Statistical Prediction of ENSO from Subsurface Sea Temperature Using a Nonlinear Dimensionality Reduction

Abstract

Abstract Numerous statistical and dynamical models have been developed in recent years to forecast ENSO events. However, for most of these models predictability for lead times over 10 months is limited. It has been hypothesized that the tropical Pacific thermocline structure may have critical information to permit longer lead ENSO forecasts. Models that use subsurface sea temperature information have already been known to produce better long lead forecasts. Here, a two-stage statistical ENSO forecasting model is developed and demonstrated using the spatially distributed depth of the 20°C isotherm (D20) as a proxy for the thermocline. In the first stage, a nonlinear dimension reduction method [maximum variance unfolding (MVU)] is used to decompose the D20 data into canonical modes. The leading spatial patterns as well as lagged values of Niño-3 are then used as predictors in a set of linear regression models to predict the Niño-3 index at lead times of up to 24 months. Cross-validated forecasts using this methodology are shown to have higher skill than those that use a dimension reduction of the same thermocline data using principal component analysis (PCA). The first three modes of the D20 data as revealed by MVU account for 89% of the variance of the data, as compared to only 48% of the variance if PCA is used. The spatial patterns of the MVU modes partition the data field in a different way than the PC modes, even though some similarities exist as to the main regions that are active. These patterns and their temporal structure are discussed here, with a view to understanding the possible source of the longer-range predictability of ENSO using the MVU modes. The skill of the PCA- and the MVU-based forecasts of Niño-3 varies depending on the starting month of the forecast for short lead times (5–10 months). However, for the lead times longer than 1 yr, the MVU-based forecast skill is not seasonally variable, while the PCA-based models do not provide significant skill at these lead times irrespective of the starting month of the forecast. Similar conclusions are obtained for forecast models for the Niño-3.4 and Niño-1.2 indices. The differences between the MVU- and PCA-based models are most marked for the Niño-1.2 long lead forecasts.