3.2. Data Acquisition and Instruments
The studied site was surveyed six times from UAV-mounted mass data collection tools. Two systems were utilized during the study period: a DJI Matrice 300 RTK mounted with a Zenmuse L1 LiDAR, and a MAVIC 3E.
The DJI Matrice 300 RTK equipped with a Zenmuse L1 LiDAR sensor generates centimeter-level positioning accuracy and high-density point clouds with integrated RGB information [
23]. The Zenmuse L1 integrates a high-precision Inertial Measurement Unit (IMU), enabling detailed three-dimensional surface reconstruction under GNSS-RTK conditions[
24]. A D-RTK station was deployed near the survey site during topographical surveys with the Matrice 300.
High-resolution photogrammetric surveys were carried out using a DJI Mavic 3 Enterprise RTK, generating dense point clouds through structure-from-motion processing during selected survey epochs [
24].
The stockpile was surveyed thrice per year, in spring, summer and autumn. Spring and autumn surveys were the topographical models of interests for the purpose of stability CDA. Summer surveys were collected for redundancy in the case where the fall survey could not be achieved on time before sudden snow fall. Flight missions were conducted at altitudes of approximately 60 m above ground level with at least 70% lateral and longitudinal overlap for photogrammetry data collection and 30% overlap for LiDAR mapping.
Mission planning and initial processing were performed using DJI Terra and Pix4D, while advanced point-cloud processing was carried out in CloudCompare V 2.13 [
5,
25,
26]. Custom Python scripts were used to support automated processing workflows, numerical analysis, and the extraction of derived point-cloud attributes. Scripted processing are further detailed in the results sections.
Table 1 present a summary of all models surveyed with relevant specifications, timestamps and achieved resolutions.
3.3. Noise Removal and Sub-Zoning
Vegetation was removed using the
CANUPO multiscale dimensionality classifier implemented in CloudCompare [
27]. This method distinguishes natural surfaces based on their local geometric dimensionality computed across multiple spatial scales, enabling effective separation of the planar quarry ground from the volumetric structure of vegetation. The classifier was trained using representative samples of ground and vegetation points selected from the same datasets to capture the typical geometric variability of each class. Once trained, the model was applied consistently to all epochs at each site, ensuring reproducible filtering through time. Points identified as vegetation were excluded from the datasets, leaving only the ground surface for subsequent registration and change-detection analyses.
The studied slope of the stockpile was subdivided in three separate study area to avoid color scale saturation with localized high CDA results and facilitate computer processing time.
Figure 7 illustrates the subdivision of the study area. This segmentation was performed prior to vegetation filtering to reduce computation time and increase the precision of subsequent analyses, particularly during registration and change detection. Each zone was processed independently following the same classification and analysis workflow to ensure consistency across the entire site.
Vegetation filtering was performed after subdividing the study area to ensure that only the ground surface was retained for registration and change detection analyses. For each site, the CANUPO classifier was trained once using representative samples from the initial dataset and then applied consistently to all subsequent epochs to maintain classification uniformity over time.
Representative training samples were manually selected to define the two target classes, vegetation and bare soil, from areas visually identified in the point cloud and orthophoto to capture the full range of surface geometries across the pile. The classifier was trained using 20000 core points per class, and dimensionality descriptors were computed over multiple scales ranging from 0.05 m to 1.0 m, with a step of 0.05 m.
The training results showed well-separated class distributions with minimal overlap between vegetation and soil clusters. The balanced accuracy (BA) reached approximately 0.94–0.95, and the Fisher Discriminant Ratio (Fdr) exceeded 4.5, confirming a strong discriminative performance. The same classifier parameters were then applied to all subzones and epochs to ensure comparability through time.
Figure 8a–c illustrates an example of the vegetation-filtering process using the CANUPO classifier.
Figure 8a shows the initial point cloud containing both vegetation and ground points,
Figure 8b displays the vegetation class, and
Figure 8c presents the ground class retained for subsequent registration and change detection analyses.
3.4. Scan Placement
Consecutive datasets were co-registered within a common coordinate reference system using the Recursive Iterative Closest Point (R-ICP) method proposed by [
21]. This method extends the conventional ICP framework described by [
28] by introducing a recursive bias filtering process designed to improve alignment robustness in complex and partially overlapping three-dimensional datasets. Conventional ICP placement aligns two point clouds by iteratively minimizing the distance between corresponding points. The quality of ICP placement is affected by occlusions and changes within the selected placement sub-set. R-ICP integrates a recursive filtering step to reliably exclude occlusions and changes in the landscape and capture a robust representation of placement quality.
R-ICP is initialized by coarse placement derived from sensor-based orientation and translation parameters. ICP placement is then conducted using in this work CloudCompare V2.13. The resulting absolute C2C distances are computed and statistically analyzed. A perfectly random C2C distribution will feature a symmetrical curve, while occlusions, non-overlapping zones and changes in the landscape introduce asymmetrical bias to the distribution. A filtering threshold at each iteration is defined based on a confidence interval (CI) computed from the C2C distance distribution (here set at 95%). Points exceeding the upper limit of this CI are considered outliers and filtered out. ICP and filter loops are repeated until a suitable error is reached or CI stops decreasing.
Figure 9 presents an example of the evolution of the C2C distance distribution curves through successive R-ICP iterations for Zone 1 of
Figure 8 between the November 2024 and May 2025 datasets. At each iteration, after identifying and removing non-corresponding regions from the previous step, a new ICP adjustment was performed to reposition the compared scan relative to the reference scan. The updated alignment was reassessed using C2C distance measurements, followed by another filtering stage to exclude remaining non-overlapping areas and occlusions. During the first iteration, the C2C distribution was highly skewed toward larger values, indicating the presence of significant non-corresponding regions between the two epochs. As the recursive filtering and realignment progressed, the distributions became narrower and more symmetric, reflecting the gradual refinement of point correspondences and the convergence of the alignment toward an optimal solution.
Figure 10 illustrates the spatial evolution of the C2C absolute distance maps throughout the successive R-ICP iterations for the same example zone. The color scale represents the magnitude of point-to-point distances between the reference and compared scans, with warmer colors indicating higher discrepancies. At the first iteration, large, misaligned regions are visible, particularly along the upper and lateral parts of the pile. As the recursive alignment proceeded, these high-distance zones gradually diminished and the overall color distribution became more uniform, reflecting improved spatial correspondence between the two epochs. By the final iteration, most of the compared surface exhibited low C2C values, confirming the effectiveness of the recursive approach in achieving precise alignment prior to change detection analysis.
The statistical evolution of the alignment process is plotted in
Figure 11. The figure plots the variation of mean, standard deviation, and 95 percent confidence interval (CI) of C2C distances over successive R-ICP iterations for the same example zone. The progressive decrease and subsequent stabilization of these parameters after approximately five iterations indicate convergence of the recursive alignment procedure. This confirms that the iterative filtering and ICP adjustments effectively minimized residual misalignments, resulting in an optimal registration suitable for reliable change-detection analyses.
At the end of the recursive alignment process, the final transformation matrix of the compared scan (recorded explicitly by CloudCompare in the user interface) incorporates all translation and rotation adjustments accumulated throughout the iterations. This transformation was then applied to the original, unfiltered point cloud to ensure that the complete dataset was accurately aligned with the reference scan. Once all subzones from successive surveys were registered within a common coordinate system, the datasets were ready for the subsequent step of change detection analysis.
R-ICP algorithm was applied to all consecutive data sets to ensure consistent and precise scan placement before change detection analysis. The resulting transformations minimized spatial bias and guaranteed that the detected variations between successive datasets represented genuine topographic evolution rather than residual registration error.
3.5. Level of Detection
The Level of Detection (LoD), expressed in meters (m), was determined following the formulation presented in [
21], which defines the smallest measurable elevation difference that can be considered statistically significant for a given confidence interval (see also [
6,
29]). This approach accounts for uncertainties associated with instrument precision, scan placement, beam incidence angle, and local surface roughness
The Level of Detection between two epochs
A and
B is defined as:
where
regA,B, expressed in meters (m), is the registration error between the two epochs,
represent error components for scans and
represents the confidence interval. The LoD provides an estimate of the individual point change precision resulting from the combined instrumental and alignment errors. It is empirically validated by evaluating the variability of change values within stable zones where no deformation was detected.
For two registered point clouds, the registration uncertainty at a selected confidence interval
CI is given by:
Where F(CI) is the confidence factor derived from the standard normal distribution (for example, 1.96 for 95 percent confidence) and
σC2C (m) is the standard deviation of the cloud-to-cloud distance distribution, which reflects the spread of registration residuals between the two datasets.
The combined measurement uncertainty of a given data set is expressed as:
In the case of RTK assisted drone-based acquisitions, either using LiDAR or photogrammetry,
σC2C can be considered negligible (≈ 0), as individual scans within each dataset are internally aligned through automated processing workflows and do not require manual scan registration. In this study, scan registration is only performed between different time steps, and therefore the associated uncertainty is accounted for separately in the inter-epoch comparison rather than within a single dataset.
The parameter
σₜₒₒₗ (m) represents the intrinsic repeatability of the sensor’s measurement system and captures the baseline instrument-specific uncertainty contributing to the overall
LOD. For the
LiDAR sensor (Zenmuse
L1),
σₜₒₒₗ corresponds to the laser ranging precision, which reflects the repeatability of individual distance measurements along the laser beam. At an effective measurement distance of approximately 100 m, this precision is typically on the order of 2–3 cm (
RMS,
1σ) under standard operating conditions [
24].
For the photogrammetric system (DJI Mavic 3 Enterprise with RTK),
σₜₒₒₗ relates to the RTK-assisted camera positioning accuracy, which governs the repeatability of the camera’s exterior orientation parameters. Manufacturer specifications indicate RTK fix positioning accuracy of approximately 1 cm + 1 ppm horizontally and 1.5 cm + 1 ppm vertically [
30]. In addition to this georeferencing uncertainty, the spatial resolution of the derived point cloud also contributes to the measurement uncertainty. In this study, the average point spacing of the photogrammetric point cloud is approximately 3 cm. Therefore,
σₜₒₒₗ for photogrammetry is approximated as the combination of RTK positioning accuracy and point cloud resolution, resulting in a value of approximately 4.5 cm.
The beam-footprint error (
σB) m represents the distortion of the laser footprint induced by the incidence angle between the incoming laser beam and the reflective surface, (see [
6,
21,
31]), where the footprint elongation is modeled as a function of range, beam divergence, and incidence geometry. The Zenmuse
L1 employs the Livox Avia LiDAR module, which has an asymmetric elliptical beam divergence of 0.28° (horizontal) and 0.03° (vertical) as reported in the manufacturer specifications [
24]. Assuming an elliptical beam geometry, these values correspond to an equivalent average angular radius of approximately 0.09° from the center of the beam to its boundary, resulting in an average full angular divergence of approximately 0.18°. This representative value was adopted for
σB to account for the anisotropic nature of the beam and varying incidence conditions across the slope and along oblique flight paths, providing a conservative estimate of beam-footprint-induced uncertainty.
Figure 12 illustrates the influence of surface geometry and acquisition conditions on
σB, where panel (a) shows the spatial distribution of surface dip, and panel (b) demonstrates that
σB increases with both distance and incidence angle due to footprint elongation under oblique viewing conditions. Under the survey geometry considered in this study, with a maximum range of approximately 50 m and an incidence angle of about 45°,
σB remains limited to approximately 6 cm. This value therefore represents a conservative upper bound on the geometric component of LiDAR measurement uncertainty for the acquisition conditions considered.
For photogrammetry, the geometric beam-footprint term used in LiDAR does not apply because photogrammetric 3
D reconstruction is derived from image triangulation rather than active ranging. In this case, the dominant measurement uncertainty arises from pixel-level feature localization, represented by
σPM (m). Automated tie-point extraction typically exhibits a precision of approximately 0.5 to 2 pixels depending on surface texture, illumination, and viewing geometry, as reported in multiple UAV photogrammetry studies [
32,
33,
34]. This image-space uncertainty propagates directly into object space through the Ground Sampling Distance (GSD), such that the photogrammetric measurement error can be approximated by the widely used relationship
σPM = k · GSD, where
k is an empirical factor accounting for image correlation performance and ranges between 1.0 and 3 for high-quality aerial imagery [
35,
36]. Based on the flight altitude and camera specifications of the DJI Mavic 3E, the ground sampling distance (GSD) was calculated as approximately 2 cm/pixel. Assuming a conservative value of
k = 3, the resulting pixel-matching uncertainty was taken as
σPM ≈ 0.06.
can therefore be computed for photogrammetry data sets by replacing
σB by
σPM in equation (3). The corresponding uncertainty components are summarized in
Table 2.
To quantify
σC2C, a representative stable area was selected to evaluate registration residuals using cloud-to-cloud (C2C) distance analysis. The standard deviation of C2C distances was computed for all pairs of survey epochs and used to estimate
σC2C as summarized in
Table 3.
The Level of Detection was then determined for each dataset by combining the three sensor-specific uncertainty components described in the preceding section. For LiDAR datasets, the resulting uncertainty was further reduced by accounting for the number of repeated measurements (N = 3) acquired with the Zenmuse L1, such that ECI was divided by N. This yields a Level of Detection on the order of 7–11 cm, depending on the specific time frame considered. For epoch pairs including at least one photogrammetry dataset, the corresponding LoD is higher, on the order of 20 cm, reflecting the combined effects of lower geometric precision and additional uncertainty associated with photogrammetric reconstruction.
Surface changes smaller than the computed LoD were considered indistinguishable from noise and were therefore excluded from interpretation. This filtering ensured that only statistically meaningful elevation differences were retained, improving the reliability of subsequent change detection and precursor identification.
3.6. Chande Detection Analysis
Change detection analysis was performed to quantify topographic variations between consecutive survey epochs for the studied pile. The analysis was carried out between each time step with respect to the previous one, allowing the identification of progressive surface evolution through time.
For each pair of successive epochs, the shortest distance between corresponding points was calculated, generating a dense field of elevation differences across the pile. This method enabled the identification of localized erosion and accumulation zones and provided a clear initial overview of the magnitude and spatial distribution of surface changes. The C2C-derived distance maps were then interpreted to delineate zones of material loss and gain, as well as subtler deformation features that may indicate emerging instability. A lower bound for detectable changes was imposed based on the Level of Detection (LoD) determined for each epoch pair (see previous section), such that variations below this threshold were considered insignificant and attributed to measurement noise.
It should also be noted that the density and resolution of the compared point clouds may influence the apparent extent of low-magnitude detected changes represented by blue color. In some comparisons, the aligned scan has a finer effective resolution than the reference scan. As a result, more small-scale distance variations may be detected and displayed on the compared surface. Although a resolution homogenization step was applied to reduce this effect and bring the compared datasets to a similar spatial resolution, minor differences in point spacing and surface sampling may still remain. These residual differences can generate scattered low-magnitude changes, many of which are likely related to measurement noise or sampling effects rather than meaningful surface deformation. However, this limitation does not significantly affect the main objective of the analysis, which is to identify reliable zones of actual surface change. In the context of instability monitoring, the priority is to detect true positive deformation zones, while isolated low-magnitude detections should be interpreted cautiously and not overemphasized.
The C2C absolute distance results shown in
Figure 13,
Figure 14,
Figure 15,
Figure 16 and
Figure 17. capture the progressive surface evolution of the pile from fall 2022 to spring 2025. The comparison between fall 2022 and spring 2023 (
Figure 13) corresponds to the main failure event at the site and exhibits the largest magnitude of surface change, with absolute C2C distances reaching up to approximately 3 m. Significant displacements are concentrated in the central section of the pile, clearly delineating the primary failure zone, while adjacent areas remain comparatively stable.
Between spring 2023 and fall 2023 (
Figure 14), the dominant measured C2C distances reflect localized material accumulation at the upper part of the slope in Zone 2 which is associated with quarry operational activities and does not indicate slope deformation or failure-related movement.
Subsequently, between fall 2023 and spring 2024 (
Figure 15), localized secondary movements are detected within the previously identified failure zone in the central section of the pile (Zone 2). These changes are spatially confined and of relatively low magnitude (up to approximately 1.32 m), indicating continued surface adjustment and internal reorganization within the main failure area rather than renewed large-scale instability.
The spring to fall 2024 comparison (
Figure 16) shows a deformation pattern, with localized movements occurring primarily in Zone 1. Displacement magnitudes in this zone reach up to approximately 2 m, while the central failure zone exhibits limited additional change during this interval. The observed movements remain spatially restricted and do not propagate across the entire slope.
The most recent comparison between fall 2024 and spring 2025 (
Figure 17) indicates a further reduction in deformation intensity. Only small, localized surface changes are detected along the failure zone, with measured displacements not exceedingly approximately 1.19 m. This trend suggests a progressive attenuation of surface activity following the main failure, although localized deformation remains detectable within the previously affected zones.