Viewing Soil Moisture Flash Drought Onset Mechanism and Their Changes Through XAI Lens: A Case Study in Eastern China

Abstract

Soil moisture flash droughts often pose significant challenges to humans and ecosystems, with wide‐ranging socioeconomic consequences. However, the underlying mechanisms of flash droughts and their changes remain unquantified. Taking China as a case study, we present a novel framework that combines machine learning with interpretable and cluster techniques to investigate flash drought mechanisms from 1980 to 2018. We first quantified the temporal contribution of drivers and further identified different mechanisms during drought onsets. We subsequently investigated the temporal changes in different mechanisms and classified drought event types. We identified four driving mechanism types triggering drought: Concurrent precipitation, Antecedent‐concurrent precipitation, Antecedent temperature‐concurrent precipitation, and Antecedent transpiration‐concurrent precipitation. The total effects from vegetation transpiration contributed to around 50% of the impacts for mechanisms involving antecedent transpiration and concurrent precipitation, highlighting the non‐neglectable role of vegetation water consumption in drought occurrences. Remarkably, about 60% of flash drought onsets exhibited close association with the antecedent anomalies, which contribute approximately 50% of overall effects, emphasizing the importance of the cumulative effects of drivers. Moreover, driving mechanisms associated with temperature and transpiration increased significantly over time, implying an elevated influence of these factors on droughts. Our classification of drought events reveals that nearly 70% of events were driven by at least two mechanisms, underscoring a complex time‐varying pattern of driving factors during drought events. The proposed holistic framework not only sheds insight into the multifaceted mechanisms driving flash droughts within China but also extends its potential applicability to broader geographical contexts.