Determining the provenance of limestone in historical contexts is an ongoing methodological challenge, especially for micritic limestones. Micrites (limestones composed solely of lithified carbonate mudstone) lack diagnostic features such as fossils, grains or sedimentary structures, making provenance determination particularly challenging.

We have addressed this problem by analysing Roman stone products. Using traditional petrological methods, we were previously unable to determine the provenance of 72 % wares due to the presence of micrites. With the advances in statistics, powerful statistical learning methods have been developed for a variety of purposes - from predicting cancer recurrence to stock market movements or weather forecasts. Advanced statistical methods were applied in our case to classify the provenance of micritic limestone.

We have shown that the accuracy of origin determinations improves when the number of variables is reduced and the model hyperparameters are fine-tuned. We further demonstrated that variable importance rankings in statistical classifiers can identify the most informative variables for determining provenance; in our dataset, the most useful ratio is 87Sr/86Sr. We also provided guidelines for the use of statistical learning in provenance classification. The approach is not limited to limestone. It can also be adapted to other rock types (e.g. dolomites, hornfels, gypsum or marble) and other materials (e.g. tesserae, ceramics and mortars).

Summarised from the article BRAJKOVIČ, Rok, KOSELJ, Klemen. Statistical learning improves classification of limestone provenance. Heritage. 2025, vol. 8, no. 11, 21 p. ISSN 2571-9408. DOI: 10.3390/heritage8110464. [COBISS.SI-ID 256913155]