Development of marine heatwave prediction CNN model based on dynamic mechanisms

Marine heatwaves (MHWs), extremely high sea surface temperatures (SST) occurring several days to weeks, have devastating effects on the marine ecosystem and fisheries. Since MHWs are becoming more frequent, prediction of MHW is necessary to reduce socioeconomic damage. According to Lee et al. (2020), East Asian MHWs are driven by various atmosphere-ocean processes, such as large-scale atmospheric circulation systems like Pacific-Japan and Circumglobal Teleconnection patterns, or contrasting SST patterns due to El Niño-Southern Oscillation. In this study, an East Asian MHW prediction model is constructed based on a deep learning pattern recognition method (convolutional neural network, CNN), considering the MHW dynamics analyzed by Lee et al. (2020). To identify the performance of the CNN model considering the MHW dynamics, the "Control" model using only SST as an input variable and the "Dyn-CNN (Dynamic-CNN) model trained by selecting variables and input regions related to MHW dynamics are compared. The results showed that the Dyn-CNN model outperformed the Control model in the 7-day ahead prediction, resulting in a root mean square error (RMSE) of 1.25 °C for and 1.01 °C for Control and Dyn-CNN, respectively. Also, we confirm that Dyn-CNN well predicts the MHW in the coastal region.