Towards Sub-Seasonal Flash Drought Prediction Using a Vision Transformer Approach

Unlike traditional droughts that evolve gradually, flash droughts (FD) are characterized by rapid intensification, leading to sustained dry conditions with disproportionately high impacts on ecosystems. Despite substantial progress in short- to medium-range weather forecasting, predicting these events remains a significant hurdle for both early warning systems and physically-based subseasonal-to-seasonal (S2S) prediction frameworks.

To address this challenge, we present a deep learning framework leveraging a Vision Transformer with explicit temporal attention for the prediction of soil moisture anomalies (SMA) over Europe. The model employs a dual-stream attention mechanism that disentangles temporal dynamics from spatial dependencies: temporal self-attention with rotary positional embeddings captures lead-time-dependent evolution at each location, while spatial attention encodes cross-regional relationships. This architecture enables to learn multi-scale representations, ranging from synoptic variability to persistent anomaly patterns. Furthermore, the model supports probabilistic forecasting, estimating the full conditional distribution of soil moisture anomalies to provide principled uncertainty quantification, a critical requirement for operational early warning systems. 

Additionally, the framework employs a multitask learning approach that exploits the relationship between continuous soil moisture anomalies and discrete flash drought characteristics derived from the Flash Drought Intensity Index (FDII). This index integrates both the rate of soil moisture decline and drought severity into a unified indicator.  A shared encoder learns representations that capture the coupled dynamics of soil moisture evolution and flash drought emergence, while task-specific prediction heads accommodate the distinct statistical properties of each target variable.

The results indicate that our approach achieves predictive skill competitive with more complex spatio-temporal models while maintaining computational efficiency suitable for operational deployment. Evaluation against standard baselines, including climatology and persistence, as well as state-of-the-art deep learning models, demonstrates the framework’s ability to resolve the rapid intensification dynamics typical of flash drought onset. This work lays the foundation for interpretable, scalable, and probabilistic prediction of rapid-onset drought events at S2S timescales.