Predicting groundwater and surface water dynamics in semi-arid regions is challenging due to strong temporal variability, data noise, and nonlinear interactions among climatic, hydrological, and human factors. This study evaluates the performance of seven machine learning models including eXtreme Gradient Boosting (XGBoost), Random Forest (RF), Support Vector Regression (SVR), Support Vector Machine (SVM), Decision Tree (DT), Radial Basis Function Network (RBFN), and Bayesian Model Averaging (BMA for jointly predicting groundwater levels and river discharge in the Khanmirza watershed of southwestern Iran. Using nearly three decades of raw, non-transformed hydrometeorological data (1994–2023), the analysis provides an operationally realistic comparison free from the effects of signal decomposition or preprocessing. Results show that XGBoost consistently achieves the highest predictive accuracy across multiple groundwater and surface water monitoring stations. However, the BMA ensemble, which probabilistically integrates XGBoost and RF outputs using Gaussian process–based optimization, attains statistically comparable accuracy while offering lower bias, improved stability, and more consistent performance across varying hydrological conditions. Overall, the findings highlight two key insights for water-stressed semi-arid basins: (1)advanced gradient-boosting models such as XGBoost can effectively capture complex hydrological behavior when trained on long-term observational records, and (2)probabilistic ensemble approaches like BMA provide a reliable and generalizable alternative that enhances robustness an essential attribute for operational water management in data-scarce environments.
ebrahimzadeh et al. (2026) studied this question.