Evaluation of Flash Drought Identification with Machine Learning Techniques

Droughts are extreme dry events that decrease an ecosystem’s and society’s availability to water resources, leading to impacts on vegetation health and agricultural production and food shortages. Of particular note are droughts that develop on a more rapid timescale (about 1 month), termed flash droughts. Flash droughts have gained increasing attention in the past decade, because they can result in more rapid desiccation, or deterioration in crop health, than what would normally be expected. Research into flash drought events have found certain key variables, such as soil moisture and evaporation from the soil and plants, and potential evaporation are among the key variables driving flash drought events. Varying approaches have resulted in the creation of multiple methods for identifying and quantifying flash droughts, each using different variables and thresholds (for the rapid intensification) to define them.
Machine learning (ML) techniques have been growing in popularity in the environmental sciences due to their ability to accurately represent different environmental phenomena, including drought. However, the use of applying ML to rapidly developing droughts remains largely unexplored. Thus, this dissertation aims to investigate the ability of various ML techniques to identify flash drought phenomena, using multiple different methods for identifying flash drought. Because ML use in flash drought is largely unexplored, this dissertation explores how multiple different ML algorithms, such as random forests, support vector machines, different types of artificial neural networks, can identify flash drought events. The ML algorithms were trained on key variables known to drive flash drought events – soil moisture, evaporation, potential evaporation, temperature, precipitation, and the change in soil moisture, evaporation, and potential evaporation. Lastly, feature importance (the importance of each variable to the ML algorithms) was determined from Shapely values and permutation importance methods to give the ML algorithms interpretability and explainability. Results showed ML is capable of representing flash drought events, with boosted trees and recurrent neural networks showing the most skill. However, the ML algorithms thought flash droughts were more active in the hotspot regions, and the more likely in the late growing season, than observations actually show. Feature importance in the ML algorithms showed that the algorithms were relying heavily on soil moisture, precipitation, and potential evaporation to predict flash drought, explaining why they over-emphasized the seasonality. Global representation of flash drought will also be investigated to determine how well the ML models generalize and represent flash drought over vastly varying ecosystems, and to examine rarely examined flash drought events. The skill ML has shown can allow us to not only push our understanding of flash droughts forward, but can also help represent flash droughts in numerical models and allow us to identify flash droughts in real time, and potentially lead to flash drought predictions.