ILRIS 3-D scanners are typically operated through Optech's graphical controller software. To operate the scanner, the user manually defines a scan area as well as scan parameters such as optical camera setting, vertical and horizontal resolution, pulse interception (first or last) and the location to save the data. Before the data can be further processed, Optech's parser must be applied. All of the previous steps can be applied using Optech's ICL application (Teledyne Optech, 2014b). It is an executable program that reads a text file with pre-set scan parameters, runs the scanner once and outputs a ASCII-formatted point cloud (x, y, z, intensity). We applied the ICL application using task scheduling software to make it operate the automated data collection task. Our processing workflow then monitored the output folder and intercepted the incoming point cloud for further processing.

The ICL application does not apply a proprietary process known as automated scan correction (ASC), which is part of the graphical controller software. This process is normally used to compensate range and angular measurements for temperature drift within the ILRIS itself (Wang and Lu, 2009). To compensate for the lack of ASC in the ICL application, we developed our own temperature correction process described in Sect. 3.2.2.

The first step in pre-treatment is the removal of unwanted points within the point cloud. Typical change detection workflows consist of the removal of vegetation points, removal of points outside the target and removal of outlier points (e.g. Abellán et al., 2014). In our workflow no specific algorithm for vegetation removal was applied because our test area was mostly clear of vegetation and we removed the effect of vegetation on point cloud registration through a rejection scheme (Sect. 3.2.3). By including vegetation, this also allowed us to monitor changes in vegetated areas on the slope, which can be important to study the effect of vegetation on rockfall triggering, for example (Krautblatter and Dikau, 2007), or used as a means to track the 3-D displacement of the landslide using object tracking methods (Monserrat and Crosetto, 2008; Oppikofer et al., 2009).

We applied two filters to the data, a statistical outlier removal and a pass through filter, available in the PCL filter class (Rusu and Cousins, 2011). The statistical outlier removal was used to remove areas with low point densities and sparse outliers, such as artefacts from multipath or edge effects (Lichti et al., 2005). By removing these points, errors in calculating surface normals, in registering the point cloud and in change detection are reduced. The outlier removal algorithm calculates for each point the distance to all its neighbours and removes points whose distances are outside of the point cloud's global mean and standard deviation. The pass through filter is used to remove points outside of a specified target area. For example, these may include points in the foreground or background or densely vegetated areas. This is done by defining limits in each dimension where points falling outside are to be removed.

The next pre-treatment step is querying the total number of points acquired in the point cloud. If the number of points does not meet a pre-defined threshold, the entire point cloud is rejected, no output is generated and the processing is queued until the next point cloud is intercepted. The purpose of this is to remove point clouds heavily affected by poor atmospheric conditions. These clouds suffer from low point density, are difficult to register and do not produce meaningful change detection results.

The last pre-treatment step consists of an atmosphere correction algorithm and was conceived due the restriction on Optech's ASC mentioned in Sect. 3.2.1. This step was applied retroactively and has now been implemented into the system for automatic correction. Atmosphere corrections are applied as a scale factor and usually compensate for the varying speed of a laser at a given wavelength as it passes through varying refractions of air as a function of temperature, pressure, humidity, and CO2 content, e.g. Ciddor correction (Ciddor, 1996). Due to the lack of ASC, the internal system temperature drift had a larger effect on the range measurements than the refraction of the atmosphere, so we opted for a target-based correction. This correction may not be required for other scanner types that automatically apply this correction during data collection.

To conduct the atmosphere range correction, we used a network of pre-existing stable targets on the slope. The targets were measured independently using a total station during the monitoring period and showed non-significant displacement. We programmed the algorithm to automatically identify the targets based on the point cloud intensity values. The algorithm calculates the distances between the centroids of every target for the reference scan and for the target scan being corrected. The ratio of target distances of the reference scan and of the scan being corrected is then calculated. This ratio, or scale factor, was then applied to the point cloud being corrected. Application of this algorithm resulted in centimetre-level range corrections at the 1000 m range.

A registration pipeline is necessary since we cannot assume that the position and orientation of the scanner remains constant over time and because there are time-dependent measurement errors resulting from non-instrumental factors (e.g. environmental factors) that may not be accurately modelled. Even when a TLS scanner is in a fixed position, Lichti and Licht (2006) found there is a home position random bias which causes the measured position and orientation of the instrument to change over time. We found that repeated laser scanning, without moving the position of scanner, produced misaligned point clouds over different scan epochs. To decrease the overall processing time of the registration pipeline, we implemented an initial alignment stage. This provides that ICP algorithm with a better starting fit and consequently reduces the number of iterations required for convergence of the best-fit algorithm.

Time-dependent errors can also vary during a single data collection for slower scanners causing distortions of the scan. To reduce this effect, we opted to collect more frequent shorter scans so that the scans are taken with more consistent environmental conditions. To increase measurement certainty, we prefer to repeat point cloud acquisitions for this study design, rather than do repeated point measurements within the same scan, which results in point clouds that take longer to collect and are more affected by time-dependant errors. Our preference is for shorter scans to reduce distortions occurring within a scan in favour of errors in point cloud home position for scans collected at different epochs. The latter can be corrected using point cloud registration.

We designed our registration pipeline to consist of two main steps, an initial alignment stage and a fine alignment stage (Fig. 5), using the PCL registration application programming interface (API; Holz et al., 2015). The purpose of this design was to improve overall convergence time of the registration and to align clouds that are far apart, in cases where the scanner was moved, for example. In typical workflows the initial alignment stage involves manually selecting corresponding points between point clouds of successive epochs (e.g. Oppikofer et al., 2008). In our approach this is done automatically using descriptor matching (Holz et al., 2015).

The initial alignment step is performed using a subset of points known as keypoints. Keypoints consist of points in a point cloud that are both distinctive and repeatable. That is, they are unique points that can be found even if the point cloud was collected using different scanners or scan positions. To define these keypoints, we use the intrinsic shape signatures (ISS) algorithm (Zhong, 2009), which uses a pruning step to discard points with similar spreads along principal directions and includes points with large variations along each principal direction. At each keypoint we define feature descriptors using the fast point feature histogram algorithm (Rusu et al., 2009). For each keypoint, the relative orientation of normals and distances between all point pairs within a specified search radius are calculated. Correspondences are estimated between features in scans from difference epochs, using a nearest-neighbour search in feature space, using a fast approximate k-d tree neighbourhood search algorithm known as Fast Library for Approximate Nearest Neighbors (FLANN; Muja and Lowe, 2009). We use the Random Sample Consensus algorithm (RANSAC; Fischler and Bolles, 1981) to estimate the best rigid translation and rotation between the reference and data clouds completing the initial alignment. The idea of the initial alignment stage is to get the two point clouds close enough that the fine alignment algorithm converges quicker. The initial alignment stage can also successfully align points from different positions, e.g. if the scanner was moved or the orientation of the scanner changed. Parameters for the ISS pruning step, feature definition and RANSAC algorithm were empirically derived for our study case prior to the commencement of near-real time monitoring and can be found in the Supplement.

In the fine alignment stage, we use all the points in the point cloud as input to optimize the alignment. The correspondences are then trimmed down to include only stable areas using a rejection scheme. We designed our own iterative correspondence algorithm using the PCL registration API (Holz et al., 2015). The algorithm consists of an iterative process where we cycle through the following steps until a convergence criterion is met:

Matching step → find correspondences between data and reference point clouds.

For the matching step, we find correspondences from points in the reference cloud to points in the data cloud using a normal shooting method (Chen and Medioni, 1992). We use a combination of correspondence rejection algorithms applied in series to filter out poor or erroneous matches. First we apply the RANSAC algorithm to eliminate outlier correspondences, as in the initial alignment step, followed by a surface normal filter and finally by a median rejector. The application of the RANSAC algorithm within the iterative framework keeps the algorithm from converging into a local minimum (Holz et al., 2015). The normal rejector filters out correspondences that have an incompatible normal and the median rejector filters out correspondences that are greater than a factor times the median for each iteration. It thus adapts during each iteration, becoming smaller as the point clouds become more closely aligned. In the alignment step, we find optimal rigid transformation by applying the Levenberg–Marquardt nonlinear solver (Levenberg, 1944; Marquardt, 1963) to minimize the error between the reference and data cloud using a point-to-plane error metric (Chen and Medioni, 1992). The three main steps – matching, rejection and alignment – are repeated until a predefined convergence/termination criterion is met. The convergence criteria consist of a maximum number of iteration absolute transformation threshold, a relative transformation threshold, maximum number of similar iterations, relative mean square error and absolute mean square error. Parameters for the ICP algorithm and rejectors applied in this study were empirically derived prior to monitored and can be found in the Supplement.

We use a four-dimensional (4-D; space and time) algorithm described in Kromer et al. (2015b) to detect change between successive point clouds and filter random noise due to surface roughness and instrumental error using neighbourhood distance values in both space and time. We apply an empirical calibration step to subtract systematic errors that are a result of using the same reference scan for all distance calculations from the reference scan (e.g. Fig. 6). Point cloud to point cloud distances are averaged using neighbourhood distance values in space and through time. A balance between spatial and temporal averaging should be optimized for the signal being studied, as discussed in Kromer et al. (2015a), to avoid spatial or temporal smoothing of the distance values. The combined total of spatial and temporal neighbours used for averaging also determines the reduction in uncertainty of the calculated mean distance values; for example, by averaging more distance samples in space and time, the uncertainty in the mean distance value will be reduced by a factor of (Eq. 1) 1NN∗Tstep, where NN is the number of spatial neighbours used and Tstep is the number of temporal scans used for averaging.

The algorithm is described in detail in Kromer et al. (2015b). Here we summarize the main steps of the algorithms as they pertain to the near-real-time monitoring system. Each point cloud that is acquired first passes through the pre-treatment stage and registration pipeline. The initial point clouds collected are part of the calibration stage and this continues until the specified number of calibration point clouds is reached. Following the calibration phase, an accumulation phase begins. In this phase, points clouds are processed up until the number of point clouds used for temporal filtering is reached, defining the time step (Tstep). Once enough clouds have accumulated, temporal filtering begins. In this stage, for each point cloud, 4-D filtering is applied using the previous Tstep point clouds and the calibration distances are subtracted.

The 4-D algorithm calculates distances between point clouds using a slope-dependent normal, similar to that of the M3C2 algorithm described by Lague et al. (2013). Based on our experience with the system on a real slope in adverse atmospheric conditions, we made several minor changes to the 4-D algorithm's distance calculation step. In the distance calculation step described in Kromer et al. (2015b) we project a set number of points on to the local surface normal vector and take the average distance along the normal as the raw distance (Fig. 6). Here we added a limitation as to how far the points can be found away from the local surface normal vector. This limitation is a specified factor of the mean point spacing of the slope. For example, if set to a factor of 1.5, points outside 1.5 times the mean point spacing will not be projected on to the local normal vector for raw distance calculation. Additionally, to prevent averaging distances using spatial neighbours that are too distant from the target point, we apply a hybrid range and nearest-neighbour search. The hybrid approach firstly does a range search surrounding the target point, then checks whether the number of points found meets a minimum threshold. This threshold was set to (Eq. 2) 16πR2, where R is the range search radius. If the threshold is not met, a not-a-number (NaN) value is assigned to the target point. With these modifications, the number of points used to calculate the raw distance and for spatial averaging and the distance uncertainty will be variable. The spatial variability in the uncertainty is calculated using the spatio-temporal confidence interval (Sect. 3.2.5).

We define a spatio-temporal level of detection to account for errors that vary through space and time. Factors such as variable target distance (and thus footprint size), variable point density, incidence angle, variable reflectivity, atmospheric conditions and variable roughness all contribute to spatially variable errors on the slope (Lague et al., 2013). Additionally, changing atmospheric conditions, scanner temperature, slope reflectivity and misalignment errors can change through time.

Lague et al. (2013) estimated statistically significant change between two point clouds of a complex topography using a spatially variable confidence interval. We added the temporal component to the confidence interval because both spatial and temporal averaging is conducted in the 4-D algorithm. The spatial–temporal confidence interval is calculated at the 95 % confidence level and represents an estimate of distance uncertainty for a specific point in space at a specific moment in time. As in Lague et al. (2013) and Fey and Wichmann (2017), we define the confidence interval at 95 % or the level of detection at 95 % (LoD95%) to represent an estimate of the minimum detectable change.

To estimate the spatial–temporal confidence interval, we first calculate the distribution of distances using all the comparisons from the reference cloud to the calibration clouds and the distribution of distances from the comparison of the reference cloud to all the data clouds within the specified temporal averaging window (Tstep). The two distributions, reference to calibration cloud distances and reference to Tstep cloud distances, are assumed to be two independent Gaussian distributions with independent variances, variances (σcal,σdata), as in Lague et al. (2013). The two distributions have means μcal and μdata and have sizes of ncal (NN× calibration clouds) and ndata (NN×Tstep), respectively. The confidence interval at 95% (Zscore 1.96) is then calculated using a Z test formulation for the difference between means μcal and μdata for ncal and ndata greater than 30 in Eq. (3). LoD95%=±1.96σcal2ncal+σdata2ndata+σreg To account for changing systematic errors, misalignment errors and remaining errors in the total error budget, we define an empirical registration term to the level of detection estimate as in Lague et al. (2013) and Fey and Wichmann (2017). This is estimated by calculating distancing using the 4-D algorithm at stable target and assumed stable locations on the slope. The standard deviation of the distance measurement is then used for the registration term in the level of detection calculation.

In our change detection design, positional uncertainties between the reference point cloud and the true slope surface are not propagated, as all subsequent scans are registered and compared to the reference scan. For this reason, we do not include a positional uncertainty term as in Fey and Wichmann (2017). Generally, for landslide and rockfall early warning monitoring, the absolute accuracy of the distance measurement is of less importance than being able to confidently detect whether displacement or changes in displacement have occurred.

We designed the monitoring system so that both RAW and processed point clouds can be visualized in the field or through a remote connection to the field computer. This was done to avoid large data transfer to a remote server and so results could be directly visualized and interpreted in the field. To support visualization and interpretation, we store point clouds with mapped raw distances, filtered distances and confidence intervals in point cloud libraries binary pcd format. Because all distances are mapped onto the reference point cloud, we also stored all of the measured distances and confidence intervals over time in a database mapped to the index points of the reference point cloud. This allows time series of distances and confidence to be extracted by point picking on the slope. We programmed a basic point cloud visualizer using the PCL's visualization class (Rusu and Cousins, 2011). The visualizer can be initiated after each point cloud is processed. We used CloudCompare to visualize and create some of the figures in this paper.