Predicting non-mixing river flow using data-driven approaches: A case study in the Menindee region in Australia

Area of Science:

• Environmental Science
• Hydrology
• Data Science

Background:

• Thermal stratification in rivers impedes mixing, affecting oxygen and nutrient distribution, and can lead to fish kills.
• Understanding and predicting river non-mixing conditions are vital for aquatic ecosystem health and water resource management.

Purpose of the Study:

• To develop and assess a novel hybrid data-driven model for classifying river flow mixing and non-mixing conditions.
• To investigate the influence of meteorological, hydrological, and process-based model data on river stratification.
• To utilize explainable artificial intelligence (XAI) for understanding model predictions.

Main Methods:

• A hybrid Support Vector Machines (SVM) model was developed using data from the Darling River, Australia.
• The model integrated meteorological data, hydrological factors, and outputs from the LAKEoneD model.
• Supervised machine learning and XAI techniques were employed for classification and analysis.

Main Results:

• The hybrid SVM model, enhanced with LAKEoneD data, outperformed other models in predicting non-mixing conditions.
• Explainable AI analysis identified minimum air temperature and relative humidity as key predictors of non-mixing flow.
• Maximum air temperature was also significant, especially preceding fish death events.

Conclusions:

• The proposed hybrid model serves as a valuable tool for predicting river stratification and its impact on aquatic life.
• Findings offer critical insights for environmental authorities to enhance water quality management strategies in river systems.
• This research provides a scientific approach to forecasting fish health based on river flow dynamics.