Skillful subseasonal soil moisture drought forecasts with deep learning-dynamic models.

Area of Science:

• Climate science
• Artificial intelligence in meteorology
• Hydrological forecasting

Background:

• Deep neural networks excel at short-term weather forecasting but struggle with long-lead predictions of soil moisture and droughts.
• Predicting extreme events like droughts beyond two weeks remains a significant challenge for traditional dynamic models.

Purpose of the Study:

• To develop a hybrid model combining deep learning and dynamic forecasts for enhanced subseasonal predictions.
• To improve the accuracy and lead time of root zone soil moisture and flash drought forecasts.

Main Methods:

• Utilized a recursive deep learning model (RISE-UNet) integrated with subseasonal forecasts from dynamic models.
• Developed a hybrid approach combining RISE-UNet with dynamic model outputs and antecedent reanalysis data.

Main Results:

• Achieved skillful forecasts of root zone soil moisture up to four weeks in advance.
• The hybrid model significantly outperformed existing dynamic models (ECMWF, GEFS) and reanalysis-driven deep learning models.
• Demonstrated high skill in predicting flash droughts, outperforming ECMWF and GEFS for major events in the US, China, and Australia.

Conclusions:

• Combining deep learning with dynamic model forecasts substantially improves subseasonal prediction skill beyond two weeks.
• The hybrid approach shows particular promise for predicting root zone soil moisture and flash drought events.
• The inclusion of initial two-week dynamic forecasts and antecedent soil moisture is key to extending forecast skill to weeks three and four.