GSA Connects 2021 in Portland, Oregon

Paper No. 91-2
Presentation Time: 9:00 AM-1:00 PM


MASON, Aria and COX, Ronadh, Geosciences Department, Williams College, 18 Hoxsey St., Williamstown, MA 01267

PebbleCounts (Purinton & Bookhagen 2019) is a Python-based clast-counting algorithm that was developed for analysis of gravel river beds. We refined the PebbleCounts parameters for use with boulder populations, and tested it at the Minard boulder beach, County Kerry, Ireland.

From digital orthoimages (which we generated using Agisoft Metashape), PebbleCounts identifies edges and delineates clast outlines. It has two modes: semi-automated KMS, "k-means analysis with manual selection" (user decides which identified grains to include in the count); and fully automated AIF "automatic with image filtering" (user sets parameters and algorithm does the rest). PebbleCounts was optimised for smaller gravel (pebble size range) that had substantial colour contrast with matrix sediment. Thus we found that several default parameters—controlling denoising, edge detection, and the lower size cutoff (in pixels) for included grains—needed to be changed so that the algorithm could provide accurate clast detections at larger sizes as well as in a lithologically uniform population.

PebbleCounts unsurprisingly generated more accurate population statistics than a control manual point count (by removing the inherent bias toward coarser clasts). The KMS and AIF methods produced comparable outputs. The AIF runs captured 10x more clasts than KMS from the same field of view (103 vs. 102). Population D50 were very similar, but the AIF captured many more very large and very small clasts, thus better defining the fine and coarse tails of the distribution. Folk & Ward measures returned "well sorted" for the KMS and "moderately well sorted" for the AIF. Visual inspection of the AIF picks showed that the algorithm made correct delineations of the smallest clasts (6 cm pebbles in this instance). It is the inclusion of these pebbles (as well as the inclusion of a greater number of large individual boulders at the coarse end) that results in the less-well-sorted classification for the AIF, and post-analysis inspection shows that this represents the population well.

One caveat, and it applies to all 2D grain population analyses, is that not all clasts were oriented with their X-Y planes parallel to the image plane. In particular, some smaller clasts may have settled into larger intergranular spaces with their X axes near vertical; thus the Z axis may in some cases have been misidentified as the Y axis. This initial result sets a basis for ongoing analysis of boulder-beach clast populations, which will be tested in coming field seasons.