Comparative analysis of machine learning based dissolved oxygen predictions in the Yellow River Basin: The role of...

Dissolved oxygen (DO) is a key water quality indicator reflecting river health. Modeling and understanding the spatiotemporal dynamics of DO and its influencing factors are crucial for effective river management. Machine learning (ML) models have gained popularity in water quality prediction; however, their accuracy strongly depends on the predictor variables. Predictor availability varies considerably, prompting the inquiry of whether readily accessible catchment attributes, when combined with ML, specifically a Random Forest (RF) model, can effectively predict river DO dynamics in the Yellow River Basin (YRB)? Is there a necessity to collect additional water quality data to improve model performance? To address this, we collected approximately 10,800 monthly DO measurements (2016-2022) from 135 monitoring sites in the YRB. We categorized predictor variables into three groups: catchment attributes and meteorology (CAM), water quality parameters (WQPs), and a combination of both CAM + WQP. The RF models achieved satisfactory performance, with Nash-Sutcliffe efficiency exceeding 0.35 at 67 %, 61 %, and 73 % of sites for CAM, WQP, and CAM + WQP, respectively. Overall, the CAM model effectively predicted DO in the YRB, generally outperforming the WQP model. This is primarily because the CAM model not only includes core WQP variables (substituting water temperature with air and soil temperature) but also incorporates additional environmental variables that influence DO (such as atmospheric pressure), thereby improve its prediction performance. However, in regions with significant anthropogenic emissions and water diversion/transfer, CAM alone struggled to predict DO accurately. In these areas, the WQP model, which better captures anthropogenic influences, proved more effective. Despite improved performance with the integrated CAM + WQP model, limitations remain, including difficulty in predicting extreme values, lack of dynamic anthropogenic variables, and sparse monitoring coverage in high-altitude regions. This comparative modeling analysis provides valuable insights into the relative strengths of catchment- and water-quality-based predictors, informing the design of more accurate DO prediction frameworks for large-scale river management.