A machine learning framework for sea surface salinity prediction at short time scales

Sea surface salinity (SSS) is an integral ocean state variable that influences the global weather patterns. The conventional approach to forecast SSS maps over a basin at a lead time of a few days has been through dynamical models. However, the dearth of daily freshwater fluxes induces errors in short-term SSS predictions from these dynamical models. In this study, using only past collocated SSS data, we develop an univariate machine learning framework based on long short term memory for short-term prediction of SSS and test its veracity over Bay of Bengal—one of the most stratified ocean basins in the world. We show that the 5-day daily SSS predictions are superior to daily SSS reanalysis when compared against blended daily SSS observations. The root-mean-squared errors are typically less than 0.5 psu over the entire basin with excellent phase alignment. The SSS forecasts at 5-day lead time are on par with daily SSS reanalysis when compared against Argo SSS observations. Using hourly buoy SSS data, we show that this framework is equally adept at forecasting high-frequency SSS variability with errors in the range of 0.02 psu at a lead time of 5 h. Using the test case of supercyclone Amphan, we also show that this framework is resilient during extreme events too. This study shows that it is indeed feasible to issue a 5-day forecast of daily SSS map over an entire basin with skills on par with reanalysis.