Abstract
Reliable water quality prediction can improve environmental flow monitoring and the sustainability of the stream ecosystem. In this study, we compared two machine learning methods to predict water quality parameters, such as total nitrogen (TN), total phosphorus (TP), and turbidity (TUR), for 97 watersheds located in the Southeast Atlantic region of the USA. The modeling framework incorporates multiple climate and watershed variables (characteristics) that often control the water quality indicators in different landscapes. Three techniques, such as stepwise regression (SR), Least Absolute Shrinkage and Selection Operator (LASSO), and genetic algorithm (GA), are implemented to identify appropriate predictors out of 28 climate and catchment-related variables. The selected predictors were then used to develop the Random Forest (RF) and Boosted regression tree (BRT) models for water quality predictions in selected watersheds. The results highlighted that while both algorithms provided reasonable results (based on statistical metrics), the RF algorithm was easier to train and robust to model overfitting. Partial dependence plots highlighted the complex and nonlinear relationships between the individual predictors and the water quality indicators. The thresholds obtained from partial dependence plots showed that the median values of total nitrogen (TN) and total phosphorus (TP) in streams increase significantly when the percentage of urban and agricultural lands is above 40% and 43% of the watershed area, respectively. Furthermore, when soil hydraulic conductivity increases, the reduction in runoff results in decreased Turbidity levels in streams. Therefore, identifying the key watershed characteristics and their critical thresholds can help watershed managers create appropriate regulations for managing and sustaining healthy stream ecosystems. Besides, the forecasting models can improve water quality predictions in ungauged watersheds.