An Inter-comparison of Neural Net, Regression, and Physical Inversion Approachs for Retrievals of Atmospheric Temperature and Moisture Profiles from a Combination of Advanced Microwave Sounding Unit (AMSU) and Microwave Humidity Sounder (MHS) Data

Statistical regressions and Neural Nets methods have been used for many geophysical remote sensing applications. An evaluation of these methods versus a physical inversion approach for retrieval of atmospheric temperature and moisture profiles from a combination of Advanced Microwave Sounding Unit (AMSU) and Microwave Humidity Sounder (MHS) measured brightness temperatures is presented here and evaluated for the combined AMSU and MHS, AMSU alone, and AMSU 50 GHz sounding channels. The statistical inversion methods include a linear and a quadratic regression in brightness temperature. These approaches assume that the geophysical parameters of interest which determine the observables can be represented by a linear combination of those observables. The weights corresponding to each of the observables are calculated using a training set of geophysical state vectors and their corresponding radiance vectors. In this application, the geophysical state vector consists of atmospheric profiles of temperature, water vapor, and cloud liquid water, as well as surface skin temperature and spectral surface emissivity. The radiance vector is composed of a number simulated brightness temperatures which are calculated at the AMSU and MHS frequencies using the Optimal Spectral Sampling (OSS) forward radiative transfer model developed by AER, Inc. This radiative transfer model assumes a one dimensional, homogenous, non-scattering atmosphere with only oxygen, water vapor, and liquid cloud water droplets as the absorbers. In addition to providing radiance vectors for the regression training set, this forward model is also used in a physical inversion algorithm. The physical inversion algorithm used here is the Unified Retrieval (UR) developped by AER, Inc. The physical inversion method attempts to accommodate non-linear effects by solving the inversion using through Gauss-Newton iteration. In addition to the requirement of providing retrieved temperature and moisture profiles, it is necessary to retrieve additional parameters (stated above) because one or more of the AMSU/MHS channels is sensitive to each parameter and those parameters can not be sufficiently specified a priori. When the non-linear effects of water vapor and cloud water are significant, the physical iterative method should outperform the statistical regression method, but when non-linear effects are small, the two retrieval methods should perform equally well. It is an objective of this work to illustrate under what conditions and to what degree the various algorithms differ.