Metaheuristic approaches for prediction of water quality indices with relief algorithm-based feature selection

Abstract

Monitoring and assessing groundwater quality are important for sustainable water resource management. Therefore, the present study aimed to analyze and predict the water quality indices using metaheuristics algorithms. Groundwater samples were acquired from 26 tube wells periodically, and laboratory analysis was performed in the soil and water laboratory, Water Technology Centre (WTC), ICAR-Indian Agricultural Research Institute (ICAR-IARI), New Delhi, India. Water quality parameters such as pH, electrical conductivity (EC), calcium (Ca2+), magnesium (Mg2+), sodium (Na+), potassium (K+), chloride (Cl−), carbonate (CO32−), bicarbonate (HCO3−) and sulphate (SO42−) were used for the computation of water quality indices. The relief algorithm was used for the selection of the best input combinations. Seven meta heuristics algorithms viz., Reduced Error Pruning Tree (REPTree), Random Forest (RF), Bagging, Additive Regression (AdR), M5Preuning Tree (M5P), Random Subspace (RSS), and Artificial Neural Networks (ANN) were trained and tested for Water Quality Index (WQI), and the best four candidate models, i.e., M5P, ANN, RF, and RSS were validated for the prediction of Kelly ratio (KR), sodium adsorption ratio (SAR), permeability index (PI), and percent sodium (%Na). The performance of models was assessed using statistical indicators, i.e., mean absolute error (MAE), root mean squared error (RMSE), Willmott index (WI), Nash-Sutcliffe efficiency (NSE), coefficient of correlation (R), and visual graphics (i.e., Taylor diagram and radar chart). On a quantitative comparative scale, the M5P model performed better for predicting WQI (WI = 0.997, NSE = 0.988, MAE = 7.173, RMSE = 10.769 and R = 0.999) and % Na (WI = 901, NSE = 0.692, MAE = 3.817, RMSE = 5.005 and R = 0.853). The ANN model had a significant edge over other models for predicting indices namely, KR (WI = 0.844, NSE = 0.625, MAE = 0.121, RMSE = 0.164 and R = 0.825), PI (WI = 0.987, NSE = 0.949, MAE = 2.508, RMSE = 3.269 and, R = 0.981), and SAR (WI = 0.981, NSE = 0.916, MAE = 0.314, RMSE = 0.385 and R = 0.976). The result indicated that the M5P model has the potential to predict the water quality indices even under limited input parameters scenarios in the sub-tropical climatic condition. Also, study findings revealed that the relief algorithm based input feature selection could be used for assessing water quality indices. The outcome of the present study would be of interest to policymakers or irrigation planners to achieve groundwater sustainability in terms of water quality suitability for distinct applications and remedial measures to enhance groundwater quality.