Scrutinizing different predictive modeling validation methodologies and data-partitioning strategies: new insights...

Area of Science:

• Environmental science
• Hydrogeology
• Data science

Background:

• Groundwater salinity impacts water quality, agriculture, industry, and public health.
• Reliable predictive models are essential for effective groundwater resource management.

Purpose of the Study:

• To validate a Group Method of Data Handling (GMDH)-based model for predicting groundwater salinity.
• To evaluate the performance of three validation methods (hold-out, k-fold cross-validation, leave-one-out) with various data partitioning strategies.

Main Methods:

• Developed a GMDH model to predict groundwater salinity in a coastal aquifer.
• Applied hold-out (last and random), k-fold cross-validation, and leave-one-out validation methods.
• Utilized diverse data partitioning strategies (e.g., 40% random hold-out) and assessed performance using RMSE, MSE, and R².

Main Results:

• The hold-out random strategy with 40% data partitioning yielded the most accurate predictions (lowest RMSE) for monitoring wells 1, 2, and 3.
• GMDH model performance varied significantly across different validation methodologies and data partitioning approaches.
• Different validation strategies provide distinct insights into model behavior, bias, and variance.

Conclusions:

• The choice of validation methodology and data partitioning strategy critically influences groundwater salinity prediction model accuracy.
• Employing multiple validation techniques in conjunction is recommended for a comprehensive understanding of model performance.
• Optimized validation strategies enhance the reliability of groundwater salinity forecasts for informed resource management.