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

Deep neural networks that learn from climate reanalysis data have produced skillful weather forecasts within ten days. However, it is still a great challenge for dynamic models to predict soil moisture, droughts, and other extreme events with lead times beyond two weeks. Here, we combine a recursive deep learning model (namely RISE-UNet) and subseasonal forecasts from dynamic models and achieve skillful forecasts of root zone soil moisture up to four weeks in advance. Our hybrid model, combining RISE-UNet and dynamic model forecasts, outperforms reanalysis-driven RISE-UNet models, while both methods show significantly higher performance than the latest European Centre for Medium-Range Weather Forecasts (ECMWF) and Global Ensemble Forecast System (GEFS) dynamic models, the postprocessed ECMWF or GEFS subseasonal forecasts by RISE-UNet, or ensemble model output statistics. The hybrid model shows skill in predicting flash droughts, which is higher than ECMWF and GEFS models in most cases, as demonstrated for major events in the United States, China, and Australia. The forecast skill of the hybrid modeling approach from weeks three to four is mainly due to the inclusion of the first two-week dynamic model forecasts and antecedent root zone soil moisture reanalysis. Our results indicate that combining deep learning with dynamic model forecasts can substantially improve the skill of subseasonal predictions beyond two weeks, particularly for root zone soil moisture and flash drought events. This study presents a hybrid approach that combines deep learning with dynamic models for soil moisture forecast. This hybrid approach significantly improves soil moisture and flash drought forecasts beyond two weeks, offering a promising path to earlier warning of climate extremes.