Predicting Groundwater Quality Parameters Using Supervised Machine Learning

Abstract

Studies relating to groundwater have asserted that groundwater quality assessment is difficult, time-consuming and costly. An easy, vigorous, cost and time-effective tool is needed to predict water quality. The study employed supervised learning algorithms (decision tree regression and polynomial regression techniques) to build a model for assessing and predicting groundwater quality using easily measured parameters. The study employed experimental design (factorial design) and random sampling technique (multistage sampling technique) for the data collection. Model performance determinants such as R2, RMSE and d-statistics were used to compare the performance of the model with aqueous geochemical models such as Visual Minteq, Phreeq C and Wateq4F. ANOVA was used to determine the significance mean differences in the predicted groundwater chemical parameters of the study regions. The estimated R-square, RMSE and d-statistics for Visual Minteq (0.997, 16.97 and .987), Phreeq C (0.999, 33.16 and 0.960), Wateq4F (0.972, 15.33 and 0.988) and machine learning model (0.999, 1.690 and 1.00) indicated that the model developed has high predicting ability over the aqueous geochemical models. Model validating tools like accuracy (0.96), RMSE (1.690) and R2 (> 95%) demonstrated that the model could be used to forecast groundwater quality with high accuracy using easily measured parameters. ANOVA test demonstrated significant mean differences in the predicted groundwater chemical parameters of the study regions (P-value = 0.00 < 0.05). It is recommended that artificial intelligence tools, such as supervised learning as an easy, time and cost-effective way of predicting water quality should be encouraged in groundwater assessment.