Ensemble machine learning method for δ¹⁸O prediction in groundwater.

Area of Science:

• Hydrogeology
• Environmental Science
• Machine Learning

Background:

• Interpreting groundwater oxygen isotopes (δ¹⁸O) in arid zones is complex due to intricate recharge patterns, salinization, and sparse data.
• Existing methods often struggle with the heterogeneity of arid environments.

Purpose of the Study:

• To evaluate the efficacy of ensemble machine learning models for predicting groundwater δ¹⁸O in the mid-Nile Valley.
• To identify key hydrochemical and spatial predictors of groundwater δ¹⁸O.
• To assess the reliability and limitations of machine learning predictions in data-scarce arid regions.

Main Methods:

• Applied Random Forest, Lasso, and Gradient Boosting models to predict δ¹⁸O using hydrochemical and spatial variables.
• Utilized both random splitting and spatial cross-validation for model performance assessment.
• Conducted sensitivity analysis by excluding spatial variables to isolate the influence of hydrochemical factors.

Main Results:

• Random Forest outperformed other models under random splitting (R² = 0.83).
• Model performance significantly decreased under spatial cross-validation, indicating the importance of spatial structure.
• Longitude, chloride, potassium, and magnesium were identified as primary predictors.
• Hydrochemical variables alone explained approximately 40% of the δ¹⁸O variance.
• Prediction uncertainty increased substantially for samples outside the data-rich range.

Conclusions:

• Ensemble machine learning offers a valuable tool for groundwater isotope prediction in data-limited arid regions.
• Spatial effects are crucial and must be considered for accurate δ¹⁸O interpretation.
• Model predictions are most reliable for site-specific analyses within the established data range.
• Further research requires collecting more end-member samples to improve predictions in underrepresented isotopic ranges.