Bayesian prediction of monthly precipitation on a fine grid using covariates based on a regional meteorological model

Abstract

In this article a Bayesian hierarchical model (BHM) for observed monthly precipitation is proposed. This BHM incorporates covariates based on an output on a fine grid from a regional meteorological model. At the data level of the BHM, the observed monthly precipitation is transformed using the Box–Cox transformation, and each month is modeled separately. To capture spatial correlation at the data level, a Gaussian field with Matérn correlation function is used. It is assumed that the data are subject to measurement error. The location and log‐scale parameters at the latent level are also modeled with Gaussian fields with Matérn correlation functions. An output from a regional meteorological model on a fine grid is used to construct spatial covariates for the latent parameters of the BHM for each month of the year. These covariates are then projected onto each of the observed sites for each month and incorporated into the BHM. Markov chain Monte Carlo simulation is used for posterior inference and Bayesian kriging is used to predict the latent parameters on the grid. This BHM was applied to observed data on monthly precipitation, which come from forty sites across Iceland from the years 1958 to 2006. The data were corrected for wind, wetting, and evaporation loss. An output from a linear model of orographic precipitation defined on a 1 km by 1 km grid over Iceland was used to construct the covariates for the BHM. Copyright © 2015 John Wiley & Sons, Ltd.