From Drought Monitoring to Forecasting: A Statistical-Dynamical Modeling Framework

Although research on drought prediction has shown some improvement over the years, accurate provisions of drought information in a timely manner is still a challenge. Both statistical and dynamical drought prediction methods have been widely used in research and practice. While these approaches have yielded skillful predictions in specific case studies, some limitations still restrict their use. One of the main limitations is the deterministic treatment of the initial conditions. This motivates development of a drought prediction system that is based on full characterization of the initial condition. The framework employs a data assimilation (DA) method based on particle filter (PF) to quantify the uncertainties associated with antecedent land surface condition. The initial condition at each forecast step is probabilistically sampled from the ensemble of initial conditions characterized by data assimilation and through a multivariate approach based on copula functions the  probabilistic drought prediction is developed. Large computational demands are overcome by developing a modular parallel particle filtering framework (PPFF) which facilitates large ensemble sizes in PF applications. The framework is examined on both synthetic and real case studies in the Western U.S. basins. Both case studies demonstrate that the proposed framework improves the drought prediction skill significantly. The ability of the framework to predict the declared droughts in the area in a timely manner is also assessed.