Prospects for improved seasonal Arctic sea ice predictions from multivariate data assimilation

Predicting the summer Arctic sea ice conditions a few months in advance has become a challenging priority. Seasonal prediction is partly an initial condition problem; therefore, a good knowledge of the initial sea ice state is necessary to hopefully produce reliable forecasts. Most of the intrinsic memory of sea ice lies in its thickness, but consistent and homogeneous observational networks of sea ice thickness are still limited in space and time. To overcome this problem, we constrain the ocean–sea ice model NEMO-LIM3 with gridded sea ice concentration retrievals from satellite observations using the ensemble Kalman filter. No sea ice thickness products are assimilated. However, thanks to the multivariate formalism of the data assimilation method used, sea ice thickness is globally updated in a consistent way whenever observations of concentration are available. We compare in this paper the skill of 27 pairs of initialized and uninitialized seasonal Arctic sea ice hindcasts spanning 1983–2009, driven by the same atmospheric forcing as to isolate the pure role of initial conditions on the prediction skill. The results exhibit the interest of multivariate sea ice initialization for the seasonal predictions of the September ice concentration and are particularly encouraging for hindcasts in the 2000s. In line with previous studies showing the interest of data assimilation for sea ice thickness reconstruction, our results thus show that sea ice data assimilation is also a promising tool for short-term prediction, and that current seasonal sea ice forecast systems could gain predictive skill from a more realistic sea ice initialization