PCMDI has conducted exploratory studies of the potential usefulness of various linear and nonlinear statistical methods for model/model and model/data intercomparisons.
Sengupta and Boyle (1993) applied the technique of common principal components (CPCs) to identify similar modes among vertically averaged atmospheric temperature fields as observed by the Microwave Sounding Unit (MSU) on satellites, as simulated by the ECMWF global atmospheric model, and as represented by ECMWF analyses (a combination of data and model). They concluded that the use of CPC is a potentially promising approach for rigorously identifying model/data and model/model similarities, especially when the analysis is restricted to regions (e.g., the equatorial Pacific) where a dominant mode of variability (e.g. the ENSO) is likely to occur.
Sengupta and Boyle (1995) considered the relative merits of applying nonlinear principal component analysis (NLPCA) based on auto-associative neural networks vs the more traditional linear analysis (PCA) to identify the dominant modes of precipitation variability over the U.S. They found that NLPCA effected somewhat greater data reduction and that the leading nonlinear mode captured the seasonal cycle of precipitation more clearly than the leading linear mode. They concluded that use of NLPCA should be considered especially when linear PCA fails to uncover physically meaningful patterns in climatological data analysis.
References
Sengupta, S.K., and J. S. Boyle, 1993: Statistical intercomparison of global climate models: A common principal component approach. PCMDI Report No. 13, 41 pp.Abstract
Sengupta, S.K., and J. S. Boyle, 1995: Nonlinear principal component analysis of climate data. PCMDI Report No. 29, 21 pp. Abstract
UCRL-MI-119847