Evaluating the role of weather forecast skill in seasonal drought prediction

Drought is a slowly developing natural phenomenon, which can result in large economic losses. Drought losses could potentially be mitigated through the use of skillful forecasts. Hydrological drought prediction skill is related both to knowledge of initial hydrologic conditions (primarily soil moisture) and climate forecast skill. Climate forecast skill is, in most cases (aside from regions for strong ENSO signals, for instance) modest. For this reason, the traditional approach to seasonal hydrological forecasts has used the Ensemble Streamflow Prediction method (ESP), which in its most common application resamples ensembles of future weather from observations, implicitly assuming that climate forecast skill is no better than climatology. However, weather forecasts have demonstrable skill for lead times of at least several days. We evaluate a drought forecast scheme as a modification of ESP that uses weather forecasts taken from reforecasts produced by the NCEP Global Forecast System for the first 15 days, merged with climatology thereafter. In particular, we report the results of three experiments. In each experiment the ESP method was used to predict soil moisture percentiles out to 6 months lead, on the 1st and 15th of each month during 1971-2000. The first experiment used a 30-member ensemble weather forecasts for the prediction period (6 months) sampled from gridded observations for the period 1971-2000. In the second experiment, the 15-day ensemble mean GFS forecast was substituted for the first 15 days of the 6-month ESP forecasts. The third experiment is similar to the second but gridded observations were used in lieu of the GFS ensemble mean forecast. Drought prediction skill in the first experiment is derived solely from the knowledge of initial condition, in the second and third experiment from the knowledge of initial conditions and imperfect/perfect knowledge of the first 15 days weather conditions. Forecasted soil moisture (SM) percentiles are evaluated with respect to SM percentiles derived by forcing the model with gridded observations for the 1971-2000 period over the Continental US.