| This thesis uses an optimal statistical method for averaging global surface temperatures. The method minimizes the standard error budget resulting from statistical sampling errors due to station gaps and random data errors due to instrumental and human factors. Empirical orthogonal functions are used to represent the inhomogeneous covariance structure of the temperature field, and are computed from the following datasets: the Jones' land and United Kingdom Meteorological Office (UKMO)'s ocean 5° x 5° data (1949--1998), the National Center for Environmental Prediction/National Center for Atmospheric Research Reanalysis data (1949--1998), the Climate Prediction Center's optimally interpolated sea surface temperature data (1982--1999), and the National Climatic Data Center's blended data from Global Historical Climatology Network and Special Sensor Microwave Imager (1992--1999). The Jones' box-data (1856--1998) over the land and UKMO's box-data (1856--1998) over the ocean are used as the observations, which are optimally averaged by our optimal averaging method. Using this method, one can generate not only the global average surface temperature of the monthly mean, annual mean, or decadal mean, but also an estimate of the error in the averaging process. The errors resulting from our method are only about 30% of those in the earlier assessments of the same quantities. The behavior of global averages is dominated by the sea surface temperature, whose spatial average has smaller standard errors compared with land data. Global warming trends during 1920--1944 and 1978--1998 are obvious. The standard error decreases from 0.065°C prior to 1900 to 0.03°C in recent time. These errors mainly result from random errors in the observations rather than sampling errors due to station gaps. |
