A Spatial Time Series Framework for Modeling Daily Precipitation at Regional Scales

Estimates of precipitation at regional scales constitute one of the most important input parameters for hydrologic impact assessment studies. At these scales, Limited Area Models (LAMs) provide an emerging means for enhancing the accuracy of precipitation predictions (Giorgi and Mearns, 1991; Kim and Soong, 1996; Miller and Kim, 1996; Kim et al., 1998). Dynamic downscaling using LAMs yield precipitation predictions which are physically and dynamically consistent with other atmospheric variables produced in the downscaling procedure. Dynamical downscaling, however, is computationally expensive and not errorfree due to limited spatial resolution and model parameterizations. Stochastic characterization of rainfall fields based on rain gauge data and ancillary information, e.g., terrain elevation, still provides one of the basic tools for constructing rainfall maps at regional scales (Bras and Rodríguez-Iturbe, 1985; Seo et al., 2000; Kyriakidis et al., 2002), even though the physical and dynamic consistency of such maps is not guaranteed. Time domain approaches for modeling daily precipitation typically involve vectors of time series, e.g., multivariate autoregressive (AR) models. Such models exploit the typically better informed time domain, but are limited to predictions only at rain gauge locations (Wilks, 1998; von Storch and Zwiers, 1999). This limitation hinders the all important task of spatiotemporal mapping. More recently, time series approaches have been generalized to a continuous spatial domain and maps of precipitation levels are constructed at any arbitrary location via interpolation of time series model parameters (Johnson et al., 2000). In this paper, a framework for stochastic spatiotemporal modeling of daily precipitation in a hindcast mode is presented. Observed precipitation levels in space and time are modeled as a joint realization of a collection of spaceindexed time series, one for each spatial location. Time series model parameters are spatially varying, thus capturing space-time interactions. Stochastic simulation, i.e., the procedure of generating alternative precipitation realizations (synthetic fields) over the space-time domain of in-