Conclusions

Building on the work of Koldunov et al. (2017), we provide here an Arctic ocean-sea ice reanalysis for the period 2007-2016. By applying a smoothing-algorithm to the adjoint sensitivities, thereby eliminating local random spikes in the respective fields, we were able to increase both the assimilation window to 4-year and the number of iterations performed over each window. Through the filter process, the adjoint model may underestimate the real sensitivities of the cost function to control parameters related to ocean-sea ice processes (not shown). Nevertheless, remaining adjoint sensitivities appear still effective during the optimization process in reducing the model-data misfits. Through the increased number of iterations, the optimization achieves a significantly larger cost function reduction, i.e., improvement, than reported previously by Koldunov et al. (2017). In particular, the data assimilation approach improves the spatial sea ice distribution, reducing the total SIC in the winter season. Together with SIC, SST is also significantly improved. However, despite the significant improvement in SIT estimates, we note that residual SIT, and thus sea ice volume errors, remain substantial.
Comparing INTAROS-opt with the TOPAZ4 and PIOMAS reanalyses, we see that all three products reproduce the SIC variations well, regarding their spatial pattern. However, SIT differences between different renanlyses remain large with the total SIC in INTAROS-opt matching the satellite observations best as compared to the PIOMAS and TOPAZ4 reanalyses. Overall, INTAROS-opt and the TOPAZ4 reanalysis have smaller RMSEs (0.40 m and 0.41 m) than the PIOMAS reanalysis (0.46 m).
Besides using additional hydrographic and sea ice thickness data, we also used the WOA18 atlas to constraint the model climatology. However, although we increased the weighting of the cost contribution of in situ profiles, the resulting model-data misfits remain large. We speculate that this is because of the sparse temporal and spatial coverage of the hydrographic observations. Because of this, the ocean climatology remains a crucial source of hydrographic data for reducing the model bias. Arctic hydrological climatology datasets need to be further improved to reduce their differences, especially over the Arctic marginal shelves. By assimilating the WOA18 atlas, we have increased the mean freshwater content in the Canadian Basin. The freshwater is mainly added through adjusting the initial salinity, which is then redistributed to the Canadian Basin by the mean circulation. Changes of circulation within the Arctic Ocean, similar to Morison et al. (2012), are consistent with a contribution to the freshening. Together with the circulation changes, mean transports through the Fram Strait and around Greenland are changed. Generally, discrepancies in the freshwater content, transport across the Fram and Davis straits remain large between INTAROS-opt and TOPAZ4, supporting the need for improving coverage of hydrographic observations in the Arctic Ocean.
Compared with the other filter-based ocean-sea ice reanalyses, our product is dynamically consistent. The data could be used for understanding the causes and consequences of the Arctic sea ice changes. The results above encourage us to use, in future applications, a single assimilation window and extend the reanalysis from the year 1991 upto date.