Multi-Scale Decomposition for Skillful All-Season MJO Prediction With Deep Learning

Abstract

The Madden-Julian Oscillation (MJO) is a key intraseasonal atmospheric pattern in the tropics, significantly influencing global weather and extreme events. Accurate subseasonal forecasts depend on reliable MJO prediction, and recent artificial intelligence (AI)-based models have shown great promise for enhanced prediction of MJO index. However, existing AI models have used only MJO anomalies as input data, overlooking the broader atmospheric context such as seasonal and interannual variability. Here, we present a new deep learning framework that explicitly incorporates background fields as inputs together with MJO anomaly variables. Through the multi-scale decomposed inputs, our model improves the MJO prediction skill up to 26 days in boreal winter and 29 days in boreal summer. The detailed model interpretation analyses reveal the important role of background information in long-range MJO forecasts, highlighting utilizing background fields in the development of next-generation sub-seasonal forecasting models, which could help advance global weather predictability and disaster preparedness.