Logit transforms in forecasting precipitation type

The Meteorological Development Laboratory (MDL) has recently completed an updated suite of Model Output Statistics (MOS) guidance for the Global Forecast System (GFS). As part of this new package, equations for the conditional probability of freezing, frozen, and liquid precipitation types were redeveloped for the contiguous U.S. (CONUS) and Alaska, for the 0000 and 1200 UTC cycles every 3 hours out to 192 hours, and out to 84 hours for the 0600 and 1800 UTC cycles. Seven cool seasons (15 September – 15 May) of METAR present weather observations and GFS model forecast data was available to develop the equations.

We developed generalized and regionalized operator equations using data from 1243 METAR sites (CONUS and Alaska) which report present weather reliably. For each station and time of day (i.e., 0000, 0300, …, 2100 UTC), the logit model was used to calculate “50%” values for several variables forecast by the GFS known to be good discriminators of precipitation type. These include, for example, 850 hPa temperature, 1000-850 hPa thickness, and 1000-500 hPa thickness. Due to the rarity of freezing and frozen precipitation events in some parts of the country (i.e., the southern U.S. and parts of California), stable estimates for the 50% values could not be obtained for some stations, and had to be estimated from each variable's dependence on station elevation. Next, “transformed” predictors (or deviations from the 50% values) were calculated for each station and each variable, by subtracting the 50% values from the model forecast of that variable. These transformed predictors account for climatological differences between stations within a given region. Finally, the station 50% values, as well as climatological relative frequencies of each precipitation type, were analyzed to a 5-km grid. This allows values to be interpolated to any station not included in the development and for which precipitation type forecasts are desired.

Other predictors offered to the regression analysis include several thicknesses, temperature and wet-bulb temperature at various levels, u- and v-wind components and speed, and sine and cosine day of the year. A “k-fold” cross-validation procedure was used whereby each season of data was withheld as an independent sample and equations developed on the remaining seasons. Additional developmental details and verification results for the independent data will be included in the extended abstract and presented at the conference.