Point cloud processing

Each scan consisted of a cloud containing up to 107 data points, with a mean point spacing of 2x10-3 m. A DEM of Difference (DoD) method was used to calculate cliff change over each scan interval, following (Rosser et al., 2005). A simplified data processing workflow to produce erosion statistics from the raw data is shown in Figure 4.2. The processing steps and data quality and uncertainty are explained in detail in following sections.

Figure 4.2: Workflow of TLS data processing steps. The left column contains the software packages used; the right column outlines the processing sequence within each software package.

4.2.2.1. Point cloud matching and quality control

The raw point cloud data were imported into RiScan. Some sections of each scan, such as the distal points and upper section of the cliff (>20 m above the cliff toe) were deleted at the outset to improve processing times in the following stages by reducing the point cloud size. The first scan at each site was reoriented in a local coordinate system whereby the cliff face was normal to the Y-axis, cross-shore parallel to the X-axis and vertically parallel to the Z-axis. This allowed the subsequent data to be projected with the cliff face itself parallel to the plane [x, y], creating a DEM where the cell values were depth normal to the cliff.

Each successive scan was then manually registered to the previous scan using a minimum of four common reference points on each scan. This minimum was used to constrain all degrees of freedom in a 3D rototranslation, enabling a fit error to be derived. The relatively small surface change between scans over the total scan area meant large areas of the cliff remained unchanged between scans, allowing reference points to be easily located. Once the scans were approximately aligned, points were removed from outside the area of interest and erroneous points, such as birds in flight, raindrops and the shore platform, were further trimmed (Figure 4.3A).

I undertook the final alignment of point clouds using RiScan’s multi-station adjustment (MSA) tool. To initiate the scan matching, a plane patch filter was used to identify planar surfaces within known constraints. For comparison with previous studies, the parameters are outlined in Table 4.2. The location and normal of the resultant surfaces were matched using a multi-station adjustment (MSA) algorithm. This is a considerably more precise means of aligning scans compared to using the entire point cloud, and is less prone to error resulting from the influence of variable point-cloud density and small-scale noise in the data (Haas et al., 2012).

During MSA, scans were iteratively matched to a maximum root mean square error separation (RMSE) of < 5 x 10-3 m (

Table 4.3). The proximity of the scanner to the cliff and the high point-cloud density from these scans meant that this a considerable improvement on the ~10-2 m achieved in other recent studies at this site (Rosser et al., 2013), and hence is an appropriate resolution to investigate smaller-scale change. After visual confirmation of matching, each scan was exported as a set of coordinates with a precision of 1 x10-3 m to an ASCII file.

Table 4.2: RiScan plane patch filter parameters for comparison with other studies

Parameter Value

Maximum plane error (m) 0.01

Minimum points per plane 10

Minimum search cube size (m) 0.016

Maximum search cube size (m) 2.048

Table 4.3: RiScan multi-station adjustment MSA) parameters for comparison with other studies

Parameter Value

Start radius: first iteration (m) 1

Start radius: last iteration (m) 0.2

Maximum tilt angle: first iteration (°) 10 Maximum tilt angle: last iteration (°) 2

Minimum change of error 1 (m) 0.25

Minimum change of error 2 (m) 0.25

Outlier threshold 1

Final standard deviation error (m) < 0.005

A number of sources of error were introduced during the data collection and processing of TLS point cloud data. During data collection, the return strength of the laser is affected by a number of surface factors including wetness, colour and mineral constituents of the target cliff face, alongside weather and

atmospheric conditions (Abellán et al., 2014). Surface morphology is also important, particularly where uneven, rough surfaces scatter signal reflections and protruding features obstruct the cliff face behind (occlusion) (Schürch et al., 2011). As such, uncertainty was approached statistically and is outlined in Section 4.2.2.3.

Changes in deposits at the cliff toe (e.g. boulders and cliff talus) between successive scans proved problematic. False erosion events were produced when mobile material was present at the cliff toe in one scan but was removed before the next. To combat this, I clipped out the lower sections of some scan areas, which produced non-rectangular cliff maps. I also clipped scan edges, as surfaces which were angled such that the laser beam has a high incidence angle also produced unreliable and inconsistent returns and hence point positions.

4.2.2.2. Raster and DoD generation

Rasters of each point cloud were generated in the geospatial imagery processing software ENVI v5.3. The imported ASCII point clouds were converted to a triangular irregular network using Delaunay triangulation, then linearly resampled onto a raster at 5 x10-3 m resolution (Wheaton et al., 2010). Here, the cliff face was re-projected on to the plane [x, y] (Figure 4.3B). This resolution was chosen such that the mean point density across the sites is higher than one point per cell. Rasters were stacked and resampled to ensure that the grid and cell extracts were coincident.

To calculate change through time, I subtracted each raster image from the previous scan, to create a DoD. In addition, the total change over the survey period from the first and final scan was calculated using the same approach. It is possible that some cumulative error may have accrued during matching with successive scans, but using the same unchanged matching features for each scan minimised this error source. Uncertainty in volumetric changes can arise from the interpolation methods used when gridding the data. J. G. Williams et al., (2017) reported error as a function of detachment area and perimeter length, where higher perimeter to area ratios increase uncertainty. However, a sufficiently conservative estimate of the minimum change detection depth

(Section 4.2.2.3) provides sufficient precision to observe the volumetric changes at this scale alongside robust management of the data uncertainty. The RMSE separation between the matched scan surfaces for each time-step was below the grid resolution (5 x 10-3 m) and therefore fulfilled this requirement (Lim et al., 2010).

Slope and hill-shade images were obtained from the DEM using a topographic modelling function (Lim et al., 2011). These were used to generate cliff images over which to drape the erosion data to enhance context when viewing. These rasters were exported in ERDAS Imagine *.img format.

4.2.2.3. DoD detachment statistics

In ArcGIS, I imported the DoDs and clipped them to the final area of interest (AOI). This extended up to 10 m above the cliff toe and as wide as the data quality at the scan edges would allow, usually 25 – 35 m of cross-shore distance at the cliff toe. Changes were classified as either positive or negative, with the former representing surface apparent accretion seaward. These were removed as they were confirmed visually to be attributable to bird nests, localised cliff talus accretion and boulder movement between the scanner location and the cliff toe.

In order to address the uncertainty and determine the precision threshold above which change in cliff morphology could be detected, I identified a segment of cliff which was unaltered throughout the survey period. The DoD from each scan interval (Figure 4.4A) was then plotted as overlain kernel density functions to identify the error between each scan. The combined function (Figure 4.4B) for successive scans from all intervals forms a logistic distribution with a modal value of ~0 m. These errors were stable and persistent over time, implying systematic range uncertainty from the scanner (Figure 4.4C). Following Abellán et al. (2009), twice the standard deviation of error equals 0.006 m, thus quantifying scanner accuracy. Negative change was masked below 2 x10-2 m, so can be considered a conservative estimate of the minimum threshold. This approach was undertaken to maximise the precision of pairwise change detection between successive scans, particularly where global registration precision was less important (Schürch et al., 2011). With a raster grid resolution

of 5 x10-3 m and a minimum event depth of 2 x10-2 m the smallest detectable detachment volume was 5 x 10-7 m3 (0.5 cm3).

Individual rock detachments were delimited by polygons and the volume of material removed in each was calculated (Figure 4.3C). Low point densities, zones of the raster where point density was lower than the grid resolution, occurred predominantly where occlusion produced poor matching between successive scans. These areas were masked using an 80° threshold topographic filter. Such areas constituted a small (< 1%) percentage of the AOI.

Figure 4.3 (previous page): Point cloud processing stages. A) Raw point cloud of the lower cliff at Site 5 coloured by reflectivity (green = low, red = high). Black areas are occluded zones where point density is low. B) Detail of rasterised xyz coordinates from the point cloud in ENVI gridded at 0.005 m resolution, shown as topography with shaded relief. The fissile, flaky shale texture and some joint surfaces are evident (1). C) Detail of difference raster for the full annual surface change coloured by erosion depth draped over surface topography from the final scan. Two classifications of detachment can be seen: the larger central event represents a discrete block removal (2); smaller, shallower and more frequent events show shale fragmentation (3). A low point density, high-angle slope removed in the slope filter can also be seen (4).

Figure 4.4: A) Sample of an unchanged cliff section of the Site 2 DoD between March 2016 and April 2016. B) Combined kernel density function of the pixel error between sequential scans in the unchanged cliff section shown in A. C) Time series of modal error across all scans at Site 2 (solid line) alongside twice the standard deviation error of 0.006 m (dotted lines).

To calculate the vertical distribution of erosion at the lower cliff, the scan area of each site was partitioned into vertical bins of 0.1 m relative to ODN. The centroid of each detachment was taken and binned according to its elevation. Rock detachment volume (Dv) and frequency (Dq) were calculated for all

erosion events within each elevation bin. Since each site scan area does not have a constant width, all values were normalised for scan width for each elevation bin.

4.3. Results and analysis: Erosion of the lower cliff