Quantifying user variability in the traditional X‐ray ice‐rafted‐debris counting technique

Abstract

The flux of icebergs is most commonly constrained in the palaeo‐record by manually counting gravel clasts in sediment core X‐ray images—a trusted method in Quaternary science and palaeoceanography for nearly 40 years. In this study, we explore the reproducibility (user consistency and between‐user variability) of the traditional clast‐counting technique, something that has rarely been quantified. A participatory study was undertaken whereby 14 individuals counted gravel clasts (>2 mm) on X‐radiographs in cm intervals of two different glaciomarine sediment cores: (i) a matrix‐supported subaqueous diamicton facies and (ii) a clast‐poor to clast‐moderate laminated facies. Pairwise comparisons of all 14 users' counts using Spearman's rho indicate statistically significant relationships in all but one case (p < 0.05). A strong correlation (rho = 0.895; p < 0.001) was also identified when comparing individual users' average clast counts on both images, indicating high internal counting consistency by users but significant differences in terms of overall counts. Substantial variation was also observed between different users' counts on the same image. In the diamicton facies, the standard deviation of all users' counts per cm was 1.91 (range: 0–4.3), while in the laminated facies, the standard deviation per cm was 1.28 (range: 0–3.67). Additionally, several users recorded a lower average clast count in the clast‐rich diamicton facies and a higher count in the clast‐ poor laminated facies relative to other participants. These results may have significant implications when it comes to interpreting past iceberg flux and ice‐marginal dynamics. As a way forward, we present a new rigid protocol to improve transparency, consistency and reproducibility for studies implementing this technique. Our findings have implications regarding how manual clast counts should be interpreted and implemented but also strengthen the case for automated clast‐counting algorithms to reduce user bias and improve the consistency of IRD detection using X‐radiographic images.