Relative Performance of Self-Organizing Maps and Principal Component Analysis in Pattern Extraction from Synthetic Climatological Data

Abstract

As a contribution toward improving our ability to identify robust patterns of variability in complex, noisy climate datasets, we have compared a relatively new technique, Self-Organizing Maps (SOMs), to the well-established method of principal component analysis (PCA). Recent results suggest that SOMs offer advantages over PCA for use in climatological and other studies. Here each analysis technique was applied to synthetic datasets composed of positive and negative modes of four idealized North Atlantic sea-level-pressure fields, with and without noise components, to identify the predefined patterns of variability. PCA, even with component rotation, fails to adequately extract the known spatial patterns, mixes patterns into single components, and incorrectly partitions the variance among the components. The SOMs-based analyses are more robust and, with a sufficiently large set of generalized patterns, are able to isolate all the predefined patterns with correct attribution of variance. With PCA, it is difficult, if not impossible, to detect pattern mixing without prior knowledge of the patterns being mixed.