Using long short term memory networks to predict daily-averaged sea level anomaly and surface currents

Abstract Short-term prediction of sea level anomaly (SLA) and ocean surface currents over a domain are usually carried out using ocean general circulation models. In this study, we present a univariate machine learning framework using long short term memory networks that forecasts daily-averaged SLA with a 3 d lead time over the north Indian Ocean using historical satellite altimetry data at a spatial resolution of ∼13 km. Assuming that SLA reanalysis from state-of-the-art systems represent the pinnacle of performance skill of forecasts from dynamical models, we show that the SLA forecasts from our model exhibits superior skills through comprehensive analyses. The errors are typically less than 0.04 m over most of the domain and the correlations are close to unity. The skills of the daily-averaged forecasted currents with a 3 d lead, estimated using geostrophic and Ekman theory, are comparable with the best available reanalysis when compared to in-situ observations both in the open ocean or shelf. Treating these forecasted surface currents as synthetic observations, we show that assimilating them can significantly improve the instantaneous subsurface currents during forecasts. The subsurface correlations turn significant with 99% confidence level across depth which were otherwise mostly insignificant and the errors reduce by 0.1 m·s −1 . We demonstrate that the short-term forecast of daily-averaged SLA and surface currents can be approached as a collection of localized low-dimensional independent univariate systems thereby reducing computational costs by large margins. This machine learning framework heralds a paradigm shift in the realm of ocean forecasting.