Complementary Precipitation Isotope Models Reveal Young and New Water Fractions in New Zealand Rivers

Improved knowledge of catchment transit times can enhance our understanding of hydrological and biogeochemical processes occurring in the subsurface, and thus prediction of catchment responses to land use and global change. However, a lack of suitable precipitation tracer data in many areas worldwide hinders transit time assessment at large spatial scales. We evaluated variation in young water fractions (F_yw—the fraction of streamflow less than 2.3 ± 0.8 months old) across 79 New Zealand catchments representing ~46% of New Zealand's total river discharge. F_yw was calculated using two precipitation isotope models: a seasonal kriging model and a machine learning model (PINZ) trained to predict seasonal and non-seasonal variation. We also used data from PINZ to estimate monthly new water fractions (F_new—the fraction of streamflow derived from precipitation that fell within the past month) using ensemble hydrograph separation. Our primary goal was to assess the reliability of the two precipitation isotope models for transit time estimation across New Zealand, where non-seasonal variation in precipitation stable isotope values is prominent in some regions. To evaluate whether F_yw and F_new derived from the two precipitation isotope models captured meaningful variation in catchment hydrology, we tested for consistency of their associations with two established predictors of storage in the subsurface: catchment geology and baseflow recession constants. Our results across all sites indicate that an average of 18% of river flow was younger than ~2.3 months, with 11% younger than 1 month. F_yw and F_new were similarly related to catchment geology and gradients of baseflow recession constants. Kriging provided more accurate F_yw estimates than PINZ, which tended to underestimate extreme precipitation isotope values, leading to overestimates of F_yw. However, the PINZ model offered reliable estimates of F_new when robust estimation was used to reduce the influence of outliers; this held for sites where seasonal cycles were poorly defined, highlighting the potential for machine learning precipitation isotope models to support transit time estimation in regions with weak seasonal isotope cycles (e.g., in tropical or marine climates).