A probabilistic approach to provenance studies using whole object elemental composition: Chert (flint) as a case study

Provenancing archeological artefacts and “whole rock” fragments, such as cobbles and pebbles, is important both for studying past societies and reconstructing geomorphological and sedimentary processes. Determining the provenance of archaeological artefacts, for example, is helpful in the study of various aspects of human society such as mobility, trade routes, raw material procurement strategies and societal development, whereas whole rock fragments are useful indicators of sediment transportation routes and landscape evolution. In this paper, we introduce a novel probabilistic approach for exploring the limitations and potential of provenance studies utilizing trace element composition of archaeological artefacts and bulk rock fragments through a set of statistical analyses, demonstrated using a dataset of chert (flint) samples from the eastern Mediterranean as a case study. Because our general goal is to identify sediment provenance at the highest possible resolution, and attribute it to specific outcrops, the robustness of such classifications must consider the inherent variability of the studied samples, which poses a complex problem when an entire suite of elements is considered. In the current study, a scheme for evaluating the possible resolution and potential of provenance studies of flint artefacts is illustrated by applying a set of statistical techniques. The data is investigated across the different labelling hierarchy using Spearman correlation and the Kruskal-Wallis test, and a synthetic dataset that follows the statistical properties of the original data is generated using multidimensional copulas that mimic the correlation of the different elements within each subset. The synthetic dataset is then iteratively sampled, filtered and passed into a classifier that is evaluated using a set of classification metrics (e.g., accuracy, recall, etc.). Using this probabilistic approach, the power of supervised classification is determined for the different hierarchy levels and classes based on their distributions, and the number of specimens required for distinguishing sources at different levels (e.g., era, age, formation and site) is evaluated. We show that 10–20 specimens are needed to determine the geological era and age of the strata from which flint was collected with reasonable certainty (accuracy >0.9), whereas specific outcrops cannot be reliably identified even when a large number of specimens are considered. This approach has the potential to enable the classification of items of unknown origin and the associated uncertainty by comparing them against a dataset of possible sources that can be expanded over time.