The stability of riverbanks composed of sensitive clay presents a significant geotechnical challenge. These materials are prone to highly retrogressive, flow-type landslides. The failure mechanism typically initiates with a localized disturbance at the toe of the slope, followed by a progressive retrogressive failure surface that expands upslope. The strong strain-softening and brittle behavior of sensitive clays, coupled with their low remolded strength, complicates the identification of hazardous slopes and can dramatically increase the area affected by a landslide event. The advancement of digital photogrammetry and low-altitude remote sensing platforms has revolutionized the monitoring of surface processes, including landslides. It is now feasible to acquire centimeter-level accuracy, high-resolution imagery and distributed point cloud data over large areas. This technological progress provides a viable method for measuring pre-failure deformations in sensitive clay slopes, which is crucial for early warning. Therefore, this study employs a combined methodology of Unmanned Aerial Vehicle (UAV) photogrammetry and a geodetic network monitored by a total station to analyze a specific case study, leveraging the capabilities of UAV drones for high-temporal-resolution data capture.
1. Introduction and Site Characterization
The study focuses on a riverbank slope historically subject to intermittent retrogressive failures. While the specific landslide behavior has regional peculiarities and does not represent all sensitive clay landslides, its episodic nature provides a unique advantage for monitoring. The relatively long intervals between successive failure stages increase the likelihood of capturing morphological changes using a field monitoring program. The core study area was selected based on a synthesis of existing displacement data, visual signs of slope deformation, and unfavorable stability conditions. At this location, the bank is approximately 11 meters high with an average inclination of about 18.4°. A distinct scarp, up to 1.2 meters high, is present at the crest. Fluvial erosion at the toe has reached the top of a lower, weak clay layer, promoting instability along this near-horizontal plane of low strength.
To establish a high-precision reference framework, a monitoring network of 108 Ground Control Points (GCPs) was installed in September 2019. The network was arranged along five parallel transects spaced approximately 5 meters apart, concentrating on the slope and the plateau area behind the crest. Each GCP consisted of a 1-meter long, 10M threaded steel rod driven about 0.9 meters into the ground, ensuring visibility above low grassland vegetation. A steel cap was mounted on top. This design positioned the entire GCP within the frost depth, aiming for the rod to experience similar freeze-thaw cycles as the surrounding soil. The cap served a dual purpose: enhancing safety and improving visual recognition in imagery due to its 60mm diameter.
2. UAV Data Acquisition and Photogrammetric Processing
2.1 Flight Planning and Image Capture
Image acquisition throughout the study period was conducted using a GPS-guided octocopter UAV drone with a diameter of 1 meter. The platform operated in autonomous flight mode following pre-defined waypoints. A two-person team managed operations. The UAV was equipped with a motorized gimbal integrated into its central chassis, allowing for precise camera angle control and effective stabilization against aerial platform movements during flight. The mission utilized a calibrated camera to capture image sequences along planned flight lines. Achieving complete coverage of the landslide area required specific forward overlap within a flight line and side overlap between adjacent parallel lines. Systematic flights by UAV drones are essential for generating consistent time-series data.
2.2 Photogrammetric Processing Pipeline
The processing of UAV imagery is a systematic, algorithm-driven workflow transforming 2D image sequences into high-precision 3D geographic information. The pipeline can be delineated into four consecutive technical stages.
Stage 1: Image Preprocessing and Feature Enhancement. Processing begins with image quality optimization. UAV imagery is susceptible to variable illumination and lens distortion. Radiometric correction is applied using adaptive histogram equalization to adjust local contrast in over- or under-exposed regions. The core operation can be expressed as:
$$I_{out}(x, y) = \frac{T(I_{in}(x, y))}{N_{max}} \cdot 255$$
where \( T() \) is a gray-level mapping function based on sub-region histograms and \( N_{max} \) is the maximum allowable gray level. Geometric correction is performed using the Brown-Conrady model, iteratively solving for radial distortion coefficients \( k_1, k_2 \) and tangential distortion coefficients \( P_1, P_2 \) to calibrate pixel coordinates, with typical residuals controlled within 0.3 pixels. This preprocessing can improve the Signal-to-Noise Ratio (SNR) by up to 12 dB.
Stage 2: Dense Matching and Point Cloud Generation. Following sparse point cloud generation, an enhanced Semi-Global Matching (SGM) algorithm is used for sub-pixel disparity calculation. A Census transform is incorporated to improve the matching cost function:
$$C(p, d) = \sum_{q \in W_p} \text{XOR}(\text{Census}(q), \text{Census}(q_d))$$
where \( W_p \) is a (9×9) pixel window and \( q_d \) is the corresponding point in the right image offset by disparity \( d \). A multi-baseline matching strategy is employed to enhance reconstruction in vegetated areas, yielding a final point cloud density of 300–500 points/m² with a planimetric accuracy better than 3 cm.
Stage 3: Digital Surface Model (DSM) Refinement. Point cloud post-processing employs a multi-level filtering strategy. First, Statistical Outlier Removal (SOR) eliminates noise (standard deviation multiplier \( k=2.0 \)). Second, an improved Cloth Simulation Filter (CSF) separates ground from non-ground points, with a rigidity parameter of 0.65. Surface modeling uses Moving Least Squares (MLS) for fitting:
$$\min \sum_{i=1}^{n} \theta(\|x – x_i\|) \cdot (f(x) – z_i)^2$$
where \( \theta(\|x – x_i\|) = e^{-\|x – x_i\|^2 / h^2} \) is a Gaussian weight function with bandwidth \( h=0.3 \) m controlling weight decay, \( f(x) \) is the fitted surface function, and \( z_i \) is the actual elevation at point \( x_i \). The resulting DSM has a grid resolution of 5 cm with a vertical RMSE ≤ 8 cm relative to GCPs.
Stage 4: 3D Modeling and Multi-Resolution Texture Mapping. A watertight mesh is generated using Poisson surface reconstruction (depth parameter \( d=10 \)), preserving terrain features at a 0.1 m scale. Texture mapping employs a view-dependent weighted blending technique, calculating visibility weight for each triangular face in each image:
$$w_i = \frac{\cos \theta_i}{\|X – C_i\|^2}$$
where \( \theta_i \) is the angle between the camera line-of-sight and the face normal, and \( C_i \) is the camera center coordinate. The final textured model supports Levels of Detail (LOD), automatically switching to 2K texture maps at a 50 m viewing distance, improving storage efficiency by 40%.
2.3 Integrated Survey Analysis and Accuracy Assessment
The extensive GCP network served two primary functions: continuous model accuracy assessment and providing redundancy against anticipated high GCP loss during landslide activity. Seven GCPs were designated as control points for bundle adjustment, while the remainder served as check points. Accuracy was evaluated by comparing total station-surveyed coordinates with photogrammetrically derived coordinates. The initial model (November 2019) showed image space Root Mean Square Error (RMSE) of 0.23 pixels, leading to spatial RMSE errors of 8 mm (X), 9 mm (Y), and 14 mm (Z). The use of steel-capped GCPs, however, introduced challenges. Three damage mechanisms were observed: 1) pedestrian impact driving the rod through the cap, 2) animal chewing deforming the cap’s ring, and 3) partial fracture of internal locating lugs within the cap—a defect difficult to detect visually but introducing systematic positioning error. The accuracy metrics for all epochs are summarized below.
| Epoch | # Control Points | # Check Points | Estimated Accuracy (mm) | # Images | RMSEX (px) | RMSEY (px) |
|---|---|---|---|---|---|---|
| 2019/11 | 7 | 73 | 5 | 18 | 0.2285 | 0.2203 |
| 2021/11 | 7 | 77 | 5 | 22 | 0.2228 | 0.2420 |
| 2022/04 | 7 | 55 | 5 | 22 | 0.2272 | 0.2116 |
| 2022/11 | 7 | 61 | 4 | 46 | 0.2312 | 0.2634 |
| 2023/04 | 7 | 33 | 4 | 33 | 0.1886 | 0.2237 |
| Epoch | RMSEX (mm) | RMSEY (mm) | RMSEZ (mm) | Total RMSE (mm) |
|---|---|---|---|---|
| 2019/11 | 8 | 9 | 14 | 18 |
| 2021/11 | 13 | 12 | 24 | 30 |
| 2022/04 | 7 | 12 | 23 | 27 |
| 2022/11 | 13 | 22 | 13 | 29 |
| 2023/04 | 7 | 8 | 19 | 22 |
3. Results: Deformation Analysis from UAV-Derived Data
3.1 Differential Digital Elevation Models (DoDs)
The analysis of change was primarily conducted through Difference of Digital elevation models (DoDs). In these visualizations, mid-gray tones indicate no elevation change, dark gray to black represents lowering (erosion/subsidence), and light gray to white indicates uplift (deposition/heave). The 3D representation of the slope geometry clearly shows the existing scarp forming the downslope boundary of the flat plateau area. The magnitude of scarp offset diminishes towards increasing X-coordinates, indicating that the left portion of the study area (smaller X) experienced greater downslope displacement than the right.
Notably, visual observations indicated no significant change in slope geometry between 2019 and 2020. Consequently, elevation differences in the initial monitoring period were minimal and concentrated in the no-change range, demonstrating that vegetation clearance prior to UAV drone flights effectively controlled related errors. The DoD for the period encompassing the 2022 landslide event reveals dramatic changes: up to 2 meters of uplift (white tones) from accumulated debris at the toe, and a maximum subsidence of approximately -4 meters (black tones) in the landslide scar. The textured area indicates channel reformation. While DoDs effectively identify zones of deformation, 3D displacement vector fields provide a more precise quantitative description of the pre-failure process.
3.2 Surface Displacement Vector Field and Kinematic Interpretation
Analysis of surface displacement profiles along cross-slope transects further reveals the kinematics of failure. A target transect crossing the approximate center of the two subsequent flow events serves as an indicator for the upslope propagation of the failure surface beyond the initial crestal scarp. Incremental displacement profiles characterizing three main stages of slope behavior were constructed from the UAV drone photogrammetry models aligned via the stable GCP network.
Stage 1: Incipient Deformation (Pre-2022). Over a two-year period with no visible macroscopic change, displacement magnitudes were minimal but measurable. A near-horizontal displacement component at the toe indicated sliding of a wedge along the lower clay layer, representing initial localized toe failure.
Stage 2: Post-2022 Failure. Following the first major flow slide, GCPs located immediately downslope of the crestal scarp showed continued displacement along the inferred failure direction. The displacement vectors indicated a coherent block movement.
Stage 3: Post-2023 Failure. After the second event, which exposed a new scarp further back, displacements in the plateau area behind the original crest confirmed that the progressive failure surface had extended deeper into the slope mass. The cumulative horizontal displacement of points on the plateau showed an initially near-linear increase over time, which can be modeled as:
$$D_h(t) = v \cdot t + D_0$$
where \( D_h(t) \) is the cumulative horizontal displacement, \( v \) is the average displacement rate, \( t \) is time, and \( D_0 \) is an initial offset. This linear trend characterizes the tensile strain generated as the slope mass translates along the near-horizontal basal shear surface.
4. Discussion: UAV Photogrammetry as an Early Warning Tool
The five-year monitoring program demonstrated the high potential of repeat UAV drone surveys coupled with a robust ground control network for detecting pre-failure deformation in sensitive clay slopes. The key finding is the measurable, gradual accumulation of displacement directed towards the crestal scarp that preceded the 2022 and 2023 flow-type failure events by months to years. The detection of this precursor deformation hinged on the high georeferencing accuracy achieved through the GCP network, despite the challenges posed by GCP damage.
The observed displacement patterns provide direct insight into the mechanics of progressive failure. The near-linear accumulation of horizontal displacement on the plateau behind the crest is a critical signature. It represents the surface expression of deep-seated shear strain along the weak layer. Monitoring this rate, \( v \), could be a quantitative early-warning indicator. An acceleration in this rate, potentially following a power-law or exponential trend as failure approaches, could be described by:
$$v(t) = A \cdot (t_f – t)^{-n}$$
where \( v(t) \) is the instantaneous velocity, \( A \) and \( n \) are constants, and \( t_f \) is the time of failure. While such a clear acceleration phase was not fully captured in this dataset prior to the major collapses, the established linear creep phase is itself a vital warning sign of active, deep-seated sliding.
The study also underscores the complementary nature of different data products. DoDs offer an excellent overview of volumetric change after an event, but 3D displacement vectors derived from multi-epoch point cloud comparison are superior for quantifying subtle pre-failure movements. The integration of frequent UAV drone surveys (e.g., monthly or seasonal) with real-time kinematic (RTK) positioning or PPK capabilities could further enhance accuracy and reduce dependency on physical GCPs, which are vulnerable in active landslide settings.
5. Conclusion
This research implemented and evaluated a methodology based on UAV photogrammetry and a total station-monitored GCP network to assess pre-failure displacement in a sensitive clay landslide. The five-year dataset revealed that significant flow-type failures in 2022 and 2023 were preceded by a slow, cumulative deformation directed towards the slope crest. Monitoring points located on the plateau behind the crest exhibited progressive, initially near-linear downslope displacement. This linear horizontal displacement trend is interpreted as the surface manifestation of tensile strain resulting from translation along a sub-horizontal basal shear zone within the lower sensitive clay layer. The synthesis of displacement vector fields indicates that by the end of the monitoring period in 2024, the progressive failure surface had retrogressed beyond the original crestal scarp, leaving the slope at risk of further retrogression. The study conclusively demonstrates that high-frequency, high-accuracy monitoring using UAV drones can capture precursor movements to large slope failures, providing a scientifically grounded and technically feasible approach for early-warning systems in areas prone to sensitive clay landslides. Future work should focus on automating processing pipelines and establishing velocity thresholds that trigger different levels of alert.