2.1 Study Area

The study area is Sangneung Village in Sancheong-gun, Gyeongsangnam-do, Republic of Korea, a rural mountain community where steep slopes adjoin residential areas and farmland. The upper part of the area consists of relatively steep mountain terrain, whereas the village and agricultural fields occupy the lower part. During intense rainfall, soil and rock released from the upper slopes can therefore move rapidly downstream along valleys and topographic depressions and directly affect inhabited areas.

Intense rainfall triggered a landslide on the upper mountain slope. Eroded soil and rock, fallen trees, and surface debris moved downslope and entered and accumulated in parts of the residential and agricultural areas. The upper failure zone contains exposed erosional surfaces and numerous trees that were overturned or displaced by the landslide, while continuous transport pathways and depositional landforms occur across the middle and lower slopes. Mobilized material was concentrated particularly on gentler lower slopes and in pre-existing topographic depressions, locally forming deposits several metres thick. The site therefore exhibits a typical landslide-debris-flow pattern in which an upper source area, a transport zone, and a lower depositional area can be clearly distinguished.

Pre-landslide conditions were identified using aerial imagery acquired on April 27, 2024, with a spatial resolution of approximately 0.24 m, whereas post-landslide conditions were surveyed by UAV on August 14, 2025. A DJI M350 RTK equipped with a Zenmuse L2 sensor was used to acquire imagery and LiDAR data simultaneously at a flight altitude of approximately 120 m. The imagery had a ground sampling distance (GSD) of approximately 0.038 m, and ground control points were surveyed using GNSS-VRS to ensure positional accuracy. The high-density LiDAR point cloud was used to identify ground points in mountain terrain containing both trees and exposed ground and to reconstruct detailed post-landslide topography. The availability of pre- and post-landslide geospatial data makes the area suitable for quantitatively analysing elevation change, erosion and deposition extents, movement direction, and volumetric change associated with slope failure.

Fig. 1. Orthophotos of the Study Area in Sangneung Village, Sancheong-gun: (a) Pre-landslide Aerial Orthoimage Acquired on April 27, 2024, (b) Post-landslide UAV Orthoimage Acquired on August 14, 2025

2.2 Data

Heterogeneous pre- and post-event terrain datasets were collected to analyse landslide-induced topographic change. The pre-landslide dataset consisted of an aerial orthoimage acquired in April 2024 (GSD 0.24 m) and contours and spot elevations from the national 1:5,000 digital topographic map. According to the metadata, the map was produced in 2024 using 2023 aerial photography. It therefore predates the July 2025 landslide by about one year; comparison with the April 2024 orthoimage showed no notable changes in roads, buildings, or cultivated-field boundaries. The post-landslide dataset was acquired by UAV LiDAR surveying in August 2025. A LiDAR sensor mounted on a DJI M350 RTK platform collected the point cloud at a flight altitude of approximately 120 m. The mean point spacing was approximately 0.03 m, and a simultaneously acquired RGB orthoimage (GSD 0.038 m) was used to detect fallen trees. Ground control points were established by VRS surveying based on continuously operating GNSS reference stations. All datasets were transformed to a common coordinate reference system. Horizontal coordinates were standardized to KGD2002/Central Belt 2010 (EPSG:5186), and UAV LiDAR heights were converted to orthometric heights using KNGeoid18. Consistency between the vertical datum of the digital topographic map and that of the UAV LiDAR data was additionally evaluated from pre- and post-event elevation residuals over stable road sections.

2.3 Terrain Data Generation and Preprocessing

2.3.1 Pre-Landslide DTM Generation

Because no high-resolution pre-event terrain survey was available, the pre-landslide digital terrain model (DTM) was constructed from contours and spot elevations in the 1:5,000 digital topographic map. ArcGIS Topo to Raster (ANUDEM) and triangulated irregular network (TIN) methods were compared for generating a gridded DTM from the contour data. Topo to Raster produced departures from the measured contour elevations during hydrologic enforcement and yielded poorer agreement with the post-event DTM. In contrast, the TIN method more faithfully preserved the input elevations and produced smaller residuals over stable areas. The pre-landslide DTM was therefore generated using a TIN and resampled to the same 0.5 m grid as the post-landslide DTM (Fig. 2).

Fig. 2. Pre-landslide DTM Generated from 1:5,000 Digital Topographic Map Contours and Spot Elevations Using TIN Interpolation (Magenta: Source Contours and Spot Elevations)

2.3.2 Post-Landslide DTM Generation and Preprocessing

The post-landslide DTM was generated from the UAV LiDAR point cloud. Taking advantage of the high point density, with a mean spacing of approximately 0.03 m, minimum-value binning was applied at a 0.5 m interval by assigning the lowest elevation return within each grid cell as the ground elevation. This process is effective for excluding upper returns from the canopy and understory and representing the forest-floor surface. However, buildings and other non-ground objects misidentified as ground could remain as local elevation anomalies in the initial DTM. Such areas were identified through joint interpretation of the orthoimage and DTM, masked, and corrected using surrounding ground elevations.

Fig. 3(a) shows the initial 0.5 m DTM generated by minimum-value binning, together with residual buildings and non-ground objects. Areas highlighted in orange exhibit anomalously high elevations relative to the surrounding terrain or were identified as non-ground through orthoimage interpretation. If retained, these objects could be interpreted as deposition in the pre- and post-landslide DTM difference, thereby overestimating the deposited volume.

Building footprints and residual non-ground areas were therefore removed manually and converted to NoData cells. Each removed area was then interpolated and filled using the mean elevation of valid surrounding ground cells. Fig. 3(b) shows the final DTM after correction, in which local non-ground elevation anomalies were removed and a terrain surface continuous with the surroundings was restored. The final DTM was aligned to the same 0.5 m grid as the pre-landslide DTM and used to generate the DoD. Manual editing was confined to visually identifiable building footprints and residual non-ground anomalies within the ROI; unchanged terrain cells were not modified. The corrected surface was validated by joint inspection of the orthophoto, DTM, and profile continuity in Fig. 4.

Fig. 3. Post-landslide DTM Generation and Correction Using UAV LiDAR Data: (a) Initial 0.5 m DTM with Residual Buildings and Non-ground Objects Highlighted in Orange; (b) Final DTM after Removing the Residual Objects and Filling the Masked Cells Using the Mean Elevation of Neighboring Ground Cells

To verify the terrain continuity of the corrected post-landslide DTM, a profile line was defined on the orthoimage and an elevation profile was extracted along the same path (Fig. 4). The left panel shows the approximately 141 m profile path overlaid on the UAV orthoimage, and the right panel shows the terrain profile after elevations associated with buildings and forest canopy were removed and the resulting gaps interpolated from valid surrounding ground elevations. The corrected profile excludes local protrusions caused by building roofs and canopy returns and reproduces a continuous ground surface along the terrain gradient, confirming that the final DTM represents the ground terrain rather than a surface model.

The concept of minimum-value binning is illustrated in Fig. 5. Because the lowest return within each grid cell is assigned as the ground elevation, returns from the canopy and understory are naturally excluded and the forest floor is reconstructed. In contrast, mean-value binning produces elevations biased upward by canopy returns. At this site, the high-density point cloud, with a mean spacing of approximately 0.03 m, provides numerous ground returns within each 0.5 m cell, allowing the terrain surface to be reconstructed reliably by minimum-value binning without separate ground-point classification.

Minimum-value binning nevertheless introduces a potential negative bias because the lowest point in each cell is taken as the ground elevation. The magnitude of this bias is proportional to grid spacing and terrain slope. At a 0.5 m grid spacing, this effect may increase locally on steep terrain. Its magnitude was not assumed to be negligible a priori; rather, the aggregate uncertainty arising from preprocessing and terrain representation was incorporated into the empirical 0.84 m LoD derived from stable areas, and changes below this threshold were excluded. Local negative bias may nevertheless remain in very steep cells.

Fig. 4. Cross-Sectional Verification of the Corrected Post-Landslide DTM: (a) Profile Path Overlaid on the UAV Orthoimage, (b) Terrain Profile Obtained after Removing Buildings and Forest-Canopy Returns and Interpolating the Resulting Gaps from Neighboring Ground Elevations

Fig. 5. Concept of Minimum-Value Binning. The Lowest Return Within Each Cell Is Taken as the Ground Elevation, Excluding Canopy Returns. Mean-Value Binning Is Shown for Comparison

3.1 Asymmetry of the Terrain Datasets

To quantify landslide-induced terrain change, a DEM of Difference (DoD) was generated by subtracting the pre-event DTM from the post-event DTM, and erosion and deposition volumes were calculated from the resulting grid. In this study, erosion denotes areas where elevation decreased because soil and rock were removed or exported by the landslide (negative DoD), whereas deposition denotes areas where material accumulated and elevation increased (positive DoD). Because DoD values contain both actual terrain change and the errors inherent in the two DEMs, failure to distinguish between them can overestimate or underestimate change. This issue is particularly important here because the pre-landslide terrain was reconstructed from digital-map contours and spot elevations, whereas the post-landslide terrain was acquired by UAV LiDAR; consequently, the two datasets have structurally different vertical accuracies.

The procedures used to generate and preprocess the pre- and post-landslide DTMs were described in Section 2. The pre-landslide DTM was generated from digital-map contours using a TIN, whereas the post-landslide DTM was produced by minimum-value binning of the UAV LiDAR point cloud. Minimum-value binning was selected to exploit the high-density UAV LiDAR data containing ground returns from the forest floor.

3.2 Co-registration Accuracy Assessment Using Stable Areas

To determine whether a systematic vertical offset existed between the two DTMs and to establish a basis for the LoD, elevation differences were analysed over stable road areas unaffected by terrain change. The raw differences extracted from 14 road sections covering 1,487 m² ranged from -4.35 to +2.48 m and included values that were difficult to explain physically. These values were attributed to data mismatches near image boundaries and to roadside retaining walls and cut slopes and were removed using the interquartile range (IQR) method. As shown in Table 1, the resulting mean, median, standard deviation, and RMSE were -0.04, +0.11, 0.84, and 0.84 m, respectively.

After outlier removal, the mean difference of -0.04 m was small relative to the residual standard deviation of 0.84 m, indicating that a global vertical offset would not exert a dominant influence on the analysis. No uniform vertical correction was therefore applied. The residual dispersion nevertheless reflects both the limited ability of the contour-derived pre-landslide DTM to reproduce local terrain form and the effects of slope and curvature. Residuals were approximately 0.2 m on gentle slopes and increased to approximately 1 m in steep, highly curved terrain.

3.3 Error Propagation and Level of Detection (LoD)

When each DEM has an independent elevation uncertainty ($\delta$), the uncertainty of the elevation difference is combined as the root sum of squares of the two errors.

Here, $\delta_{new}$ and $\delta_{old}$ represent the independent elevation uncertainties of the post- and pre-landslide DTMs, respectively. The standard deviation of the differences observed over stable roads, 0.84 m, is an empirical residual that combines errors from both DTMs; it was therefore applied directly as a fixed LoD in the primary analysis. The 0.84 m value was selected for the primary analysis because it directly represents the observed site-specific dispersion between the two final DTMs, whereas 1.67 m was used only to test whether the volume estimates remained robust under a deliberately more conservative 95 % threshold. In a conservative comparison scenario, this value was treated as a proxy for pre-landslide DTM uncertainty ($\delta_{old} = 0.84\text{ m}$) and combined with the nominal post-landslide LiDAR DTM uncertainty ($\delta_{new} = 0.15\text{ m}$), yielding $\delta_{DoD} = 0.85\text{ m}$ and a 95 % threshold of 1.67 m. Rather than a rigorous estimate based on independently separated error components, the 95 % value was interpreted as a conservative upper threshold that accounts for the possibility that post-landslide DTM error is already partly represented in the residuals.

3.4 Region of Interest (ROI) Definition and Sediment Budget

The definition of the analysis domain has a decisive influence on the interpretation of terrain-change estimates. A debris flow is an open-system mass-transfer process in which material is released from an upper slope, transported downslope, and deposited. Both the source and depositional areas must therefore be included for the sediment budget to be interpreted physically. To evaluate the influence of domain definition, results from an initial deposition-centred region were compared with those from an expanded region (122,106 m²) that included the landslide source area.

3.5 Correction of Deposition Overestimation Caused by Fallen Trees

In parts of the post-landslide depositional area, fallen trees covered the ground and caused LiDAR returns to represent the wood surface rather than the underlying terrain. The post-landslide DTM therefore included tree thickness above the actual debris surface, leading to overestimation of deposition. Fallen-tree areas were identified by combining colour information from the RGB orthoimage with DoD values. Among pixels excluding live vegetation (green) and exposed soil and debris (brown), dark-gray pixels with deposition exceeding 2 m were classified as fallen trees (Fig. 6), and a tree thickness of 1.5 m was subtracted from their deposition values. Because this correction was applied only to the final DoD rather than to the post-landslide DTM itself, the actual debris deposited beneath the trees was preserved. The DoD > 2 m criterion restricted classification to pronounced positive anomalies, and 1.5 m was adopted as a representative wood-surface correction based on visual interpretation of the orthophoto. Because tree thickness was not measured independently, this value was treated as a practical fixed correction.

Fallen trees were detected by a rule-based classification combining RGB colour characteristics with a DoD height criterion. Because withered fallen trees appear as low-saturation dark gray, the following colour indices were calculated from normalized pixel values (R, G, B in [0,1]).

The indices were used to classify three terrain-cover types according to the following rules. Live vegetation was identified by green dominance, exposed soil and debris by red dominance and high saturation, and fallen trees by their near-achromatic appearance with low saturation and brightness.

Live vegetation: Greenness > 0.04

Exposed soil and debris: Redness > 0.05 AND S > 0.30 AND (not live vegetation)

Dark gray: S < 0.28 AND V < 0.45 AND (neither live vegetation nor exposed soil and debris)

Colour information alone caused shadows, dark leaves, and mixed pixels along debris boundaries to be misclassified as gray. To reduce these false positives, the colour rule was combined through a logical AND operation with a DoD height criterion, so that only pixels showing elevation increases greater than plausible debris deposition were confirmed as fallen trees.

Fallen tree (detected) = Dark gray AND DoD > 2 m

This combination excluded shadows and boundary pixels that appeared gray but showed no deposition, retaining only areas where wood accumulated above debris and produced a large elevation increase. This combination yielded a conservative extraction of visually identifiable fallen trees (magenta areas in Fig. 6). A tree thickness of 1.5 m was subtracted from the deposition value in each detected area to estimate the actual debris deposition.

The colour thresholds used in the classification (Greenness 0.04, Redness 0.05, saturation 0.28-0.30, and brightness 0.45) were not theoretically fixed values but were empirically adjusted for the orthoimage of the study area through comparison with visual interpretation. Because these thresholds may vary with illumination, shadows, and sensor characteristics at the time of acquisition, they cannot be transferred directly to other sites. However, the final correction combined the colour conditions with the quantitative height criterion DoD > 2 m through a logical AND operation, reducing the influence of moderate variations in the colour thresholds on the final detection. Accordingly, the correction was confined to conservatively identified fallen-tree areas, reducing its influence on the overall terrain-change volume.

Fig. 6. Fallen Trees Detected by Combining Orthoimage Colour Indices with the DoD Threshold (Magenta)

3.6 Flow-Aligned Longitudinal Profile Analysis

DoD maps and volumetric estimates summarize change over the entire analysis domain and do not directly show the depth of deformation at individual locations. Profile-based examination is also necessary to assess spatially how sensitive the estimated volume is to the selected level of detection (LoD). A longitudinal profile was therefore established along the debris-flow path, and the longitudinal distribution of elevation change was analysed.

The profile orientation was defined from the flow direction identified in the field. Comparison of the positions and orientations of individual houses in the pre- and post-landslide orthoimages indicated southeastward overturning and displacement, showing that material moved from the upper northwest toward the lower southeast. The profile was therefore aligned with this direction at an azimuth of 129 degrees and extended 252.8 m from above the source area to the lower end of the depositional area. Its spatial location is shown in Fig. 9(a).

Elevation change was extracted from the 0.5 m DoD grid at 0.5 m intervals along the profile, yielding 506 observations. Positive values represent elevation increases due to deposition, whereas negative values represent elevation decreases due to erosion.

4.1 Co-registration Accuracy and Spatial Distribution of Terrain Change

Over the stable road areas, the mean elevation difference between the two DTMs was -0.04 m, with a standard deviation and RMSE of 0.84 m. Because the mean was close to zero, no systematic vertical offset was identified, and the dispersion of 0.84 m was adopted as the LoD. Fig. 7 presents the DoD for the expanded analysis domain and shows a typical debris-flow pattern in which material was removed from the upper source area (red; negative DoD) and accumulated across the broad lower depositional area (blue; positive DoD). The figure shows the final DoD after fallen-tree correction and application of the 0.84 m LoD; its extent is the expanded ROI, with movement from the upper northwest source area toward the lower southeast depositional area.

Fig. 7. DEM of Difference between the Pre- and Post-Landslide DTMs (Red: Erosion, Blue: Deposition); Final DoD after Fallen-Tree Correction and 0.84 m LoD; Extent: Expanded ROI; Flow: NW to SE

4.2 Improvement of the Sediment Budget by Expanding the Analysis Domain

Table 3 and Fig. 8 show changes in erosion, deposition, and net volume according to ROI definition and correction stage. In the initial ROI, erosion was less than half of deposition, resulting in a net change of +84,836 m³ and a deposition-to-erosion volume ratio of 2.21. This excessive dominance of deposition occurred because the source area lay outside the analysis domain. After the domain was expanded to include the source area, erosion increased by approximately 63 %, from 69,852 to 113,964 m³, whereas deposition changed very little. Consequently, net change decreased by approximately half to +42,939 m³, and the volume ratio was adjusted to 1.38.

The greater deposition volume relative to erosion can be explained by the physical behaviour of debris flows. As compacted ground is disturbed during detachment and transport, the material expands in volume through bulking, so the deposited volume may exceed the eroded volume even at equal mass. A volume ratio of approximately 1.4 in the expanded ROI is physically plausible when both bulking and the export of some material downstream are considered ^{(Rengers et al., 2021)}. However, because neither the exact bulking ratio nor the volume exported beyond the domain was measured independently, the volume ratio alone was not used to claim closure of the sediment budget.

4.3 Fallen-Tree Correction and Final Terrain-Change Volumes

Correcting deposition elevations by 1.5 m within the detected fallen-tree areas reduced the net change to +41,033 m³. After applying the final LoD of 0.84 m to exclude minor changes within the data-error range, erosion, deposition, and net change were estimated as 105,391, 148,522, and +43,131 m³, respectively, with a volume ratio of 1.41. Net change remained stable at +41,000 to +44,000 m³ as the LoD varied from 0 to 1.0 m (Table 4). Even under the 95 % probabilistic threshold of 1.67 m derived from error propagation, erosion, deposition, and net change were 85,088, 135,521, and +50,432 m³, respectively. Thus, the estimates showed limited sensitivity within the tested LoD range.

Fig. 8. Erosion, Deposition and Net Change by Region of Interest and Correction Stage

4.4 Longitudinal Distribution of Elevation Change Along the Flow Direction

Fig. 9(a) shows the profile location and flow direction overlaid on the post-landslide orthoimage, and Fig. 9(b) shows the longitudinal distribution of elevation change extracted along the profile. Section characteristics are summarized in Table 5. The profile is clearly divided into a short initial depositional section, a middle-to-upper erosional section, and a lower depositional section.

From the profile origin to 15.5 m, elevation increased, with a mean deposition thickness of 1.65 m and a maximum of 3.60 m. Examination of this section in the post-landslide orthoimage showed that it lay outside the upper boundary of the source area, where material released from upslope was temporarily retained during the initial stage of movement. The elevation increase was therefore interpreted as temporary material accumulation rather than residual pre-failure terrain.

Between 15.5 and 163.7 m, a 148.2 m section was dominated by erosion. Mean erosion depth was 4.16 m, and the maximum erosion depth of 9.38 m occurred 79.1 m from the profile origin. Erosion depth gradually decreased in both the upstream and downstream directions from this maximum, indicating that the largest failure occurred near this location.

From 163.7 m to the profile endpoint, elevation increased over an 89.1 m section, indicating deposition. Mean deposition thickness was 4.61 m, with a maximum of 7.53 m at 233.8 m. In contrast to the erosional section, deposition thickness tended to increase downstream, indicating that transported material was retained and accumulated in the lower part of the flow path.

The absolute magnitude of elevation change is noteworthy. Mean changes in the erosional and lower depositional sections were 4.16 and 4.61 m, respectively, approximately five times the adopted LoD of 0.84 m. The maximum change of 9.38 m exceeded the LoD by more than eleven times and was 5.6 times the 95 % probabilistic threshold of 1.67 m. These magnitudes indicate that metre-scale erosion and deposition remained detectable within the tested LoD range. However, smaller or spatially localized changes remain uncertain because of the contour-derived DTM.

Although the erosional section of the profile (148.2 m) was longer than the lower depositional section (89.1 m) and their mean changes were similar, deposition over the full analysis domain was approximately 1.4 times erosion. The profile was intentionally positioned through the centre of the source area and therefore emphasizes erosion. When the analysis is extended across the full domain, the lateral transport zone and the broad depositional area are included, resulting in deposition dominance. The profile and volumetric results are thus not contradictory; rather, they are complementary indicators of local deformation magnitude and the domain-wide sediment budget, respectively.

Fig. 9. Location and Elevation-Change Distribution of the Longitudinal Profile Aligned with the Flow Direction: (a) Profile Location Overlaid on the Post-Landslide Orthoimage; (b) Elevation-Change Distribution Along the Profile (Azimuth 129 Degrees, Total Length 252.8 m). Red Denotes Erosion, Blue Denotes Deposition, and the Dashed Lines Indicate the Level of Detection of Plus or Minus 0.84 m

4.5 Interpretation and Applicability of the Results

The contour- and spot-elevation-derived pre-landslide DTM used in this study has greater uncertainty than the UAV LiDAR-derived DTM. An LoD of 0.84 m estimated over stable areas was therefore applied to exclude minor changes within the data-error range. Accordingly, the proposed method is more appropriate for estimating terrain change associated with large landslides having erosion and deposition depths of several metres than for detecting subtle surface displacement.

The fallen-tree correction was applied only to areas clearly identifiable in the post-landslide orthoimage, and material transported beyond the analysis domain was not included in the volume estimates. The estimated erosion, deposition, and net volumes should therefore be interpreted in relation to the defined analysis domain and the fallen-tree interpretation criteria. Unlike repeated high-resolution surveys, this study combines a legacy contour-derived DTM with UAV LiDAR and applies a fixed empirical LoD; spatially variable uncertainty was not assessed.