Ensemble Logistic Regression to Forecast and Monitor Different Phases of Drought over Contiguous United States

This study explores the usage of logistic regression to forecast and monitor the drought onset and variability over Contiguous United States (CONUS) using ensemble forecasts of the climate forecast system model (CFSv2) on monthly time scales. The precipitation forecasts produced by the general circulation models (GCMs) are well known to have poor forecast skill beyond one lead month, which results in raw forecasts that are not useful for drought forecasting. Several bias correction methods are proposed to correct the precipitation forecasts in order to improve the skill beyond a month. However, the reliability and efficacy of the bias correction methods to produce usable forecasts is not uniform across the seasons and regions over the CONUS. In this study, we explore a novel way to forecast and monitor a drought event by examining the changes in the model forecasted standardized precipitation index (SPI) using logistic regression for each ensemble member. The logistic regression uses the historic precipitation and temperature observational record as the training set and addresses the predicted event: "what is the probability of an ensemble member producing the reduction of SPI by -0.5 in the forecast given the changes in historical precipitation and temperature in consecutive months". The forecast probability in conjunction with the real-time observations provides excellent guidance to forecast the predicted state of drought phase and severity. This method produces encouraging results for Brier skill scores and other metrics when applied to (i) the onset of flash drought over the Northern Plains in the late spring of 2017, which was not detected by the North American Multimodel Ensemble system (NMME) forecast and (ii) growing and decay phases of the mega drought that occurred over the Texas and Mexico regions in 2011.