Deep Learning Improves Prediction of the Boreal Summer Intraseasonal Oscillation Using Predictive Source Analysis

Abstract The Boreal Summer Intraseasonal Oscillation (BSISO) is a prominent tropical intraseasonal variability during summer. It propagates northeastward from the northern Indian Ocean to the western North Pacific and has a more complex structure than the winter‐dominant Madden‐Julian Oscillation, which makes its prediction challenging. We propose a real‐time prediction framework for the bimodal intraseasonal oscillation index using deep learning (DL) models for 1‐month forecasting. We show that the BSISO prediction skill (Cor = 0.6) reaches about 29 days during June to August, outperforming existing numerical models. Using explainable artificial intelligence, we quantitatively evaluated DL model prediction sources. Precipitable water (PW), sea surface temperature (SST), and 200 hPa zonal wind significantly contributed to forecasts over 15 days. For 15–20‐day predictions, the heatmap‐based analysis indicated that PW mainly contributes to eastward propagation over the maritime continent. The Indo‐Pacific SST warm pool also exhibits large predictive contributions, suggesting a role in triggering northward propagation.