Predicting Flash Droughts Using Transformers: Understanding Surface and Root Zone

Abstract

Flash droughts, characterized by their rapid onset and devastating agricultural and ecological impacts, pose a growing threat in a changing climate. Accurate and timely predictions are crucial for implementing mitigation strategies and minimizing their widespread consequences. This research presents a novel transformer-based forecasting system designed to predict soil moisture with a focus on detecting the early warning signs of flash droughts in North America. This study integrates the concepts of the two main soil moisture zones, surface and root zone, to provide a comprehensive understanding of drought dynamics. The research leverages the NLDAS (North American Land Data Assimilation System) simulation dataset, offering high-resolution spatiotemporal information crucial for accurate modeling. The transformer-based architecture is employed to capture complex temporal dependencies and non-linear relationships inherent in soil moisture variations, enabling accurate predictions of both surface and root zone moisture content. This approach supports the development of a robust forecasting model capable of capturing sudden and intense decreases in soil moisture characteristic of flash droughts. The system considers the relationship between surface and root zone soil moisture, acknowledging their distinct roles in impacting vegetation health, water availability, and overall ecosystem resilience. Through rigorous evaluations and comparisons with existing forecasting methods, the system’s performance in capturing spatiotemporal variability and providing lead time for proactive mitigation strategies is assessed, highlighting the transformative potential of deep learning for flash drought prediction.

Key Contributions

• Presented a transformer-based prototype for dual-zone (surface and root zone) soil moisture forecasting targeting flash drought early warning.
• Introduced NLDAS reanalysis data as a high-resolution input source for transformer-based drought modeling over North America.
• Demonstrated the model’s ability to capture sudden, intense soil-moisture depletion characteristic of flash drought onset.
• Disseminated preliminary spatiotemporal skill results to the international geosciences community at the EGU General Assembly 2024.

Recommended citation: Chang-Silva, R., & Park, S. (2024). "Predicting Flash Droughts Using Transformers: Understanding Surface and Root Zone." EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-3793. https://doi.org/10.5194/egusphere-egu24-3793