Do Small and Large River Floods Change Differently under Future Climate Change?

Future fluvial flood risk can be estimated by applying change factors (CFs) to present day fluvial flood hydrology. At the global scale, CFs can be derived from the outputs of GCM-GHM ensemble pairs, such as ISIMIP3b. CFs are often calculated based on the proportional change of an index flood (e.g. median annual flood) between the future and present day, and subsequently applied over all return periods. Whilst this approach is statistically robust given the limited number of samples in ISIMIP3b, it assumes that rare and regular floods change proportionally. Scientific literature disputes this, suggesting that future changes may vary significantly based on flood rarity. Calculation of return period specific change factors can be achieved with stationary extreme value analysis on baseline and future periods, however the limited samples from climate ensembles result in noisy change factors for extreme events. Moreover, the analysis applied using different ensemble members results in greater uncertainty. A potential method to increase sample size to reduce noise is to apply non-stationary extreme value analysis across an ensemble’s entire time series.

This work compares multiple approaches to derive return period dependent CFs, including stationary flood frequency analysis, and non-stationary GAMLSS, across multiple distributions and ensemble members. We demonstrate and assess the differences in CFs between 2y (regular) and 100y (rare) floods, using Europe as a test domain. We find that CFs can be modeled as return period dependent, with clear spatial patterns present, and significant differences in changes in rare and regular floods in many locations. The variation of CF with return period depends on region, driving GCM - GHM pair, and the selected distribution fitting approach. In some locations, GCM-GHM pairs largely agree on the degree and direction of change between rare and regular floods, yet in others, GCMs-GHM pairs can give very different estimates of this (including opposing directions). Stationary flood frequency analysis for CFs results in greater noise, and CFs with greater magnitude. The GAMLSS approach significantly mitigates spatial noise, but reduces the intensity of changes in CFs as a function of RP. Our study overall highlights the importance of considering how CFs vary by return period, and how this variation itself is critically dependent on the driving GCM-GHM ensemble used.