Machine learning methods to predict Sea Surface Temperature and Marine Heatwave occurrence: a case study of the Mediterranean Sea

This repository contains the code to produce the results of  " Machine learning methods to predict Sea Surface Temperature and Marine Heatwave occurrence: a case study of the Mediterranean Sea" a discussion paper by Giulia Bonino, Giuliano Galimberti , Simona Masina, Ronan McAdam , and Emanuela Clementi submitted to Ocean Science journal.

Abstract of the manuscript:

Marine heatwaves (MHWs) have significant social and ecological impacts, necessitating the prediction of these extreme events to prevent and mitigate their negative consequences and provide valuable information to decision-makers about MHW-related risks. In this study, machine learning (ML) techniques are applied to predict Sea Surface Temperature (SST) time series and Marine Heatwaves in 16 regions of the Mediterranean Sea. ML algorithms, including Random Forest (RForest), Long short-term memory (LSTM), and Convolutional Neural Network (CNN), are used to create competitive predictive tools for SST. The ML models are designed to forecast SST and MHWs up to 7 days ahead. For each area, we performed 15 different experiments for ML techniques, modifying the training and testing periods. Alongside SST, other relevant atmospheric variables are utilized as potential predictors of MHWs. Datasets from the European Space Agency Climate Change Initiative (ESA CCI SST) v2.1 and the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis from 1981 to 2021 are used to train and test the ML techniques. For each area the results show that all the ML methods performed with minimum Root Mean Square Errors (RMSE) of about 0.1°C at a 1-day lead time and maximum values of about 0.8°C at a 7-day lead time. In all regions, both the RForest and LSTM models consistently outperformed the CNN model across all lead times. Importantly, the ML techniques show results similar to the dynamical Copernicus Mediterranean Forecasting System (MedFS) for both SST and MHW forecasts, especially in the early forecast days. For MHW forecasting, ML methods compare favorably with MedFS up to 3-day lead time in 14 regions, while MedFS shows superior skill at 5-day lead time in 9 out of 16 regions. All methods predict the occurrence of MHWs with a confidence level greater than 50% in each region. Additionally, the study highlights the importance of incoming solar radiation as a significant predictor of SST variability along with SST itself.

This repository contains:

• MLtechniques.ipynb is the code to build ML techniques (e.g. LSTM, CNN, RForest) to predict seven days of SST over the 16 selected regions.