Probabilistic Seasonal Prediction of Drought Using Multivariate Observations

A probabilistic ensemble method using the bootstrap is developed to assess the predictability of the future state of the standard precipitation index (SPI) commonly used for drought monitoring. The methodology is data driven, and has the advantage of being easily extended to use more than one variable as predictors. Using 110 years of monthly observation of precipitation, surface air temperature, and 60 years of Nino 3.4 index, the method was employed to assess the impact of the different variables in enhancing the prediction skill. A predictive probability density function (PDF) is produced for future six-month SPI and a log-likelihood skill score is used to cross-compare various combination scenarios using the entire predictive PDF and with reference to the observed values set aside for validation. The results suggest that the multivariate prediction using complementary information from three and six month SPI and initial surface air temperature significantly improves seasonal prediction skills for capturing drought severity and delineation of drought areas based on observed six month SPI. The improvement is observed across all seasons and regions over the continental United States compared to the other prediction scenarios that ignore the surface air temperature information. Furthermore, we explored two other variables: relative humidity and vapor pressure deficit and demonstrated that they contain invaluable predictive skills that can increase the lead time of prediction according to the reference drought severity indices from the U.S. Drought Monitor (USDM).