Integrating Climatic and Physical Information in a Bayesian Hierarchical Model of Extreme Daily Precipitation

Charlotte Love,John England,Gregory Karlovits,Angela Duren,Amir Aghakouchak,Brian Skahill

doi:10.3390/w12082211

Abstract

Extreme precipitation events are often localized, difficult to predict, and available records are often sparse. Improving frequency analysis and describing the associated uncertainty are essential for regional hazard preparedness and infrastructure design. Our primary goal is to evaluate incorporating Bayesian model averaging (BMA) within a spatial Bayesian hierarchical model framework (BHM). We compare results from two distinct regions in Oregon with different dominating rainfall generation mechanisms, and a region of overlap. We consider several Bayesian hierarchical models from relatively simple (location covariates only) to rather complex (location, elevation, and monthly mean climatic variables). We assess model predictive performance and selection through the application of leave-one-out cross-validation; however, other model assessment methods were also considered. We additionally conduct a comprehensive assessment of the posterior inclusion probability of covariates provided by the BMA portion of the model and the contribution of the spatial random effects term, which together characterize the pointwise spatial variation of each model’s generalized extreme value (GEV) distribution parameters within a BHM framework. Results indicate that while using BMA may improve analysis of extremes, model selection remains an important component of tuning model performance. The most complex model containing geographic and information was among the top performing models in western Oregon (with relatively wetter climate), while it performed among the worst in the eastern Oregon (with relatively drier climate). Based on our results from the region of overlap, site-specific predictive performance improves when the site and the model have a similar annual maxima climatology—winter storm dominated versus summer convective storm dominated. The results also indicate that regions with greater temperature variability may benefit from the inclusion of temperature information as a covariate. Overall, our results show that the BHM framework with BMA improves spatial analysis of extremes, especially when relevant (physical and/or climatic) covariates are used.

Highlights

The ability to estimate the magnitude and frequency of extreme precipitation events is an essential part of infrastructure planning and flood prediction [1–3]
Each generalized extreme value (GEV) spatially dependent parameter is defined by a general linear model (GLM) of the covariates plus a spatial random effects term (τ) that accounts for residual spatial association not captured by the covariates
We argue that these top performing while XYZPT1 and XYZPT3 were the top for eastern Oregon (EOR) (Figure 2a)

Summary

Introduction

The ability to estimate the magnitude and frequency of extreme precipitation events is an essential part of infrastructure planning and flood prediction [1–3]. Extreme events directly affect infrastructure and residents of a region, while having an economic effect [3–5]. Extreme events can be localized and difficult to predict, especially when gauge data is sparse [6]. In regions that do have relatively dense observational networks, integrating spatial information (i.e., multiple extremes across space) for frequency is not straightforward [6]. Following the early contributions by Gumbel [7], which expanded upon the work of Fisher and Tippett [8], frequency analysis became widely used for estimating rainfall or streamflow corresponding to different return periods [5,9–11]. The fitted distribution is used to estimate the probability of occurrence of different events (e.g., precipitation and/or streamflow) [14]

Objectives

Methods

Results