Comparing a Spatial Propinquity Extreme-Value Model with a Simple Univariate Generalized Pareto Approach for Precipitation

Extreme weather and climate events such as heavy precipitation, drought, heat waves and strong winds can cause extensive damage to the society in terms of human lives and financial losses. As climate changes, it is important to understand how spatial distribution of extreme weather events may change as a result.

Two approaches are employed to represent spatial dependence of extreme field. First, traditional temporal generalized Pareto (GP) model is applied at individual locations with quantile-based thresholds. Second, rather than considering extreme values at individual locations and their temporal dependence, we consider an overall spatial field that is conditioned on being extreme by utilizing the conditional extreme value model introduced by Heffernan and Tawn (2004) (HT2004) and applied to identify severe storm environments in Gilleland (2013).

We apply these models to an observed precipitation dataset over CONUS. The dataset is stratified by decades and model forecasts are compared by mean error distance metric and by three geometric indices defined in AghaKouchak et. al (2010). Ultimately, we use supervised machine learning (random forest) classification model to rank importance of the geometric quantities resulted from univariate GP and multivariate HT2004 models.

This work is a first step to utilize a novel approach in statistical model comparison to help to deepen our understanding on added value (i.e. connection between individual grid cells) that multivariate spatial model such as HT2004 represents for spatial dependence analysis of extremes fields that are not modeled in traditional univariate case. We intend to expand this approach to compare climate and statistical model forecasts of extremes in our future research.