Modelling of arsenic concentration in multiple water sources: A comparison of different machine learning methods

Abstract

Arsenic contamination is increasingly a serious global health concern, especially in developing countries like Ghana where the intensification of mining activities has resulted in adverse pollution of various water sources within mining communities. To harness the early detection of arsenic contamination in multiple water supply systems, this work aimed to assess and compare the prediction efficiency of different machine learning methods. Novel machine learning algorithms such as extreme gradient boosting (XGB), light gradient boosting (LGB) and generalized regression neural network (GRNN), which are yet to be explored in modelling arsenic concentration in various water supply systems (e.g. surface water and groundwater) were considered in this study. These models were compared with state-of-the-art machine learning methods of decision tree (DT), multivariate adaptive regression spline (MARS), multilayer perceptron neural network (MLP) and random forest (RF). The applied methods were evaluated using the coefficient of determination (R2), Nash–Sutcliff efficiency (NSE) and mean squared error (MSE). The statistical analysis showed that the newly tested models (XGB, LGB and GRNN) produced satisfactory predictions which were comparable to the state-of-the-art methods. Thus, all the methods applied including the newly introduced models have proven efficient in the arsenic modelling task with R2 > 0.93, NSE > 0.92 and MSE <2.35 E−06.