Multimodal Deep Learning for Two-Year ENSO Forecast

Predicting the onset of the El Niño Southern Oscillation (ENSO) in the current rapidly changing climate could help save thousands of lives annually. Since the variability of this phenomenon is increasing, its prediction is becoming more challenging in the post-2000 era. Hence, we present a novel Multimodal ENSO Forecast (MEF) method for predicting ENSO up to two years for the post-2000 condition. The model receives a Sea Surface Temperature (SST) anomaly video, a heat content (HC) anomaly video, and an augmented time series to predict the Niño 3.4 Index. We utilize a multimodal neural network to elicit all the embedded spatio-temporal information in the input data. The model consists of a 3D Convolutional Neural Network (3DCNN) that deals with short-term videos and a Time Series Informer (TSI) that finds the base signal in long-term time series. An Adaptive Ensemble Module (AEM) ranks the 80 ensemble members based on uncertainty analysis, discarding outliers and calculating a weighted average to reach the final prediction. We successfully tested the model against observational data and the state-of-the-art CNN model for a long and challenging period from 2000 to 2017. For almost all target seasons, MEF’s skill is higher than that of the state-of-the-art CNN method, with correlation values exceeding 0.4 for all lead months. Moreover, the proposed method captures nearly 50% of all El Niño and La Niña events, even for 23-month lead times. The results ensure the MEF’s validity as a reliable tool for predicting ENSO in the upcoming Earth’s climate.