Prediction and Assessment of Groundwater Quality in Geographic Information System Environment Using Machine Learning Methods (Semi-Arid Regions)

Abstract

In recent decades, the population of Iran has steadily increased, while water resources are limited, and excessive use of groundwater resources has led to irreparable damages to aquifers. This research, utilizing geographic information system (GIS) models and machine learning techniques, aims to predict and classify groundwater quality parameters using existing qualitative variables. The accuracy of various methods has been assessed in the Birjand plain. Input data were based on water quality sampling from 2011 to 2020, collected from 18 observation wells.Parameter assessments from a statistical analysis perspective revealed that acidity exhibited the least variability (2.31%), while magnesium, with a coefficient of variation of 70%, showed the highest variability. Geostatistical analyses indicated that for TDS and pH parameters, the Inverse Distance Weighting (IDW) model performed better, and for Ca, Mg, Na, and SO4 parameters, ordinary kriging yielded the best results with the lowest RMSE. Evaluation of machine learning model performance demonstrated that Random Forest (RF), Decision Tree (DT), and Support Vector Machine (SVM) models, during the training phase, achieved an average R2 exceeding 94%, compared to an average R2 exceeding 85% in the testing phase for most parameters, providing satisfactory results. A comparison between GIS and machine learning models indicated that machine learning models outperformed in estimating groundwater quality parameters. Therefore, in the absence of field investigations on groundwater quality, data-driven methods are considered reliable for monitoring water quality.