On the Harvest of Predictability From Land States in a Global Forecast Model

Various land surface treatments in a suite of subseasonal-to-seasonal forecasts are applied to diagnose the degree to which potential predictability from the land surface is harvested, where breakdowns occur in the process chains that link land surface states to atmospheric phenomena, and the role played by memory in the climate system. Version 2 of the Coupled Forecast System (CFSv2) is used for boreal summer simulations spanning 28years. Four types of retrospective forecasts are produced: those where land surface initial states are from the same date and year as the initial atmosphere and ocean states; ensembles where initial land states come from different years than the atmosphere and ocean; simulations where soil moisture is specified from an observationally constrained analysis; and simulations where an alternative triggering mechanism for convection is employed. The specified soil moisture allows estimation of an upper bound for land-driven predictability and prediction skill in boreal summer. Realistic land initialization represents the best possible case with this model in forecast mode, while the simulations with initial land states from different years isolate the impact of atmosphere and ocean initialization on forecasts. Harvested predictability is calculated, and its relationship to memory of initial anomalies is estimated. The pathway of land surface information through the energy and water cycles to the atmosphere, and ultimately its effects on precipitation, is traced, showing a robust propagation of useful signal through land surface fluxes, near-surface meteorological states, and boundary layer properties, but largely disappearing at precipitation, implying problems with the convective parameterization. Plain Language Summary The performance of the National Weather Service's operational climate forecast model is examined to see how the land surface, namely, moisture in the soil, affects the skill of forecasts. We estimate the potential skill derived from the best possible initialization and prediction of land surface states and how much of that potential skill can be realized by the current version of the forecast model. Additionally, we trace the signal of information in the model from the land surface into the atmosphere and find that while good soil moisture information greatly extends the duration of useful temperature and humidity forecasts, much information appears to be lost at the point in the model where clouds and precipitation are simulated. This result suggests that the model could be improved to make better use of land surface data to produce more skillful precipitation forecasts.