Enhancing Deep Learning Soil Moisture Forecasting Models by Integrating Physics-based Models

Accurate soil moisture (SM) prediction is critical for understanding hydrological processes. Physics-based (PB) models exhibit large uncertainties in SM predictions arising from uncertain parameterizations and insufficient representation of land-surface processes. In addition to PB models, deep learning (DL) models have been widely used in SM predictions recently. However, few pure DL models have notably high success rates due to lacking physical information. Thus, we developed hybrid models to effectively integrate the outputs of PB models into DL models to improve SM predictions. To this end, we first developed a hybrid model based on the attention mechanism to take advantage of PB models at each forecast time scale (attention model). We further built an ensemble model that combined the advantages of different hybrid schemes (ensemble model). We utilized SM forecasts from the Global Forecast System to enhance the convolutional long short-term memory (ConvLSTM) model for 1–16 days of SM predictions. The performances of the proposed hybrid models were investigated and compared with two existing hybrid models. The results showed that the attention model could leverage benefits of PB models and achieved the best predictability of drought events among the different hybrid models. Moreover, the ensemble model performed best among all hybrid models at all forecast time scales and different soil conditions. It is highlighted that the ensemble model outperformed the pure DL model over 79.5% of in situ stations for 16-day predictions. These findings suggest that our proposed hybrid models can adequately exploit the benefits of PB model outputs to aid DL models in making SM predictions.