Multivariate Methods for the Detection of Greenhouse-Gas-Induced Climate Change

Abstract

This investigation considers whether observed changes in surface air temperature are consistent with GCM equilibrium response predictions for a doubling of atmospheric CO 2 . The model considered is a version of the Oregon State University (OSU) atmospheric general circulation model (AGCM). The study consists of three stages. In the first stage we examine the spatial structure of changes in the annual mean and annual cycle for surface air temperature, mean sea-level pressure (SLP) and precipitation rate. Signal-to-noise (S/N) ratios or equivalent test statistics are then computed (using the 1×CO 2 and 2×CO 2 data) in order to identify variables most useful for detection purposes. Changes in both means and variances are considered as possible detection parameters. The highest S/N ratios are obtained for annual-mean and winter surface air temperature, and the lowest S/N ratios are obtained for SLP. There are significant increases in the temporal and spatial variability of precipitation, and significant decreases in the the temporal and spatial variability of surface air temperature. In the second stage we examine the spatial structure of observed changes in surface air temperature and determine their statistical significance. Significance is assessed using a permutation procedure and the statistics applied in the S/N analysis. The results indicate that there have been significant observed changes in the annual mean and the means for individual months. Observed changes in temporal and spatial variability are generally in the same direction as the model-predicted results. The final stage of the investigation addresses the question of whether the model-predicted surface air temperature signal is present in the observed data. Anomaly fields are computed for the observed (1977–86 minus 1947–56) and simulated (2×CO 2 minus 1×CO 2 ) data. The simulated anomalies are scaled using results from a one-dimensional model. Global tests of the mean indicate that there are highly significant overall differences between the observed changes and the scaled simulated changes in all tests performed. These results are strongly influenced by the low spatial variance of the simulated anomaly fields, and are not sensitive to the applied scaling or to the selection of data for defining observed changes. The spatial patterns of the observed and simulated temperature changes are significantly different in all cases except February. Two possible explanations for the low degree of correspondence between observed and simulated patterns of temperature change are considered. The first is that the observed temperature signal is still too small to be detected against the background noise of natural variability. The second explanation is that the model signal may be erroneous due to model deficiencies and/or inherent differences in equilibrium and transient patterns of temperature change. Both explanations are likely to be valid.