Large-scale prediction of stream water quality using an interpretable deep learning approach

Area of Science:

• Environmental Science
• Data Science
• Water Resource Management

Background:

• Deep learning models show promise for water quality prediction due to their ability to map complex, nonlinear relationships.
• A significant knowledge gap exists in explaining the physical mechanisms behind deep learning predictions for water quality changes, limiting their practical use.
• Interpretable AI is crucial for understanding and trusting model outputs in critical applications like water quality management.

Purpose of the Study:

• To develop and validate an interpretable deep learning framework for predicting the spatiotemporal variations of water quality parameters.
• To address the lack of physical interpretability in existing deep learning models for water quality prediction.
• To identify key environmental and socioeconomic factors influencing stream water quality in a large region.

Main Methods:

• An interpretable deep learning framework was established to predict daily stream water quality parameters.
• Mereological, land-use, and socioeconomic variables were integrated as predictors across 138 sub-catchments in southern China.
• The SHapley Additive exPlanations (SHAP) method was employed to interpret model predictions and identify significant variables.

Main Results:

• The framework achieved high prediction accuracy for chemical oxygen demand (COD), total phosphorus (TP), and total nitrogen (TN), with coefficients of determination exceeding 0.80.
• Inclusion of land-use and socioeconomic data alongside hydrological variables improved model performance.
• SHAP analysis effectively identified significant predictors, including air temperature, forest cover, grain production, population density, urban area proportion, and recent rainfall, and their influence on various water quality parameters.

Conclusions:

• The interpretable deep learning framework provides a reliable and understandable approach for predicting water quality variations.
• Environmental and socioeconomic factors play critical roles in shaping stream water quality, and their influence can be effectively quantified.
• This approach enhances the practical applicability of deep learning in water quality management and scientific research.