Extended Abstract (1.8M)J5.3
A spatial time series framework for modeling daily precipitation at regional scales
Phaedon C. Kyriakidis, Univ. of California, Santa Barbara, CA; and N. L. Miller and J. Kim
A geostatistical framework is proposed for space-time modeling of daily precipitation at regional scales. Emphasis is placed on the typically better informed time domain: the spatiotemporal distribution of daily precipitation levels is modeled as a joint realization of a collection of spatially correlated time series.
Temporal non-stationarity is accounted for by establishing parametric temporal trend models at rain gauge locations; the spatial average of all precipitation profiles is used as the single predictor in establishing such local temporal trend models. Fitted temporal trend model parameters, e.g., local regression coefficients which are different from one rain gauge location to another, are then regionalized (interpolated or better simulated) in space for determining trend models at any unmonitored location. This regionalization procedure takes into account any spatial cross-correlation in the model parameters, and any relationship between these parameters and ancillary information such as elevation or large-scale atmospheric state variables from NCEP/NCAR reanalysis data.
The remaining spatiotemporal residual component, modeling any spatiotemporal variability not included in the trend component, is also viewed as a collection of spatially correlated (stationary) residual time series. Parametric time series models are then established at the rain gauge locations and the fitted parameters are again regionalized in space, accounting for their cross-correlation and any relevant ancillary information.
Stochastic conditional simulation is used for generating alternative synthetic realizations of daily precipitation fields for the period of interest; this is achieved indirectly via conditional simulation (in space) of the parameters of the trend and residual time series. All simulated precipitation fields identify the rain gauge precipitation measurements; they also reproduce the sample precipitation histogram, as well as their spatiotemporal correlation.
The spatiotemporal model outlined above can be readily extended to the mixture processes case, i.e., modified to simulate both daily precipitation intermittence and intensity in space and time, thus adding to the existing pool of stochastic weather generators. Downscaling (spatial and/or temporal) can also be viewed within this framework as the procedure of integrating information available on a coarse scale (average precipitation values over specific areas or time periods) to further condition fine-scale precipitation realizations, in addition to the rain gauge precipitation data.
The proposed spatiotemporal precipitation model is employed for generating alternative synthetic realizations of daily precipitation during the period of 11/01/1981 to 01/31/1982 over a region in northern California at 1 km resolution. Regionalization of the parameters of the temporal trend and residual time series is conditioned to coarse-scale, lower-atmosphere variables obtained from the NCEP/NCAR reanalysis and their interaction with local terrain characteristics. The alternative precipitation fields can be used for evaluating hydrologic response uncertainty due to uncertain precipitation forcing, a task of paramount importance in impact assessment studies.
Joint Session 5, Statistical Downscaling (Joint with the 16th Conference on Probability and Statistics and the 13th Symposium on Global Change and Climate Variations)
Tuesday, 15 January 2002, 4:00 PM-5:30 PM
Previous paper Next paper