Groundwater quality forecasting using machine learning algorithms for irrigation purposes

Using conventional methods to evaluate the irrigation water quality is usually expensive and laborious for the farmers, particularly in developing countries. However, the applications of artificial intelligence models can overcome this issue through forecasting and evaluating the irrigation water quality indexes of aquifer systems using physical parameters as features. This paper aims forecasting the Total Dissolved Solid (TDS), Potential Salinity (PS), Sodium Adsorption Ratio (SAR), Exchangeable Sodium Percentage (ESP), Magnesium Adsorption Ratio (MAR), and the Residual Sodium Carbonate (RSC) parameters through Electrical Conductivity (EC), Temperature (T), and pH as inputs. To achieve this purpose, we developed and evaluated Adaptive Boosting (Adaboost), Random Forest (RF), Artificial Neural Network (ANN), and Support Vector Regression (SVR) models using 520 samples of data related to fourteen Groundwater quality parameters in Berrechid aquifer, Morocco. The results revealed that the overall prediction performances of Adaboost and RF models are higher than those of SVR and ANN. However, the generalization ability and sensitivity to the inputs analyses show that the ANN and SVR models are more generalizable and less sensitive to input variables than Adaboost and RF. Globally, the developed models are valuable in forecasting the irrigation water quality parameters and could help the farmers and decision-makers in managing the irrigation water strategies. The developed approaches in this study have been revealed promising in low-cost and real-time forecast of groundwater quality through the use of physical parameters as input variables.

Automated summary

The study focuses on predicting groundwater quality using physical parameters as input, developing and evaluating Adaboost, Random Forest, Artificial Neural Network, and Support Vector Regression models. Results show that Adaboost and RF models perform better overall, but ANN and SVR models are more generalizable and less sensitive to input variables.