Deep Learning Predictions of SST at the Atlantis II Seamounts with Links to Global Teleconnections

Abstract

Located in the western North Atlantic, the subsurface temperature and salinity structure at the Atlantis II Seamounts is sensitive to fluctuations in the Gulf Stream, and historical ocean structure profiles at this location show a bimodal distribution, making seasonal climatologies a poor predictor of ocean states. Historical CMIP6 ensemble means show dynamic changes in temperature between 1980 and 2007 (0.5°C), with no corresponding change in sea surface height anomalies, suggestive of decadal changes independent of the Gulf Stream position. While previous deep learning models have been able to predict SST globally, the Gulf Stream region has some of the lowest accuracy and shortest forecast horizons. We use a deep learning (convolutional neural network) model to predict sea surface height anomalies and sea surface temperature for 1 month to 2 years, focusing on the Atlantis II Seamounts. The training dataset consists of the global 1850–2007 CMIP6 temperature, height, and mixed layer depth that capture climate modes of variability that impact the Gulf Stream. We test our model predictions with Ocean Reanalysis System 5 (ORAS5) reanalysis data. Our model can beat the climatology for up to 24 months by predicting different decadal regimes in the Gulf Stream. Using back-propagation saliency maps, we connect Atlantis II Seamounts predictions to multiple global modes of climate variability, such as the North Atlantic Oscillation, the Indian Ocean dipole, the Pacific decadal oscillation, and ENSO.