.

4.5. Accuracy and Reproducibility Review

To compare the accuracies of 3D digital models obtained by UAS and TLS, this study used 169 VCP test points on the top surface of the anchor, obtained using a total station. The coordinates of these 169 VCPs were used to track the positions of the same points in both the UAS-based 3D model and the TLS-scanned 3D model. The differences between the coordinates of the tracked points and the VCP test points were expressed using standard deviation and root-mean-square error (RMSE) values. Examined the accuracy of the following three types of 3D point clouds based on total station data. Test-A for accuracy verification between the total station and TLS scans, Test-B for accuracy verification between the total station and UAS-based 3D numerical models, and Test-C for accuracy verification between the total station and the combined analysis. The results of these tests are summarized in

Table 3. Standard deviation (STD) and root-mean-square error (RMSE) values according to the comparison test and the analysis process.

Test		A (TLS)	B (UAS)	C (Combined UAS and TLS)
X	maximum (MAX)	30	28	30
	minimum (MIN)	0	0	0
	average (AVG)	8	12	5
Y	MAX	30	28	30
	MIN	1	0	0
	AVG	14	14	9
XY	MAX	30	30	29
	MIN	1	3	1
	AVG	17	18	11
	STD	8	8	7
	RMSE	19	19	13
Z	MAX	22	50	32
	MIN	0	7	0
	AVG	8	36	6
	STD	4	12	4
	RMSE	9	38	7
(Unit: mm)

. The RMSE values for the horizontal x- and y-axes in the 3D numerical models generated by TLS (Test-A) and UAS (Test-B) were 19 mm. However, the combined 3D numerical model (Test-C) demonstrated improved accuracy with an RMSE of 13 mm. Specifically, the RMSE for the vertical z-coordinate was considerably larger in the UAS-based 3D numerical model (Test-B) at 38 mm, compared with 9 mm in the TLS-based model (Test-A) and 7 mm in the combined model (Test-C). This improvement in accuracy is attributed to the adjustment of UAS-based 3D digital models based on TLS-based models, enhancing the precision of VCP inspection point locations. The large z-coordinate error in UAS-based 3D models, particularly in vertical structures like slopes, tends to reduce model utility. However, combining TLS scans with UAS images addressed the z-coordinate error and improved accuracy, especially in uneven sites with blind spots. This integration of UAS images and TLS filled data gaps and resulted in enhanced accuracy, as demonstrated in Test-C.

Figure 10a illustrates the blind spots on the upper surface of the anchor observed during laser scanning with TLS. These blind spots have been addressed and supplemented with UAS images for the respective area (Figure 10b). The reproduction, which addresses the limitations of blind spots inherent in each data-collection method, is confirmed by the results shown in Figure 11. These confirm the effectiveness of 3D precision analysis based on combined analysis. During the reproduction analysis, the analyzed area was defined to exclude regions in the reproduced 3D numerical model where surface treatment was not feasible, such as vegetation, wires, and cables. Reproduction verification involved the calculation of the area ratio of regions with shape distortion or average resolution below the texture expression level compared with the analyzed target area. The analyzed surface area (totaling 205.54 m²) included the total area of 10 anchors and the ground surface. The areas with distortion and texture errors amounted to 10.28 m² and resulted in a reproduction accuracy of 94.99%.

5. Results and Discussion

5.1. Detection of Damage Factor

The 3D numerical model was developed using data collected in the first phase. A new 3D numerical model for the second phase was created 6 months following the introduction of artificial damage to detect changes in damage factors for anchors installed on the slope, including cracks, failures, rotational displacement, and ground adhesion.

Table 4. Marking damage factors of anchor heads on a 3D model.

No.	Marking Damage Factor	No.	Marking Damage Factor
1		2
3		4
5		6
7		8
9		10

presents the detected damage factors by comparing the 3D numerical models from the first and second phases. Red indicates cracks (C), yellow denotes destruction (D), blue represents movement (M), and orange signifies settlement (S). No damage was detected at anchors 1 and 9. Anchor 2 includes a crack at the center of the head. At anchor 3, cracks, destruction, and movement are observed. Additionally, localized occurrences of cracks, destruction, and movement are identified at anchors 4 and 8. Rotational movements are noted at anchors 2, 5, 6, and 7, with anchor 5 exhibiting the largest rotational displacement. Destruction is found at the end of anchor 6, while settlement is observed at anchors 7 and 10, as confirmed by the 3D numerical model. Quantitative values for the identified damage factors—cracks, destruction, movement, and settlement—were extracted.

5.2. Crack Analysis

Table 4 shows the locations of cracks in 10 anchors, and occurred in anchors numbered 2, 3, 4, 7, 8, and 10.

After inducing artificial damage, the actual crack width was measured using a crack gauge.

Table 5. Comparison of crack width between 3D model and field measurements.

No.	Field Measurement	3D Model	Error
3-C3-1			0.05
	0.75	0.80
3-C3-3			0.03
	0.50	0.53
3-C3-2			0.04
	0.50	0.54
4-C1-1			0.03
	0.25	0.28
4-C2-1			0.11
	0.10	0.21
7-C1-1			0.03
	0.35	0.38
8-C2-1			0.03
	0.20	0.23
8-C1-1			0.05
	0.45	0.50
(Unit: mm)

compares these measured crack width values with those obtained from the second 3D numerical model. Eight cracks were selected for width measurements. Crack measurements in the 3D numerical model images were conducted using the Acute 3D Viewer software program (version 4.0). The average error range between the actual measurements and the 3D numerical model values was ±0.05 mm. Cracks with measured widths less than 0.20 mm did not appear in the 3D numerical model. However, the crack in 4-C2-1, with a measured width of 0.10 mm, was recorded as 0.20 mm in the 3D numerical model. This discrepancy may be because images from low-altitude flights, which captured cracks as small as 0.10 mm, were processed simultaneously with high-altitude data, leading to errors in the 3D numerical model representation.

UAS images were captured at altitudes ranging from 5 m to 60 m. The first set of photographs generated a 3D digital model with a resolution of 4 K (5472 × 3648), while the second set created a 3D digital model with a resolution of 8 K (8000 × 6000).

Based on this, we can calculate GSD by substituting it into the following equation.

$$\mathrm{H}={\displaystyle \frac{\left(\mathrm{i}\mathrm{m}\mathrm{W}\times \mathrm{G}\mathrm{S}\mathrm{D}\times \mathrm{F}\mathrm{R}\right)}{(Sw\times 100)}}$$

(3)

Here, imW is the image width (pixels), FR is the camera focal length, and Sw is the camera sensor width.

With a camera focal length of 14 mm and a photographing height of 5 m, the ground sampling distance (GSD) for an 8 K resolution image was 0.567 mm/pixel. For a 4 K resolution image with the same height and a camera focal length of 8.8 mm, the GSD was 1.63 mm/pixel. To assess the error in crack width based on the resolution of the 3D numerical model, the crack widths identified in both the first and second photographs were compared with the measured and actual values for each numerical model, as shown in

Table 6. Comparison crack width based on images acquired at the resolutions of 4 K and 8 K.

Anchor No.	3D Model (4 K)	3D Model (8 K)	Measurements
3			field: 0.15 4 K: −0.18 8 K: −0.06
	0.33	0.21
7			field: 0.30 4 K: −0.21 8 K: 0.0
	0.51	0.30
(Unit: mm)

. When the measured crack width was 0.15 mm, it appeared as 0.32 mm in the 4 K resolution image and 0.21 mm in the 8 K resolution image. The crack width (measurable at 8 K resolution) was 0.21 mm when the minimum pixel size was considered; by contrast, when it was measured using a resolution of 4 K, it was 0.32 mm. The discrepancy between the measured crack widths was 0.18 mm at a resolution of 4 K, and 0.06 mm at a resolution of 8 K. For a crack width of 0.3 mm, the measurement yielded 0.3 mm at a resolution of 8 K, and 0.51 mm at a resolution of 4 K, resulting in an error of 0.21 mm.

GSD represents the distance between the centers of pixels in an image. A GSD of 1 mm/pixel implies that each pixel spans a distance of 1 mm in reality. However, this does not determine the capability to distinguish damage in the image. Shigeta et al. [47] and Wang et al. [48] have demonstrated (using indoor tests) that crack widths as small as 0.1 mm can be identified with an image pixel size of 1.5 mm. High precision is required for detecting cracks, as the anchor’s water pressure plate, constructed with concrete, demands different maintenance approaches based on crack width. Cracks widths less than 0.3 mm are typically repaired using surface treatment methods, while cracks widths of ≥0.3 mm are addressed with injection repair methods. Therefore, cracks widths of 0.3 mm or greater are managed as structural cracks from a maintenance perspective. This study confirmed that the Acute 3D Viewer program can measure crack widths less than 0.3 mm (with an error of ±0.05) in images acquired with a resolution of 8 K and a pixel size of 0.567 mm.

5.3. Analysis of Destruction

Failures of the anchor head occurred in anchors 3, 4, 6, 7, and 8, as shown in Table 4. The anchor area was 3.736 m² according to the design drawing. The 3D numerical model yielded an error ranging from 17.1% to 6.2% compared with the designed plan area, with anchor areas measured only on the strata surface, as detailed in

Table 7. Detected failure area on anchor head using the 3D model.

Anchor No.	Design Surface Area	Surface Area of 3D Model	Measured Efficiency	Failure Area	Percentage of Failure
1	3.736 m2	3.476 m2	6.9%	-	-
2		3.115 m2	16.6%	0.011 m2	0.29%
3		3.156 m2	15.5%	0.147 m2	3.93%
4		3.100 m2	17.0%	0.098 m2	2.62%
5		3.098 m2	17.1%	-	-
6		3.181 m2	14.9%	0.107 m2	2.86%
7		3.349 m2	10.4%	0.050 m2	1.34%
8		3.504 m2	6.2%	0.083 m2	2.22%
9		3.121 m2	16.5%	-	-
10		3.382 m2	9.5%	-	-

, ranging from 3.098 to 3.5004 m². The error in the area can be attributed to distortion caused by reduced overlap due to interference from vegetation, soil, and gravel in the ground structures, as illustrated in

Table 8. Causes of reproducibility error of anchor area based on the 3D model.

Anchor No.	Photograph	3D Model
3
2

. For areas with high overlap, such as anchor 3, the failure area is visible. However, the failure area on the side of anchor 2, which has low data overlap, is difficult to access due to terrain variations. Additionally, the data quantity in that area was relatively low, and vegetation was present nearby. Consequently, when the 3D numerical model was constructed, vegetation was also included, leading to lower data quality. Section 4.5 found that the overall reproduction rate of the slope was approximately 95%. However, in uneven ground structures like the contact area between the anchor and the ground, areas with distortion and the influence of vegetation and soil prevented accurate surface treatment, resulting in high distortion.

The percentage of failure relative to the designed area of the anchor head for anchor 3, with failures observed at four locations, was maximized at 3.93%, as shown in Table 7. While this may be considered a minor failure in terms of area, detecting quantitative values for the failure volume is crucial due to the 3D nature of structures.

Table 9. Detected failure volume of anchor head based on the 3D model.

Anchor No.	Designed Volume	Failure Volume				Total
		B1	B2	B3	B4
2	538,973	130.16	-	-	-	130.16
3		1792.84	21.11	264.53	108.52	2187.00
4		1258.10	25.84	261.84	-	1545.78
6		2256.20	32.49	-	-	2288.69
7		1812.69	-	-	-	1812.69
8		2125.00	1488.98	-	-	3613.98
(Unit: cm3)

presents the detected failure volumes for the anchors. The B-1 point of anchor 6 exhibited the largest failure volume, while anchor 8 had the highest overall failure volume. Although anchor 3 had the largest failure area, the failure volume was low, indicating that the failure was confined to the surface of anchor 3.

The occurrence of failure in the anchor head was assessed according to qualitative evaluations based on visual observations given the difficulty in measuring damage factors, such as cracks and ground adhesion, from a maintenance perspective. Owing to the lack of precise quantitative criteria, it is challenging to assess the extent of failure based solely on qualitative evaluation results. However, this study demonstrated that the implemented 3D numerical model enabled the quantitative detection of the failure area and volume based on orthographic projections. While there are currently no established quantitative evaluation metrics for failure, accumulating such foundational data is expected to provide a basis for future quantitative maintenance assessments.

5.4. Analysis of Ground Adhesion

The purpose of the water pressure plate in an anchor is to distribute evenly the anchor’s tension to the ground, and it functions effectively when it is tightly adhered to the surface. The performance evaluation-based slope reinforcement method assesses the adhesion between the anchor’s head and the ground using three qualitative metrics: “no gap,” “some gaps”, and “overall gaps”. When overall gaps are present, compensation and reinforcement are typically applied. However, when only some gaps are observed, the decision to apply compensation and reinforcement may depend on the inspector’s judgment. Relying solely on qualitative evaluation data to assess the occurrence of ground adhesion gaps is challenging. Therefore, it is essential to determine numerically the precise location and extent of gap occurrence to manage the time history effectively.

The slope under study features residual soil in highly weathered conditions, which increases the likelihood of local ground adhesion loss owing to its irregular surface. In this study, artificial ground settlement was induced, and blind spots occurred at the anchor’s bottom and ground surface positions in the UAS 3D numerical model, preventing accurate measurements of ground adhesion. Figure 12a shows the UAS-based 3D numerical model as a mesh. Figure 12b presents point cloud data from TLS. Figure 12c depicts the mesh processing results of the TLS point cloud data, Figure 12d illustrates the results obtained by combining UAS images with TLS point cloud data, where the point cloud data were inserted based on coordinates, processed, and meshed to create a 3D numerical model. The TLS point cloud data in Figure 12b do not show ground adhesion for the entire anchor owing to blind spots. Mesh representation of the TLS point cloud data in Figure 12c was challenging owing to occluded areas. However, in Figure 12d, the 3D numerical model was created by aligning and reprocessing TLS coordinates, resulting in a more accurate model that compensates for image distortion, blind spots, and coordinate errors in the TLS point cloud data.

The 3D numerical model created with UAS images yielded a higher error range regarding the z-coordinate compared with the errors of the x- and y-coordinates; it also contained blind spots owing to the photographic angle and altitude. Specifically, blind spots frequently occur in the contact area between the anchor and the ground due to uneven structures. To address this issue, the potential for detecting measurement values for ground settlement was verified using a 3D numerical model combined with terrestrial TLS, as shown in

Table 10. Detected settlement based on the 3D model.

Anchor No.	Detected Points	Settlement
1		A	B	C
		0.007	0.017	0.004
		D		E
		0.024	-	0.018
		F	G	H
		0.034	0.017	0.014
7		A	B	C
		0.064	0.018	0.028
		D		E
		0.044	-	0.024
		F	G	H
		0.155	0.040	0.119
10		A	B	C
		0.013	0.009	0.027
		D		E
		0.011	-	0.014
		F	G	H
		0.179	0.018	0.152
(Unit: m)

.

Ground settlement occurred at anchors 1, 7, and 10, while rotation displacement was observed at anchors 2, 5, 6, and 7. At anchor points 7 and 10, ground settlement of at least 0.1 m was noted, and the adhesion between the anchor head and the ground surface was measured. The elevation differences between the ground and the 20 anchor points were calculated and averaged to represent the degree of adherence, considering the irregular and uneven ground structures. The results are presented in

Table 11. Detected settlement below the anchor head based on the 3D model.

Anchor No.	Detection of 20 Settlement Points	Average Settlement
7		0.081
10		0.126
(unit: m)

. For anchor 7, an average ground settlement of 0.081 m was observed, while for anchor 10, an average ground settlement of 0.126 m was confirmed.

5.5. Analysis of Rotational Displacement

The damage factor associated with the ground adhesion of the anchor head can also account for rotational displacement. The anchor head was directly subjected to the vertical force of the tensile load, while the support force was influenced by the surface flatness. Although performance evaluation metrics typically do not consider rotational displacement, it is important to evaluate it in relation to ground adhesion. Rotational displacement can indirectly indicate fine gaps between the ground and the anchor head or a decrease in vertical force due to the relaxation of anchor tension, leading to a reduced tensile load. The movement of the anchor’s rotator was assessed by combining UAS and TLS to create a 3D numerical model. The results showed that considerable displacements, such as that at anchor 5, were easily identified. However, subtle movements, as observed at anchors 2, 6, and 7, were more challenging to detect. In the 3D numerical model used for combination analysis of UAS and TLS from the first and second phases (see

Table 12. Detected rotational displacement of anchor head based on the combined 3D model.

Anchor No.	Before Damage	After Damage	Rotational Displacement
2				Rotation angle =1.31° Displacement = 0.005 m
5				Rotation angle = 13.58° Displacement = 0.064 m
6				Rotation angle = 0.89° Displacement = 0.011 m
7				Rotation angle = 0.97° Displacement = 0.042 m

), anchors were extracted and mapped based on their coordinates. The analysis revealed slight movements at anchors 2, 6, and 7. Displacement angles were measured by comparing them with the anchor cap’s center as the reference point. The most pronounced displacement was observed at anchor 5, with a rotational angle of 13.58° and a displacement of 0.064 m. Detection was also possible for rotational angles less than 1°.

5.6. Discussion

Combination analysis was performed using UAS images by merging 3D data with TLS 3D point cloud data to reconstruct a TIN, incorporating damage factors such as cracks, failures, displacements, and ground settlements. This approach enabled the detection of numerical values through the 3D numerical model and confirmed the feasibility of applying strategic performance evaluation-based maintenance.

Crack width measurements may vary in accuracy depending on the resolution of the UAS camera. However, cracks narrower than 0.3 mm were successfully detected. Additionally, failure areas and volumes were identifiable. Nevertheless, UAS images contained data gaps in the adhesive area between the anchor and the ground, which prevented the detection of displacement in the fine anchor head and the amount of ground subsidence. While UAS image data allows for the rapid acquisition of large-scale 3D data, its accuracy is influenced by image overlap, resolution, and camera calibration errors. In contrast, TLS data provide high-coordinate accuracy but lead to limited area analysis owing to its data characteristics, yielding lower precision outcomes in crack and failure analysis compared with image analysis outcomes. Despite the data gaps caused by blind spots on the anchor’s upper surface during laser scanning, high-density 3D point cloud data enabled the detection of displacement and ground settlement numerical values. However, detecting cracks and failures remained challenging owing to the resolution of the images and errors in camera orientation.

In detecting damage factors in a 3D numerical model of a slope, image-based 3D data demonstrated high potential for area analysis, such as crack and failure detection. However, TLS point cloud data yielded higher accuracy and precision regarding the analyzed displacement and ground settlement outcomes.

This study conducted a testbed analysis by combining UAS images and TLS data to detect damage factors on slope anchors in a 3D numerical model. The goal was to minimize the impact of vegetation, a major interfering factor in ground structural analyses. Collecting UAS image data with high overlap and acquiring TLS point cloud data can be challenging owing to vegetation and surrounding obstacles in actual road slopes. In this study, both UAS and TLS were utilized. However, by incorporating additional equipment, such as drone TLS and ground cameras for image acquisition, the impact of these interfering factors can be minimized, and 3D data collection can be improved. Additional research is needed to implement 3D numerical models based on combination analysis using various methods. The results of this study indicate a clear distinction between the detection of damage factors based on image analysis and point cloud data. Combining these two methods to create a numerical model with 3D data can achieve high accuracy and precision in identifying damage factors for maintenance purposes.

6. Conclusions

To evaluate the effectiveness of performance assessment based on images for the strategic maintenance of anchors installed on slopes, a 3D numerical model was developed by combining UAS images and TLS. The accuracy and reproduction of the model were subsequently verified. Additionally, quantitative measurements were performed to detect cracks, failures, displacements, and ground adhesion in the anchor’s head. These findings were then analyzed from a maintenance perspective, leading to the following conclusions:

• In vertical structures, such as slopes, the z-coordinate error of the 3D numerical model based on UAS images was relatively larger than the x- and y-coordinate errors. This issue is particularly pronounced in uneven structures with installed anchors, where data gaps arise owing to blind spots caused by the photographic angle. To address this problem, a 3D numerical model was created by combining TLS scan data. This model improved the accuracy of the z-coordinate through the mutual complementation of the two datasets;
• By constructing a 3D numerical model, the accuracy difference due to resolution was assessed for crack detection. In the 3D numerical model with 8 K resolution, cracks smaller than 0.3 mm were detected with an error range of ±0.05 mm. This led to important findings regarding maintenance;
• The 3D model achieved an approximate reproduction accuracy of 95%. However, distortion in the anchors installed on slopes, caused by factors such as vegetation and gravel, resulted in an error ranging from 6.2% to 17.1% compared with the designed area. Additional research is required to address these interfering factors;
• Numerical values for area and volume of failures were detected, with the failure area of the anchor water pressure plate ranging from 0.29% to 3.93% compared with the design, indicating minor damage. Although squantitative evaluation metrics for anchor failure have not yet been established, accumulating numerical data from 3D numerical models could provide a basis for quantitative maintenance evaluations;
• For important anchors where z-coordinate data are crucial, ground adhesion detection was not possible with the UAS 3D numerical model owing to the low overlap from a low shooting angle and blind spots from uneven structures. However, the combined analysis with TLS point cloud data enabled the measurements of ground elevation differences. Ground subsidence was 0.081 m at anchor 7 and 0.126 m at anchor 10. Rotational displacement of the anchor head was observed at anchors 2, 5, 6, and 7, with fine rotation angles measurable within 1°;
• This study confirmed that image-based 3D numerical models offer considerable potential for area analysis, such as crack and failure detection. However, for damage assessment involving displacement and ground subsidence, TLS point cloud data provides higher accuracy and precision compared with images. Therefore, the combination of UAS and TLS data to create a 3D numerical model was validated, showcasing the advantages of each method while addressing their respective limitations.

This study provides an efficient and objective technical means for slope safety maintenance, especially suitable for areas where traditional methods are difficult to cover, promoting the practical application of digital twin technology in infrastructure management. Despite the impact of vegetation occlusion and terrain complexity, the study proves the advantages of the combined technology in damage assessment through data fusion and algorithm optimization, can provide an important reference for subsequent research and engineering practice.