Parameter estimation and uncertainty quantification of rainfall-runoff models using data assimilation methods based on deep learning and local ensemble updates

Rainfall-runoff (RR) modeling is crucial for flood preparedness and water resource management. Accurate RR model predictions depend on effective parameter estimation and uncertainty quantification using observed data through data assimilation (DA). Traditional DA methods often struggle with challenges such as non-Gaussianity and equifinality. To address these challenges, this study introduces two ensemble smoother methods, i.e., ESDL with a deep learning-based update, and ESLU with a local ensemble update, aiming to enhance the calibration of RR models. To demonstrate the effectiveness of our proposed methods, we conduct a comprehensive analysis involving various DA techniques applied to parameter estimation of RR models. We compare these methods with traditional approaches, evaluating deep neural network architectures, iteration numbers, and measurement errors. The results unequivocally showcase the consistent reliability of ESDL and ESLU, especially the latter one, across diverse scenarios, establishing them as promising approaches for the effective calibration and uncertainty quantification of RR models.