Optimizing artificial neural network models for predicting the irrigation potential of groundwater sources

Abstract

Seasonal variability in groundwater characteristics often poses significant challenges in accurately estimating irrigation potential. To address these challenges, recent attempts have endeavored to utilize hybridization of machine learning models. However, one major challenge with hybridized models is that the interplay of different components in a hybrid model makes it challenging to identify which features or configurations contribute meaningfully to overall performance without proper feature reduction and tuning. Therefore, this study optimized the predictive performance of Artificial Neural Networks (ANNs) by applying feature reduction techniques and varying model architecture by increasing the number of neurons (n) in the hidden layers in multiples for n = 1, 2, 4, 8, and 16. Findings from the study through correlation analysis revealed high interdependence among groundwater parameters, with the strongest being a 0.9672 correlation between Total Dissolved Solids (TDS) and Electrical Conductivity (EC). This correlation suggests shared geochemical processes and weathering patterns among these features. Feature ranking through the Minimum Redundancy - Maximum Relevance (MRMR) method identified bicarbonate (HCO₃⁻) as the most influential variable, with an importance score of 0.48 in improving the predictive accuracy of the ANN models. By reducing the number of groundwater features as well as increasing the architecture of ANN, the predictive accuracy of the ANN models improved, with R², Mean Squared Error, and Percentage Bias values ranging from 0.83 to 0.91, 2.343 to 3.9325 and -0.16 to 1.30%. Comparison of theoretical and ANN-predicted irrigation potentials was consistent, confirming the suitability of groundwater sources for irrigation in the study area. The study demonstrates the potential of optimized ANN models for predicting irrigation suitability, providing a more practical approach to groundwater management, particularly in regions where seasonal variations influence water quality.