A deep machine learning approach to predict Sea Surface Temperature and Marine Heatwaves occurrences over the Mediterranean Sea

Marine Heat Waves (MHWs) have significant social and ecological consequences. There is a growing need to predict these extreme events to prevent and possibly mitigate their negative impacts and to better inform decision-makers on MHWs-related risks. In this context, we applied Long Short-Term Memory (LSTM) networks to predict Sea Surface Temperature (SST) time-series and, in turn, MHWs, over the Mediterranean Sea. LSTM networks are types of recurrent neural networks capable of learning order dependence in sequence prediction problems, and they have been widely applied in temperature forecasting problems. The model is a multi-step LSTM model, which means that it predicts seven time-steps of SST into the future. In order to build an efficient prediction model, as input times-series, LSTM exploits SST together with other relevant atmospheric variables (e.g. Geopotential Height, Incoming solar radiation), selected as potential MHWs drivers. The datasets used to train and to test the model are the European Space Agency (ESA) Climate Change Initiative (CCI) Sea Surface Temperature (SST) v2.1 for SST and ERA5 reanalysis for the atmospheric components from 1981 to 2016.  Preliminary results over target areas suggest that, besides the SST itself, the incoming solar radiation has the highest predictive skill on SST variability with respect to the other atmospheric variables. The model is accurate in predicting the occurrence of MHWs in the test dataset at the earliest days of forecast. In addition, the root mean square error analysis between predicted and actual SST time-series shows that LSTM models errors compare favorably with respect to the Copernicus Mediterranean Forecasting System (MedFS, i.e. dynamical model) errors, at least at the earliest days of forecast.