Abstract

Dissolved Oxygen (DO) represents a key operational constraint for river-dependent aquaculture, and short-term DO forecasting can support practical agricultural water management in systems that directly abstract river water. Relying on daily physicochemical observations from the Halda River, Bangladesh (April 2024–March 2025), four Machine Learning (ML) models were evaluated for DO prediction, i.e. Linear Regression, Artificial Neural Network, Support Vector Machine, and Random Forest. Among the latter ML models, Random Forest achieved the highest predictive performance (R² = 0.9015, RMSE = 0.0833, and MAE = 0.0574), substantially outperforming the considered baseline Linear Regression model (R² = 0.4469, RMSE = 0.1892, and MAE = 0.1478). DO was generally stable (about 6.3–7.0 mg/L) while ionic variables, e.g., chloride, conductivity, and Total Dissolved Solids (TDS), showed pronounced dry-season peaks, thus indicating conditions where oxygen stress can become more likely and less predictable. Explainable ML interpretation developed on feature importance and SHapley Additive exPlanations (SHAP) consistently identified water temperature and ionic concentration proxies, especially chloride, alongside conductivity, and TDS, as the dominant drivers of DO variability. Relying on the latter findings, an operational decision-support workflow is proposed in which forecasts based on the Random Forest ML model are paired with low-cost monitoring of temperature and chloride to trigger time-graded farm actions, such as targeted aeration, temporary intake closure, and stocking/feeding adjustments. The reported results in the developed study demonstrate a transferable pathway for moving from accurate and interpretable ML forecasts to actionable agricultural water management protocols that reduce risk of critical losses in river-based aquaculture.