The Effects of Salinity and Stratification on Rapid Sea Ice Advance in the Arctic Ocean

The annual sea ice minimum extent in the Arctic Ocean has decreased almost two-fold since the advent of satellite observations in the 1970s, leaving more open water before the fall freeze-up.  Here, we leverage a combined dataset from the 2022 NASA Salinity and Stratification at the Sea Ice Edge (SASSIE) field program to elucidate the central hypothesis that drove SASSIE: Does surface salinity stratification due to sea ice melt, precipitation, and riverine inputs lead to changes in the rates or extent of autumnal sea ice advance? The SASSIE study region in the Beaufort Sea is stratified both by melting sea ice in the summer and riverine discharge. We leverage measurements of oxygen isotopes as well as colored dissolved organic matter (CDOM) to trace the origins of fresher water at the surface.

In addition to an in-depth analysis of in situ data, we use the General Ocean Turbulence Model (GOTM) for individual profiles as well as the Regional Ocean Modeling System (ROMS) initialized and forced with observations from the SASSIE field campaign. These observations include temperature and salinity from the salinity snake instrument at 1-2cm depth, shipborne thermosalinograph (4m) and underway conductivity-temperature-depth (uCTD) measurements (5-100m), acoustic Doppler current profiler (ADCP) data, as well as meteorological and net heat flux observations. In realistically forced runs, we re-create the observations during the month-long cruise. We then modify the stratification to both increase and decrease salinity stratification to assess the importance of salinity stratification on the autumnal sea ice advance. We compare these model outputs to satellite-derived freeze-up data as well as in situ observations from autonomous platforms in the area. Preliminary results show a strong control of salinity on rapid sea ice advance, in which areas that are highly stratified freeze significantly faster than areas of deeper or weaker stratification.

Based on this hypothesis, we present a novel way of modelling the autumnal Arctic Sea Ice advance using a Convolutional Long-Short-Term Memory (LSTM) Neural Network model. In this machine learning approach, we demonstrate that the inclusion of the experimental merged salinity OISSS v3 dataset derived from the Soil Moisture and Ocean Salinity (SMOS) and Soil Moisture Active Passive (SMAP) satellites significantly improves forecast accuracy of sea ice concentration in our study area, which encompasses the East Siberian, Chukchi, and Beaufort Seas. The model is based on 8 years of training data and tested using 3 years of evaluation data. Using this 60-day forecast, we show that the spatial forecasting pattern of sea ice concentration is significantly improved. This is further illustrated in an ablation study, in which we find sea surface salinity to be the 4^th most important predictive term after sea surface temperature, net heat flux, and sea ice concentration.

Through these studies, we show the connection between the terrestrial water cycle, oceanic freshwater fluxes, and sea ice formation in the Arctic, and present a novel technique of sea ice prediction that will become increasingly useful as the Arctic becomes more ice free.