A surrogate-aided approach for accelerated Bayesian calibration of hydrologic models

Bayesian calibration of hydrologic models effectively addresses parameter uncertainties and improves predictions, but the joint inference of hydrologic and error model parameters often suffers from slow convergence due to high-dimensional interactions. To overcome this, a surrogate-aided error model is introduced that decouples their inference. This method uses support vector regression (SVR) as a surrogate, to estimate error model parameters conditioned on hydrologic parameters. This approach accelerates convergence (requiring 50% fewer samples) and improves predictive accuracy, consistently improving or maintaining Continuous Ranked Probability Scores across a range of test models. These advantages are demonstrated through an application to the GR4J model across 12 MOPEX watersheds. The reduced computational demand makes this particularly valuable for large-scale hydrologic modeling when computational resources are limited. • Proposes a new framework using a surrogate-aided model to help efficiently decouple the inference of hydrologic/error parameters • Achieves 50% faster MCMC convergence compared to common methods across 12 MOPEX watersheds • Reduces CRPS to 0.25, outperforming benchmarks using an SVR-optimized error model