Flash droughts: an AI model outperforms the CEPMMT’s sub-seasonal forecasts for European soil moisture, but not for their onset

Researchers have shown that the way the problem is framed matters just as much as the model: predicting root moisture in physical units, using residual learning, makes all the difference.

Predicting soil moisture several weeks in advance is crucial to anticipating flash droughts, those episodes when soils dry out in a matter of weeks. A team led by Noelia Otero trained a Vision Transformer-type model, with a separate focus on time and space, to predict root zone moisture over Europe, at the sub-seasonal scale.

The target matters as much as the model

First lesson: without residual learning, the model does no better than simple persistence (assuming that tomorrow looks like today). Second lesson: this advantage only appears if humidity is predicted in physical units rather than in standardized anomalies. A probabilistic variant, obtained by adding quantile heads, provides well-calibrated distributions. Tested over 2021-2022 against other deep learning models and ECMWF operational forecasts, the system achieves the best scores at all time frames and reliably identifies abnormally dry soils (below the 20th percentile).

The limit is that the start of a flash drought, defined by rapid intensification over several pentades, remains out of reach, for this model as for all current systems.