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

Abstract

Abstract The effect of machine learning and other enhancements on statistical–dynamical forecasts of soil moisture (0–10 and 0–100 cm) and a reference evapotranspiration fraction [evaporative stress index (ESI)] on subseasonal time scales (15–28 days) are explored. The predictors include the current and past land surface conditions and dynamical model hindcasts from the Subseasonal to Seasonal Prediction project (S2S). When the methods 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 ESI. Improvements for both soil moisture and ESI 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. The physical reasons for the nonlinear machine learning improvements are also explored. Significance Statement Rapidly intensifying droughts pose extra challenges for predictability. Here, dynamical forecast model output is combined with nonlinear machine learning methods to improve forecasts of rapid changes in soil moisture and the evaporative stress index (ESI).