An ensemble learning framework for streamflow reconstruction incorporating aleatoric and epistemic uncertainty

Study Region South Sulawesi, Indonesia Study Focus Machine learning (ML) models have been utilized to reconstruct streamflow data over data-sparse regions, but their aleatoric and epistemic uncertainty have been limitedly assessed. To address this gap, this study focuses on a machine learning-based framework that assesses propagated multiple uncertainties in reconstructing long-term streamflow at 22 stations in South Sulawesi, Indonesia. Three decision-tree and three neural network architectures are integrated via a weighted-stacking ensemble approach, incorporating simulated streamflow by a hydrologic model and ten regional climate indices as predictors. The ensemble performance is evaluated with observational data-related (aleatoric) and model-related (epistemic) uncertainty sources. New Hydrological Insight for The Region Results show that the decision tree-based ensemble approach outperforms other models, reaching NSE of 0.54 and KGE of 0.48 on median. Incorporating spatial information and large-scale climate indices, such as Pacific Warm Pool Region and Quasi-Biennial Oscillation, improve the predictive performance up to NSE of 0.72 and KGE of 0.68 on median. The aleatoric uncertainty remains as a primary source, driven by inherent variability of streamflow. Results show significant trends only in high streamflow over South Sulawesi from multi-type MK tests, suggesting a high risk of unprecedented floods in a changing climate. This study highlights the potential of AI in reconstructing streamflow data with a comprehensive assessment of multiple uncertainty sources and hydrological variability and trend in data-sparse regions.