Analysis of non-stationary extreme events in a climatological context

M. Nogaj, P. Yiou, D. Dacunha-Castelle and Ph. Naveau

One prevalent question rising in the scientific community concerns the 
change in occurrence and amplitude of abrupt and widespread climate 
events with major impacts in the past decades. It is conceivable that 
the anthropic forcing^ of climate change could increase the probability 
of extreme events, such as floods or heat waves. Using Extreme Value 
Theory, we analyze non-stationary time-series of temperature and 
precipitation. The probability of an extreme event under non-stationary 
conditions depends on the rate of change of the parameters of the 
distribution as well as on the rate of change of the frequency of their 
occurrence. In this study, we use the NCEP reanalysis data (1948-2004) 
of temperature and precipitation over the extended region of the North 
Atlantic. The data being highly dependent, a pre-processing by 
declustering (elimination of data aggregates) is needed. We then 
investigate the distribution of extremes over a given threshold, so that 
the resulting dates of exceedances follow a non-stationary Poisson 
process and the associated peaks are fitted by a Generalized Pareto 
Distribution with time-dependent scale parameters (sigma). These 
conditions are checked with likelihood tests. Within this framework, the 
concept of the return period is altered, since the return level is 
highly dependent on the extrapolated period of consideration. Moreover, 
the scale and Poisson intensity parameters are explained by different 
covariates, such as time, the North Atlantic Oscillation (NAO) or the 
Greenhouse gas content.

In summary, this research brings us a step further in the estimation of 
climate change impacts on abrupt climate events. Indeed, amplitude and 
frequency variations of extreme temperatures and heavy precipitation are 
evaluated at different locations.