The population of Iran has steadily increased in recent decades. However, water resources in the country are limited, and the excessive use of groundwater resources has irreparably damaged aquifers. This study predicted and classified groundwater quality parameters based on existing qualitative variables by using geographic information
system (GIS) and machine learning models. The accuracy rates of multiple methods were assessed in relation to the
Birjand plain. The input data in this study were based on water quality sampling results for samples collected from 18
observation wells between 2011 and 2020. Parameter assessments revealed that acidity exhibited the lowest variability(2.31%), whereas Mg content exhibited the highest variability (70%). In addition, geostatistical analyses indicated that the inverse distance weighting model had more favorable performance for predicting total dissolved solids and pH,whereas ordinary kriging had more favorable performance for predicting Ca, Mg, Na, and SO4 contents with the lowest root mean square error. A performance evaluation of machine learning models demonstrated that the random forest,decision tree, and support vector machine models achieved an average R2 exceeding 94% in the training phase, compared with an average R2 exceeding 85% for most parameters in the testing phase. Machine learning models outperformed GIS models in estimating groundwater quality parameters. Accordingly, data-driven methods can be considered reliable for monitoring groundwater quality in the absence of field investigations. The limitations of this study included its restricted geographic scope, its reliance on a limited number of observation wells, and its consideration of only a specific set of water quality parameters.
Keywords: Decision tree, Support vector machine, Random forest, Groundwater pollution.