Fig 1.
Locations of samples used for developing and validating the bioavailable 87Sr/86Sr isoscape.
Fig 1 developed in ArcGIS Pro and includes regional boundaries and coastlines feature layers (sourced from Natural Earth) and hillshade feature layer (created using GNS DEM 8m shapefile). Map projected to the New Zealand Transverse Mercator (NZTM 2000).
Fig 2.
Stacked plot illustrating the distribution of the collected 87Sr/86Sr variability by substrate.
The animal substrate (n = 8) samples in pink display a tighter range of 87Sr/86Sr values compared to the plants in green (n = 314) and topsoils in blue (n = 92) samples that are all used to construct and calibrate the bioavailable 87Sr/86Sr isoscape model.
Fig 3.
Bioavailable 87Sr/86Sr isoscape (A) and interquartile range raster (B) for Aotearoa created using random forest regression. The bioavailable 87Sr/86Sr isoscape (A) (R2 = 0.53, RMSE = 0.00098) demonstrates the predicted 87Sr/86Sr values, ranging from 0.70567 to 0.71118, for the entire country including the Chatham Islands. Log-transformed 87Sr/86Sr values of the bioavailable data were used to construct quartile-1 and quartile-3 regression models that were then subtracted from one another (Q3-Q1) to create the final interquartile range raster (B) with values ranging from 0.0001 to 0.002. Figure developed in ArcGIS Pro with a coastlines feature layer (sourced from Natural Earth) and projected to NZTM 2000.
Fig 4.
Random forest regression performance plots.
Plots demonstrate (A) model variable importance plot and (B) n-fold cross validation with best fit line in red.
Fig 5.
Partial dependence plots with predictive variables (x-axis) and predicted bioavailable 87Sr/86Sr (y-axis).
Units of measurement are provided on the x-axis for each variable. For r.toprock, the x-axis represents the numbered categories of rock type (S2 Table 2 in S2 File and S2 Fig 1 in S2 File) obtained from the New Zealand Land Resource Information System online portal. Refer to the Supplementary Information, S2 Table 1 in S2 File, for descriptions of all variables used in the RF model and their sources.
Fig 6.
Observed cow milk 87Sr/86Sr (x-axis) versus predicted 87Sr/86Sr values (y-axis) from the bioavailable isoscape.
Error bar values were extracted from the uncertainty raster for each cow milk sample location. The red line represents the line of best fit and the shaded region represents the 95% confidence interval (CI). The black line shows the 1:1 line between predicted and observed 87Sr/86Sr values. Sample points are color-coded by region (regions depicted in Fig 1). All cow milk samples fall within the 95% CI, except sample Cow 7 labeled on plot.
Table 1.
Cow milk sample data.
Table 2.
Descriptive statistics for cow milk top 33%, 20%, & 10% quantile predictions.
Fig 7.
Diagrams showing simplified (A) basement terrane geology and (B) bioavailable 87Sr/86Sr isoscape highlighting regions with low, intermediate, and high 87Sr/86Sr values. The geology in (A) follows Edbrooke et al. [117]. The Alpine Fault is shown as a solid black line; the division between Eastern Province and Western Province basement rocks is shown by a black dashed line; and the Taupo Volcanic Zone (TVZ) outline is shown by a dotted red line. Low, intermediate, and high Sr value ranges for simplified isoscape (B) were determined using the “Equal Intervals” symbology function in ArcGIS Pro. Black denotes regions of low 87Sr/86Sr values (0.70570–0.70750), dark grey denotes intermediate 87Sr/86Sr values (0.70750–0.70930), and light grey denotes regions with high 87Sr/86Sr values (0.70930–0.71120). Figure developed in ArcGIS Pro using a coastlines feature layer (sourced from Natural Earth) and projected to NZTM 2000.