Data-driven prediction of dissolved oxygen to identify fish kill factors: Case study in Redbank weir at Murrumbidgee River, Australia

Dissolved oxygen (DO) is critical in aquatic ecosystem health, prompting the need for precise predictive models. This study introduces a novel Convolutional Neural Network-Gated Recurrent Unit (CNN-GRU) framework to predict DO levels at Redbank Weir on the Murrumbidgee River, Australia. Considering a definition of stratification that accounts for thermal stratification, alongside satellite data to obtain an estimation of chlorophyll content in waters. The models were trained using data from 2017 to 2024. The CNN-GRU model demonstrates superior performance, with validation metrics indicating a Mean Absolute Error (MAE) of 0.6157 and Root Mean Square Error (RMSE) of 0.7521 mg/L for daily predictions, and MAE = 0.6391 and RMSE = 0.6356 for hourly predictions. Furthermore, the explainable AI analysis reveals strong correlations between DO levels and surface water temperature, while stressing the significant impacts of water level and evaporation on overall model explainability. The analysis highlighted the importance of intricate interactions among environmental variables associated with low DO levels during fish kill events. These factors interact in complex ways: for example, higher temperatures reduce oxygen solubility while simultaneously accelerating microbial respiration, which further consumes DO. This research emphasises not only predictive accuracy but also the importance of explainability in machine learning models. Ultimately, providing stakeholders with insights for informed decision-making regarding aquatic ecosystem management and water quality regulation. • The model combines deep learning and explainable AI to predict dissolved oxygen. • Water temperature is a key predictor in our model for predicting dissolved oxygen. • Water level and evaporation significantly affect the overall model explainability.