HydroPredictor a hybrid machine learning model for addressing data scarcity in groundwater prediction.

Area of Science:

• Hydrogeology
• Machine Learning
• Environmental Science

Background:

• Groundwater prediction is challenging in data-scarce regions due to limited data and complex environmental factors.
• Existing models struggle with spatial heterogeneity and nonlinear hydrogeological processes.

Purpose of the Study:

• To develop a robust hybrid machine learning framework, HydroPredictor, for accurate groundwater potential prediction.
• To address the limitations of conventional models in data-limited and environmentally sensitive areas.

Main Methods:

• A hybrid framework integrating CatBoost and a regularized Multi-Layer Perceptron (MLP).
• Training on a geo-referenced dataset (315 samples) from Feija Basin, Morocco, using ten environmental predictors.
• Optuna-based hyperparameter optimization and 5-fold cross-validation for model validation.

Main Results:

• HydroPredictor achieved 89.23% testing accuracy, with an F1-score of 0.8937 and AUC > 0.90.
• The model significantly outperformed Random Forest, SVM, and standalone MLP.
• Demonstrated superior generalization compared to existing literature models with minimal overfitting (∆Accuracy = 0.35%).

Conclusions:

• HydroPredictor offers a scalable and robust solution for groundwater classification in data-limited settings.
• The hybrid approach effectively combines interpretable tree-based embeddings with deep neural representations.
• Provides a reproducible tool for sustainable groundwater resource management amidst environmental uncertainty.