Comparing machine learning isoscapes of 87Sr/86Sr ratios of plants on the island of Sardinia: Implications for the use of isoscapes for assessing the provenance of biological specimens

Strontium isotope analysis is widely used to evaluate the provenance and mobility of biological specimens. Frequently applied in archaeology, palaeontology, ecology, forensics, and food science, strontium isotope analysis compares the ⁸⁷Sr/⁸⁶Sr ratio of a specimen against a strontium isoscape - a representation of expected ⁸⁷Sr/⁸⁶Sr ratios across a landscape - to identify areas that are more and/or less likely to be the source of the specimen. Strontium isoscapes are built using different methods, but all approaches start with empirical ⁸⁷Sr/⁸⁶Sr ratios sampled from areas with known coordinates and use them to assign likely ⁸⁷Sr/⁸⁶Sr ratios to unknown areas. Following the publication of Bataille et al., 2018 and Bataille et al., 2020, machine learning using a random forest algorithm has become a common method of producing strontium isoscapes. Despite the recognition that this method requires training with local ratios, especially in geologically complex regions, very little work has evaluated machine learning isoscapes' accuracy. This study compares and evaluates two previously published machine learning isoscapes of Sardinia against new empirical data provided by the project ZANBA. The ZANBA data is then used to create a third machine learning map of Sardinia, which is tested against previously published empirical data. The three isoscapes show different levels of predictive accuracy, with more primary data points leading to more correct predictions. However, a densely sampled landscape did not create an isoscape that gave substantially more accurate predictions than a moderately densely sampled landscape when tested against primary data from outside the original sampling areas. Areas of an isoscape with low root mean squared error (RMSE), which is often interpreted as indicating accuracy, did not necessarily give more correct predictions. Finally, a machine learning isoscape of Sardinia that incorporates both new and previously published empirical data is presented.