In the preprint "Millimeter-wave precipitation observations versus simulations: sensitivity to assumptions", submitted to the Journal of Atmospheric Sciences (JAS) by C. Surussavadee and D. H. Staelin and appended here, it is shown that AMSU retrievals of surface-precipitation rates exhibit useful performance above ~1 mm/h, and exhibit significantly better performance for hydrometeor water paths because most hydrometeors reside at higher and more satellite-accessible altitudes.  The neural-network-based retrieval algorithm was trained with the MM5/TBSCAT model described in the JAS paper and tested against independent pixels generated by the same model for 122 globally and seasonally distributed storms.  Thirteen frequencies 23-190 GHz were observed, including all 8 AMSU-A frequencies below 56 GHz, and all 5 AMSU-B frequencies.

When the true (MM5) values were divided into octaves it was found that the rms retrieval accuracies for snow water path varied between 0.05 mm for MM5 paths of 0.125-0.25 mm, and 0.18 mm for MM5 paths of 1-2 mm.  The corresponding rms water-path retrieval performances for graupel, cloud ice, and rain water for the same two octaves were 0.19-0.93, 0.10-0.41, and 0.16-0.78 mm, respectively.  The rms retrieval accuracies for 15-minute surface precipitation rates varied between 0.4 mm/h for MM5 rates 0.125-0.25 mm/h, and 21 mm/h for MM5 rates 32-64 mm/h.  These surface precipitation–rate errors were correlated (~0.3) with the presence of virga.  The rms errors for retrieved cell-top altitudes for snow/graupel were less than ~1.2 km for cell-top altitudes 0 – 20 km. 

As part of this study it was found that the sensitivity of AMSU retrieval accuracies to model assumptions was small and therefore future system performance limits should be reliably predictable.  This sensitivity was tested by varying many aspects of the models, such as ice factors, loss tangents, back-scattering fractions, hydrometeor size and altitude distributions, and production rates for snow and graupel.  Since it was shown in this paper that the brightness temperatures are much more sensitive to these same assumptions, and particularly to the back-scattering fraction, achieving such retrieval accuracies will require that discrepancies between predicted and observed radiance spectra be small.  The paper suggests that these radiance discrepancies may already be approaching acceptability.