Application of Logistic Regression to Forecast and Monitor the Onset of Drought over the Contiguous United States

Abstract:

This study analyzes the probabilistic precipitation forecasts produced by the climate system forecasting model (CFSv2) using a linear logistic regression approach over the Contiguous United States (CONUS) for the onset and growth of drought in 2012 and 2017. The precipitation forecasts are bias-corrected using BCSD approach before analyzing, as the raw precipitation forecasts produced by the general circulation models (GCMs) exhibit biases. The North American Multi-Model ensemble system (NMME) experiments show that the GCM forecasts exhibit poor forecast skills beyond the first lead month over the CONUS even after the bias-correction, and this is a stumbling block for the seasonal drought forecasting. This study explores the application of a linear logistic regression in conjunction with climate pooling by enhancing the training data set at a particular location with the data in the surroundings based on the similarity of precipitation-surface temperature correlation. The logistic regression examines the changes in the standardized precipitation index (SPI) for delta SPI <-0.5 and -1.0 instead of the SPI itself as the changes in the precipitation forecasts are less susceptible to systematic biases. Our results indicate that the logistic regression-produced ensemble mean probability provides a reliable system in identifying the onset and growth of the flash drought in the 2012 and 2017 years. We evaluated this approach with the Brier skill score, where the simple probability based on the CFSv2 ensemble members used as reference probability for several drought severity categories. The areas of drought onset and growth are detected well by the logistic regression approach, with more than 75% of the drought-affected regions over the CONUS show realistic SPI changes compared to the straightforward approach based on the ensemble probabilities.