Abstract: This paper proposes a way to explore the relationship between the combination of water quality indicators and the accuracy of river dissolved oxygen prediction. First,the XGBoost model is used to calculate the water quality index feature importance score,and then based on the greedy rule and the water quality index feature importance score,8 water quality index combinations are arranged. Finally,the BP neural network is used to predict dissolved oxygen for the 8 water quality index combinations. Experimental results show that pH,water temperature,conductivity,and ammonia nitrogen are the four key indicators that affect the prediction of dissolved oxygen. Among the 8 combinations of water quality indicators arranged,pH,water temperature,conductivity,ammonia nitrogen,turbidity,and CODmn are the most accurate combinations of input indicators for the prediction of dissolved oxygen. Experimental analysis by exhaustively enumerating all water quality indicator combinations proves that the method is effective and feasible with lower time complexity,and can be used to select a combination of input indicators with high accuracy of dissolved oxygen prediction to improve the accuracy of dissolved oxygen prediction.