Surface water quality prediction in the lower Thoubal river watershed, India: A hyper-tuned machine learning approach and DNN-based sensitivity analysis.

The accurate and efficient monitoring, assessment, and prediction of water resources is crucial for the sustenance of life and health of the environment. Traditional methods for assessing water quality can be laborious and time-consuming, but the use of machine learning algorithms can improve the speed and precision of these predictions. The study aimed to evaluate the quality of surface water in the lower Thoubal river using a comprehensive approach that analyzed 12 physico-chemical properties of the water samples collected from 16 different sites. Four machine learning algorithms namely, deep neural network (DNN), gradient boost model (GBM), generalized linear model (GLM), and random forest (RF) were applied and compared for the prediction efficiency of a water quality index (WQI). Sensitivity and uncertainty analysis of the parameters were carried out using a DNN-based model. The study incorporated model evaluation using learning curve along with five key performance assessors. The findings revealed marked variation in the WQI in the watershed. Of the samples tested, about 70 % were found to have good water quality, 18 % had excellent water quality, 11 % had poor water quality, 1 % had very poor water quality, and less than 1 % were deemed unsuitable. Turbidity, BOD, and COD were found to be the most influential parameter towards WQI prediction. The effectiveness of the predictive models ranked as GLM>DNN>RF>GBM, GLM>GBM>DNN>RF, and GLM>DNN>RF>GBM; based on testing, training and 5-fold cross-validation. The GLM model consistently showed superior performance with the lowest RMSEs of 0.14783 in training, 0.15936 in validation, and 0.26115 in 5-fold cross-validation. Conversely, during the training, RF displayed the worst RMSEs of 5.38354, 4.79754 by GBM in validation, and 7.04887 by GBM in 5-fold cross-validation. The analysis using the Taylor diagram and learning curve further supported that GLM was the most effective model for predicting surface water quality. Biological oxygen demand, chemical oxygen demand, electrical conductivity, turbidity and total hardness were found to be the key parameters influencing the water quality. Unregulated and illegal sand mining from the riverbed, garbage dumping into the water bodies, runoffs from agricultural fields have significantly deteriorated the water quality in the study area. The scalable approach of the study and its results can benefit the local water managers and water research community. [Display omitted] • Traditional water quality assessment methods are tedious. • Machine learning and deep learning can predict water quality with precision. • Generalized linear model was found to be the most effective predictive model. • Turbidity, conductivity and biological oxygen demand affect water quality negatively. • Unregulated sandmining and garbage dumping are the major water pollution source. [ABSTRACT FROM AUTHOR]