Rockfall is a mass-movement hazard often found in mountainous environments, posing risk to human lives and the safe operation of infrastructure. In Canada, rockfall hazard is particularly problematic for transportation infrastructure, where traffic corridors have been constructed in steep river valleys, adjacent to natural and cut slopes. In many of these valleys, right-of-way space is limited, forcing operators to construct highways and railways with minimal clearance to potentially unstable rock masses. The proximity and extensive presence of fractured and weathered rock slopes, combined with the seemingly unpredictable nature of rockfall events, makes it difficult for operators to manage this hazard.

Like any geohazard, we are interested in knowing how frequently we can expect events to occur, where they are most likely to come from, and whether or not they will reach and cause harm to our elements at risk. For rockfall this could involve identifying lithologies which are more susceptible to rockfall (e.g. Rosser et al., 2005; van Veen, 2016) or correlating rockfall frequency to triggering events like severe rain storms (e.g. D'Amato et al., 2016; Pratt et al., 2018). In order to build these relationships, we first need knowledge of previous rockfall in the area. This is one of the main challenges in the study of rockfall. Due to challenging terrain in mountainous areas, and safety concerns where rockfall activity is high, site access to the slopes of interest is often limited, prohibiting direct measurement.

In order to circumvent these obstacles, modern remote sensing techniques, such as lidar, have seen widespread use in the study of landslide hazards like rockfall (Jaboyedoff et al., 2012). Terrestrial laser scanning (TLS) in particular has been increasingly applied in recent years to the study of rock slope instabilities (Abellán et al., 2014; e.g. Lato et al., 2009, 2015; Kromer et al., 2017). TLS provides a detailed 3-D model of the slope geometry, supporting a range of geotechnical and geomechanical analyses. From a single scan we can extract slope angle measurements, map rock mass structure and look for evidence of past rockfall (Telling et al., 2017). Using multiple successive scans of the same slope, progressive changes to the slope surface can be measured. Multi-temporal scanning enables researchers to detect rockfall events over time and build detailed magnitude–frequency relationships for slopes, supporting the study of processes such as coastal cliff erosion (e.g. Rosser et al., 2007; Williams et al., 2018), as well as hazard and risk assessments for linear infrastructure such as railways (e.g. van Veen et al., 2017). The identification of rockfall events using sequential scans also permits the back analysis of rockfall-triggering factors and failure mechanisms. A study by Kromer et al. (2015) used TLS change detection to investigate the occurrence of a 2600 m3 failure above a section of the Canadian National (CN) Railway in western Canada, including an analysis of structural constraints, pre-failure deformation, and precursor rockfall leading up to failure.

While the application of TLS to rock slope monitoring is often focused on determining how likely future rockfall is or where the rockfall may come from, it is also important to determine how likely it is that the rockfall material will reach an element at risk should a fall occur. In order to answer that question, rockfall modelling programs can be used to simulate the runout of falling rock material using numerical models (Turner and Duffy, 2012). A rockfall simulation requires a representation of the slope surface and rockfall mass being modelled in 2-D, 2.5-D, or 3-D. TLS can support rockfall simulation by providing a high-resolution 3-D model of the slope surface and, in cases where a rockfall event has been identified using change detection between multiple scans, the volume and location of the rockfall mass. In order to make use of the quality and quantity of 3-D point cloud data being collected as part of a rock slope monitoring program in western Canada, a 3-D rockfall simulation technique using game-engine technology has been developed (Ondercin, 2016; Sala, 2018). The technique uses the video game development platform Unity 3D (Unity Technologies, 2018) and is capable of simulating rockfall runout using fully 3-D meshes built from TLS point cloud data.

One of the strengths of this simulation technique is its ability to simulate rockfall runout using numerous interacting bodies. This capability allows us to produce simulations of fragmental rockfall events, which are defined by the presence of multiple mobile fragments of rock during runout (Hungr and Evans, 1988). Conventional rockfall modelling programs (e.g. RocFall, Rocscience Technologies, 2016; Rockyfor3D, Dorren, 2015; RAMMS:ROCKFALL, Bartelt et al., 2016) simulate single boulder trajectories at a time and therefore are unable to model fragmental rockfall runout. Efforts to model fragmental rockfall processes, including the disaggregation of falling rock masses along pre-existing discontinuities, or the breakage of intact blocks during impact, have been demonstrated by previous authors. Cuervo et al. (2015) utilized a discrete element technique to model the disaggregation of a 1000 m3 rockfall event in southern France comprised of numerous mobile fragments. Wang and Tonon (2011) developed a discrete element code which models the fragmentation of a falling rock at impact, taking into consideration the effect of impact velocity, ground condition, energy loss, and fracture properties such as persistence. The GIS-based tool RockGIS, presented in Matas et al. (2017), incorporates both breakage along pre-existing discontinuities and the generation of new fragments during impact into a 3-D runout modelling program and has been utilized to model a 10 000 m3 fragmental rockfall in the eastern Pyrenees mountains. The goal of the work presented in this paper is to demonstrate the capability of our novel game-engine-hosted simulation technique to model a series of rockfall events detected using TLS change detection at two rock slopes in south-central British Columbia. Emphasis is placed on the ability of the technique to be used for fragmental rockfall runout simulation, including a discussion of how these types of simulation could support mitigation design.

The rock slopes of focus for this paper are part of the Ashcroft subdivision of the CN Railway. The subdivision follows sections of the Thompson and Fraser River valleys, located between the towns of Ashcroft and Lytton, British Columbia. This region serves as an important transportation corridor for Canada, with the presence of the CN Rail mainline, as well as sections of the Canadian Pacific Railway, and the Trans-Canada Highway. A previous study by Piteau (1977) identified that rock cuts in this region are subject to slope instability issues due to a combination of lateral erosion from river activity, over-steepening from blasting during the construction of the railways, and a lack of adequate rockfall catchment areas.

Two sites in the Ashcroft subdivision will form the basis for this research, White Canyon and Goldpan. The location of these sites along the Thompson River and proximity to the town of Lytton, BC, can be seen in Fig. 1. The monitoring of these slopes is part of the Railway Ground Hazard Research Program, a collaborative research initiative focused on the characterization and assessment of mass movement hazards affecting Canadian railways.

Data collection in the Ashcroft subdivision first took place in 2012, and regular monitoring at both sites has been ongoing since 2014, with data collection taking place on a seasonal basis, approximately every 3 months. Point cloud data collection consists of TLS scans acquired using an Optech ILRIS 3D-ER (2012–2017) or Riegl VZ-400i (2018) lidar system. Additionally, aerial laser scan (ALS) data coverage for both sites, with an average point spacing of 0.3 m, was acquired in 2014 and 2015. Site photos are collected using a Nikon D700 or similar camera with a 135 mm prime lens. A series of photos of the slope are taken using a GigaPan EPIC Pro robotic camera mount and stitched together into a single high-resolution site panorama using GigaPan Stitch (GigaPan Systems, 2013).

Goldpan is located on the north side of the Thompson River, approximately 26 km upstream from the town of Lytton. Present at the base of the slope is a section of the CN Rail main line. The rock slope spans 800 m, rising up to 65 m vertically above track level. The average slope angle at the site is 55–60∘. TLS data are collected from scanning vantage points across the river from the slope, accessed via Goldpan Provincial Park. Scan distances range between 170 and 230 m, producing an average point spacing of approximately 6 cm.

White Canyon is located on the north side of the Thompson River, approximately 5 km upstream from the town of Lytton. Present at the base of the slope is a section of the CN Rail main line. The rock slope spans 2.4 km, rising up to 375 m vertically above track level. The average slope angle at the site is 40–45∘. TLS data are collected from scanning vantage points across the river from the slope. Scan distances range between 400 and 600 m, producing an average point spacing of approximately 10 cm.

Goldpan and White Canyon are located in the Intermontane belt of the Canadian Cordillera. Both sites belong to the Quesnellia volcanic arc terrane, which is composed of Carboniferous to mid-Jurassic volcanic, sedimentary, and plutonic rocks (Struik, 1987; Monger and Nokelberg, 1996). The rock mass at Goldpan is largely massive and is predominantly composed of volcanics of the mid-Cretaceous Kingsvale–Spences Bridge Group. The main rock types of this group are basaltic to andesitic flows interspersed with volcaniclastic sandstones, shales, and conglomerates (Brown, 1981). The comparatively large rock slope of White Canyon belongs to the Mt Lytton plutonic complex (Greig, 1989). Locally, the dominant rock type is an amphibolite grade quartzofeldspathic gneiss. Amphibolite banding is also present and gneissic layering is crosscut throughout the canyon by intrusive phases of gabbro, tonalite, and granodiorite (Brown, 1981). Preferential weathering around these intrusions often results in the formation of vertical rock spires which act as source zones for rockfall.

Mitigative measures at both sites have been installed in response to frequent rockfall activity impacting railway operations. At Goldpan this includes rockfall wire mesh draped over parts of the slope, extensive sections of shotcrete, and four concrete rock sheds. In the eastern half of White Canyon there are two timber rock sheds and one concrete shed, as well as wire mesh rockfall nets and concrete lock blocks adjacent to the track. The western half of White Canyon also has wire mesh rockfall nets and lock block retaining walls, as well as an additional three concrete rock sheds and one timber rock shed. Slide detector fences are present at both sites, comprised of horizontal wires strung between upright telephone poles, and provide warning by switching the track signal to stop when broken.

Five rockfall events were selected from a database of rockfalls which have been identified using change detection at White Canyon West, White Canyon East, and Goldpan since 2012 (Kromer et al., 2015). These events have masses ranging from 2 to 170 m3, exhibiting a combination of structurally controlled failure modes including wedge sliding, planar sliding, toppling, and overhanging blocks. Images of the five slope sections where the rockfall events of interest took place can be seen in Fig. 2.

Each event showed material accumulation below the source zone in the change detection results. This is essential information for comparison with our simulation results. Each event is also close to track level, with the highest fall occurring 46 m vertically above the track. These events were selected because the shorter distance from source to a notable accumulation of material presents a simpler trajectory for back analysis and reduces potential confusion in interpreting whether the material gain is due to the selected rockfall or another mass-movement nearby. In each case, change was detected using the Multiscale Model to Model Cloud Comparison (M3C2) point–point distance calculation (Lague et al., 2013) in CloudCompare (2018), between the pre- and post-fall TLS scans. A summary of each of the rockfall events including change detection results, and site photos before and after the event, is provided in Figs. 3–7. Discontinuity and slope angle measurements used for the stereonet failure mode analysis shown in each figure were completed using the Compass tool in CloudCompare.

For each of the selected rockfall events from our field sites, we can set up a simulation in the Unity game engine. The main input parameters for simulation setup are

coefficient of friction,

A more detailed description of the material coefficients and their influence on the rotational and translational velocity of simulated rock fragments, along with a parametric investigation into the effects of simulated block shape and release orientation, can be found in Sala (2018).

The geometry used to build simulations is comprised of slope and rockfall elements. The slope geometry is a 3-D mesh produced from the pre-fall TLS scan of the slope, generated using Poisson reconstruction in CloudCompare. The rockfall geometry is a 3-D mesh produced from the pre- and post-fall TLS scans of the source region, generated using Poisson reconstruction and connected into a single hull in Blender (Blender Foundation, 2017). By extracting the rockfall source volume directly from the same TLS data used to generate the slope geometry, the release position and orientation of the source volume relative to the slope surface is well defined. This 3-D hull is then segmented into a collection of convex hull fragments, using Voronoi fracturing in Blender. This method subdivides the source rockfall volume into a specified number of fragments selected by the user, with control over the size and shape of fragments produced. An example of various Voronoi fracturing results, including recursive fracturing, and fracturing with preferential shape constraints can be seen in Fig. 11. The end result is a simulated rockfall source volume which reflects the shape and fractured nature of the rockfall event detected from the field.

A detailed explanation of the steps necessary to move from field remote sensing data to a functioning simulation in Unity, including the workflow from CloudCompare to Blender to Unity, can be found in Sala (2018). One of the key strengths of this modelling technique is its ability to simulate the collision between multiple moving rigid bodies at once. Support for multi-body collisions allows us to simulate the disaggregation of heavily jointed source rockfall volumes. This enables us to study the types of rockfall runout expected to occur at our field sites, taking into consideration the implications of fragmental runout discussed above. Snapshots of an example fragmental rockfall simulation for the 170 m3 Goldpan event (Site A) can be seen in Fig. 12.

When developing or attempting to calibrate any new simulation technique, it is necessary to compare simulation outputs to real observations of the phenomena you are modelling in order to test how well the simulation performs. Where direct observation of rockfall events is possible, such as in rockfall drop tests, (e.g. Ushiro et al., 2006; Vick et al., 2015; Volkwein et al., 2018), this information could include translational and rotational velocity, mapped runout trajectories, or pass height above the slope. For the five rockfall cases selected from our field sites, this information is not available as the events were not observed directly. Instead the location and magnitude of change present in the change detection results for each event serve as the basis for simulation comparison.

Simulations of each of the five selected rockfall events were run and change maps of the simulation results were produced. In order to generate change detection maps for the simulations, the positions of the post-simulation fragments were merged with the slope geometry to create a post-simulation 3-D mesh. Similarly, the fragmented rockfall geometry, pre-simulation, was merged with the slope geometry to create a pre-simulation 3-D mesh. Each mesh was then converted into a point cloud in Blender. The conversion of the mesh data to a point cloud allows us to use the M3C2 point–point distance calculation for both the actual and simulated rockfall events. The transition from simulation mesh to point cloud is done by selecting only visible vertices in the mesh, using a similar vantage point to the direction of the original TLS scan. The selection of only visible vertices in the mesh is necessary as the 3-D rockfall fragments, pre- and post-simulation, have vertices across their entire surfaces, both front and back facing. This results in multiple layers of points in the pre-fall source zone and post-fall accumulation zone, adversely affecting point cloud normal calculations and producing errors in the distance measurements between what should be distinct pre- and post-fall surfaces. The extraction of visible vertices from the mesh geometry, using a specified vantage point, ensures that only a single layer of points is present in the simulated point cloud.

The results of the change detection analysis for each simulated rockfall event are shown in Figs. 13–17. In each case the game-engine-based simulation prototype was able to produce rockfall runout which compared well in terms of the location and magnitude of the measured actual change. The percentages of source rockfall volume retained at different locations on the slope is shown. Volume measurements using the actual event point cloud data were completed using the 2.5-D Volume tool in CloudCompare. Simulated volume percentages were determined using the 3-D mesh data for each of the convex hull fragments simulated in Unity 3D. For each event, the simulated volume percentages for each of the slope locations was within 5 % of the actual percentage, with the majority of the rockfall material ending up in the track-side ditches for every rockfall event except the Goldpan wedge slide. In the Goldpan case, Site A, we see two distinct areas of accumulation, on the mid-slope bench and on top of the rock shed. In this case the simulation results produced similar accumulations in both locations, with the majority of the material leaving the slope over the edge of the rock shed.

The objective of this work was to demonstrate the ability of our rockfall simulation workflow to model complex rockfall source geometries comprised of many moving fragments. While this process is a step forward in the simulation of fragmental rockfall events, a discussion of current limitations is important. The two main limitations of this work are fragmentation timing and the distribution and size of rockfall fragments used.

Five rockfall simulations were run based on TLS change detection inputs from events identified at our field sites in south-central British Columbia. The simulations utilized high-resolution 3-D meshes generated from 6–10 cm TLS point clouds, and complex rockfall volumes, extracted from pre- and post-event scans. The fracturing of source rockfall volumes enabled the modelling of fragmental rockfall runout, taking into consideration the interaction of fragments with each other as well as slope-stopping features such as gullies and ditches. The results of this work demonstrate the ability of our rockfall modelling prototype to produce runout material accumulations which agree well with observed change from the actual events. With no direct observations of rockfall runout, or on-site measurement, the conversion of post-simulation mesh data in order to produce simulated change detection presents a novel way to perform rockfall runout comparisons in the absence of other data types.

The application of the technique to mitigation design was also discussed, demonstrating the potential of the model to be used for investigating the effectiveness of rockfall catchment ditches. This type of analysis, visualizing the accumulation of rockfall runout volume downslope, could also be applied to the assessment of other countermeasure options such as lock block retaining walls or rock sheds. The ability to perform these types of analyses is limited in industry standard rockfall modelling programs, as they are often able to simulate only a single moving block at a time. While these models are very useful for single block falls, and regional runout assessments, they are unable to capture the build-up of rockfall material on the slope, which can affect the path of subsequent rockfall fragments and the effectiveness of retaining structures. The simulation of these falls as multiple moving bodies in our technique is able to capture this interaction between rockfall material and the finite capacity of the structures used for rockfall protection.