Dissolved oxygen concentration predictions for running waters with different land use land cover using a quantile regression forest machine learning technique

Modeling dissolved oxygen (DO) in running water represents a challenge due to complex interactions among various processes affecting its concentration and the intricacy of using process-based water quality models. In this study, a quantile regression forest (QRF) machine learning technique was used to develop data-driven models for predicting DO levels in three rivers that drain watersheds with distinctly different land use and land cover characteristics in different geographical regions. Water quality data, spanning 2007 to 2019, was used to develop and validate the models. Key DO drivers were first identified based on the variable importance index, and models were constructed for different combinations of the identified drivers as the input variables. Each model was calibrated for each input scenario using 80% of the data and validated by predicting the DO concentrations using the remaining 20% of the data. Excellent model performance was obtained with water temperature, pH, specific conductance, and chemical oxygen demand (COD) as input variables across the stations with water temperature and pH as the top predictors. The developed models outperformed multilayer perceptron neural network (MLPNN) and U.S. Environmental Protection Agency models in explaining data variance as well as giving lower errors in predictions. The commonality of the top-ranked predictors for the three geographically distant rivers suggests the possibility of building parsimonious models with a minimal number of predictors for in-stream DO predictions. These predictors are among the common physio-chemical water quality parameters of existing ambient water quality monitoring programs and are readily available for the model development.