283 / 2024-02-28 23:25:24

A dissolved oxygen prediction model in Shenzhen River estuary based on wavelet transform and random forest

dissolve oxygen,Shenzhen river estuary,wavelet transform,Random Forest,SHAP

Estuaries around the world have undergone different transformations, like hypoxia, which was driven by a multitude of stressors. However, only a few studies have quantified the effect of driving factors in estuary hypoxia and the interaction pattern has yet to be understood. This study aims to investigate the process of hypoxia transformation in the Shenzhen River estuary as well as to construct a suitable machine-learning method for driving factor identification and dissolved oxygen prediction in the estuary. 60 multi-dimensional stressors on DO, such as physical and chemical water quality monitoring indicators, tidal data, wastewater treatment plant discharged monitoring indicators, and meteorological data, were investigated. The Pearson analysis was used to examine the linear relationship between indicators, while the wavelet transform was used to analyze the nonlinearities as well as the multi-time scale correlation in both the time domain and frequency domain. During the missing value interpolation, we compared the influence of nearest neighbor interpolation, linear interpolation, and cubic polynomial spline interpolation on monitoring data modeling. The inputted data was then partitioned into training and test sets in a 7:3 ratio. By comparing the impact of different machine learning models with evaluation metrics including R-square, Mean Squared Error, Root Mean Squared Error, Mean Absolute Error, and Mean Absolute Percentage Error, a suitable model for DO prediction was obtained. Additionally, the effect of DO periodicity on the model was studied by introducing its hysteresis features. Following this, the interpretable Shapley additive explanations method (SHAP) was used to quantitatively characterize the attribution of indicators to DO and identify its driving factors. The results showed that (1) the DO showed a 7-12 month oscillation period; (2) the adjustment of size-feature ratio had the most significant impact on MAPE, while the cubic polynomial spline interpolation efficiently mitigated model errors and enhanced their learning ability from input data; (3) the Support Vector Regression model exhibited poor performance, whereas the Random Forest model effectively reduced errors and minimized reliance on input indicators; (4) the nonlinear relationship between each stressor and DO was demonstrated by cross wavelet transform, while SHAP quantified the contribution of each indicator, which highlighted the influence of wastewater discharge and precipitation on changes in DO. Taken together, the analysis of the driving factors of estuarine deoxygenation, coupled with the successful establishment and application of the interpretable wavelet transform-random forest model, would have important implications for water ecosystem health governance.