Evaluation of Flash Drought Identification with Machine Learning Techniques, Part 2: Common Deep Learning Algorithms

Flash droughts (FDs) are drought conditions that develop on a rapid time scale (~1 month). As such, traditional drought metrics such as the Palmer drought severity index and standardized precipitation index, which focus monthly to annual drought impacts, have been unable to properly represent FD events. Thus, studies have focused on developing indices that can capture rapidly developing drought. These indices have focused on soil moisture, evapotranspiration (ET), and potential ET (PET) as the key drivers. To mitigate impacts, many studies have focused on identifying and predicting drought events. Furthermore, machine learning (ML) has emerged in recent decades as a useful tool for drought identification and prediction, and ML applications using long-term metrics can successfully identify seasonal scale droughts and have shown some skill at drought prediction. In addition, while much work has been conducted to investigate and identify FD events, yielding multiple methods to identify FD, limited work has been completed using ML techniques. Thus, this study seeks to use several popular NN architectures to identify flash drought such as artificial neural networks (NNs), recurrent NNs (RNNs), and combined convolutional and recurrent NNs (CRNNs). The NNs were trained on a set of flash drought variables (soil moisture, ET, PET, precipitation, and temperature) and additional analyses were performed to determine how much each variable contributes toward rapid intensification and flash drought identification.