The application of Unmanned Aerial Vehicle (UAV) photogrammetry in geohazard investigation represents a significant technological leap. While existing literature often demonstrates the capability of UAV drones for rapid data acquisition and three-dimensional (3D) modeling of landslide bodies, systematic and in-depth research focusing on fine-grained deformation analysis and the revelation of internal mechanisms remains relatively scarce. Many studies are confined to technical demonstrations or macroscopic change detection, lacking a comprehensive, multi-parameter analytical framework to decipher the subtle yet critical signs of slope instability. This research gap hinders the transition of UAV drone technology from a powerful mapping tool to an indispensable analytical system for understanding landslide kinematics and triggering factors. Therefore, this study aims to conduct a meticulous, short-to-medium-term deformation monitoring campaign on a specific, active landslide body. By leveraging the high spatial and temporal resolution data provided by UAV drones, we employ a Structure-from-Motion (SfM) based workflow to process multi-temporal imagery. The analysis extends beyond simple visual comparison, delving into seven distinct but interrelated aspects: distance change, crack propagation, volumetric calculation, slope gradient, aspect orientation, cross-sectional profile evolution, and correlation with precipitation. This multi-faceted approach is designed to achieve a fine-grained understanding of the landslide’s deformation behavior, thereby validating the efficacy of UAV drone photogrammetry in supporting detailed geomechanical analysis and early-warning-oriented monitoring.
The core of modern photogrammetric processing from UAV drone imagery is the SfM-MVS (Multi-View Stereo) pipeline. Unlike traditional photogrammetry that requires known exterior orientation parameters, SfM automatically solves for camera positions and the sparse 3D structure of the scene by identifying and matching common feature points across overlapping images. The general workflow can be summarized as follows:
- Feature Detection & Description: Key points (e.g., SIFT, SURF) are extracted from each image.
- Image Matching: Feature descriptors are matched across all image pairs to find correspondences.
- Sparse Bundle Adjustment (SBA): This step simultaneously refines the 3D coordinates of the feature points, camera positions, and orientations, as well as intrinsic camera parameters (focal length, principal point, lens distortion), by minimizing the total reprojection error. The optimization can be represented as:
$$
\min_{\mathbf{P}_i, \mathbf{X}_j} \sum_{i=1}^{m} \sum_{j=1}^{n} v_{ij} \, d(\mathbf{x}_{ij}, \mathbf{P}_i \mathbf{X}_j)^2
$$
where $ \mathbf{P}_i $ is the projection matrix for camera $ i $, $ \mathbf{X}_j $ is the 3D coordinate of point $ j $, $ \mathbf{x}_{ij} $ is the observed 2D image coordinate, $ d() $ is a distance function (e.g., Euclidean), and $ v_{ij} $ is a binary variable indicating visibility. - Dense Cloud Generation (MVS): Using the estimated camera geometry, dense matching algorithms (e.g., Patch-based MVS) compute depth maps for each image, which are fused into a dense point cloud.
- Mesh & Texture Generation: The dense point cloud is triangulated to create a 3D mesh model, which is then draped with image textures to produce a photo-realistic 3D reconstruction.
For geospatial accuracy, Ground Control Points (GCPs) are essential. They are surveyed with high-precision GNSS (e.g., RTK) and used to scale, rotate, and translate the SfM-derived model into a real-world coordinate system. The error model for a GCP is:
$$
\sigma_{total}^2 = \sigma_{GNSS}^2 + \sigma_{identification}^2 + \sigma_{SfM\_model}^2
$$
where $ \sigma_{GNSS} $ is the surveying error, $ \sigma_{identification} $ is the error in marking the GCP in the imagery, and $ \sigma_{SfM\_model} $ is the internal consistency error of the SfM cloud at that point.
The study area is a small, active landslide body exhibiting clear signs of retrogressive erosion and tension cracking. The local geology consists of unconsolidated Quaternary deposits, primarily silty clay with gravel, overlying weathered bedrock. The climate is characterized by distinct wet and dry seasons, with intense rainfall concentrated during the monsoon period. A DJI Phantom 4 Pro V2.0 UAV drone was deployed for data acquisition. This platform is equipped with a 1-inch 20-megapixel CMOS sensor and a mechanical shutter, minimizing motion blur—a critical feature for high-accuracy photogrammetry. Four survey missions were conducted between June 2023 and June 2025, strategically timed to capture conditions at the onset and end of rainy seasons, as well as during the dry period. Flight planning ensured high overlap (80% frontlap, 70% sidelap) for robust 3D reconstruction. The specifications of the data collection campaigns are summarized below:
| Survey Epoch | Acquisition Date | Number of Images | Approximate Ground Sampling Distance (GSD) | Primary Weather Condition |
|---|---|---|---|---|
| Epoch 1 (E1) | 2023-06-11 | 348 | ~2.5 cm | Post-rain, moist ground |
| Epoch 2 (E2) | 2023-11-04 | 151 | ~3.5 cm | End of rainy season |
| Epoch 3 (E3) | 2025-04-04 | 185 | ~3.0 cm | Dry season |
| Epoch 4 (E4) | 2025-06-30 | 216 | ~3.0 cm | Early rainy season |
The raw imagery from the UAV drone was processed using a combination of specialized software to leverage their respective strengths. ContextCapture (Bentley Systems) was primarily used for generating high-fidelity, textured 3D mesh models and Digital Surface Models (DSMs), owing to its superior mesh quality and stability in measurement tools within its viewer. Agisoft Metashape was employed for the efficient generation of Digital Elevation Models (DEMs) and Digital Orthophoto Maps (DOMs), which served as the base data for slope, aspect, and profile analyses. The integrated processing workflow is illustrated below.
Data Processing & Analysis Workflow:
- Image Alignment & Sparse Cloud Generation (SfM): All images from an epoch were aligned using tie points.
- Georeferencing: The sparse cloud was scaled and oriented using surveyed GCPs.
- Dense Cloud & Mesh Generation: A dense point cloud was built and converted into a 3D mesh.
- Product Generation: High-resolution textured 3D models, DSMs (from ContextCapture), and DEMs/DOMs (from Metashape) were exported.
- Multi-temporal Analysis: The 3D models and raster products from all four epochs were compared and analyzed quantitatively.
Results and Multi-Dimensional Analysis
1. Distance-to-Failure Scarp Analysis
Using the integrated measurement tool in ContextCapture’s viewer, the horizontal distance from a stable feature point (e.g., a building corner) to the distinct head scarp of the landslide was measured in the 3D models of E1, E2, and E3. A consistent measurement protocol was applied to ensure comparability. The results show a clear reduction in distance, indicating the retrogressive nature of the landslide, where the scarp is moving backwards (upslope).
- E1 Distance: 16.58 m
- E2 Distance: 13.03 m (Change ΔE1-E2 = -3.55 m)
- E3 Distance: 11.93 m (Change ΔE2-E3 = -1.10 m)
The most significant retreat occurred between E1 and E2, coinciding with the peak rainy season. The continued retreat between E2 and E3, albeit smaller, confirms ongoing instability even during drier periods.
2. Crack Propagation Mapping
The high-resolution 3D models revealed the development and propagation of tension cracks near the landslide crown. In E1, only minor, incipient cracking was visible along the scarp. By E2, a major continuous tensile crack had formed, approximately 40 meters in length with an average width of 24 cm. This feature was precisely digitized from the orthophoto and mapped onto the DEM. The formation of such a well-defined crack is a critical precursor to further failure, as it indicates the development of a discrete failure plane and the loss of tensile strength in the soil mass. The ability of UAV drone photogrammetry to detect and quantify these millimeter-to-centimeter scale features non-invasively is a key advantage over conventional survey methods.
3. Volumetric Change Assessment
Volumetric change was calculated using the “volume by mean plane” method within the 3D analysis software. A consistent polygonal boundary, delineating the main depletion zone (source area) of the landslide, was used across all four epochs. The software calculates the volume between the 3D surface of the landslide and a reference plane defined by the boundary points. The results, detailed in the table below, show complex dynamics of erosion and deposition.
| Epoch | Surface Area (m²) | Cut Volume (m³) | Fill Volume (m³) | Net Volume Change ΔV (m³)* | Primary Process Inferred |
|---|---|---|---|---|---|
| E1 | 5,047.3 | 6,875.6 | 6,766.7 | Baseline | – |
| E2 | 4,938.8 | 6,914.5 | 7,735.8 | +969.1 (Net deposition) | Significant erosion from scarp followed by deposition in lower parts. |
| E3 | 5,035.6 | 7,110.3 | 7,339.4 | -396.8 (Net erosion) | Minor scarp retreat and surface erosion, less deposition. |
| E4 | 5,056.6 | 7,580.9 | 6,301.8 | -1,037.6 (Net erosion) | Major erosive event, likely triggered by early rains. |
*Net Volume Change ΔV = (Fillepoch_n – Fillepoch_{n-1}) – (Cutepoch_n – Cutepoch_{n-1}). A positive value indicates net deposition within the bounded area, a negative value indicates net erosion.
The volumetric analysis reveals that the landslide body is not simply losing material; it undergoes cycles of erosion from the head and sidewalls, transport, and temporary deposition within the source area itself. The net erosion from E3 to E4 is particularly noteworthy and aligns with increased slope angles observed in other analyses.
4. Slope Gradient and Aspect Analysis
DEMs from each epoch were analyzed in a GIS environment to calculate slope gradient and aspect. The slope histogram consistently showed that the most prevalent slope angle within the actively failing region fell within the interval of 31° to 45°. This range is widely recognized in geomorphology as highly susceptible to failure in cohesive soils. More importantly, the analysis indicated a slight but consistent shift of the dominant slope class towards steeper angles over time, as summarized below:
| Epoch | Dominant Slope Class Interval | Class Width | Percentage of Landslide Area in Class |
|---|---|---|---|
| E1 | 31.65° – 39.73° | 8.08° | ~9% |
| E2 | 33.78° – 42.21° | 8.43° | ~9.3% |
| E3 | 34.11° – 42.89° | 8.78° | ~9.6% |
| E4 | 35.90° – 45.04° | 9.14° | ~10.2% |
The aspect analysis revealed that the landslide facets predominantly face East and Southeast. This orientation is significant as it influences solar insolation, drying rates, and potentially the direction of subsurface water flow, all of which can affect shear strength. After the major deformation event between E1 and E2, the aspect distribution showed a stronger concentration towards the East (67.5° – 112.5°), suggesting a directional bias in the failure mechanism, possibly related to geological structure or cross-slope hydrological gradients.
5. Cross-Sectional Profile Evolution
To investigate internal deformation patterns, longitudinal profiles (A-A’, B-B’, C-C’) parallel to the slide direction and one transverse profile (D-D’) were extracted from the DEM series. The longitudinal profiles vividly illustrate the progressive steepening of the head scarp and the development of a steeper main body. For instance, the slope angle of the scarp in profile A-A’ increased from 76° (E1) to 128° (E4), a cumulative increase of 52°. The transverse profile D-D’ showed relative stability in the central part of the landslide but indicated more pronounced lowering (erosion) along the lateral margins. This pattern suggests that the failure is not a simple planar slide but involves more complex processes, including lateral spreading and preferential erosion along pre-existing weaknesses or gullies. The deformation rate, calculated from profile change, was highest between E1 and E2 (wet period) and continued, though at a reduced rate, between E3 and E4 (dry-to-wet transition). The rate of crest retreat $ v_r $ can be estimated as:
$$
v_r = \frac{\Delta d}{\Delta t}
$$
where $ \Delta d $ is the horizontal retreat distance and $ \Delta t $ is the time between epochs.
6. Correlation with Precipitation
The timing of the surveys was explicitly designed to investigate the hydro-meteorological forcing on landslide activity. Historical climate data and recorded precipitation during the monitoring period were analyzed. The region receives over 70% of its annual rainfall between May and November. The largest deformations—significant scarp retreat (3.55 m), major crack formation, and substantial volumetric change—were all recorded between E1 (June, start of rains) and E2 (November, end of rains). The deformation continued in the subsequent dry and early wet periods (E2 to E4), as evidenced by ongoing scarp retreat and net erosion, but at comparatively lower magnitudes. This pattern strongly suggests that intense or cumulative rainfall is the primary triggering and accelerating factor, potentially through mechanisms like:
- Rise in pore-water pressure reducing effective stress: $ \tau = c’ + (\sigma – u) \tan\phi’ $, where $ u $ is pore pressure.
- Increase in soil unit weight (surcharge).
- Surface erosion and infiltration softening the soil.
However, the persistence of deformation during drier months implies the presence of other contributing factors, such as progressive strain-softening of the soil, internal seepage with delayed groundwater response, or anthropogenic activities (e.g., uncontrolled surface water runoff from upslope). The UAV drone time series effectively captures this coupled climatic-geomechanical response.
Discussion and Conclusion
This study demonstrates that UAV drone photogrammetry, underpinned by robust SfM algorithms, transcends its role as a mere topographic mapping tool. It establishes itself as a comprehensive platform for fine-grained, multi-parameter landslide deformation monitoring. The integrated analysis of distance, cracks, volume, slope, aspect, profiles, and environmental context provides a holistic view of the landslide’s behavior. Key findings include the confirmation of retrogressive failure, the quantification of pre-failure crack development, the identification of cyclic erosion-deposition dynamics within the source area, and the clear correlation between deformation rates and rainfall intensity.
The research underscores several critical advantages of using UAV drones for this application:
- High Resolution & Safety: Ability to capture centimeter-scale features without requiring personnel to access unstable ground.
- Comprehensive Data: Generation of directly measurable 3D models, DEMs, and orthophotos from a single survey.
- Temporal Flexibility: Enables cost-effective, on-demand monitoring at critical times (e.g., pre- and post-rainy season).
- Mechanistic Insight: The rich dataset allows for the analysis of deformation patterns (e.g., lateral vs. central erosion) that inform hypotheses about the underlying failure mechanism.
Future work should focus on integrating data from UAV drones with other sensors, such as terrestrial or airborne LiDAR, for enhanced penetration through vegetation and comparison of surface vs. terrain models. Furthermore, coupling the high-resolution geometrical data from UAV drones with numerical slope stability models (e.g., finite element method) would be a powerful next step to back-analyze material properties and simulate future failure scenarios under different rainfall conditions. The automation of change detection algorithms and the establishment of threshold-based early warning systems using UAV drone derived parameters (like crack width or scarp retreat rate) represent promising directions for transforming this technology into an operational tool for landslide risk management.