Evaluation of flash drought identification with machine learning techniques. Part II: Neural network algorithms

Abstract

Flash droughts (FDs) describe the rapid onset of drought conditions and can have numerous impacts on agriculture and ecosystems due to rapid desiccation of the landscape. Recent studies into FDs have shown that the surface variables that drive these events include soil moisture, evapotranspiration, and potential evapotranspiration. However, studies that use machine learning (ML) to examine flash drought are still limited, despite the skill they have shown at identifying seasonal-scale droughts. This study is the second part in an investigation of FD representation and prediction using ML methods. Part I investigated traditional ML techniques, while this part investigated the ability of neural network–based methods to identify FD events. Key variables (soil moisture, evaporation, potential evaporation, temperature, and precipitation) were incorporated to train the ML models. Artificial neural networks (ANNs) struggled to identify FDs because they could not learn the surface interactions that characterize FD events. Convolutional U-networks were more capable of learning FD patterns but still overemphasized spatial patterns. Recurrent neural networks (RNNs), using a long short-term memory layer, most frequently had the highest skill score when predicting FD patterns and reduced much of the overprediction of FDs that occurred with the standard ML algorithms. RNNs could also recreate regions where FDs were more frequent as well as the seasonality climatology of FDs better than standard algorithms for some FD identification methods (but struggled more with the seasonality for other identification methods). However, the RNNs still struggled with the precise timing of FD events. Overall, RNNs showed some skill and promise in representing FDs on climatological time scales.