Spatio-temporal pattern analysis of MOPITT total column CO using varimax rotation and singular spectrum analysis

In this paper, we apply varimax Empirical Orthogonal Function (EOF) analysis to Measurement of Pollution in the Troposphere (MOPITT) total column CO retrievals to answer the question of whether or not it is possible to disentangle the dominant CO sources associated with inferred modes of variability at regional to global scales. This study aims to highlight the strengths and limitations of EOF analysis, specifically their usage in the field of atmospheric chemistry. In particular, we highlight an instance where varimax rotation fails to reveal additional spatial structure of a dataset. Additionally, we emphasize to the reader that the success of EOF analysis to determine linearly independent physical modes relies on both the statistical distribution of the dataset as well as its temporal covariance structure. We analyzed daily MOPITT Version 8 Level 3 joint (TIR-NIR) products from 2005 to 2018, aggregated every 8 d on a 1° by 1° grid. Our findings show that EOF patterns of MOPITT CO are consistent across various regional subdomains, demonstrating that these spatial patterns are independent of the chosen domain. A comparison of the eigenvalue spectrum reveals that unrotated EOF analysis yields three distinct modes, while varimax rotation reduces these to two. The power spectra of the principal components indicate that the first two unrotated modes are primarily driven by annual and semi-annual cycles, while the third mode reflects seasonal variations occurring over roughly three months. To further isolate these modes, we employed singular spectrum analysis (SSA) at each grid point to generate long-term, seasonal, and residual EOF patterns. The power spectrum analysis of the principal components shows that the long-term EOFs replicate the original two dominant modes, while the seasonal EOFs reveal significant variations over 2 to 3 months, and the residual modes exhibit time scales of 2 months or shorter. By plotting the mean skewness field, we show the dataset is non-Gaussian, leading us to conclude its principal components are time-dependent despite being uncorrelated. The periodic decay observed in the temporal autocorrelation function for each time series suggests a classification of wide-sense cyclostationary behavior. The non-stationarity of each time series and the temporal dependence of modes lead us to conclude that EOF analysis alone cannot fully disentangle individual CO sources. Careful consideration is therefore required when interpreting EOF patterns of other composition datasets with similar characteristics.