Deep-Learning Model for Sea Surface Temperature Prediction near the Korea Peninsular

Abstract

Recently, the sea surface temperature(SST) near the Korea peninsular has been increased rapidly due to the global warming, and these phenomena can cause the high water temperature and extensive damage to Korea fish farms. In order to reduce the damage caused by the high water temperature events, it is necessary to predict the events in advance. In this study, therefore, we propose a method for predicting the high water temperature phenomenon in advance to prevent the high water temperature damage. The SST have time-series data characteristics, and the long short-term memory(LSTM) network specialized for time-series data prediction among artificial neural network models is used. First, using the SST prediction model to predict the SST. And if the predicted SST is over 28oC, which is the Korean government standard for issuing the high water temperature warning, it is judged as high water temperature. In here, European Centre for Medium-Range Weather Forecasts(ECMWF) ERA5 sea surface temperature data near the Korea peninsular were used. To evaluate the SST prediction accuracy of the prediction model, the 1-day to 7-day prediction model was evaluated using the R2, root mean square error (RMSE) and mean absolute percentage error (MAPE) evaluation indicators were used. In the 1-day prediction SST model achieved results of 0.993, 0.12°C and 0.35%, R2, RMSE and MAPE, respectively. The 7-day prediction SST model was 0.917, 0.422°C and 1.24%. We also performed F1 score to evaluate the high water temperature classification accuracy. The F1 score for the 1-day and 7-day SST prediction models were 0.965 and 0.882.