Articles · Open Access Policy
Abstract
Rivers are important for aquatic biodiversity, fisheries, drinking water supply, and agriculture, but are suffering from continuous pressure due to the discharge of untreated sewage, industrial discharge, and farm runoff. Large amounts of records are generated as a result of routine monitoring, but turning these records into timely and readily understood statements of pollution risk is not easy. This study aimed to formulate a machine learning model to predict the water quality index (WQI) based on routinely measured water quality parameters and categorize each water sample according to the level of pollution. The water records of the river water were used from more than 2800 monitoring sites in India for 11 years from 2006 to 2016. The following indicators were used: temperature, pH, conductivity, biochemical oxygen demand (BOD), nitrate, fecal coliform, and total coliform. Data screening and gap filling were performed, and a weighted WQI was then computed and further divided into four risk levels (Low, Moderate, High, and Severe). Seven prediction methods were trained and tested. The gradient boosting model was selected as the primary model for the framework because it had the highest R² (0.998) and the lowest mean error (0.41 index units) between the measured and predicted values of WQI, and achieved 98.3% accuracy rates in risk classification and 97.7% accuracy when tested on the subsequent non-training years. 90% of samples were in the Moderate, High, or Severe categories, with BOD and fecal coliform contributing almost 78% of the water quality variation explained. The omission of the total coliform test resulted in a minimal decrease in accuracy, but the omission of both bacterial tests resulted in a 76.1% decrease in accuracy. The framework provides a practical and cost-effective method of early warning and prioritization of monitoring effort in aquatic ecosystems, identifying hotspots.