Application of semi-supervised models for groundwater level simulation in arid regions with small sample sizes

Area of Science:

• Hydrology
• Machine Learning
• Environmental Science

Background:

• Groundwater level monitoring is vital for sustainable aquifer management but faces challenges due to data scarcity in arid regions.
• Traditional models struggle in basins with sparse monitoring networks, limiting predictive accuracy.

Purpose of the Study:

• To develop and evaluate a semi-supervised learning model for predicting spatial groundwater levels in the Tarim River Basin, addressing data scarcity.
• To identify key environmental covariates influencing groundwater dynamics using advanced model interpretation techniques.

Main Methods:

• A semi-supervised learning architecture was employed, integrating limited groundwater observations with unlabeled environmental data.
• Adversarial pseudo-label generation and variational feature extraction were utilized to enhance prediction accuracy.
• SHAP (SHapley Additive exPlanations) analysis was performed for model interpretability.

Main Results:

• The semi-supervised model achieved high performance (R² = 0.87, RMSE = 0.75 ± 0.03), outperforming traditional methods.
• Key predictors identified include distance to river, hydraulic conductivity, groundwater storage anomalies, and NDVI.
• Significant groundwater depletion was observed in the northern Tarim River Basin, while the southern region showed a rising trend.

Conclusions:

• Semi-supervised learning, particularly with physics-informed pseudo-labeling, effectively overcomes data scarcity in groundwater prediction.
• The developed framework provides a transferable solution for data-sparse arid regions, enabling better water resource management.
• Despite localized recovery, groundwater resources face ongoing pressure, necessitating careful water allocation and ecological preservation strategies.