A predictive model integrating multimodal meteorological information and spatiotemporal evolution of MJO

The Madden-Julian Oscillation (MJO), as the primary mode of tropical intraseasonal variability, plays a critical role in improving subseasonal prediction skill. However, due to its multi-scale evolutionary characteristics and highly nonlinear dynamical processes, existing prediction methods still struggle to effectively capture the complex spatiotemporal structure of MJO. To address this issue, we propose a novel prediction model named MISM (Multi-modal data and Integrated Spatiotemporal features for MJO prediction), which integrates multimodal inputs and spatiotemporal feature extraction. The model jointly leverages historical Real-time Multivariate MJO (RMM) indices and multiple meteorological variables as inputs. It constructs a spatial feature extraction module based on Squeeze-and-Excitation (SE) blocks, convolutional layers, and the Swin Transformer, as well as an autoregressive attention mechanism for nonlinear temporal modeling. Experimental results demonstrate that the MISM model extends predictive skill to beyond 30 d and shows overall superior performance compared with traditional dynamical and statistical methods in long-lead forecasts beyond 25 d. Furthermore, saliency maps are utilized to analyze the contribution regions of meteorological factors. The results reveal that the western Pacific and the Indonesian archipelago consistently exhibit high sensitivity across different lead times, with oceanic regions generally contributing more than land areas. Water vapor and sea surface temperature anomalies play a more prominent role in short- to medium-term forecasts, while low-level wind fields and convective activity contribute more significantly in longer-term forecasts. High-level circulation exerts a stable influence across all lead times, highlighting the model’s ability to capture the mechanisms of MJO evolution.