39HYDRO Investigation of Machine Learning Algorithms in Representing Global Flash Drought Patterns

Flash droughts describe droughts that develop over a rapid time scale (~1 month). Focus on FDs over the recent years has uncovered several key variables in driving FD events, the most important of which is soil moisture, evapotranspiration, and potential evapotranspiration, though other variables such as temperature and precipitation play an important role. Recent studies have also begun to develop ways to quantify FD events, resulting in the development of multiple FD identification methods. These FD identification methods have been used to examine their climatology, on both the regional and global scales. In addition, studies have begun to investigate the ability of machine learning techniques to represent FD events on the regional scale for the contiguous United States, and wherein some of the machine learning methods were able to represent the climatology of FD events. To expand on this, this study incorporated machine learning to predict flash drought across the globe for a global analysis, using the ERA5 reanalysis dataset for 1979 – 2021. In this study, machine learning techniques, Ada boosted trees and recurrent neural networks, were applied on the global scale to represent FD events. Results found the machine learning models were better optimized for local scale predictions while global scale predictions often focused on specific latitudes at the expense of others. This study also provided an opportunity to examine global flash drought patterns with different flash drought identification methods, and it provided an opportunity to examine various flash drought events across the globe and examine the varying societal responses to FD events via case studies.