Effects of spatial and temporal aggregation on the accuracy of statistically downscaled precipitation estimates in the upper Colorado River basin

To test the accuracy of statistically downscaled precipitation estimates from numerical weather prediction models, a set of experiments to study what space and time scales are appropriate to obtain downscaled precipitation forecasts with maximum skill have been carried out. Fourteen-day forecasts from the 1998 version of the National Centers for Environmental Prediction (NCEP) Medium-Range Forecast (MRF) model were used in this study. It has been previously found that downscaled temperature fields have significant skill even up to 5 days of forecast lead time, but there is practically no valuable skill in the downscaled precipitation forecasts. Low skill in precipitation forecasts revolves around two main issues. First, the ( intermittent) precipitation variability on daily and subdaily time scales could be too noisy to derive meaningful relationships with atmospheric predictors. Second, the model parameterizations and the coarse spatial resolution of the current generation of global-scale forecast models might be unable to resolve the local-scale variability in precipitation. Both of these issues may be addressed by spatial and temporal averaging. In this paper the authors present a diagnostic study using a set of numerical experiments to understand how spatial and temporal aggregations affect the skill of downscaled precipitation forecasts in the upper Colorado River basin. The question addressed is, if the same set of predictor variables from numerical weather prediction models is used, what space ( e. g., station versus regional average) and time ( e. g., subdaily versus daily) scales optimize regression-based downscaling models so as to maximize forecast skill for precipitation? Results in general show that spatial and temporal averaging increased the skill of downscaled precipitation estimates. At subdaily ( 6 hourly) and daily time scales, the skill of downscaled estimates at spatial scales greater than 50 km was generally higher than the skill of downscaled estimates at individual stations. For the 6-hourly time scale both for stations and for mean areal precipitation estimates the maximum forecast skill was found to be approximately half that of the daily time scale. At forecast lead times of 5 days, when there is very little skill at daily and subdaily time scales, useful skill emerged when station data are aggregated to 3- and 5-day averages.