A large-sample investigation of statistical procedures for radar-based short-term quantitative precipitation forecasting

We present the results of an extensive evaluation of radar-based quantitative precipitation forecasting techniques. Using a large data set of radar observations from the Tulsa, Oklahoma, WSR-88D radar we evaluate several techniques, including persistence, advection, and neural-network-based schemes. The scope of our study is limited to very-short-term forecast lead-times of up to three hours. We consider several spatial resolutions ranging from 4×4 km2 to 32×32 km2. Performance of the schemes is evaluated using several popular criteria that include correlation coefficient, multiplicative bias, and probability of detection. We discuss the effects of average storm intensity and rainfall intensity integration on the predictability limits. The most significant conclusions from the study are: (1) advection is the most important physical process that impacts useful predictions; (2) larger and more intense storms are easier to forecast; and (3) both spatial and temporal integration significantly extends the predictability limits.