Non-linear neural networks forecasting of sea level anomaly in the Alboran Sea

Forecasts based on artificial intelligence (AI) concepts exploit past time series of satellite images to infer near future ocean conditions at the surface by feed-forward non-linear neural networks. The size of the AI problem is drastically reduced by splitting the spatio-temporal variability contained in the remote sensing data by using empirical orthogonal function (EOF) decomposition. The problem of forecasting the dynamics of a two-dimensional surface field can thus be reduced by selecting the most relevant empirical modes, and non-linear time series predictors are then applied on the time independent amplitudes only. In the present case study, we use altimetric maps of the Mediterranean Sea and the Alboran Sea, combining TOPEX-POSEIDON and ERS-1/2 data for the period October 1992 to March 2000. The learning procedure is applied to each mode individually. The final forecast is then reconstructed from the EOFs and the forecasted amplitudes, and compared to the real observed field, the persistence and linear forecasts for validation purposes.