Simulating dissolved oxygen concentrations at the watershed scale : A machine learning approach with physical constraints

Abstract

This study focuses on simulating dissolved oxygen (DO) concentrations at the watershed scale using machine learning (ML) models, with an emphasis on incorporating domain constraints to improve prediction accuracy. The main objectives are to evaluate the performance of different ML models, assess the impact of physical and spatial dependencies, and identify the most critical features influencing DO simulation. Random Forest (RF), Extra Trees (ET), and Histogram-based Gradient Boosting (HGB) were selected for this study and trained using a set of input variables, including water and air temperature, and other hydrological information. Model performance was assessed by calculating Mean Square Error (MSE), Mean Absolute Error (MAE), and Nash-Sutcliffe Efficiency (NSE). The best model-metric combination was selected for each station, and the results were satisfactory for most monitoring stations in the basin. The feature selection analysis, run with SHapley Additive exPlanations (SHAP), was designed to capture spatial, temporal, and physical dependencies, ensuring that the models remained accurate and aligned with established physical principles. Temperature-related variables were found to be the most significant predictors of DO levels. These outcomes demonstrate the potential of ML approaches with physical constraints to effectively predict DO concentrations and contribute to better-informed water quality management in natural watersheds.