Evaluation of a Classification Method for Retrieving Atmospheric Temperatures from Satellite Measurements

Abstract

The retrieval of vertical temperature profiles from radiances measured from a satellite is a difficult inversion problem that has been solved to varying degrees of accuracy by several approaches. A method that uses a classification approach for the initial estimate is described. The initial estimate is modified by the application of a single set of regression coefficients that is valid for all latitudes and seasons. Retrievals from this method are compared to the operational retrievals obtained by the National Environmental Satellite, Data, and Information Service (NESDIS). The comparison is based on the carefully matched radiosonde-retrieval pairs that are used in the operational system to update retrieval coefficients. About 10% of the data covering the period July 1987?July 1988 was used to generate the coefficients for both the classification and the retrieval that are used in the classification method. The classification method was then used to process data covering the period July 1987?July 1989. Results from about 100 000 soundings covering the two-year period show the classification method to be consistently (0.2?0.3 K rms) better than the operational method in the lower atmosphere, the region for which the classification was optimized, and slightly worse near 100 hPa. A significant feature of the evaluation is that all the coefficients used in the classification method were constant for the two years. Most retrieval systems are periodically tuned to agree with radiosondes. Examination of the mean from the classification retrievals produced with fixed coefficients show a slight drift over the two years in the lower atmosphere, a sudden change in the upper atmosphere, (probably due to a change in one or more channels), and seasonal changes in bias at some levels. These effects are diminished in the operational results because the operational coefficients are adjusted to agree with radiosonde reports once a week. The increase in accuracy of the classification retrievals over the operational ones could be enhanced by updating the class means using a procedure similar to the one currently employed by the operational system. The evaluation period happened to include the time when a major change was made to the operational method and shows a slight improvement in the operational results after the change.