Improving Model Calibrations in a Changing World: Controlling for Nonstationarity After Mega Disturbance Reduces Hydrological Uncertainty

Abstract. Model simulations are widely used to understand, predict, and respond to environmental changes, but uncertainty in these models can hinder decision-making. The simulation of hydrological changes after a forest fire is a typical example where process-based models with uncertain parameters may inform consequential predictions of water availability. Different parameter sets can yield similarly realistic simulations during model calibration but generate divergent predictions of change, a problem known as "equifinality". Despite longstanding recognition of the problems posed by equifinality, the implications for environmental disturbance simulations remain largely unconstrained. Here, we demonstrate how equifinality in water balance partitioning causes compounding uncertainty in hydrological changes attributable to a recent 1,540 km² megafire in the Sierra Nevada mountains (California, USA). Different sets of calibrated parameters generate uncertain predictions of the four-year post-fire streamflow change that vary up to six-fold. However, controlling for nonstationary model error (e.g., a shift in the model bias after disturbance) can significantly (p < 0.01) reduce both equifinality and predictive uncertainty. Using a statistical metamodel to correct for bias shift after disturbance, we estimate a streamflow increase of 11 % ± 1 % in the first four years after the fire, with an 18 % ± 4 % increase during drought. Our metamodel framework for correcting nonstationarity reduces uncertainty in the post-fire streamflow change by 80 % or 82 % compared to the uncertainty of pure statistical or pure process-based model ensembles, respectively. As environmental disturbances continue to transform global landscapes, controlling for nonstationary biases can improve process-based models that are used to predict and respond to unprecedented hydrological changes.