A Deep Learning Model for Forecasting Global Monthly Mean Sea Surface Temperature Anomalies

Abstract

Sea surface temperature (SST) variability plays a key role in the global weather and climate system, with phenomena such as El Niño-Southern Oscillation regarded as a major source of interannual climate variability at the global scale. The ability to be able to make long-range forecasts of sea surface temperature anomalies, especially those associated with extreme marine heatwave events, has potentially significant economic and societal benefits. We have developed a deep learning time series prediction model (Unet-LSTM) based on more than 70 years of ECMWF ERA5 monthly mean sea surface temperature and 2-metre air temperature data. The Unet-LSTM model is able to learn the underlying physics driving the temporal evolution of the 2-dimensional global sea surface temperatures. The model accurately predicts sea surface temperatures over a 24 month period with a root mean square error remaining below 0.75 • C for all predicted months. We have also investigated the ability of the model to predict sea surface temperature anomalies in the Niño3.4 region, as well as a number of marine heatwave hot spots over the past decade. Model predictions of the Niño3.4 index allow us to capture the strong 2010-11 La Niña, 2009-10 El Nino and the 2015-16 extreme El Niño up to 24 months in advance. It also shows long lead prediction skills for the northeast Pacific marine heatwave, the Blob. However, the prediction of the marine heatwaves in the southeast Indian Ocean, the Ningaloo Niño, shows limited skill.These results indicate the significant potential of data driven methods to yield long-range predictions of sea surface temperature anomalies.

Predictability of sea surface temperature anomalies at the eastern pole of the Indian Ocean Dipole—using a convolutional neural network model

“…There are machine learning model architectures that only require short data records to train to achieve model stability and useful prediction skills. Taylor and Feng (2022) developed a fully convolutional network (U-Net), trained using a 70-year reanalysis product, to achieve reasonable multiseason prediction skills for the 2-dimensional monthly SST anomalies in the tropical-subtropical Pacific. A much simpler artificial neural network model based on correlations in the observational data can also achieve reasonable prediction skills for the IOD (Ratnam et al, 2020).…”

Machine Learning for numerical weather and climate modelling: a review

“…This led to a 5.8% improvement of the correlation coefficient for Nino3.4 index prediction and 13% improvement in corresponding temporal classification with a 12-month lead time compared to a 2D CNN. Taylor & Feng (2022) It is clear from the small number of (but rapidly evolving) studies in this space that there is great promise for the use of ML for seasonal and multi-year prediction of ocean variables, with many avenues to pursue to achieve potential skill gains.…”

“…The literature on the use of ML for prediction on seasonal to climate timescales is still relatively sparse compared to its use for nowcasting and weather prediction. Some examples have been covered in previous sections, such as Weyn et al (2021) on subseasonal to seasonal timescales in the atmosphere, and Ham et al (2019), Ham et al (2021), Kim et al (2022 and Taylor & Feng (2022) on seasonal to multiyear timescales in the ocean. A major cause for this sparsity is that deep learning typically requires large training datasets, and the available observation period for the earth system is too short to provide appropriate training data for seasonal to climate timescales in most applications.…”