Introduction

Geohazard monitoring technologies continue to develop, allowing users to increase the frequency, repeatability, accuracy, and extent of displacement measurements (Hendry et al. 2015; Carlà et al. 2017a; Macciotta et al. 2017; Smethurst et al. 2017; Williams et al. 2017). Remote sensing methods are now implemented as part of routine geohazard monitoring programs (Carlà et al. 2017b; Kromer et al. 2017; Riquelme et al. 2017; Cucchiaro et al. 2018; Rodriguez and Hendry 2018). Methods include technologies such as Interferometric Synthetic Aperture Radar (InSAR), Light Detection and Ranging (LiDAR), and unmanned aerial vehicle (UAV) photogrammetry. These methods have been applied on a variety of geohazards affecting cut slopes (Sousa et al. 2016; Beregovoi et al. 2017; Vanneschi et al. 2017) and natural slopes (Petschko et al. 2016; Fey and Wichmann 2017; Roque et al. 2018). These technologies allow collection of data over large areas with high spatial resolution, as opposed to only at discrete locations (e.g., slope inclinometers, survey monuments). The use of UAVs to assess geohazard kinematics is becoming more common around the world; examples of this can be found from southeast Spain (Agüera-Vega et al. 2018; Martínez-Carricondo et al. 2018), China (Fan et al. 2017; Huang et al. 2018), Italy (Mazzanti et al. 2017; Allasia et al. 2019), and North America (Al-Rawabdeh et al. 2017; Solazzo et al. 2018).

UAVs provide multiple perspectives of the geohazard from photographic or video records. Each photographic record from the UAV is stored with geotags (coordinates and orientation information) that allows for georeferentiation of a point cloud obtained from digital photogrammetry reconstruction algorithms. The point clouds can result in highly detailed topographic information to enhance conventional investigation methods. Current UAVs and UAV software, such as Pix4Dcapture by Pix4D S.A. (2018a) or DroneDeploy by DroneDeploy Inc. (f.k.a. Infatics Inc. 2020), allow defining automatic flight patterns to collect information easily and quickly and at a relatively low cost from a safe distance. This encourages more frequent data collection which can be very useful for rapidly evolving site conditions. The increased monitoring frequency enables the collected data to provide earlier warning of changing conditions, assessment of triggering factors, and integration into a geotechnical asset management program.

This paper presents a case study illustrating the use of UAVs and digital photogrammetry to monitor an unstable slope adjacent to a section of a transportation corridor. The study site is a 90-m-high, 520-m-wide slope located in Highway 837 adjacent to the south bank of the Red Deer River approximately 10 km northwest of the town of Drumheller, Alberta (Canada). The section of interest had a history of rockfall and rockslide events that have impacted the highway at the toe of the slope. The analysis presented in this paper corresponds to two events that occurred in December 2017 and May 2018. The use of digital photogrammetry provided detailed information to assess the rockfalls and debris flow failure mechanisms observed at the site. The collection of data before and after one of the recorded failures allowed insight into the stability of sensitive areas on the steep slope. A change detection analysis on the point clouds derived from digital photogrammetry indicated sources of failure and the magnitude of the event. Since erosion processes from precipitation and surface water runoff had been associated with the occurrence of instabilities at the site (Klohn Crippen Berger 2000), a digital elevation model (DEM) derived from the point cloud was utilized to investigate the drainage network of the slope. Lastly, the point clouds and the analysis of the source of failure provided information to conduct rockfall analysis in 2D (although 3D analysis would also be possible). The results presented in this paper show the effectiveness of UAVs to enhance the understanding of these instabilities, monitor, and aid in decision-making about the optimal mitigation at landslide-prone areas.

Alberta Transportation’s C018 rock slope

The study site is in a region of southeastern Alberta known as the Badlands. The Badlands are an arid to semiarid region with average temperatures ranging from − 18.6 °C in the winter to 26.7 °C in the summer, and precipitation averaging 370 mm of combined snow and rainfall per year (Government of Canada 2018). The landscape along the Red Deer River has little vegetation cover and is highly eroded and weathered. Highway 837 carries an average of 340 vehicles per day (Alberta Transportation 2017), connecting local communities between Drumheller and Bleriot Ferry and serving as a tourist route, which provides access to a large paleontology museum, the Midland Provincial Park, and the Badlands area. This 6-km section of the highway was constructed adjacent to steep valley slopes (1:1 or steeper) up to 60 m high. A 500-m section of these slopes directly adjacent to the highway (Fig. 1) has a history of erosion on riverbank slopes, rockfalls, rockslides, and debris flows that have blocked traffic, which is known by the provincial ministry of transportation (Alberta Transportation [AT]) as geohazard site C018 (as denoted in the AT Geohazard Risk Management Program). AT has been required to increase the frequency of interventions at this site over the past 20 years to ensure serviceability and safety.

Fig. 1
figure 1

Plan view of the site located next to Highway 837 near Drumheller, Alberta. The red dash line shows the extent of the study site. Black dash lines show the three current active zones in the slope (F1, F2, and F3)

Full size image

The stratigraphy of the site is part of the Upper Cretaceous Horseshoe Canyon Formation from the Edmonton Group (Fig. 2), which is characterized by interbedded sedimentary rocks mainly composed of fine-grained sandstone, bentonitic mudstone, and carbonaceous mudstone, with coal seams and bentonite beds (Prior et al. 2013). The terrain in this region is the result of glaciation deposition followed by erosion from meltwater from creeks and the Red Deer River (Borneuf 1972; Stalker 1973). The exposed bedding planes are visible in the valley slopes due to the variations of color, texture, and slope (Fig. 2); these structures have an approximately sub-horizontal inclination (Allan 1921) of 10°. The Horseshoe Formation is weak to extremely weak bedrock with a UCS of less than 5 MPa and GSI of 20 to 25 (Marinos and Hoek 2000). Previous investigations in the area describe the material near the site as highly plastic clay, overlaying sandstones, and shales (Klohn Crippen Berger 2000). These geologic characteristics cause the Highway 837 to be susceptible to different instability processes along the riverbank slopes and the slopes above the highway during precipitation events.

Fig. 2
figure 2

Slopes exposing the Horseshoe Formation in the vicinity of the study site

Full size image

Weathering of exposed bedrock and surficial material exhibits a soft soil–like behavior, with discontinuous blocks of more indurated bedrock which remains as intact blocks (Klohn Crippen Berger 2018a). Rock layers composed mainly of sand-size particles present a higher susceptibility to the weathering process; however, these layers appeared to be mixed with fine soil particles. The mixture of sand-size and fine soil particles allows the erosion process to vary throughout the slope. Analyses on moisture susceptibility from different materials in the slope resulted in a deterioration of the material structure (i.e., dispersive material). The bonding of coarse-grained particles decays during the saturation process, changing their behavior to a weaker rock by destroying its internal structure. The degradation of the materials in the presence of water further suggests swelling potential, a likely key factor for the release of rock blocks into the highway.

During the summer of 2017, two large debris flows of approximately 1300 m3 occurred in the northern section of the study site. These two large events were preceded by rainfall events of approximately 30 mm and 8 mm in 24 h, 2 days before failure. Later, on December 18, 2017, approximately 120 m3 of blocks was estimated to have been released from 5 to 30 m above the highway in a similar location to the previous events (Klohn Crippen Berger 2018a). Further site inspections revealed that during winter 2017–2018, clumps of frozen soil and bedrock had fallen (Klohn Crippen Berger 2018b). During this freezing event, rock blocks were released from the slope, bounce to the highway, and reached the frozen surface of Red Deer River shore (Fig. 3a).

Fig. 3
figure 3

Slope failures at the C018 slope: a rock blocks found on the frozen Red Deer River in December 2017; b rock blocks found across the highway in May 2018; c accumulation rockfalls and debris slide behind Jersey barrier in November 2018

Full size image

On May 23, 2018, a large rockfall and debris flow at this location again blocked the highway. The event was preceded by a 15-mm rainfall event in 48 h that occurred 6 days prior to failure. This latest slide exposed a large rock outcrop approximately 40 m high and 50 m wide above the highway, on the northern side of the slope (Klohn Crippen Berger 2018a). After the later event, a variety of rock block sizes were found on the roadbed, from 4 mm gravel size to cobbles of 0.5 m in equivalent diameter (Fig. 3b). Some of these larger particles appear to disaggregate while falling or upon impact with the roadbed. Following the May 2018 event, AT partially closed one lane section and temporarily installed concrete barriers to stop material from reaching the road. The barriers were Jersey barriers approximately 80 cm in height (Fig. 3c). The barrier extended for 90 m along the highway.

According to Cruden and VanDine’s (2013) classification system, the study site appears to show a complex failure mode that is weather-dependent. During spring, summer, and fall, the erosion process triggers rockfalls. In addition, preceding precipitation events weaken the slope materials and increase the susceptibility for debris flows (Fig. 3c). During the winter, the materials harden, and rockfalls become the predominant mode of failure.

Application of UAV technology at the study site—methods

The study site required a rapid and accurate assessment of the slope instabilities to reduce the risk to the public road users. However, the location and geometry of the site imposed a challenge to observe the location and conditions of the slope high above the highway. UAVs are equipped with a digital camera to capture photos along a flight path. The application UAV technology at the study site allowed capturing images of the slope not achievable from the highway. UAV data used in this paper was provided by AT (December 2017 and May 2018, after the slope failure events). Moreover, a third survey was done in November 2018 by the authors to continue monitoring the slope. The sequence of surveys allowed assessing the progressive erosion process that is deteriorating the slope surface through change detection techniques (comparing the surface of successive UAV surveys allows quantifying the changes of the surface over time). The information collected from the three UAV surveys was analyzed following six steps, as shown in Fig. 4. This methodology derived from processes was recommended by Pix4D (Pix4D S.A. 2018b) and LiDAR processing (Lague et al. 2013; Esposito et al. 2017).

Fig. 4
figure 4

Methodology followed with UAV data collected on three surveys at the study site

Full size image

This section describes the fundamentals of the first five steps for the implementation of UAVs at the study site.

Capturing UAV photos

The use of digital photogrammetric algorithms requires a set of pictures with enough overlap to reconstruct the terrain and to produce a point cloud with a resolution that captures the scale of significant features. This is achieved by allowing a minimum overlap of 60% between images, collecting at least three photos per feature, and using high-resolution cameras. The UAVs used in the surveys had a 12-MP camera for the first and third flight and a 17-MP camera for the second flight. The 12-MP camera had a sensor 1/2.3″ CMOS (6.3 mm width and 4.7 mm height), a 94° field of view, a 20 mm focal length, and aperture f 1:2.8. The 17-MP camera had a sensor 1/2.3″ CMOS (6.3 mm width and 4.7 mm height), an 82° field of view, a 44 mm focal length, and aperture f 1:3.3. The photos were stabilized using a gimbal attached to the camera to reduce the vibrations.

UAV systems are a quick and easy tool to deploy for surveying slope instabilities with difficult access. The UAVs were equipped with an internal GPS+GLONASS dual positioning module to allow precise control of the location during the flight. The internal GPS has a hovering accuracy of ± 1.5 m. Different software allows creating automatic flight plans to optimize collecting photos using UAVs; for this analysis, the flight path was created using Pix4DCapture (Pix4D S.A. 2018a). An example of the flight path and photo locations is available as a supplementary material to this paper. The automatic survey is established for different drone parameters such as (i) grid pattern, (ii) flight height, (iii) flight speed, and (iv) camera angle. These parameters facilitate the collection of multiple photo records covering the study site.

The first two photogrammetry surveys were captured using a step stair sequence (different elevation) with a ground base elevation between 50 and 90 m from the highway, with over 600 photos collected per survey. The third survey was performed at a single elevation of approximately 50 m measured from the crest of the slope, with 900 photos. The resulting ground sampling distance (GSD) was between 1 and 3 cm per pixel. The UAVs used on the surveys are light quadcopters (less than 1.5 kg) that can reach speeds up to 20 m/s. Although high speeds reduce the flight time and battery consumption, the speed was kept below 7.2 m/s to avoid capturing blurry images during the survey. Lastly, the camera angle can provide oblique images to capture more details on very steep slopes. Capturing photos at oblique angles allows adequate coverage of steep slopes. Care should be taken to avoid angles and orientations that result in direct sunlight on the lens that may result in the overexposure of the images. This can be mitigated through proper flight planning (time of the day and flight path). The photos from the surveys were captured at an angle between 0 and 10° from the vertical. During the surveys, the internal GPS and gimbal provide the location and orientation of each photo taken during the survey (also known as geotags). The geotag collected on each photo allows having local precision within a few meters. This information allows determining the position, orientation, and scale of the capture objects to generate a point cloud in each survey. Nevertheless, the low precision of the internal UAV GPS required having ground control points (GCPs) whose coordinates are measured.

The use of GCPs allows optimizing the absolute accuracy of the point cloud georeferentiation obtained from the geotags; this resulted in a georeferentiation with root mean square error between 0.01 and 0.06 m. The coordinates of the GCP were measured with an RTK GPS (traditional survey methods). The site conditions limited the distribution of the GCPs due to the steepness of the slope. The GCPs were distributed along the highway and crest of the slope. The first two surveys had 5 GCPs near the highway with a precision of ± 2 cm in the horizontal direction and ± 2 cm in the vertical direction, while the last survey used 10 GCPs near the highway and the crest of the slope with a precision of ± 10 cm in the horizontal direction and ± 10 cm in the vertical direction. Finally, a visual inspection of the captured photos from the site was made to remove low-quality photos reducing errors in the point cloud reconstruction (as suggested by the software manufacturers such as Pix4D and Adam Technology).

Digital photogrammetric reconstruction

The photos from each survey are used to generate point clouds. The point cloud requires having a resolution that captures the surficial details to allow analyzing the changes of relevant features. The events that occurred prior to the UAV surveys are over 10 m wide, with a depth between 30 cm and 2.0 m. The process to reconstruct the topography at the study site using UAV photos was fully automated with commercial software (Pix4Dmapper Pro, Pix4D S.A. 2018b). This software uses structure from motion (SFM) algorithms (Westoby et al. 2012), in combination with the camera correction process, photographic georeferencing, and private algorithm processes to find common points (key points) between photos. The key points have known coordinates in X-, Y-, and Z-axis that allow reconstructing the topographic ground surface through point clouds (Küng et al. 2012a). Pix4D mapper divides the reconstruction process of the point cloud in two steps; the first step consists in the calibration of camera internal and external parameters. The internal camera parameters are the temperature, vibrations, focal length, and shutter mode (Vautherin et al. 2016) to reduce distortion and noise in the photos. The external parameters are the orientation and position of the photos (Vallet et al. 2011). Next, sample key points are automatically generated in each photo and matched to other overlapping photos. At the study site, between 40,000 and 60,000 points were extracted per photo and matched between 17 and 23% of points per photo. The coordinates of the key points are calibrated based on the GCPs. The calibration required manual selection of key points on 8 to 15 photos per GCP. The calibration using GCP improves georeferencing and scale of the point cloud. The relative difference between the initial position and the corrected position from the three surveys was between 2 and 6 cm. The second step is a densification of the point cloud based on the key points from the first step. Pix4D mapper pro uses key points as input to apply a clustering multi-view stereo algorithm (CMVS) and patch-based multi-video stereo algorithm (PMVS) to densify the point cloud (Furukawa and Ponce 2009; Küng et al. 2012b). The size of the original photo was reduced to one half of its original size to increase computational efficiency while maintaining an optimal point cloud density that reflects the roughness different features (Pix4D S.A. 2018b). Reducing the scale allowed computing one 3D point every 8 pixels. The validity of each 3D point was defined by having 3 or more re-projections into the photos. In the third survey, 5 or more re-projections were required due to the homogeneity of the ground surface of the farm field above the slope. Details of point cloud processing of each UAV survey are presented in Table 1.

Table 1 Hardware characteristic, processing time, and points from densification
Full size table

Following the densification of each point cloud, outlier points were filtered without altering the overall roughness of the point cloud. The filter was done using the open-source software CloudCompare V2.9 (CloudCompare 2011). Outlier points are identified as those outside the confidence interval defined by the average distance to a common point of 10 neighboring points plus “n” standard deviations of the distance. This was achieved in a two-step process using the Statistical Outlier Removal (SOR) tool in the software. The first step removed points outside with “n” of 3 standard deviations, followed by “n” of 2. This process ensures a precise outlier filter.

Change detection

Change detection analyses using point clouds require accurate alignment between clouds from different surveys. The UAV surveys had different precisions associated with the GSD, the geotags and GCPs that translate into point cloud precision. Variability in the location of the points is referred to as registration error. Aligning the point clouds reduces this registration error. The fine alignment of the point clouds is achieved by using known stable areas outside of the area of analysis. This process was performed using the first survey as a reference for alignment for the other two. The alignment areas were selected to the north and south sides of the study site, as shown in Fig. 1. The fine alignment iterates to create a transformation matrix of 10,000 points randomly sampled until a maximum root mean square error (RMSE) of 1.0e−5 m is achieved. The final transformation matrix is applied to the entire point cloud to proceed with the cloud-to-cloud comparison. The alignment and cloud-to-cloud comparison were computed using CloudCompare V2.9 (CloudCompare 2011).

There are multiple methods available to evaluate the differences between point clouds (Lague et al. 2013; Kromer et al. 2015). The two methods used in this paper are as follows: the closest point method (C2C) to provide a first estimate with lower computational requirements (Girardeau-Montaut et al. 2005); and the multi-scale model-to-model cloud comparison (M3C2) based on a more robust statistical analysis (Lague et al. 2013). This methodology is based on Esposito et al. (2017) who also used C2C to quantify the registration error and M3C2 to assess landslides on a coastal region in Italy. The registration error was quantified by selecting three zones outside the study site and assumed to be unaltered.

The C2C method computes change as the absolute distance from a point on the first point cloud to the nearest point on the subsequent point cloud; this method is limited as it is affected by the point cloud roughness (Girardeau-Montaut et al. 2005). The reader should be aware that the roughness is affected by the density of points in the point cloud. Consequently, variation in density would induce errors in the implementation of the C2C method. The computation of the distance using the C2C method was computed using a quadratic regression on the reference cloud for points within a 0.3-m radius.

The M3C2 method allows computing the total displacement of two point clouds in a positive (material gain) or negative (material loss) direction. The direction of the displacement is quantified based on the calculation of normal vector on the reference cloud (i.e., December 2017 survey) using the software’s triangulation model. The total displacement is measured by selecting a point cloud as a reference and calculating the average distance of each point to a second point cloud. The points selected from the second point cloud are those points located within a circular projection from the reference point. In the analysis, a circular projection of 0.3 m in diameter was used. The orientation of the projection from the reference point is with respect to the normal vector of the original cloud, calculated within the 0.3-m-diameter circle.

Also, the algorithm in the M3C2 method computes the level of detection (minimum distance change that can be measured) within a 95% confidence interval by accounting the roughness of the reference point cloud (also calculated within the algorithm) and the registration error from the cloud determined using the C2C method. Details of the procedure can be found in Lague et al. (2013).

Drainage network

Site observations have shown that slope instability events at the study site have been preceded by precipitation events. In addition, the slope maps from the third survey showed that the farm land (Fig. 1) above the slope drains towards the study site. An analysis of the drainage network (watershed) of the ground surface allows determining the preferential path runoff water takes from the top of the slope to the highway. The paths allow assessing the relationship between runoff water and the slope instability at the site. In addition, this analysis allows quantifying the contributing area for water to flow through the preferential paths of the slope surface. The drainage network was calculated using ArcGIS Desktop V10.5 (Environmental Systems Research Institute Inc. 2016).

The drainage network was estimated based on the changes in elevation of the ground surface. The process proposed by Jenson and Domingue (1988) was used to derive the accumulated flow path of a DEM generated from the third survey. The resolution achieved with the UAV point clouds provides the opportunity to analyze the influence of preferential drainage paths (watershed) on the active zones at a small scale. Then, the point cloud was transformed into a DEM with a cell size of 20 cm by 20 cm using a rasterize tool in CloudCompare V2.9 (CloudCompare 2011) from the May 2018 point cloud. The direction of flow is calculated in each cell by comparing the elevation difference with the surrounding eight cells. Next, the cumulative flow is calculated for each cell by summing the number of contributing cells. The cumulative flow is transformed into a contributing area by multiplying the quantity of contributing cells on each cell by the cell size. The result is a dense drainage network that accounts for any difference elevation at the scale size. The main drainage lines were determined by filtering the drainage network. The filter was applied to remove drainage lines with less than 125 m2 of contributing area. The minimum value accounts for 0.78% of the total contributing area in the DEM. This process has shown to be scale-dependent (Tarboton et al. 1991); a larger extension of the DEM could increase the values of the contributing area.

Rockfall trajectory modeling

Using the surface reconstruction from the UAV, 2D rockfall simulations were used to estimate the run-out distance, bouncing height, and kinetic energy generated by falling, bouncing, and rolling blocks. The rockfall hazard was predominantly located on the northern active zone of the study site. The rockfall simulations were conducted using Rocscience’s RocFall (Rocscience Inc. 2018), using a Monte Carlo–based method to derive a statistical distribution from 1000 rockfall events. RocFall simulates the rocks as solid blocks with parabolic trajectories and notes the modeled location impacts, subsequent rebounds, and rolling. Energy losses during impacts are modeled using coefficients of restitution which range from 0 if all of the kinetic energy is lost and 1 if no energy is lost.

The trajectory of rockfalls from the rock outcrop on the northern side of the study site (Fig. 1, F1) was modeled using the cross-section A-A. The source of rockfalls has a length of 50 m equal to 10% of the entire section of the study site. The cross section was generated from the point cloud of the second UAV survey following the rockfall and debris flow. The rock outcrop exposed after the rockslide became a concern to the highway due to the susceptibility of eroding the base layer and releasing other large blocks into the highway. A 2.5D rockfall trajectory simulation was conducted to evaluate if the geometry of the slope results in rockfall trajectories that deviate significantly from those represented from a 2D analysis. A lower density DEM with cell size of 1 m by 1 m from the May 2018 point cloud was generated for the 2.5D analysis. Rockfall Analyst (Lan et al. 2007, 2010), an add-on for ArcGIS Desktop V10.5, was used for modeling the 2.5D rockfall trajectories (Macciotta and Martin 2019). The analysis used a lumped mass approach, which models the blocks as points without considering the shape and volume of the blocks. The rockfall seeders (location of rockfall initiation) were located at the crest of the slope at the elevation of material observed to result in rockfalls which consists of stronger materials overlying erosion features. The higher potential corresponded to the layer of stronger materials overlying eroded areas. Initial velocity was set at 0.1 m/s.

Initial input parameters for the models in Rockfall Analyst and RocFall were estimated based on field measurements and observation of the predominant size of blocks and location after falling. The initiation of the rockfall trajectory models in RocFall was distributed between 32 and 46 m upslope from the highway, consistent with the observed rockfall source areas on site. The analysis assumes an initial horizontal velocity of 0.1 m/s, no initial vertical and angular velocity at three starting points within the observable source area. This reflects the observations on site where blocks tend to reach the road and approximate the river. Also, these low initial velocities reflect what is expected at the initial movements of the fall after the block detaches from the slope. The size of the blocks assumed a 0.5-m-diameter block with a mass of 896 kg released 1 m above ground. The rockfall modeling followed the rigid-body methodology that accounts for the shape of the block. The shape was simulated as a cube with rounded edges and specified in the software model as a “superellipse of the fourth order.” The modeling was not capable of accounting for the disaggregation of the blocks and was made with the largest sizes found on the highway.

Typical coefficients of energy restitution for the slope materials (talus, uniform debris, sedimentary rock, and asphalt) were used initially and then modified such that the simulated rockfall fit with the distribution of material observed on the highway (Pfeiffer and Bowen 1989; Giani 1992; Hoek 2018). The final coefficient of restitution parameters are presented in Table 2 and are similar to values from the literature for similar areas (Pfeiffer and Bowen 1989; Giani 1992; Hoek 2018).

Table 2 Input surface parameters used in the 2D rockfall simulations
Full size table

Results

Digital photogrammetric reconstruction

The use of digital photogrammetric techniques had several sources of errors from the collected data and post-processing methods. The resulting point cloud density of the first survey was an average of 60 3D points per cubic meter; the second survey resulted in an average density of 100 3D points per cubic meter, and the third resulted in 280 3D points per cubic meter. The height and slope of the study area resulted in a lower resolution at the bottom of the slope. This variability can result in a lower accuracy near the toe of the slope. An example of the changes in GSD in cross section from the study for a flight path 50 m above the crest of the slope used in November 2018 is shown in Fig. 5.

Fig. 5
figure 5

Changes in GSD from a single flight path along the slope surface in the November UAV flight

Full size image

The change in accuracy from the UAV surveys allowed reaching an average GSD between 2.0 and 6.0 cm/pixel, being highest at the base of the slope. The point cloud scale and global position were improved using different GCP targets on the area. The RMSE of the GCP coordinates with respect to the 3D point cloud reconstruction are 0.02 m, 0.01 m, and 0.06 m for December 2017, May 2018, and November 2018 respectively.

Fine alignment of the two 2018 point clouds to the reference point cloud from 2017 allowed reducing the relatively local and global uncertainty of the point cloud with respect to each other. The alignment of 10,000 random points at each side of the slope and three registration zones with a maximum RMSE of 1.0e−5 m resulted in a maximum distance frequency of 0.05 m for May and November 2018. The registration error for May and November point cloud resulted in 0.08 m and 0.13 m respectively, based on the alignment of the cloud and registration zone analysis. This local variability is also increased by vegetation cover, snow cover, and the slope deterioration between each point cloud. The removal of points associated with the small coverage of snow and vegetation, predominantly located on the upper section of the slope, was not considered in the point cloud. However, the SOR filter allowed smoothing the point cloud, reducing the point density from 13 to 11 million for December 2017, from 7.5 to 6 million for May 2018, and from 27 to 23 million for November 2018.

The point clouds from the three surveys allowed measuring the dimension and obtaining a closer perspective of the slope instabilities. Figure 6 shows the three point clouds for the largest active zone on the slope and the extent of the failure zone. The rock outcrop is approximately 50 m wide and located 30 m above the highway. The extent has not changed between December 2017 and November 2018; however, the upper scarp is becoming steeper as the material erodes. The material that is being deposited on the highway is the result of further erosion within the active zone limits. The erosion is more pronounced to the northern side of the active zone than the southern side.

Fig. 6
figure 6

Point cloud on the largest active zone F1 northern side of the study site: panel a shows the photo taken from the UAV after the May event; panel b shows the point cloud from first UAV survey; panel c shows the point cloud from the second UAV survey; panel d shows the point cloud from the third UAV survey

Full size image

The photograph from the second survey (Fig. 6a) shows the areas of rockfalls and debris being generated from the upper section, near the scarp. The December 2017 point cloud (T1, Fig. 6b) shows the initial conditions of the rockfall zone from the start of the analysis presented in this paper. The point cloud from May 2018 shows the progressive erosion process towards the northern side (T2, Fig. 6c). On the third survey, it was noticeable that the southern side of the active zone continues to release debris to the highway (T3, Fig. 6d).

Change detection

Use of historical point cloud data allowed estimating the changes on the surface between December 2017 and November 2018. The change detection analysis of the first survey with respect to the second and third surveys shows the progressive erosion process that occurred at the surface. The analysis using the M3C2 method resulted in 8.5% of the points in the reference point cloud had a significant change (movement greater than the level of detection) at 95% confidence level (Fig. 7). Areas with significant change showed a small cluster located in the active zone.

Fig. 7
figure 7

Zones of significant change (red points) derived from the change detection analyses using the M3C2 method; panel a shows the comparison between the first and second surveys; panel b shows the comparison between the first and third surveys

Full size image

The comparison between December 2017 point cloud and May 2018 point cloud resulted in an average level of detection of 0.2 m. The comparison between December 2017 point cloud and November 2018 point cloud resulted in an average level of detection of 0.3 m. After filtering the points with movement lower than the level of detection from both slope analyses, the results provide the magnitude and direction (i.e., material gain or loss) of the areas with measurable topographic change (Fig. 8). Assessing the cumulative slope changes from December 2017 to November 2018 in a two-step process required to homogenize the level of detection of both change detection analyses, which resulted in a minimum resolution of 0.3 m.

Fig. 8
figure 8

Change detection analysis from the three UAV surveys using M3C2 method: panel a shows the comparison between the first and second surveys; panel b shows the comparison between the first and third surveys. Warmer colors indicate a gain in material and cooler colors indicate a material loss

Full size image

The estimated distance change from the three point clouds shows the highest material loss to be located in the northern area (Fig. 8a, F1). This area of the slope also indicates the accumulation of material near the base of the slope and along the ditch of the highway. The results also show additional movements in the same area in November (Fig. 8b, F1). The analysis shows more material accumulation along the ditch between December 2017 and May 2018 than between December 2017 and November 2018, which is associated with highway maintenance and cleanup operations, following recommendations by KCB (Klohn Crippen Berger 2018b). Minor movements were measured near the road on other sections of the slope towards the south. This movement on the bottom sections of the slope agreed with the field observation.

The analysis reveals contrasting movements along the slope, increasing towards the center and northern sides of the slope. Conversely, fewer movements are found near the crest of the slope. The change detection analysis shows several zones on the upper and middle sections of the slope with an average material loss from 0.18 to 0.58 m in depth. The loss of material occurred in May 2018, accompanied by an average accumulation of material along the ditch of the highway from 0.09 to 0.69 m. In November, less material is shown to be loss on the upper and middle sections of the slope. These changes correspond to a loss of 738 m3 of material and a gain of 323 m3 at the toe of the slope in the northern area (Fig. 8b, F1). The difference in volume (loss and gain) corresponds to material being removed from the ditch along the highway prior the UAV flight in December 2017. At the center and eastern sides of the slope, the two smaller instabilities had a loss of material of 25 m3 (Fig. 8b, F2 and F3).

Lastly, a comparison of the cross-section A-A extracted from the December 2017 point cloud and May 2018 point cloud allowed visually assessing if the rock outcrop had moved between the two UAV surveys (Fig. 9). The cross section revealed a zone of material loss after the rockslide, and the remaining material deposited on the ditch of the road; unfortunately, the scan on May 2018 was made 1 day after the clearing of the highway. Thus, the collection of the total deposition zone is not complete.

Fig. 9
figure 9

Cross-section A-A for the rockfall trajectory analysis at the study site

Full size image

Drainage network

The computation of the drainage network using the raster generated from the November 2018 point cloud allowed to quantify the drainage above and at the face of the slope. Figure 10 shows the location of the preferential drainage lines from the farm crop to the toe of the slope. The figure shows the drainage network classified in 10 classes ranging from 125 to 15,800 m2 of contributing area. The lines are color-coded to indicate the amount of contributing area along the drainage lines. Thus, it is expected higher values are located near the toe of the slope. Still, one drainage line at the top of the farmland reached 10,800 m2 of contributing area, equal to 38% of the total mapped area of the farm crop. This drainage line flows to the face of the slope in between zone F1 and zone F2, the largest active zones in the slope (Fig. 10). This distribution could be related to the higher activities located to the center and the northern section of the study area.

Fig. 10
figure 10

Drainage network of the study site using DEM from point cloud from the November 2018 UAV survey

Full size image

The three active zones (F1, F2, and F3, shown in Fig. 10) have drainage lines distributed on the surface. Active zone A shows drainage lines with up to 1100 m2 of contributing area at the center of the active zone. The exposed outcrop on May 2018 generated a small bench at the middle of the active zone that increased the contributing area to approximately 800 m2. This increment in zone A coincides with the material loss between the point cloud from May and November 2018, as shown in points T1, T2, and T3 in Fig. 6.

Active zone F2 (Fig. 10) has a high drainage line of approximately 3000 m2 of contributing area starting from the upper section of the slope. This high-intensity drainage line has eroded the upper section of the slope. In the southern area of the active zone F2, there is another high-intensity drainage line with approximately 800 m2 of contributing area. This area in zone F2 started to increase the border of the erosion process similarly to active zone F3 (Fig. 10).

The analysis of the drainage network allows assessing runoff water as one contributing factor to the erosion process at the study site. High-intensity drainage lines are converging south of active zone F3 (Fig. 10) which could suggest future stability problems. The result gives an indication of the relation between runoff water and the active zones. However, this analysis does not account for water infiltration or evaporation. Although not conclusive at this site, the analysis provides important insight into drainage factors of the slope as they relate to the activity observed.

Modeled rockfall trajectories

The results from the 2.5D rockfall analysis were not used to calculate the length of the rockfall trajectories or to compare the kinetic energies from both techniques, but to evaluate if the geometry of the slope results in rockfall trajectories that deviate significantly from those represented from a 2D analysis. The results confirm that the trajectories along the slope do not notably deviated from an idealized 2D trajectory, and therefore, the critical block in the northern side of the slope can be analyzed with a 2D approach, as shown in Fig. 11, F1. The 2D simulation of rockfall trajectories considered blocks of equivalent diameter to 0.05 m, 0.1 m, 0.2 m, 0.3 m, 0.4 m, and 0.5 m. The corresponding masses are shown in Fig. 12. The analysis of the trajectory of potential rockfalls estimates rock blocks reaching the highway, resting on the side berm of the highway, and crossing the road towards the Red Deer River. The simulation of 1000 possible falling blocks can reach a height of approximately 5 m (Fig. 12b) at the edge of the highway with a total kinetic energy of up to 110 kJ (Fig. 12c). Figure 12 shows the overall distribution of trajectories and how all block size trajectories overlap.

Fig. 11
figure 11

2.5D rockfall simulation for trajectory of blocks generated with two lines of seeders located at the crest of the slope

Full size image
Fig. 12
figure 12

2D rockfall simulation on section A-A. a Trajectories from 1000 blocks using the Monte Carlo method. b Bounce height from 95% percentile of rockfall trajectories. c Total kinetic energy from 95% percentile of rockfall trajectories

Full size image

The calibration of the coefficients of restitution resulted in trajectories with similar characteristics to those observed during the site visits. These modeled trajectories resting near the highway and near the riverbank matched these observations. These sizes resemble cube blocks with a side length of 0.08 m, 0.16 m, 0.32 m, 0.6 m, 0.65 m, and 0.8 m as measured during the site visits. The detachment locations for falling blocks were modeled as a continuous line, located between the head scarp and the location of the large block (also according to observed rockfall scars with the higher potential energies).

The simulation using RocFall (Rocscience Inc. 2018) showed modeled trajectories of the different rock block sizes identified from the May 2018 survey. The simulation shows the impact of different block sizes and sources on the energy of the block and bouncing height when reaching the ditch adjacent to the road. These are key parameters for decision-making regarding rockfall protection (e.g., height of lock block walls) (Macciotta and Martin 2019).

Discussion

Rock slope failure processes and their relationship with weather conditions

The geology of the slope contributes to a differential erosion process and variability in the weathering of the materials on the slope. The formation of preferential flow paths tends to be developed in areas with a higher content of sand-size particles. The different geological layers are distributed horizontally, and the perpendicular direction of the preferential flow paths crosses all different rock layers. Consequently, the flow paths intensify the erosion processes in the vicinity of these channels resulting in areas with more susceptibility to failure. This erosion process increases near the base of the slope as sedimentary layers are less distinctive and layers are a mixture of sand-size particles with different levels of fine particles.

The results from the change detection analysis in Fig. 8 and site observations resulted in the identification of the following three modes of failure.

The first mode of failure is the development of debris flows. This is postulated to be the result of layers of the argillaceous rocks rapidly eroding with seasonal precipitation, and freeze-thaw cycles. This erosion generates a layer of loose clayey soil that is prone to flowing down the slope once it has a high enough moisture content to sufficiently lower its strength. These events are more probable during the spring due to melting snow and higher levels of precipitation. The event in May 2018 was an example of the mode of failure (Fig. 8a, F1).

The second mode of failure is rockfalls. This was observed to be a result of the differential weathering of the sedimentary rock. Layers of sedimentary rock more prone to weathering are eroded from beneath more competent rock layers, leaving the competent rock unsupported (Fig. 8, F1), leading to rock detachment and rockfall events. The rockfall events shown in Fig. 3b were an example of the mode of failure after the exposure of the large rock blocks following the debris flow in May 2018.

The third mode of failure is the detachment and falling of blocks of frozen soil from the surface of the slope. This mode of failure was identified from the observation of frozen blocks of soil and bedrock over the highway and the frozen river and the identification of the source of the frozen material on the slope (Fig. 3a).

Mitigation strategies

The susceptibility of the slope to generate rockfalls after precipitation events, and particularly after the large event in May 2018 that exposed a large outcrop with the potential to release large blocks to the highway, requires a detailed analysis of the rockfall hazard that users are exposed to. The point cloud derived from the UAV survey provided the necessary means to identify and analyze in detail the critical area of the slope prone to rockfalls. Rockfall trajectories can be better estimated by extracting detailed topographic information in a point cloud. The rockfall trajectory analysis provided the information on which to base decisions and design for rockfall hazard mitigation. Nonetheless, this information must be integrated with field observations to adjust and calibrate the characteristics of the event such as block size and shapes, location, deposition zone, and surface conditions.

The use of sequential topographic reconstructions, field observations, and the comparison of two cross sections from the two UAV surveys in December 2017 and May 2018 (Fig. 9) allowed for an informed assessment of the potential hazards associated with this slope. The information gained in this analysis lead to proposing a combination of a catchment net near the base of the slope and slope maintenance (e.g., scaling) to decrease the risk to highway users from rockfalls on the northern side of the slope (Klohn Crippen Berger 2018a). This catchment net was selected to provide retention of 95% of the blocks with a total dissipation energy of 150 kJ as indicated by the model. Options to manage the risk to debris flows included weather-based hazard notifications and detection fences that, when triggered, alert the users about a potential blocked section of highway. These last are still under development at this site (Klohn Crippen Berger 2018a).

Conclusion

UAV and digital photogrammetric techniques are very useful to characterize inaccessible areas. Creating a historical registry through point clouds becomes a useful methodology for monitoring purposes and assess changes, as well as to obtain comparison maps from 3D point clouds or section profiles. Other sub-products like DEM in raster format may be used for drainage analysis or rockfall simulations.

The analysis of rock failure process and the mitigation strategies have taken advantage of UAV photos and digital photogrammetry. Risk management of ground hazards on large extents of linear infrastructure limits the resources that specific problems might require. The use of low-cost UAV becomes a useful tool that provided detail information of the surface ground conditions, especially on sites with difficult access. At the C018 site, the slope inclination and height became an obstacle for visual inspection after the sliding events; however, the use of low-cost point cloud during several inspections provides a new perspective that can optimize the ground hazard assessment. The use of digital photogrammetric techniques based on pictures obtained with UAV devices allows quantifying the magnitude and extent of the slope instabilities.

The low-cost point cloud provides the opportunity to generate historical records that show different perspectives on the rockfall hazards at the study site. The use of change detection analysis in combination with visual inspections provided the information to determine the location of critical zones, magnitude, and extent for the implementation of mitigation strategies.

Ultimately, the use of digital photogrammetric techniques based on pictures acquired with UAV became a tool that helped in the risk management process on the C018 site after the latest events. The information gathered from the UAV survey would continue to influence future decisions on the site as it captured a larger picture of the processes that otherwise might be difficult or not viable to obtain.

The geologic characteristics of the slope at the study site in combination with water have always been the driven factor of the historical failure process that has impacted the highway. Analysis of the drainage network has given an indication of the potential critical areas that can be impacted during a period of heavy rainfall.

The geometry of the slope make the access difficult for reaching the active zone source and other weathered features upslope. The implementation of UAV technology to assess the conditions of the slope allowed having a large perspective on the conditions during each survey. Monitoring and investigation of the C018 slope have provided insight into the use of low-cost UAV for monitoring and assessing rock slopes with multiple failure mechanisms associated with weather conditions. This tool provides the engineer with multiple perspectives of ground hazards with complex modes of failure and difficult access, which translates in more information to improve the assessment of the slope and optimized mitigation strategies. The point cloud in combination with the change detection analyses allowed quantifying the dimensions and showing the evolution of the erosion process at the study site. Low-cost UAV surveys should not replace or undermine the importance of site inspections; rather, it enhances the information to make better decisions.