Sea level prediction using ARIMA, SVR and LSTM neural network: assessing the impact of ensemble Ocean-Atmospheric processes on models' accuracy

This study aims to integrate a broad spectrum of ocean-atmospheric variables to predict sea level variation along West Peninsular Malaysia coastline using machine learning and deep learning techniques. 4 scenarios of different combinations of variables such as sea surface temperature, sea surface salinity, sea surface density, surface atmospheric pressure, wind speed, total cloud cover, precipitation and sea level data were used to train ARIMA, SVR and LSTM neural network models. Results show that atmospheric processes have more influence on prediction accuracy than ocean processes. Combining ocean and atmospheric variables improves the model prediction at all stations by 1- 9% for both SVR and LSTM. The means of R accuracy of optimal performing LSTM, SVR and ARIMA models at all stations are 0.853, 0.748 and 0.710, respectively. Comparison of model performance shows that the LSTM model trained with ocean and atmospheric variables is optimal for predicting sea level variation at all stations except Pulua Langkawi where ARIMA model trained without ocean-atmospheric variables performed best due to the dominating tide influence. This suggests that performance and suitability of prediction models vary across regions and selecting an optimal prediction model depends on the dominant physical processes governing sea level variability in the area of investigation.