Prediction of sea surface temperature and high water temperatures near the Korean Peninsula based on a long short-term memory algorithm

Abstract

Climate change caused by global warming has resulted in high water temperatures under the influence of continuous increases in sea surface temperature (SST). The severity of this phenomenon is rapidly increasing around the Korean Peninsula. SSTs play an important role in Earth systems. SST increases affect marine habitats and damage the marine environment, resulting in changes in fish species and damage to aquaculture and fisheries industries in coastal waters. As SST rises, seawater expands, accelerating sea level rise and causing coastal flooding; this can lead to loss of human life. Therefore, in this study, we designed a model to estimate SSTs and predict high water temperatures near the Korea Peninsula. To estimate SSTs, we used a long short-term memory (LSTM) algorithm, which is an artificial neural network model for time series data prediction. We used the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 SST data product for the region around the Korean Peninsula. The model was trained to estimate SSTs for 1 and 7 days. SSTs at ≥ 28°C were defined as high water temperatures. We evaluated the accuracy of the model for predicting SSTs using the coefficient of determination (R2), root mean square error (RMSE), and mean absolute percentage error (MAPE), and evaluated high water temperatures using the F1 score and receiver operating characteristic (ROC) analysis. The 1-day prediction model was used to estimate SSTs and high water temperatures on August 8, 2019 because high water temperatures are common in summer. The results showed that the predicted SSTs were consistent with observed data (R2 = 0.993, RMSE = 0.122°C, MAPE = 0.347%). The 7-day prediction model had lower accuracy than the 1-day model (R2 = 0.917, RMSE = 0.422°C, MAPE = 1.244%), with a difference between models of -0.28°C. Thus, SST underestimation increased with the length of the prediction period. The F1 scores of the 1- and 7-day prediction models were 0.965 and 0.882, respectively, and ROC data points were concentrated in the upper left corner for both models, indicating good high water temperature prediction accuracy.