Deep Ensemble and Encoder-Decoder neural networks for ENSO forecasting
Summary
This study introduces two novel statistical models for the prediction of the El Niño-Southern Oscillation
(ENSO). The first model is termed Deep Ensemble (DE). This neural network method was recently
developed by Lakshminarayanan et al. (2017). A DE is applied for the first time in the prediction of the
ENSO. Instead of only predicting one value for Oceanic Niño Index (ONI), as it is done by other statistical
ENSO models, the DE predicts the mean and the standard deviation of a Gaussian distribution. In this
way, the forecast comes with an estimation of the predictive uncertainty, which is a novel feature in the
prediction of the ENSO by statistical models. The predictor variables for the DE are chosen such that
they represent different aspects of the ENSO dynamics. The memory component of the subsurface is
included by the use of the warm water volume (WWV) of the equatorial Pacific. As a driver of more
stochastic effects, the area averaged zonal wind stress anomaly in the west Pacific (WP) is added to the
predictor variables. To include information about the interdecadal changes in the background state of
the Pacific, the amplitude of the leading empirical orthogonal function (EOF) of the 5-year running-mean
sea surface temperature anomaly (SSTA) field is used. Other predictor variables which are used are the
dipole mode index (DMI) of the Indian Ocean Dipole (IOD), two variables from the theory of evolving
complex networks as well as the ONI itself. The trained DEs show similar prediction skills as other
statistical ENSO models do that are currently used in operational ENSO forecasting. As expected, the
DEs assign low predictive uncertainties to forecasts with a small lead time. In contrast, for very long lead
times, the DEs predict the climatological distribution of the ONI. Unfortunately, decadal variations in
the estimated predictive uncertainty are not clearly visible within the forecasts. This can be attributed
to the low amount of available training data.
The second model is a so-called Encoder-Decoder (ED) model which is inspired by the architecture of
Autoencoders (AEs). The ED is used to predict the entire SSTA field in the Pacific ocean between
30◦S-30◦ N and 120◦ E to 80◦ W. Therefore, the prediction provides information on the spatial pattern of
the anomalies. The model shows generally weaker skills than the DE regarding the ONI. Hence, it has a
weaker prediction skill in comparison to other statistical models. However, it is still remarkable that the
ED can make skillful predictions given the little amount of data which was available for the training in
respect to the complexity of the model. This shows that the bottleneck architecture of the ED effectively
prevented overfitting.
Both models are analyzed onto their predictive skill during different decades and for different seasons.
It is proven that the models have a good predictive skill for long lead times when the background state
of the Pacific is relatively warm (El Niño-like decades), whereas they perform considerably worse for a
colder background state (La Niña-like periods). Based on the findings of this study and of a literature
review, a hypothesis is proposed to explain why it is possible to make meaningful prediction beyond the
spring predictability barrier during El Niño-like periods whereas in La Niña-like periods this is usually
impossible.