Superensemble seasonal forecasting of soil moisture by NMME

Abstract

ABSTRACTSoil moisture affects hydro‐climate processes by altering water and energy exchanges between land surface and atmosphere. Understanding of the predictability of soil moisture is not only important for a skillful forecasting of seasonal hydro‐climate, but also for agricultural drought early warning. This paper assesses seasonal forecast skill and potential predictability of soil moisture directly produced by climate models, and investigates an optimal combination of different models over China. A set of 29‐year hindcasts for soil moisture from six North American Multi‐model Ensemble (NMME) models are verified against ERA Interim reanalysis. Results show that soil moisture predictability, which is defined by anomaly correlation under a perfect model assumption, is higher than forecast skill in all models, suggesting that soil moisture prediction may have a room for improvement. Except the CESM model, NMME climate forecast models with higher predictability also have higher forecast skill, where predictability is positively correlated with forecast skill with p < 0.01 across different lead times. Soil moisture forecast skill from NMME simple arithmetic mean is higher than any individual models, and the skill is further improved through an optimization of model weights with a cross validation procedure. As compared with simple ensemble mean, the optimized superensemble mean reduces root mean squared error by 19 and 7% for seasonal mean soil moisture forecast during winter and summer seasons, respectively, and increases correlation by about 10%. This study suggests that soil moisture forecasts directly produced by climate models, when combined appropriately, can provide useful information for climate service.