An ensemble of machine-learning models, explained with SHAP and network diagrams, points to temperature and water inflow and outflow conditions as key drivers of dissolved oxygen changes in the reservoir.
Dissolved oxygen (DO) reflects how well reservoir water can support organisms, but its behavior can be hard to simulate because of intertwined physical, chemical, and biological processes. The study focused on the Javeh Reservoir in Iran and built an AI-based prediction framework that combines six ensemble learning approaches.
The best model, an Extra Trees model, was evaluated against observed dissolved oxygen using common error and efficiency statistics. To make the results easier to interpret, the team applied SHAP to identify the dominant factors behind the model’s predictions, and added a chord-diagram network view to summarize how dissolved oxygen connects to top predictors.
Best model and key drivers
The study tested six optimized models—Random Forest, XGBoost, Hist Gradient Boosting, Extra Trees, a multilayer perceptron (MLP), and support vector regression (SVR)—for predicting dissolved oxygen in the Javeh Reservoir using five-day averaged water quality flux, hydrological, hydraulics, and meteorological data. Extra Trees produced the best predictive performance, with MAE = 0.3465 mg/L, RMSE = 0.5997 mg/L, and NSE = 0.9655. SHAP analysis of the best model identified air temperature and the inflow fluxes of dissolved oxygen, phosphate, and ammonium, along with outflow discharge, as dominant drivers of dissolved oxygen variability. The researchers also used a chord-diagram connectivity visualization to summarize dissolved-oxygen linkages with the top predictors.
Model testing and interpretation limits
This is an evaluation of machine-learning predictive performance and model interpretability using the Javeh Reservoir dataset described in the study, with results reported using MAE, RMSE, and Nash–Sutcliffe efficiency (NSE). The abstract does not specify how many monitoring sites or time periods were used, how the data were split into training and testing, or whether external datasets were used to test transferability, so it’s unclear how broadly the findings and identified drivers would hold beyond the reported setting.
// Source
Scientific Reports · 2026 · DOI: 10.1038/s41598-026-64406-x
Authors: Omid Yazdan Panah, Motahareh Saadatpour
Institutions: Iran University of Science and Technology
