Bioavailable strontium isotope ratios (87Sr/86Sr) distribution across the landscape mainly follow the underlying lithology, making 87Sr/86Sr baseline maps (isoscapes) powerful tools for provenance studies. 87Sr/86Sr has already been used in Eastern Canada (EC) to track food and human remains origins, or to reconstruct animal mobility. While bioavailable 87Sr/86Sr isoscapes for EC can be extrapolated from global datasets using random forest modelling (RF), no regionally-calibrated isoscape exists. Here, we produce a regionally-calibrated bioavailable 87Sr/86Sr isoscape by analysing plants collected at 136 sites across EC, incorporating updated geological variables and applying a novel ensemble machine-learning (EML) framework. We generated and compared isoscapes generated by the traditional RF and the EML approaches. Adding local bioavailable 87Sr/86Sr to a global dataset significantly improved the model prediction with a drastic increase of predicted 87Sr/86Sr and increased spatial uncertainty in the northern Canadian craton. EML produced similar 87Sr/86Sr predictions but with tighter spatial uncertainty distribution. Regionally-calibrated RF and EML isoscapes significantly outperformed the global bioavailable RF isoscape, confirming the requirement for collecting local data in data-poor regions. This isoscape provides a baseline in EC to monitor and manage the movements and provenance of agricultural products, natural resources, endangered/harmful migratory species, and archaeological human remains and artifacts.

- Data collection and analyses

We completed the original bioavailable ⁸⁷Sr/⁸⁶Sr global database from Bataille et al. (2020) with plant samples from Eastern Canada collected during two independent campaigns between 2018 and 2022: moss, lichen and grass were collected at 28 sites across taiga and tundra habitats; Balsam fir (Abies balsamea (Mill.)) and spruce needles (Picea sp.) were collected at 107 sites across the boreal forest. Samples were analysed for ⁸⁷Sr/⁸⁶Sr by Multi-Collector Inductively Coupled Plasma Mass Spectrometer (MC-ICP-MS).

- Mechanistic ⁸⁷Sr/⁸⁶Sr bedrock model

The mechanistic ⁸⁷Sr/⁸⁶Sr bedrock model (Bataille et al. 2014) predicts the modern ⁸⁷Sr/⁸⁶Sr values of the bedrock based on the age and the nature of the geological units. The model assumes that all rocks of a given lithology come from a common parent material with ⁸⁷Sr/⁸⁶Sr value changing over time following a three-stage history: (1) the initial ⁸⁷Sr/⁸⁶Sr in undifferentiated mantle, (2) the enrichment in ⁸⁷Sr through ⁸⁷Rb decay of the parent crustal rock reservoir from its differentiation to the formation of the modern rock, (3) the enrichment in ⁸⁷Sr of the modern rock since its formation.

We updated the global mechanistic bedrock model (median, 1^st quartile, 3^rd quartile) and the associated global geological unit age rasters (min, max, mean) used in Bataille et al. (2020), for the province of Québec, using SIGEOM (Système d’Information Géominière, Ministère des Ressources naturelles et des Forêts [MRNF], Québec) fine scale geological data (Thériault and Beauséjour 2012). Update of the geological data (age, nature) was done in ArcGIS 10.3. The mechanistic bedrock model was run in R 4.3 (script provided). We provided the updated geological age rasters (min, max, mean) and the predicted ⁸⁷Sr/⁸⁶Sr from bedrock (median, 1^st quartile, 3^rd quartile), used in the random forest and ensemble machine learning analyses to produce the bioavailable ⁸⁷Sr/⁸⁶Sr isoscape for Eastern Canada.   

- Bioavailable ⁸⁷Sr/⁸⁶Sr isoscape

We generated the bioavailable ⁸⁷Sr/⁸⁶Sr isoscape for Eastern Canada using the global bioavailable ⁸⁷Sr/⁸⁶Sr sample dataset from Bataille et al. (2020) updated with samples from Eastern Canada and geological, topographic, climate, soil, and aerosol deposition variables extracted at each sampling sites.

We used two machine learning approaches: a random forest analysis and an ensemble learning machine analysis. The random forest grows multiple regression trees by bagging : for each tree the dataset is divided randomly by bootstrap into a training set and a validation set. Predictions are then aggregated from all trees. For the ensemble machine learning, we used the stacking approach: several different algorithms are trained on the dataset and the predictions of these models are used to train a meta-learner, generating the ensemble machine learning estimates.

Two isoscapes were generated with the random forest approach: the first one uses the global ⁸⁷Sr/⁸⁶Sr dataset without data from Eastearn Canada and the other one integrates Eastern Canada samples. The ensemble machine learning analysis use global ⁸⁷Sr/⁸⁶Sr dataset with the Eastearn Canada data.

Analyses were conduct in R 4.3 (script provided) and we provide the 87Sr/86Sr prediction rasters as well as the associated spatial error (standard deviation) rasters for both approaches.

- References:

Bataille, C.P., Brennan, S.R., Hartmann, J., Moosdorf, N., Wooller, M.J., and Bowen, G.J., 2014. A geostatistical framework for predicting variations in strontium concentrations and isotope ratios in Alaskan rivers. Chemical Geology, 389, 1–15. https://doi.org/10.1016/j.chemgeo.2014.08.030

Bataille, C.P., Crowley, B.E., Wooller, M.J., and Bowen, G.J., 2020. Advances in global bioavailable strontium isoscapes. Palaeogeography Palaeoclimatology Palaeoecology, 555, 109849. https://doi.org/10.1016/j.palaeo.2020.109849

Thériault, R., and Beauséjour, S., 2012. Geological map of Québec, DV2012-07. Ministère des Ressources naturelles et des Forêts, Québec.