Kernel and Information-Theoretic Methods for the Extraction and Predictability of Organized Tropical Convection

In this paper, we investigate both the dominant modes of variability and the large-scale regimes associated with tropical convection that can be recovered from infrared brightness temperature data using data mining and machine learning approaches. A hierarchy of spatiotemporal patterns at different timescales (annual, interannual, intraseasonal, and diurnal) is extracted using a nonlinear dimension reduction method, namely, nonlinear Laplacian spectral analysis (NLSA). The method separates very clearly the boreal winter and boreal summer intraseasonal oscillations as distinct families of modes. The predictability of the Madden-Julian oscillation (MJO) is then quantified using a cluster-based information-theoretic framework adapted for cyclostationary variables. Data clustering is performed in the space of the NLSA temporal patterns and the results show a strong influence of ENSO in the early MJO season.