Kernel Analog Forecasting of Tropical Intraseasonal Oscillations

This research examines the forecasting of the Madden-Julian oscillation (MJO) and boreal summer intraseasonal oscillation (BSISO) by applying a recently developed nonparametric empirical method to satellite-obtained global brightness temperature data. This method seeks to improve the forecasting of intraseasonal oscillations (ISOs) by using machine learning techniques to exploit the natural connection between kernel eigenmodes and the statistical forecasting of such eigenmodes. In particular, specific eigenmodes of a kernel operator are adopted as baseline definitions of ISOs of interest, which are then extended into forecasts and, ultimately, reconstructed spatio-temporal patterns. The pattern correlation of the forecasts produced in this manner remain above 0.6 for 50 days for both the MJO and BSISO when 23 years of training data is used, and 37 days for the MJO when nine years of data is used.

Defining ISOs in a consistent and objective manner is frequently an obstacle to forecasting the tropical climate. A consensus measure of the BSISO remains elusive. Moreover, although the Wheeler-Hendon real-time multivariate index (RMM) is now a commonly used standard for MJO measurement, its reliance on linear techniques makes it less than ideally suited to capturing the multi-scale nature of a phenomenon like organized tropical convection, as it requires ad hoc data preprocessing to isolate the temporal and spatial scales of interest. The recently developed technique of Nonlinear Laplacian Spectral Analysis (NLSA) seeks to redress this mismatch by using time-lagged embedding and local kernel measures of data similarity to more sensitively capture nonlinear dynamics. The MJO and BSISO emerge as eigenmodes of an NLSA kernel, and the spatio-temporal patterns extracted through this process display higher temporal coherence and stronger discriminating power between eastward and poleward propagation than patterns extracted through linear approaches. These modes are then incorporated in an analog scheme whose forecasts are weighted averages of historical data, with weights determined by the NLSA kernel. The extent to which tropical predictability benefits from this improved representation of ISOs, as well as by the kernel analog weights, is explored in this research.

The aforementioned kernel analog forecasting method is applied to 27 years of brightness temperature data as collected under the Cloud Archive User Service project (CLAUS). The chief result of this research is that the PC score for kernel analog forecasts of the MJO and BSISO can stay above a 0.6 threshold for nearly seven weeks of lead time. Context for how this result fares against other prediction methods and ISO indices is discussed, as well as extensions to how well such a method can predict more specific physical phenomena in the tropics, such as precipitation over the Indian subcontinent.