The Feasibility of Predicting Nino 3.4 Index Using a Sparse Approximation Algorithm

Abstract:

It is well established that sea surface temperature anomaly (SSTA) is one of the principle factors that have significant influence on global climate variability. Due to large mass and great thermal capacity of the oceans, oceanic conditions change relatively slowly and dominant patterns are thus easy to detect. Most of the current research on SSTA make use of PCA methods like EOF or SVD. Though such methods are effective in reducing dimensions, it is always hard to give a physical interpretation of the results and difficult to distinguish the minor eigenvectors from noises. Instead of finding patterns, we put forward a framework for the direct prediction of SSTAs, using a sparse approximation method, the least absolute shrinkage and selection operator (lasso), to reduce the noises in global SST observation. Global SSTA time series in 5°×5° resolution were used to fit each target SSTA vector and the lasso method was utilized to avoid over-fitting. Taking the Nino 3.4 Index as an example, the predictability of the lasso model was studied and the results showed a relatively satisfying prediction skill in terms of correlation coefficient and root-mean-square error compared with the results obtained from LDEO 5. Moreover, by taking other climate variables into consideration, we discovered a stable relation between the Nino 3.4 Index and the sea-ice extent anomaly in South Pole at a lead time of around 2 years. In addition, the bootstrapping method was used to resample the coefficients in the sparse regression model so that we could study their statistical property. 14 regressors were reserved suggesting 10 potential indices which have relatively strong relations with the Nino 3.4 Index. Some of the potential indices corresponded well to known climate indices while the rest indicated an undiscovered index in tropical oceans of eastern South America. In conclusion, the lasso method approved its feasibility in climate prediction at a relatively low computation cost, and its potential to discover climate indices from the sparsified results.