A PROVENANCE STUDY OF FRENCH LIMESTONE BASED ON VARIABLE SELECTION FROM COMPOSITIONAL PROFILES

The present study shows how multivariate analysis and variable selection techniques can be used in archaeological provenances studies to improve classification performances. Neutron activation analysis (NAA), in combination with stepwise linear discrimination analysis (SLDA) (capable of simultaneously performing variable selection and classification), was applied to differentiate among limestone samples from different quarries across the north of France based on their compositional fingerprint. A hierarchical classification approach was followed, aimed at progressively assigning limestone samples to more specific sources of origin, from the broadest classification units (French regions) to the narrowest ones (individual quarries). The application of the stepwise variable selection procedure to extract the most discriminating compositional variables prior to each classification development allowed us to obtain a perfect separation between the limestone categories considered at every classification stage (all samples were correctly classified and predicted in all cases). The high‐quality results obtained were even more remarkable considering the relatively small number of significant variables selected in each case using the SLDA method. An illustrative example was provided in order to demonstrate that the classification strategy proposed can actually allocate an unknown sculpture to a particular quarry of origin.