A Novel Model Based on Spatio-Temporal Dilated ConvLSTM Networks for Indian Ocean Dipole Forecasting Using Multi-Source Global Sea Surface Temperature and Heat Content Data

The Indian Ocean Dipole (IOD) is a critical coupled ocean-atmosphere oscillation system associated with significant weather anomalies in the global climate, particularly in the Indian Ocean rim countries. The paper presents a novel deep learning (DL) model, which we call the “spatio-temporal dilated ConvLSTM (STDNet) model”, for forecasting the Dipole Mode Index (DMI) using global sea surface temperature (SST) and heat content (HC) data. The model combines the techniques of dilation and fine-tuning to learn efficiently from the training data. CMIP6 historical simulation data from 5 modeling centres for 1861–2014 is used to train the model. Furthermore, the model is fine-tuned on reanalysis data from 1871–1973. During the testing period (1982–2019), the dipole correlation coefficient (DCC) was the highest when compared with state-of-the-art dynamical North American Multi-Model Ensemble (NMME) models, a convolutional neural network (CNN) and a dilated CNN. On a lead of 12 months, the DCC is 0.40 for the CNN, 0.44 for the dilated CNN, and 0.51 for the STDNet, and all the NMME models have negative correlations. The results show that the STDNet efficiently forecasts the DMI at leads of up to 12 months. The STDNet shows results to overcome the winter predictability barrier.