Nonstationary Frequency Analysis of Annual Maximum Rainfall Using Climate Covariates
The perception that hydrometeorological processes are non stationary on timescales that are applicable to extreme value analysis is recently well documented due to natural climate variability or human intervention. In this study the generalized extreme value (GEV) distribution is used to assess nonstationarity in annual maximum daily rainfall time series for selected meteorological stations in Greece and Cyprus. The GEV distribution parameters are specified as functions of time-varying covariates and estimated using the conditional density network (CDN) as proposed by Cannon (2010). The CDN is a probabilistic extension of the multilayer perceptron neural network. If one of the covariates is dependent on time, then the GEV-CDN model could perform non stationary extreme value analysis. Model parameters are estimated via the generalized maximum likelihood (GML) approach using the quasi-Newton BFGS optimization algorithm, and the appropriate GEV-CDN model architecture for a selected meteorological station is selected by fitting increasingly complicated models and choosing the one that minimizes the Akaike information criterion with small sample size correction or the Bayesian information criterion. For each meteorological station in Greece and Cyprus different formulations are tested with combinational cases of stationary and non stationary parameters of the GEV distribution, linear and nonlinear architecture of the CDN and combinations of the input climatic covariates. Climatic covariates examined in this study are the Southern Oscillation Index (SOI), which describes atmospheric circulation in the eastern tropical Pacific related to El Nio Southern Oscillation (ENSO), the Pacific Decadal Oscillation (PDO) index that varies on an interdecadal rather than inter annual time scale and atmospheric circulation patterns as expressed by the Mediterranean Oscillation Index (MOI) and North Atlantic Oscillation (NAO) indices.