Short-range forecasts of rainfall amount from an extrapolative-statistical technique utilizing multiple remote sensor observations

Extrapolative techniques have been successfully used over the years to make short-range (0-6 h) forecasts of precipitation occurrence or amount in a number of diverse geographical regions. In an attempt to forecast heavy rainfall amounts in the 0-3h timeframe, we enhanced the purely extrapolative method by combining it with the Model Output Statistics technique. In this approach, forecasts of radar reflectivity, cloud-top temperature, and cloud-to-ground (CG) lightning strike rate were made for boxes on a 40-km map grid by extrapolation of initial-time fields, for a large number of historical cases. The extrapolated reflectivity and temperature values were then treated as candidate predictors of rainfall in a statistical regression procedure. Additional candidate predictors from NCEP models, such as upper-air humidity and stability indices, were added to the remote-sensor-based set. All predictors were then collated with the statistical predictand, which is the peak rainfall amount within the grid box. Rainfall amounts were derived from WSR-88D Stage III radar/gauge estimates, which provide hourly values on a 4-km national grid.

A forward-selection screening regression procedure was used to derive equations relating the predictors to the probability that observed rainfall would exceed various thresholds. Predictors most often selected by the regression procedure included maximum radar reflectivity and maximum lightning strike rate over the grid box as forecasted by extrapolation, and stability indices such as the K index. The equations are applied to real-time extrapolation forecasts and upper-air information to produce probability forecasts, specifically probabilities that rainfall will exceed 2.5, 12.5, 25, and 50 mm during the 0-3 h period. All forecasts describe the probability of the given event at some place within each grid box of a 40-km grid covering the conterminous United States. Categorical rainfall amount forecasts are derived from the rainfall probability forecasts by comparing the probabilities to a set of predetermined threshold values.

Input to the algorithm includes a 10-km national radar mosaic, a 10-km infrared satellite temperature image, 15-minute lightning strike rates, and stability and humidity forecasts from the NCEP Nested Grid Model (NGM). The mean 700-500 mb wind field was used as the extrapolation velocity field for radar, satellite, and lightning features. Stability and humidity indices were obtained from the NGM analysis or 6-h forecast valid closest in time to the latest radar image.

A comparative verification study was made among probabilistic forecasts from the extrapolative system, 0-3 h and 3-6 h rainfall amount forecasts from the Rapid Update Cycle Version 2 model(RUC-2) valid during the same time period, and 6-12 h rainfall forecasts from the NGM and the Eta models, valid at the end of the 3-h valid period. Forecasts and verifying observations for the 2100-0000 UTC period from the 1999 warm season were studied. Note that 0-3 h RUC-2 forecasts are not actually available until some time into the valid period. However, both sets of RUC-2 forecasts proved to be superior to the NGM or Eta model forecasts. In general, the extrapolative forecasts explained about 70% more of the predictand variance than did the best numerical model forecasts, namely those from the 3-6 h RUC-2. In terms of yes/no forecasts (rainfall threshold exceeded/not exceeded), the extrapolative algorithm would produce 40-50% fewer false alarms than would the RUC-2, given that the two systems detected the same percentage of events. Smaller but still significant improvements over the RUC-2 were found for the cool season.

The algorithm is slated for operational implementation, with national dissemination of forecasts to National Weather Service field offices and to external users.