Predicting and zoning of groundwater quality using geographical information system (GIS) models and machine learning methods (case study: Zahedan plain)

In recent decades, the growing population, limited water resources, and overexploitation of aquifers have caused irreparable damage to the quantity and quality of the country's aquifers. In the current study, geographical information system (GIS) models and machine learning (ML) techniques using available groundwater quality variables were applied for the prediction and zoning of salinity and SAR of groundwater in the Zahedan plain and then the accuracy of these methods was compared.The input data were based on water quality sampling in 2018 from 59 observation wells. The study of parameters showed that in Zahedan plain, EC, SAR, and TDS parameters had high variability (CV>41%) and acidity showed low variability (CV = 4.16%). The results of the geostatistical analysis showed that the IDW model represented better results with the power value of 2 for TDS and EC parameters while for pH and SAR parameters, the ordinary Kriging method revealed the best result with minimum RMSE in the test stage. Performance evaluation of ML models showed that all three models RF, ANN, and SVM showed acceptable results with R2 above 90% and NRMSE values below 15% for all parameters (except acidity). However, better estimations were observed in the training step than in the test step. A comparison of different GIS models and ML also revealed the notable superiority of ML models in estimating these parameters. Finally, it can be concluded that under a shortage of field facilities for the assessment of groundwater quality, data-driven methods can be a reliable alternative to water quality monitoring.