Predicting spatial distribution of stable isotopes in precipitation by classical geostatistical- and machine learning methods

Summary

This methodological study evaluates geostatistical and machine learning approaches for predicting spatial distributions of stable isotopes in precipitation across Europe. The combined random forest regression method (MRRF) achieved the lowest mean squared error (2.61) and outperformed traditional regression kriging by 7.5% in hybrid error metrics. Machine learning methods demonstrated superior performance in maintaining prediction accuracy with reduced station network density, offering practical advantages for large-scale spatial domains with uneven data coverage.

UK applicability

The findings are potentially applicable to UK precipitation isotope mapping and hydrological studies requiring spatial prediction of stable isotopes, particularly where monitoring station density is variable across the country. The demonstrated advantage of machine learning methods with sparse data could support isotope-based research in UK environmental and forensic hydrology applications.

Key measures

Mean squared error (MSE), Kling Gupta efficiency (KGE), absolute differences, relative mean absolute error; spatial prediction accuracy across two European sampling networks

Outcomes reported

The study compared the performance of regression kriging and machine learning methods (random forest variants) for predicting spatial distribution of stable oxygen and hydrogen isotopes in precipitation across European sampling networks. Performance was assessed using mean squared error (MSE), Kling Gupta efficiency (KGE), and relative mean absolute error metrics.

Topic tags