A multi-temporal convolutional attention network for South Indian Ocean Dipole prediction

The South Indian Ocean Dipole (SIOD) is another critical sea surface temperature (SST) mode in the Indian Ocean besides the Indian Ocean Basin Mode (IOBM) and the Indian Ocean Dipole (IOD). It plays a significant role in modulating extreme precipitation, climate variability over Asia, Africa, and Australia, as well as global atmospheric circulation. Given its broad climate influences, reliable prediction of the SIOD is of great importance. Currently, traditional dynamical models typically limit their SIOD prediction capabilities to the next 2–3 months. To improve the prediction of the peak phase of SIOD events (i.e., January to March), this study proposes a multi-temporal convolutional attention network (MTCAN). This model uses SST and ocean heat content anomalies as inputs and extracts key signals through multi-scale temporal convolution and spatiotemporal attention mechanisms, enabling accurate long-term prediction of the SIOD index. Results demonstrate that MTCAN can effectively prolong the SIOD prediction up to 7 months in advance, outperforming multiple global dynamical models. Further analysis reveals that the model captures key precursors in SIOD evolution; that is, short-term predictions rely on local air–sea feedback, while long-term predictions primarily depend on remote modulation of ENSO and the South Atlantic SST via the atmospheric bridge and Rossby wave trains. This study proposes a novel approach for SIOD prediction, offering new pathways for significantly improving the prediction capability by combining neural networks with physical processes and helping understand the physical mechanism of SIOD evolution.