Climate change impacts parameter sensitivity - What does this mean for model calibration?

Many hydrological models that are used for long term projections require calibration of at least a few parameters. When calibrated on discharge only, a general rule of thumb is that 4 to 5 parameters can be calibrated. The general approach is to conduct a global sensitivity analysis, to determine the four to five most sensitive parameters, and to select these for calibration.

Parameter sensitivity differs over models, target variables, sensitivity analysis methods, and also over climates. This would also imply that parameter sensitivity could change in a changing climate, and that would interfere with the current standard calibration procedure for hydrological models. Therefore, the question is whether, within a plausible rate of change, climate change propagates into a change in parameter sensitivity.   

We investigated how parameter sensitivity changes as a consequence of climate change, and if and how this has consequences for the calibration strategy. We applied a hybrid local-global sensitivity analysis method to three frequently used hydrological models (SAC, VIC, and HBV) in 605 basins across the US, and link changes in sensitivity to changes in climate. Finally, we evaluated the impact on the top five most sensitive parameters.

The results show that in all three models especially snow parameters tend to become less sensitive in the future. However, the models differ in which parameters increase in sensitivity; for some models ET parameters increase, while for others deep layer parameters increase. Evaluating the top 5 most sensitive parameters per basin, we found that in 43% to 49% of the basins at least one parameter changes in the top 5 in the future, while a maximum of two parameter changes in the top 5 was observed over all basins (in 2 to 4% of the basins).

Overall, the results indicate that in about half of the investigated basins one parameter would have been chosen differently for calibration. If a particular model parameter is, within the current climate, not or hardly sensitive to discharge, it is not possible to calibrate this parameter – notwithstanding whether this parameter becomes sensitive in the future. Therefore, the consequence of these results is that for parameters that will become sensitive in the future, a range of feasible parameter values have to be sampled for future projections, thereby capturing predictive uncertainty as a consequence of changing sensitivities.