Methods for exploring spatial and temporal variability of extreme events in climate data

This study presents various statistical methods for exploring and summarizing spatial extremal properties in large gridpoint datasets. Extremal properties are inferred from the subset of gridpoint values that exceed sufficiently high, time-varying thresholds. A simple approach is presented for how to choose the thresholds so as to avoid sampling biases from nonstationary differential trends within the annual cycle. The excesses are summarized by estimating parameters of a flexible generalized Pareto model that can account for spatial and temporal variation in the excess distributions. The effect of potentially explanatory factors (e.g., ENSO) on the distribution of extremes can be easily investigated using this model. Smooth spatially pooled estimates are obtained by fitting the model over neighboring grid points while accounting for possible spatial variation across these points. Extreme value theory methods are also presented for how to investigate the temporal clustering and spatial dependency (teleconnections) of extremes. The methods are illustrated using Northern Hemisphere monthly mean gridded temperatures for June-August (JJA) summers from 1870 to 2005.