On the soil moisture memory and influence on coupled seasonal forecasts over Australia

In this study we assess impacts of land-surface initialization on sub-seasonal and seasonal forecast skills from a coupled model named ACCESS-S1 (a seasonal prediction version 1 of the Australian Community Climate and Earth-System Simulator). A series of sensitivity experiments is conducted to explore to what extent the model skill and mean bias can be affected by different land-surface initialisations. By focusing the analysis on the Australian continent, our study tries to address three questions: (1) how strong is soil moisture memory in the model and how realistic is that compared with some observational evidence; (2) how does the soil moisture memory affect surface fluxes; and (3) how these impacts on surface fluxes are translated into the impacts on forecasting rainfall, temperature and atmospheric circulation. Firstly, we run an offline experiment with the ACCESS-S1 land surface model JULES for the period of 1990-2012 using ERA-interim forcing data, with precipitation further adjusted by monthly observations. This produces 23-year soil moisture time series corresponding to observed meteorological forcing. Lagged correlations between soil moisture at different levels and at different months are compared with some in-situ observations over Murrumbidgee catchment. We show good agreement between modelled and observed soil moisture coupling between its top-layer and sub-surface root-zone in the southeast part of the continent, and notable soil moisture memory over these regions. We then use the JULES offline soil moisture data to initialize ACCESS-S1 in its 3-month hindcasts with start date of 1st May for the 23-year period. In contrast to the default ACCESS-S1 setup which uses climatological soil moisture in its land-surface initialisation, our results show significant improvements to ACCESS-S1 forecast skill of surface maximum temperature (T-max) and evapotranspiration by initialising the model with JULES offline data. It also has moderate improvements of surface minimum temperature (T-min) and precipitation forecasts. The skill gain is particularly evident over the eastern part of the continent where the modelled and observed soil moisture memory is strong. Our study demonstrates that in future forecast system development, we not only need to initialise the model with updated soil moisture anomalous conditions, but also need to make sure these anomalies are combined with correct and consistent soil moisture climatology for both hindcast and real-time forecast.