Optimal drought forecasts from a multi-model framework

Recent summers in the United States have been plagued by intense droughts that have caused significant economic impact to agriculture and to society that could be reduced through seasonal prediction. Skillful seasonal predictions of drought relevant variables are possible due to the slowly varying boundary conditions and their predictability. In particular, during the convective season, when the potential of extreme drought is the highest, the soil moisture can provide a means of predictability through land-atmosphere interactions. In particular recent work has developed a new classification of land-atmosphere interactions that was the bases for the Coupling Drought Index (CDI) that assesses the impact of coupling on drought. This metric was used to understand the predictability of land-atmosphere interactions in a NCEPS Climate Forecasts System version 2 (CFSv2) and indicated that there were strong biases in the coupling that lead to biases in the precipitation and temperature predictions. These shortcomings in the seasonal prediction motivated the development of the Coupling Stochastic Model (CSM), which utilizes the persistent coupling states to make seasonal predictions through a Markov Chain model coupled to a probabilistic prediction of precipitation and temperature. The CSM model can be used as an independent forecast model or to correct the known issues with existing climate models and was shown to improve the prediction of precipitation during drought. To this point the statistical model has relied on the initial conditions and persistence probabilities from reanalysis, however the needed variables to calculate the CDI and the associated statistical model parameters are available through satellite remote sensing. This provides a means to incorporate a more observationally based land-atmosphere product into the CSM model that could improve the statistical seasonal predictions. Although statistical seasonal predictions can provide a means of predictability, they are limited by the observations from which they are derived. In this sense both physically based and statistical models have the there advantages and disadvantages and better forecast can be achieved by incorporating both types of models into a multi-model framework that appropriately accounts for these strengths and weaknesses. In this work we present and test a framework for optimally combing forecasts from the North American Multi-Model Ensemble (NMME) and the CSM. This framework is evaluated over the U.S and optimally combines these models based on past performance at different temporal and spatial scales. The results, implications and limitations of this framework are presented.