Retrieving atmospheric temperature parameters from DMSP SSM/T-1 data with a neural network

We show that back propagation neural networks yield excellent results in retrieving air temperature profiles from the 1000- to the 10-mbar pressure level together with tropopause temperature and pressure estimates using operational brightness temperature data from the special sensor microwave imager (SSM/T-1) microwave radiometer. Networks trained and tested with matched SSM/T-1 measurements and conventional soundings collected during a 30-day period in northern hemisphere winter demonstrated rms retrieval errors substantially less than 2K from 500 to 30 mbar, significantly outperforming an operational linear-regression algorithm using the same data. Tropopause temperature and pressure retrievals showed rms errors of 2.15 K and 19.8 mbar. Retrieval accuracy of the system exceeds that of any previously published method using DMSP data and equals or exceeds that of published studies using data from other satellite-borne instruments. Retrieval accuracy under possible failure modes of the SSM/T-1 instrument are also considered, as are ways to recover from single-channel loss. The method retrieves profiles and tropopause parameters with acceptable accuracy either if the brightness-temperature of any one channel is offset 1.5 K or more or if uniform random noise with a peak value in the range (-2, 2) K is added in one channel. The performance is only slightly more impaired if all channels are simultaneously offset up to 1.5 K or if random noise in the range (-1, 1) K is simultaneously added to all channels. Under single-channel loss the retrieval error can be made small at virtually every level by retrieving with a network trained without that channel.