Retrieval of Atmospheric Temperature Profiles from AMSU-A Measurement Using a Neural Network Approach

Abstract

Backpropagation neural networks are applied to retrieve atmospheric temperature profiles and tropopause variables from the NOAA-15 Advanced Microwave Sounding Unit-A (AMSU-A) measurement based on two different data sources. The first case uses direct acquisition of 15-channel AMSU-A data over the eastern United States and western Atlantic Ocean for the months of July 1998 and January 1999, and the second case uses recorded global AMSU-A data for several days of January 2000. The corresponding global analysis data from the National Centers for Environmental Prediction are employed to build the neural network training sets. The retrievals yield excellent results in the atmospheric temperature profiles from the surface to the 10-hPa pressure level. For the more generalized global data retrieval case, the root-mean-square (rms) deviation of temperature retrieval is 3.2°C at the surface, only 1.0° to 1.2°C in the midtroposphere, less than 1.5°C around the tropopause, and between 1.0° and 1.5°C in the stratosphere. Simultaneous retrieval of tropopause temperature, height, and pressure yields the rms deviations of 1.9°C, 0.58 km, and 18.1 hPa, respectively, for these variables. Within the scope of regional data, the trained neural network results in smaller values of temperature profile rms deviations than those of the global-data case. When compared to a linear regression approach, the neural network retrieval yields significantly better results for all the atmospheric levels. The neural network with parameters obtained from the network training optimizations can be easily applied to AMSU-A retrieval operationally.