Parsimonious patterns in sea surface temperature of the tropical Pacific ocean

Abstract

A variety of spatiotemporal oscillations have been explored using principal component analysis (PCA) or rotated PCA (RPCA). Recent literature has noted many shortcomings of PCA and RPCA in the investigation of climate variability in a high-dimensional state space. The main issue is that both PCA and RPCA produce spatial patterns full of nonzero loadings, which often encumbers the physical interpretation of intrinsic signatures. To address this issue, sparse PCA (SPCA) was employed to identify parsimonious patterns in sea surface temperature (SST) of the tropical Pacific Ocean. Sparse regression analysis was also performed using the sparse principal component time series to obtain the associated spatial patterns in mean sea level pressure (MSLP) and surface wind fields. The results were compared with those of PCA and RPCA. The SPCA produced sparse structures pertinent to the variation of SST. The sparse regression successfully revealed the localized atmospheric responses partially connected with the individual eigenmodes of the SST, while the PCA did not identify the centers of variation. The RPCA failed to distinguish each eigenmode in the patial structure and power spectra of the SST anomaly. The RPCA PC time series could not produce any relevant spatial patterns in the regression analysis.