High Resolution Machine Learning Derived Global Monthly Isoscapes for Stable Water Isotopes with Pixel-Level Uncertainty Intervals

Stable oxygen and hydrogen isotope ratios in precipitation (δ¹⁸O, δ²H) are powerful tracers of water cycle processes. However, the construction of global isoscapes (i.e., gridded maps of isotopic composition) is limited by sparse and clustered station networks and by simplified assumptions about prediction uncertainty, reducing their reliability in hydrological and ecological applications.

Here, we present the development of global 0.04° (~ 4 km) monthly precipitation isoscapes using gradient-boosted trees trained on ~1900 stations (1962-2024) and > 15 environmental predictors including climate variables, topography, and regional circulation patterns. Spatial independence is ensured through geographically stratified cross-validation. Leave-one-region-out sensitivity tests demonstrate robust generalization to unsampled regions, while at the same time highlighting the importance of regional fractionation controls, that cannot be captured without adequate spatial coverage.

To quantify prediction uncertainty, we combine bootstrap ensembles with quantile random forests calibrated on out-of-fold errors. This approach achieves ~60-65 % empirical coverage of independent test stations, which is more than double compared to conventional bootstrap intervals (~24 %) and approaches the nominal 68 % target. Calibrated uncertainty maps dynamically highlight regions with sparse data or complex climate, while well-sampled regions show significantly lower uncertainties.

These spatially adaptive, calibrated uncertainty intervals combined with demonstrated transferability to unsampled regions enable downstream applications that require actionable confidence information. To our knowledge, this represents both the first application of machine learning to derive global monthly isoscapes for δ¹⁸O and δ²H, and the first framework providing explicitly calibrated, high resolution, pixel-level prediction intervals with validated transferability.