CNN-Based ENSO Forecasts With a Focus on SSTA Zonal Pattern and Physical Interpretation

Deep learning (DL) has achieved notable success in El Nino-Southern Oscillation (ENSO) forecasts. Most DL-based models focused on forecasting ENSO indices while the zonal distribution of sea surface temperature anomalies (SSTA) over the equatorial Pacific was overlooked. To provide accurate predictions for the SSTA zonal pattern, this study developed a model through leveraging the merits of the cosine distance in constructing the convolutional neural network. This model can skillfully predict the SSTA zonal pattern over the equatorial Pacific 1 year in advance, remarkably outperforming current dynamical models. Moreover, the physical interpretation of the model prediction reveals that the sources for ENSO predictability at different lead times are distinct. For the 10-month-lead predictions, the precursors in the north Pacific, south Pacific and tropical Atlantic play critical roles in determining the model behaviors; while for the 16-month-lead predictions, the initial signals in the tropical Pacific associated with the discharge-recharge cycle are essential. The El Nino-Southern Oscillation (ENSO) is the most prominent climate phenomenon in the Earth system. It significantly affects the worldwide weather and climate via teleconnections. Numerous studies have reported that the ENSO teleconnection and its impacts largely depend on the zonal distribution of SSTA over the equatorial Pacific. Thus, the ENSO forecast with the specific SSTA zonal pattern is important for anticipating the severity of ENSO-related disasters and mitigating the potential socio-economic impacts. However, current dynamical models have difficulties in accurately predicting the SSTA zonal pattern, while most of deep learning models only provide predictions of ENSO indices. Hence, we developed a deep learning model based on the convolutional neural network which can effectively predict the SSTA zonal pattern 1 year in advance. Moreover, we investigate the interpretability of this model by analyzing activation maps. The results suggest that crucial factors captured by this model at different lead times are physically reasonable, which verify the credibility of this model. We develop a deep learning model that can skillfully predict the explicit sea surface temperature anomalies (SSTA) zonal pattern over the equatorial Pacific 1 yr aheadPhysical interpretation shows that the source of 10-month-lead prediction stems from the Pacific Meridional Mode, South Pacific quadrupole, and tropical Atlantic SSTAThe main source of 16-month-lead forecast comes from discharge/recharge cycles, implying distinct prediction sources at different lead times