Precipitation Isotopes New Zealand (PINZ): improvements in precipitation isoscapes with machine learning

Environmental context Stable water isotopes δ18O and δ2H in atmospheric precipitation are valuable hydrologic tracers with varied applications, but measuring them over large scales is impractical. Precipitation isotope models (isoscapes) are instead needed to predict isotopes through space and time. Precipitation isotopes are affected by numerous variables ranging from orographic effects to phase changes to moisture mass origin, making them challenging to predict in complex geographies such as New Zealand. Improving data density together with recent advances in modelling approaches hold promise for improving accuracy of precipitation isoscapes in New Zealand and similar locations. Rationale Precipitation isotope models based on global networks, such as the Global Network of Isotopes in Precipitation (GNIP), lack sufficient spatial and temporal resolution for hydrological applications in New Zealand due to sparse data coverage. Models driven by regional datasets are needed to improve accuracy of isotope predictions to increase utility in environmental research applications. Methodology We develop Precipitation Isotopes New Zealand (PINZ) δ18O and δ2H isoscapes employing the Extreme Gradient Boosting (XGBoost) machine learning algorithm driven by a regional precipitation isotope database with high station density across New Zealand, and diverse climate and geographic predictors. We produce a national monthly precipitation isotope model at 5-km resolution. Model uncertainty associated with monthly predictions is evaluated using ‘leave one out’ cross validation (LOOCV, or jackknife) tests. Results PINZ shows significant spatial and temporal variability in isotope values, with notable patterns related to New Zealand’s complex topography and climatic influences. Model performance on test data is described by a root mean square error of 1.83‰. Discussion Although the model performs well overall, it exhibits limitations in predicting extreme precipitation δ18O values (most enriched and most depleted) and predicting δ18O of precipitation during periods of extreme weather conditions. Leveraging detailed local data and advanced algorithms, we demonstrate one way forward for improving isoscapes in complex environments.