Landslide Monitoring with RADARSAT Constellation Mission InSAR, RPAS-Derived Point-Clouds and RTK-GNSS Time-Series in the Thompson River Valley, British Columbia, Canada

Abstract

In this paper, we evaluate the effectiveness of four land-deformation measurement techniques for monitoring slow-moving landslides along a high-risk section of the national railway corridor traversing the Thompson River valley, British Columbia, Canada. The geomorphically active North Slide acts as an ideal field laboratory for testing and evaluating novel monitoring techniques and methods. We compare differential processing of Structure from Motion (SfM) products such as point-cloud elevation models and orthophotos derived from Remotely Piloted Aircraft Systems (RPAS), along with satellite based Interferometric Synthetic Aperture Radar (InSAR) deformation measurements derived from RADARSAT Constellation Mission (RCM). These results are ground-truthed with periodic real-time kinematic (RTK) global navigation satellite system (GNSS) measurements. We evaluate point-cloud comparison techniques, including the multi-scale model-to-model cloud comparison (M3C2) algorithm and digital ortho image correlation techniques. Multi-temporal RCM InSAR deformation measurements are processed using a semi-automated processing system for interferogram generation and unwrapping. Manual processing of small baseline subsets (SBAS) leads to the recovery of 1-dimensional line-of-sight (LoS) and 2-dimensional deformation measurements. Lastly, we discuss the strengths and limitations of these techniques, considerations for interpreting their outputs, and considerations for direct comparisons between InSAR, RPAS and RTK-GNSS deformation measurements.

1 Introduction

The Thompson River valley in south-central British Columbia (BC), Canada (Fig. 1a) serves as a field laboratory for testing and comparing the reliability and effectiveness of different static, dynamic, and real-time landslide monitoring technologies along a strategically important section of the national railway network (Huntley et al. 2017; Huntley et al. 2019a, b; Huntley et al. 2021a). In this paper, we present research at the North Slide (Fig. 1b) undertaken as part of International Consortium on Landslides (ICL) International Programme on Landslides (IPL) Project 202. The economic importance of the Thompson valley transportation corridor, along with the need to understand and manage the safety risk related to the landslides that threaten the national railway network, mandate the North Slide (Fig. 1b) a research priority for Natural Resources Canada (NRCAN) and the Geological Survey of Canada (GSC). Following the approach outlined in Huntley et al. (Huntley et al. 2022a, b), here, we compare differential processing of Structure from Motion (SfM) products, such as point-cloud digital elevation models (DEM) and orthophotos derived from Remotely Piloted Aircraft Systems (RPAS), with satellite based Interferometric Synthetic Aperture Radar (InSAR) deformation measurements derived from RADARSAT Constellation Mission (RCM). These results are ground-truthed with periodic real-time kinematic (RTK) global navigation satellite system (GNSS) measurements (cf. Huntley et al. 2022b).

Fig. 1

Location of the ICL-IPL Project 202 study area: (a) southwestern British Columbia showing major railway corridors. CN Canadian National Railways, CP Canadian Pacific Railways, A Ashcroft, K Kamloops, L Lytton, S Spences Bridge, V Vancouver, FR Fraser River, TR Thompson River. (b) The North Slide showing geomorphic extent in red. I—active slide toe (0.08 km²)—the “Solar Slump”; II—inactive (1880) slide main body and headscarp (0.55 km²); III—inactive (early Holocene) slide body (0.37 km²); IV—stable postglacial slopes and terraces. (c) Cross-section A to A’ across the North Slide modelled as a rotational-translational landslide in glacial deposits confined to a bedrock paleochannel or basin (after Huntley et al. 2021b)

RPAS, satellite InSAR, and other geospatial and temporal datasets will help stakeholders develop a more resilient railway national transportation network able to meet Canada’s future socioeconomic needs, while ensuring protection of the environment and resource-based communities from natural disasters related to extreme weather events and climate change.

2 Study Site: North Slide

The North Slide is located approximately 6 km south of Ashcroft on the east bank of the Thompson River (Fig. 1a). Around 9 pm on October 14, 1880, an ancient landslide (Fig. 1b) reactivated as a sudden onset, rapid retrogressive rotational-translational flow-slide (Stanton 1898; Evans 1984; Clague and Evans 2003). In common with other nineteenth Century landslides in the valley, the slope failed rapidly after months of intensive summer irrigation and while the toe slope was being excavated during railway construction.

Exposures across the slide main body and head scarp reveal glaciolacustrine sandy silt and sandy gravel outwash unconformably overlying silty till and glaciolacustrine silt and clay (Fig. 1c). A toe slope bulge on the river floodplain exposes back-tilted rhythmically interbedded layers of soft brown clay, stiff, highly plastic dark grey clay, and grey silt overlying bedrock (Huntley et al. 2021b). Borehole piezometer and inclinometer monitoring reveal preferential shearing of soft brown clay beds, with rupture zones at 264 and 269 m above sea level (asl): equivalent to 25 m and 30 m below the CP rail grade, and 5 m to 10 m below the riverbed. Piezometer data indicate hydrostatic conditions at depth below the track, and an upward groundwater gradient in the landslide toe (Porter et al. 2002).

2.1 Landslide Hazards and Railway Infrastructure

Thompson River is a major control on landslide form and function (Eshraghian et al. 2007; Eshraghian et al. 2008). Riverbed incision is contributing to changes in toe slope geometry, with 5 m-deep scour pools likely intersecting rupture zones. Air photos from 1928 and 1997 suggest historical channel-bank erosion rates averaging 70 cm yr.⁻¹ (Porter et al. 2002). Reactivation of the North Slide toe slope began in October 2000 along a 150 m section of riverbank (the “Solar Slump”), with 5 cm to 15 cm of settlement at the CP grade. Peak observed surface movement rates were approximately 15 cm yr.⁻¹, with an average rate of 3 cm yr.⁻¹ (Porter et al. 2002). Displacement along the shallow rupture surface ranged from 5 cm to 11 cm yr.⁻¹; while on the deeper slide surface, borehole inclinometers recorded a movement rate of 3 cm to 4.5 cm yr.⁻¹ (Porter et al. 2002). RADARSAT 2 (RS2) and SENTINEL-1 (S1) persistent scatterer interferometry showed the slide toe remained active between 2013 and 2015, with displacement of coherent targets indicating line-of-sight (LoS) deformation rates in excess of 5 cm yr.⁻¹ (Huntley et al. 2017; Journault et al. 2018; Huntley et al. 2021b).

Instrumental and operational measures are in place to reduce the risk of train derailment due to landslide activity. Mercury switch tip-over posts linked to rail signals were installed in 2001 to detect ground displacement. Small, incremental surface displacements contribute to minor track misalignment requiring continuous remote monitoring and short-term (seasonal) reorganization of train schedules to allow ground crews to safely add ballast and adjust the track positions. By contrast, infrequent, rapid, large and widespread ground movements are a concern for railway companies, government agencies, and local communities because they cause major track misalignment, train derailment, and considerable local environmental damage. Substantial channel and slope remediation following a major retrogressive failure will require costly reorganization of transportation and shipping schedules, leading to national socioeconomic losses.

3 Methods and Results

3.1 RPAS Surveying

Periodic RPAS surveys were flown using a DJI Phantom 4. This RPAS (Fig. 2) was equipped with a 12.4 M pixel FC330 outputting 3000 × 4000-pixel images (Huntley et al. 2022b). These images were geotagged by the onboard GPS. However, since no real-time or post-processing corrections were applied to the GPS tags, manual RTK-GNSS ground control was required.

Fig. 2

DJI Phantom 4 RPAS (NRCAN Photo 2020–845)

Digital optical imagery was processed using Pix4D to produce optical orthophotos at 2 cm ground sample distance (GSD), an elevation point-cloud, and a digital surface model (DSM) with a 5 cm ground sample distance. To mitigate seasonal effects while allowing movement to accumulate to detectable levels, two surveys separated by approximately two years (September 19, 2019, and September 28, 2021) were undertaken. Two post-processing techniques for calculating deformation from these datasets were evaluated: M3C2, and a combination of digital image correlation and DSM of difference.

3.2 M3C2 Point Cloud Comparison

The Multiscale Model to Model Cloud Comparison (M3C2) algorithm was introduced to fill a perceived gap in time-series point-cloud comparison (Lague et al. 2013). This algorithm had three relevant characteristics: (a) it operated directly on point-clouds without the need to grid or mesh one, or both point-clouds; (b) it calculated displacement along a vector normal to the local surface topography; and (c) it estimated confidence intervals for these measurements.

The algorithm functioned with two steps. A surface normal was first calculated at each point by considering a neighbourhood of surrounding points controlled by user-defined parameters. Secondly, two neighbourhoods surrounding each point, often a subset of the points, were used to define the surface normal and calculate the average distance between these points along the surface normal vector.

This algorithm was well suited for complex topographic situations such as steep or rough surfaces. Simpler methods like DEM differencing cannot accurately capture complex deformation in steep and/or rough topographic settings such as rock falls, or toppling of vertical and sub-vertical rock faces. A fundamental assumption of the M3C2 algorithm was that deformation occurs on a vector normal to the ground surface. This may be reasonable for measurements of sediment erosion and deposition. However, the algorithm may dramatically underestimate displacement where ground movement is parallel to the surface (Fig. 3), a situation common in earthflows and translational slides (Fig. 4; Rotheram-Clarke et al. 2022).

Fig. 3

Conceptual diagram of how M3C2 calculates displacement along surface normals (vectors). Note the insensitivity to surface parallel displacement

Fig. 4

Horizontal projection of the M3C2 surface normals (vectors) over M3C2 horizontal distance (raster). Hill-shade transparency is applied for context. All M3C2 surface normals are unit vectors, so shorter vectors have a larger vertical component, while longer vectors have a larger horizontal component

3.3 Digital Image Correlation

Photogrammetric processing and 3D reconstruction was completed using Micmac, a free and open-source software package distributed by The National Institute of Geographic and Forest Information (IGN) (Rupnik et al. 2017). The digital image correlation functions in Micmac were used to calculate subpixel offsets between precisely co-registered imagery (cf. Rosu et al. 2015; Galland et al. 2016). This was tested on both RPAS orthophotos and hill-shaded Digital Surface Models (DSMs) from the SfM results. Significantly, it was determined that hill-shaded DSMs produced less noisy outputs. This is likely a result of uniform input images containing few differences in dynamic range and surface texture due to earth materials and vegetation cover (Rotheram-Clarke et al. 2022). Combining these results with a simple DEM difference image was the only method tested capable of recovering the full 3D deformation vector with wide coverage (Fig. 5).

Fig. 5

3D RPAS deformation over North Slide between 2019-09-19 and 2021-09-28. Vectors indicate horizontal deformation from MicMac digital image correlation and colour scale represents difference in DSMs. Hill-shade transparency is applied for context

3.4 RTK-GNSS Surveying

Campaign RTK surveys were completed using a Spectra SP80 GNSS system. At a designated survey base station approximately 1 km south of the study site, a piece of rebar approximately 60 cm in length was driven into stable ground and marked precisely with a 2 mm punch. This site was occupied for 24 hours and the GNSS observations processed using the NRCAN PPP service (cf. Huntley et al. 2021a). Subsequent surveys reoccupied this known reference point. Large, heavy boulders embedded in the landslide blocks were marked with a cross pattern typical of airphoto ground control point (GCP) surveys, using orange survey paint that was visible in RPAS imagery (Fig. 6). The centres of these patterns were scored with a rock drill to ensure exact repeat occupation. These RTK-GNSS measurements constrained the RPAS SfM measurements, and benchmarked deformation measurements derived from RPAS and InSAR measurements.

Fig. 6

RTK-GNSS GCP measurements, showing scaled horizontal movement vectors plotted on annotated RPAS digital orthomosaic image. Location of DSM highlighted in red on inset map of North Slide

3.5 RCM InSAR Analysis

A stack of 3 m RCM descending images were tasked and acquired by the Canadian Space Agency (CSA). The SAR Toolbox in the Earth Observation Data Management System (EODMS) was used to generate interferometric pairs that were then corrected for spatial atmospheric errors, unwrapped, and masked for low coherence (cf. Dudley and Samsonov 2020). The network of unwrapped phase measurements was then manually examined. InSAR pairs that had obvious unwrapping errors or significant decorrelation noise were removed. A free SBAS processing package developed and distributed by NRCAN (MsBASv3, Samsonov 2019) was used to recover the deformation time-series history for the period of InSAR observation (Fig. 7).

Fig. 7

InSAR-derived ground surface velocities mapped with semi-transparent hill-shade applied for context

4 Discussion

4.1 Comparing Displacement Measurements

Individually, each landslide displacement monitoring approach has limitations that must be considered when comparing to other methods. Measurements vary in terms of directional sensitivities, frequency of observations, temporal coverage, and absolute–vs-differential measurements.

Two methods of comparison across datasets are employed in this study. First, RTK-GNSS measurements and 3D RPAS displacement are projected onto the RCM InSAR LoS, then time-series profiles of all three measurement techniques are compared at all GCP positions. This method only considers a small subset of point available to both the wide area RPAS and InSAR methods. The second comparison method leverages the broad spatial coverage of both the 3D RPAS measurements and InSAR by estimating an annual deformation rate of both datasets and comparing all collocated points.

Of the four deformation-measurement techniques presented, only the M3C2 algorithm is able to calculate deformation along a vector normal to the ground surface. The result is that the direction of deformation is highly variable. It is likely that M3C2 generally underestimates deformation of the “Solar Slump” to a degree that varies with how the movement vector and surface normal are aligned (cf. Fig. 3). The M3C2 algorithm can be expected to output nearly accurate estimates of ground deformation on surfaces perpendicular to the ground movement. At worst, M3C2 is insensitive to surface parallel movement. Although it would be possible to project normal vectors into the InSAR LoS vector or decompose the deformation vectors into their Cartesian components, comparisons would not be meaningful since InSAR can be insensitive to movement along a particular orientation. M3C2 has the added complexity that this orientation is highly variable, and can change dramatically in rough topography.

4.2 Comparing RTK–InSAR–RPAS Point-Clouds

Full InSAR time-series and cumulative RPAS deformation measurements were extracted for points locations at each of the 11 GCPs where InSAR measurements were available (Rotheram-Clarke et al. 2022). Differences in the RTK-GNSS positions were calculated to construct a displacement time-series. Displacements in X, Y and Z were projected to the RCM LoS for both the RTK-GNSS and RPAS to create an accurate basis for comparison. Coincidental time-series profiles for key GCPs across the “Solar Slump” are shown in Fig. 8a–f.

Fig. 8

(a–f) Selected RCM, RTK-GNSS and RPAS deformation time-series, projected to RCM LoS. See Fig. 6 for location of GCPs (after Rotheram-Clarke et al. 2022)

Merged orthophoto mosaics and DSMs (Figs. 4, 5 and 6) capture the baseline surface condition of the North Slide “Solar Slump”, along with the extent of bare earth and vegetation growth (e.g., grasses, shrubs and trees). Metre-scale anthropogenic features (e.g., train tracks, signals bungalow, solar panel array) are resolvable in the orthophoto mosaic and DSM.

Geomorphic features visible include: terraces graded to 300 m and 340 m asl, with steep river-cut scarps; ephemeral gullies draining the inactive nineteenth Century slide surface; active slide blocks, scarps and tension cracks across the “Solar Slump”; and a toe bulge in the active floodplain of Thompson River (Fig. 6).

The pattern of cumulative deformation in 2021 derived from RCM data is very similar to that seen in the RPAS-generated imagery. In Fig. 7, the colour stretch is +/−5 cm, indicating measurement of 5 cm to 6 cm of LoS deformation at GCPs NS-07 and NS-08 (Fig. 8c, d). There is also uplift between NS-07 and NS-12 where back-tilted clay-silt beds are exposed in the active floodplain (Huntley et al. 2021b). Future research efforts will be directed at comparing the two time-series and converting satellite LoS displacements to correspond with movement vectors derived from UAV and RTK-GNSS datasets.

4.3 Comparing Wide Area InSAR-RPAS

To examine a statistically significant number of points, 3D RPAS deformation measurements were projected to the InSAR LoS vector. This raster was then down-sampled and aligned to match perfectly the pixel geometry of the InSAR results. Only spatially coincident pixels present in both the RPAS results and the InSAR were considered, resulting in 20,843 points of comparison. The RPAS deformation represented a 2-year cumulative measurement without intermediate points in the time-series, while the InSAR data represented a temporally dense set of measurements over a one-year period.

For a scatter-plot analysis, the RPAS cumulative deformation measurements were plotted against the InSAR annual rate map with the axes scaled 2:1 and a trend line scaled to match (Fig. 9). Due to conflicts with other RCM users, the longest period between sequential InSAR pairs was 44 days. With the relatively small spatial footprint of this landslide, unwrapped phase values beyond 1 phase cycle were considered unreliable. This placed an upper bound for robust measurements of deformation rate. For this study, the bound was deemed to be 2.8 cm/44 days, or approximately 24 cm yr.⁻¹. This limit can be improved with a higher revisit frequency, and is not as much of a limitation when the spatial extents of the movement zone are larger and vary gradually. Figure 9 distinguishes points in excess of this rate by colour.

Fig. 9

Wide area direct comparison of co-located InSAR and RPAS deformation. InSAR measurements represent a best-fit annual rate, while RPAS measurements represent 2 years of cumulative deformation. Note the 2:1 scaling to account for the difference in temporal periods (after Rotheram-Clarke et al. 2022)

Figure 9 reveals a clear agreement between the RPAS and InSAR deformation trends. The cluster of points around the origin represents most of the surveyed area where the surface is relatively stable, and displays a spread in RPAS points that is approximately an order of magnitude larger than the spread of the InSAR. For deforming points, a general agreement in rate is observed up to nearly 20 cm yr.⁻¹. At higher movement rates, InSAR underestimates displacement when compared the RPAS deformation rates. This may be due to a combination of factors such as phase aliasing of rapid deformation, or a tendency for SBAS to smooth transient higher velocity deformation events. An imperfect comparison might also result from temporal overlaps between the two datasets. This is explained by a higher average rate in the period preceding the InSAR time-series, a lower average rate in the period following the final RPAS survey date, or a combination of both.

To arrive at the closest estimate of an annual rate given the sampling frequency, a histogram was constructed by scaling the 2-year cumulative RPAS deformation by half (Fig. 10). These points were differenced with the InSAR annual rate at each coincident point to create a difference metric. Across 20,843 points, the 1σ and 2σ were 0.025 m yr.⁻¹ and 0.051 m yr.⁻¹ respectively. As in the scatter plot analysis, this is an imperfect comparison because the temporal overlap of the two datasets is not consistent.

Fig. 10

Histogram distribution of differences in collocated InSAR and RPAS deformation rates (after Rotheram-Clarke et al. 2022)

5 Conclusions and Evaluation

The four landslide change-detection methods evaluated in this study have their individual strengths and limitations when applied to mapping slow-moving rotational-translational landslides. RTK-GNSS surveying is well accepted as being both accurate and precise. However, collecting data is laborious, and achieving wide area measurement coverage similar to InSAR or RPAS surveying, even at small sites, is impractical.

M3C2 is a simple and accessible algorithm to run. It is free and open source. By operating directly on point-clouds, the complexity of intermediate processing steps is greatly reduced. The displacement magnitude of the M3C2 output may appear simple to interpret, but the variability in measurement direction, and the insensitivity to surface parallel movement are major limitations for rotational and translational landslides. Since variation in movement direction is only measured on surface normals, M3C2 is the only measurement technique that cannot be directly compared with other methods.

InSAR is the only method not requiring physical site visits. RCM data acquisitions allowed for a set of measurements that were not only broad in spatial coverage, but also dense in temporal measurements. These data provided a level of insight into the seasonal dynamics of landslide movement not possible with other methods. The LoS limitation of InSAR added complexity to interpretation and satellite tasking.

Generally, we found good agreement between our InSAR and 3D RPAS methodologies for stationary and lower deformation rates. However, in parts of the landslide where the deformation rate approached approximately 20 cm yr.⁻¹, InSAR measurements typically reported lower rates than the 3D RPAS method. This may have been a result of an imperfect comparison of datasets with different measurement periods.

Combining RPAS digital image correlation with the DSM of difference was the only method able to provide a full 3D movement vector for each point with broad coverage across the landslide. Compared to M3C2, the processing requires more complexity and familiarity with a number of open source tools. However, the outputs are more directly interpretable, and capable of providing better insights into the landslide dynamics.