Introduction

Flash droughts are characterised by the rapid onset of extreme soil moisture deficits, typically occurring within a few weeks. These events frequently result in widespread crop failures, water shortages, and significant terrestrial carbon loss. While substantial progress has been made in understanding long-term drought trends and recovery patterns, the underlying mechanisms driving the development of flash droughts under anthropogenic forcing have remained unclear. This research utilises an explainable artificial intelligence (XAI)–based clustering framework to investigate the major driving mechanisms of global flash droughts and their temporal evolution from 1951 to 2022. By analysing three global reanalysis products and outputs from 13 Earth system models (ESMs) in the Coupled Model Intercomparison Project Phase 6 (CMIP6), the study tests the hypothesis that vegetation and temperature responses to anthropogenic forcing have shifted the primary drivers of these events.

Results

The study identifies two distinct types of flash droughts based on their climatic and biophysical drivers. The first type, a compound-dominated cluster (Ccomp), is characterised by the synergistic effect of elevated temperature, enhanced transpiration, and reduced precipitation. This type accounts for 32% of total flash drought events and is marked by more rapid development, longer duration, and greater severity. The second type, a precipitation-dominated cluster (Cprec), is primarily driven by rainfall deficits and accounts for the remaining 68% of events. Analysis reveals a substantial shift in these mechanisms over time. Prior to 2008, precipitation-dominated events occurred four times more frequently than compound-driven events, accounting for 73% of total occurrences. However, after 2008, compound-driven events increased substantially, becoming more frequent than precipitation-dominated events after 2017. This transition has contributed to an overall surge in global flash droughts.

Anthropogenic warming and vegetation greening are identified as the primary contributors to this surge. Greenhouse gas (GHG) forcing shows a significant positive trend in compound-driven events, whereas natural forcing (NAT) and aerosol forcing (AER) exhibit minor or insignificant changes. Since approximately 2008, temperature changes have contributed about 32% to soil moisture decline in compound-driven events, while vegetation greening (represented by leaf area index or LAI) has contributed approximately 18%. Together, these anthropogenic factors drive roughly 60% of the total contributions to the compound effect during the recent decade. Furthermore, decomposition of global trends shows that temperature and LAI changes contribute to approximately 88% of the increase in frequency of compound-driven flash droughts.

Hotspots for these anthropogenically driven events are detected in Eurasia, the Amazon, and Africa. The signal of increasing compound-driven events emerged beyond natural climate variability around 2017 in the Northern Hemisphere and the Amazon region. This surge has led to increased risks for human populations and ecosystems. Population exposure to compound flash droughts rose from approximately 100 million people in 2001 to around 650 million people during the 2015–2020 period. Vulnerable groups, particularly children under five years old and adults over 50, show the highest levels of exposure increase. Additionally, these events have caused a significant decline in terrestrial gross primary productivity (GPP), with a mean annual reduction of approximately 0.15 ± 0.1 petagrams of carbon over the most recent decade.

Discussion

The findings provide direct evidence that flash droughts have shifted from being predominantly driven by natural variability to being driven by anthropogenic forcing. In a CO2-enriched environment, the drivers of flash droughts combine to create more frequent and intense events that emerge outside historical natural variability. High-latitude regions are particularly affected due to Arctic amplification, which increases soil evaporation and enhances transpiration through promoted vegetation greening. While greening can contribute to climate mitigation via carbon sequestration, it can also lead to more frequent droughts by depleting deep soil moisture, thereby signalling a potential decline in ecosystem resilience. The results underscore the urgent need for a comprehensive, impact-based forecasting approach that integrates climate, ecological, and human dimensions to improve drought preparedness and risk management.

Materials and methods

The research utilises a long short-term memory (LSTM) neural network to model the relationship between potential drivers (precipitation, temperature, and transpiration) and soil moisture changes. The expected gradient (EG) method was then applied to estimate the individual contributions of these drivers during flash drought development stages. Cluster analysis (K-means) was used to identify distinct drought types. The study analysed data from 13 CMIP6 models and multiple reanalysis products, including GLDAS/VIC, GLDAS/Noah, and ERA5. All datasets were regridded to a common 1.5° × 1.5° grid for analysis. Time of emergence (ToE) was calculated to determine when anthropogenic signals consistently exceeded natural variability noise.