Flash Drought Prediction using Deep Learning

Abstract Flash droughts are sudden, short-term drought events that develop within weeks, unlike traditional droughts that unfold gradually over time. These events arise from a combination of climatic factors, such as low rainfall, high temperatures, and strong winds, which rapidly deplete soil moisture and stress vegetation, leading to severe agricultural, economic, and ecological impacts. Despite the significant challenges in defining and analysing flash droughts, only a few studies have employed machine learning techniques to predict these occurrences. The use of machine learning in this context remains in its early stages due to complications like imbalanced datasets and limited data size. This study addresses these challenges by employing Convolutional Neural Networks (CNN) to predict flash droughts in Eastern Australia - a region historically prone to these events. We identified flash droughts from 2001 to 2022, with the model training performed on data from 2001-2015, validation from 2016-2017, and testing from 2018-2022. The model's performance was assessed across different scenarios, including drought duration, spatial distribution, and seasonal variability. The CNN achieved a balanced accuracy of 80% and an Area Under the Curve (AUC) of 93%, demonstrating its capability to predict flash drought events effectively. While the model showed promising results in accurately forecasting flash droughts, it tended to overestimate the spatial extent of drought-prone regions, highlighting areas for future improvement. These findings underscore the potential of deep learning, in enhancing our understanding and prediction of flash droughts, offering a valuable tool for early warning systems and drought management strategies.