The ECMWF analysis system currently assimilates Level-4 sea ice concentration (SIC) from OSTIA (the Operational SST and Sea Ice Analysis produced by the UK Met Office). Here, we evaluate the impact of assimilating Level-3 SIC observations in the ECMWF ocean-sea ice analysis system.
Furthermore, we make use of the availability of Arctic-wide sea ice thickness (SIT) observations in the recent years to constrain the modelled sea ice thickness. Coupled forecasts of the ocean-seaice-wave-land-atmosphere are then initialized using the improved sea-ice initial conditions from the above assimilation experiments, and the predictive skill of Arctic sea ice up to lead times of 7 months is investigated in a low-resolution analogue of the currently operational ECMWF seasonal forecasting system SEAS5. Results show that the system successfully assimilates Level-3 SIC observations from the OSISAF (EUMETSAT Ocean and Sea Ice Satellite Applications Facility) product OSI-401-b.
Differences in the analysis are small and within the observational uncertainties, but the assimilation of Level-3 SIC will result in increased operational reliability. The impact on coupled forecasts is generally positive for SIC at lead month 1 and neutral for longer lead times. Statistically significant improvements are found over the ice edge and coastal seas in the Arctic mostly in the first 2 weeks for forecasts initialized in most calendar months, except for January starts, when the impact is neutral.
The positive impact persists up to week 4 for March, May, August, November and December start months. For SIT and sea ice volume, the forecast impact of Level-3 SIC assimilation is neutral in all lead months.
Using SIT information from CS2-SMOS (CryoSat2-Soil Moisture and Ocean Salinity) as an additional constraint results in substantial changes of sea ice volume and thickness in the ocean-sea ice analysis. Forecasts started from these sea-ice initial conditions show a reduction of the positive sea ice bias and an overall reduction of summer-time forecast errors compared to SEAS5. A slight degradation in skill is found in the autumn sea ice forecasts initialized in July and August. While there is improvement in the skill of autumn 2m-temperature forecast initialized in spring, a degradation in skill is found for the October forecasts initialized in August. We conclude that the strong thinning by CS2-SMOS initialization mitigates or enhances seasonally dependent forecast model errors in sea ice and near surface temperatures. Hence, changes in root-mean-square errors are predominantly due to changes in biases. Using a novel metric, the Integrated Ice Edge Error (IIEE), we find significant improvement of up to 28% in the September sea ice extent forecast started from April. Our results demonstrate the usefulness of new sea ice observational products in both data assimilation and forecast verification, and strongly suggest that better initialization of SIT is crucial for improving seasonal sea-ice forecasts.