Improving Subseasonal Soil Moisture and Evaporative Stress Index Forecasts through Machine Learning: The Role of Initial Land State versus Dynamical Model Output

Accurate subseasonal-to-seasonal (S2S) forecast of flash drought has the potential to enhance drought preparation and mitigation. For this reason, numerous studies have applied dynamically based forecast systems and statistical prediction models to flash drought prediction. In this study, the two approaches are combined in the form of statistical models that draw predictors from both observed land surface conditions and S2S Prediction Project dynamically-based forecasts. Both standard regression models and nonlinear machine learning methods are considered. When the models are enhanced with machine learning and other improvements, the increases in skill are almost exclusively coming from predictors drawn from observations of current and past land surface states. This suggests that operational S2S flash drought forecasts should focus on optimizing use of information on current conditions rather than on integrating dynamically based forecasts, given the current state of knowledge. Nonlinear machine learning methods lead to improved skill over linear methods for soil moisture but not for evapotranspiration fraction. Improvements for both soil moisture and evapotranspiration fraction are realized by increasing the sample size by including surrounding grid points in training and increasing the number of predictors. In addition, all the improvements in the soil moisture forecasts predominantly impact soil moistening rather than soil drying—i.e., prediction of conditions moving away from drought rather than into drought—especially when the initial soil state is drier than normal.