Enhanced Monitoring of Madden Julian Oscillation through Autoencoder Based Analysis

This study presents an alternative approach to analyzing the Madden-Julian Oscillation (MJO) using deep learning techniques. For decades, empirical orthogonal function analysis has been successfully applied to monitor MJO characteristics and evolution. However, this method overlooks nonlinear processes due to mathematical limitations. In this study, we employ an Autoencoder to capture the dominant tropical variability, with ERA5 outgoing longwave radiation data from the winters of 1988 to 1997 serving as input. The MJO patterns in phases 3-4, as decoded by the model, reveal distinct Rossby-wave vortices north and south of the equator—features not as clearly discernible in patterns derived from reconstructed empirical orthogonal functions. Our findings suggest that the nonlinear capabilities of deep learning can enhance the detection of intraseasonal variability over the Maritime Continent, offering a promising new method for monitoring MJO using observational data.