A methodological framework for assessing aquatic contamination using data science approaches

Abstract

Environmental contamination (EC) poses significant risk to surface water resources. Particularly, EC by potentially toxic elements (PTEs) have received greater attention due to their bioaccumulation and biomagnification capabilities. However, existing tools for assessing EC by PTEs lack reliability due to uncertain outcomes. Therefore, this research utilized PTEs data including Arsenic (As), Cadmium (Cd), Chromium (Cr), Copper (Cu), Iron (Fe), Manganese (Mn), Nicket (Ni), Lead (Pb) and Zinc (Zn), collected from different rivers and estuaries within the Sundarbans estuarine ecosystem of Bangladesh and developed the novel “Environmental Contamination Index (ECI)” model for rating EC in aquatic bodies. The ECI model's architecture consist of four components, including (i) indicator selection technique for selecting crucial environmental contaminants; (ii) a linear interpolation based sub-index (SI) function for normalizing the various environmental contaminants' information; (iii) arithmetic mean-based aggregation function for computing the ECI score; and (iv) score classification scheme for assessing the state of EC. Additionally, for the purposes of evaluating the model's performance in terms of sensitivity, uncertainty and efficiency, this study employed ten machine learning and artificial intelligence techniques. Furthermore, five hyperparameter optimization techniques were compared for assessing the influence of hyperparameters on the model performance and efficiency. To assess the generalization capability of the model(s), the research utilized an independent dataset for validation purposes. The computed ECI scores demonstrated the “fair” and “marginal” EC by PTEs in different sampling sites of the study domain. Among the ML-AI technique, the gradient boosting regression (GBR) model with the Bayesian optimizer (BO) hyperparameter optimization technique demonstrated exceptional superiority for predicting ECI score with training (Root mean squared error-RMSE = 1.08, Mean squared error-MSE = 1.18, Mean absolute error-MAE = 0.847, and Percentage of absolute bias error-PABE = 1.30), testing (RMSE = 4.12, MSE = 17.0, MAE = 3.08 and PABE = 4.62) and validation (RMSE = 3.91, MSE = 15.3, MAE = 3.15 and PABE = 5.51) datasets. Additionally, the ECI model demonstrated good generalization sensitivity (R2 = 0.71) and less than 5 % uncertainty level with the independent dataset for rating PTEs contamination. Further comparison between the ECI model and the heavy metal pollution index (HPI) model revealed that the ECI model is able to accurately rate the scenario of EC by PTEs depending on the number of PTEs breached the threshold limit suggested for the PTEs in water bodies. Cumulatively, the outcomes from this research implied that the ECI model could be an effective tool for monitoring EC by PTEs in water bodies. Although the ECI model is developed using the PTEs data, the model could be utilized for assessing EC from different environmental contaminants. Finally, the findings from this research have significant implication for protecting aquatic ecosystem from EC by accelerating rapid decision-making process.