The peaks-over-threshold (POT) approach to defining extreme observations is intuitive and less wasteful than other methods such as the block maxima approach. However, POT requires the selection of an appropriate high threshold to define an extreme observation. In spatial modeling problems, POT requires threshold selection at each location in the study. Threshold selection is usually performed using visual tools and therefore can be subjective and tedious when modeling data from many locations.

Extended generalized Pareto distributions (EGPDs), extensions of the standard POT model, are more robust to threshold selection and appear suitable for modeling extremes using low thresholds, thereby incorporating more data in the analysis and potentially eliminating careful threshold selection by location. This work will extend an EGPD spatially using a Bayesian hierarchical model to predict rainfall return levels across Nova Scotia.