Prior to starting my PhD at the University of Leeds I was fortunate to be given the opportunity to undertake a Masters by Research. Although its origins began in looking at the geomorphic and biotic effect of in-stream wood using 3D high resolution topographic modelling, my main efforts drifted more towards quantifying whether or not Structure-from-Motion (SfM) could be used to accurately estimate grain size at patch scale for a number of patch facies. Many people have experience undertaking Wolman sampling strategies to estimate grain size and hence many people know it is a relenting, time consuming and occasionally painfully dull process. For your efforts, you get an estimated grain size from one hundred pebbles across a selected area, not the most reliable! Instead, SfM could provide a time efficient, easy and data rich approach to estimate the size of every pebble in the same area selected for Wolman sampling and give an average of grain size within the area. Using SfM, or similar topographic survey methods like Terrestrial Laser Scanning (TLS) to estimate grain size requires a relationship to be established between grain size and the surface roughness. To do this, a parameter for surface roughness needs to be identified, which in previous research, has most commonly been standard deviation of elevation. This relationship is complicated by patch characteristics, for example, whether there is a variation in grain size within the patch, whether there is imbrication, how large grains are packed with smaller grains and the shape of each grain. These factors are also limiting when trying to use a summary value of particle axes such as D50 to measure flow resistance. Understandably, the surface cannot be defined fully by a few descriptive parameters, so as much as using SfM would speed up the process, there will still be a little uncertainty behind the values. However, previous research using SfM or TLS to estimate surface roughness has not quantified the effect any patch characteristics have on surface roughness values. This is where I focused my research efforts, in evaluating the differences patch characteristics have on surface roughness, and where the pebble measuring began.
Variation in the relationships between standard deviation of elevation and grain size in previous research
Measuring so many pebbles takes a team, mostly to stop yourself from going insane. Our approach was two-pronged, firstly, we brought 2000 pebbles with b-axes between 4 and 128 mm back from the field site in the Yorkshire Dales, gave them unique numbers, and measured all three of their axes. 6000 axes measurements later, and you can see how the number gets so high! Using these pebbles, I arranged small patches within a tray, georeferenced using a total station, depending on size, shape, imbrication and sorting. The measured b-axis D50 of each patch was related to the standard deviation of elevation calculated from SfM models created of each patch and relationships were established based on combinations of shape, sorting and imbrication. The pebble size differences were used to create the relationships. These relationships were compared with each other and to a relationship founded by using smooth, spherical balls, reducing the implications of patch characteristics to a minimum, and to the theory of how diameter would differ standard deviation of elevation for hemispheres.
A sad looking pebble, it must have known my feelings after measuring so many being one of the last of the 2000 collected pebbles
Understandably, manually manipulated patches are not representative of real world gravel bars or river beds. Therefore, the second part of the analysis involved two field sites, the original one where the pebbles were collected and another, downstream where the gravel was better sorted. Here, thirty-one 1m x 1m patches were established, and the corners were surveyed using the total station. An SfM photo survey was undertaken, and the surface was spray painted (so only surface gravels would be collected). Once this was complete, the measuring began once again, where all surface gravels had their three axes measured. Again, the D50 of each patch was compared to the standard deviation of elevation derived from the SfM point cloud. These field patches were split into well sorted and poorly sorted, and were compared to the manipulated patches of the same type.
A contrast in weather- mid 2016 summer heatwave versus the typical set up nearing the end of the pebble measuring experience on a very cold February day
A flow diagram of how the different patch types relate to each other
An issue with field patches was the underlying topography skewing the derived standard deviation of elevation. We investigated this by removing the form roughness through a variety of methods including fitting a plane, fitting a 2.5D quadric, and using 7.5cm and 10cm DEMs, then differencing the point cloud from them. We also looked at whether using a different metric for roughness had an effect, these metrics included summarising local standard deviation using a moving window (2σloc), and summarising the deviation from a fitted plane of all points within a given radius (rh), both of which had been used in previous research by other authors.
Our analysis showed that using SfM as a survey technique for surface roughness can provide accurate grain size estimates. Of the patch facies tested, sorting had the largest effect on whether or not SfM could accurately estimate grain size, with the relationship for manipulated patches having a reduction in R2 from 0.97 to 0.012, for well sorted and poorly sorted spherical gravels respectively. Particle shape effects were particularly pronounced from oblate-shaped particles, as to be expected due to their flat nature. Imbrication effects caused an increase in standard deviation of elevation for oblate patches, but not for the prolate patches, as the imbrication effect was balanced by the obscuring of the lower end of the particles, reducing topographic variability. In terms of form roughness, only minor improvements were seen in the use of DEM-point cloud subtraction for the moderately well sorted field patches, with no improvements being seen to the poorly sorted field patches. Regarding roughness metrics, all three tested were very similar, and were highly correlated.
To summarise, SfM or other high resolution surveying techniques may provide a method for estimating grain size, and for the first time, this paper looked at the effects different patch facies have on the ability of the technique to estimate grain size accurately. Previous work has used roughness-grain size relationships with little regard for factors that may be influencing the patch roughness and thus, many different relationships were found. Finding a method to quickly but accurately extract grain size information would be of great use to those working in river management, particularly now river restoration is highly prevalent. This research aimed to be the first to consider patch variation, but has not answered all questions, and thus, currently, work still needs to be done before SfM can be used to estimate grain size. Issues around further patch variation factors such as burial and hiding effects, as well as the interactions between patch factors need to be examined. Full validation of the relationships found will then be required, with data from many more field patches. This has the possibility of leading to a suite of empirical equations of use to river practitioners (see the example created from work done in this project), with the process of identifying the correct equation having the potential to be fully automated.
The decision tree of empirical equations found in this study as an example of what may be possible in the future
If this work is of interest, please look at the more detailed paper in Geomorphology which has been granted open access, and can be found using the URL:
Pearson, E., Smith, M.W., Klaar, M.J. and Brown, L.E. 2017. Can high resolution 3D topographic surveys provide reliable grain size estimates in gravel bed rivers? Geomorphology. 293(A), pp. 143-155 Continue reading