Improving Long-lead ENSO Prediction with Joint ENSO Transformer

Prediction of El Niño-Southern Oscillation (ENSO) events in sufficient advance time will facilitate early planning to mitigate its adverse impacts on agriculture, marine ecosystems and public safety. Recent studies show that deep learning models have great potential for ENSO prediction. However, the one year advance forecasting of ENSO remains challenging and previous task settings could not output multiple-month forecasts concurrently. In this letter, we propose the Joint ENSO transformer (JETR) to promote long-lead ENSO prediction. JETR contains pretext task to generate more data and balances the complex spatiotemporal association of ENSO in main task. By jointly training the pretext task with the main task, the model can skillfully output oceanic Niño index (ONI) for 24 months simultaneously. Experimental results also suggest that the proposed JETR model can predict long-lead ENSO with higher correlation compared with other deep learning models.