When we want to estimate climate risks, we need to model climate impacts. Our current climate tend to have biases even in the most basic variables. They are too hot or too cold, too dry or too wet. Impact models, however, typically require unbiased input. Fire is modelled as a response to available fuel and a certain level of humidity and temperature. Crop yields are modelled as response to accumulated temperatures and available moisture. Human mortality occurs when a particular threshold of a heat stress indicator is crossed. Hence, to model climate impacts, we need to apply bias adjustment to the outputs of climate models.
New paper published
Bias adjustment is typically applied to one variable at a time. Yet many impacts depend on multiple climate variables. It is unclear how well bias adjustment works in this case. A direct evaluation is difficult because we lack good data on observed impacts, which would be needed to evaluate modelled impacts. To avoid this difficulty, in our recently published paper in Earth System Dynamics we studied the effect of bias adjustment on multivariate hazard indicators. We used a heat stress and a fire risk indicator, which can both be estimated from temperature and relative humidity.
Traditional bias adjustment can fail
We found that in some cases, applying traditional bias adjustment on the climate drivers (here temperature and relative humidity) can lead to even larger biases for hazard indicators. Hence in such cases, applying no bias adjustment would be better. These cases typically occur (i) when the variability of the modelled impact truly depends on multiple drivers, (ii) when the original biases in the climate drivers are comparably small, and (iii) when the dependence between the climate drivers is not modelled correctly. We show that a bias adjustment method that also corrects the dependence structure can successfully solve these issues.
Relevance for compound events
The results are particularly relevant for modelling impacts of compound events, as we often don’t know how well climate models represent the dependencies between relevant climatic drivers. “Real” impacts are probably even more complex, involving more variables. As long as biases persist in climate models, impact modelling will require some sort of bias adjustment. When impacts depend on multiple variables, multivariate bias adjustment methods should be used.
(The photo on the top of this page was taken by Amir AghaKouchak)
