Soundings
7. Lei Shi |
|
Neural network techniques are applied to the retrieval of atmospheric temperature and humidity profiles from NOAA-15 and NOAA-16 Advanced Microwave Sounding Unit-A (AMSU-A) and Unit-B (AMSU-B) measurements. The neural network training sets are built based on the pole-to-pole global AMSU-A and AMSU-B measurements and the corresponding analysis data generated at the National Centers for Environmental Prediction. For the temperature retrieval, the AMSU-A data consist of channels with weighting functions peaking below 10 hPa, i.e. channels 1-12 and channel 15. For the humidity profile retrieval, the scheme is based on channels 1-8 of AMSU-A measurement and all five channels of AMSU-B measurement. Once the neural network is trained and the parameters are optimized, the neural network model is ready for retrieval of new profiles. The neural network scheme can be applied to either direct readout of local passes or recorded global data. During actual operations of temperature and humidity retrieval, the scheme uses only the satellite measurements and satellite geometry data, without requiring additional first-guess from modeled profiles. This gives advantage to many operational sites including those with limited internet connections.
The figures below show the root mean square (rms) values of temperature and specific humidity retrievals based on a combination of most diversified atmospheric conditions in our data sets. The data cover global pole-to-pole, land and ocean, and hot and cold seasons. It should be noted that when a neural network is trained for a local region or a single season, the statistics generally yield smaller rms values than shown here.
