Modelling non-stationarity in extreme rainfall using large-scale climate drivers

Flood estimates used in engineering design are commonly based on intensity–duration–frequency (IDF) curves derived from historical extreme rainfalls. Under global warming these extreme rainfalls are increasing, threatening the capacity of existing infrastructure to resist failure as IDF curves traditionally assume no change in rainfall magnitude. Hence, there is a need to investigate the implications of non-stationarity in extreme rainfalls used to derive IDFs across storm durations and annual exceedance probabilities (AEPs). One way of doing this is to incorporate covariates into the fitted probability distribution. However, there are few studies which examine non-stationarity in extreme rainfall using large-scale climate drivers as covariates, with little consensus on which covariate is the most appropriate. Here we evaluate non-stationarity in extreme rainfall across different durations, from the 1 in 2 AEP to the 1 in 100 AEP, using potential large-scale climate drivers including global and continental mean temperature, continental diurnal temperature range, continental dewpoint temperature, continental precipitable water, the Indian Ocean Dipole, the El Niño Southern Oscillation, and the Southern Annular Mode. These covariates are linked to the three parameters of the Generalized Extreme Value distribution to identify the most appropriate form of non-stationary probability model. We analyse 16 different storm durations from 6 min to 7 day annual maxima across 46 stations in Australia. Based on the Akaike Information Criteria, the precipitable water is the superior covariate at a large proportion of stations irrespective of storm duration. However, when the modelled quantile changes are inspected, only global temperature is able to adequately capture the variability in changes across both storm duration and annual exceedance probability. Further, when regional average values of mean temperature, diurnal temperature range, dewpoint temperature and precipitable water were used as covariates, there was no improvement in the model performance compared to continental-wide covariates, particularly for short durations. The results using a mean global temperature covariate show that stationary historical rainfall quantile estimates are underestimated by 12 % – 9 % for frequent (1 in 5 AEP) and 23 % – 13 % for rare (1 in 100 AEP) short duration (6 min – 30 min) events. Moving forward our results suggests infrastructure design needs to incorporate non-stationarity in IDFs to ensure infrastructure and flood planning levels are not under-designed.