Improving error structure in temperature profile retrievals from satellite observations

Satellite based retrievals, both conventional and neural network based, have typically used Root Mean Square Errors (RMSE) as a “goodness” metric. Conventional and neural network approaches have been used to retrieve atmospheric temperature profiles from meteorological satellite data. Although the error over all the examples and all the levels is low, the structure of the error for a given profile is not optimal from an operational perspective. Of the approaches studied, the best technique for a direct retrieval of atmospheric temperature profiles partitions the data based on the largest eigen value of the channel. The advantages of this technique for different sets of data are discussed