Satellite radiance data is biased relative to any given forecast model. There are two main sources of bias: scan bias, and air-mass bias. For cross-track scanning instruments such as the Advanced Microwave Sounding Unit (AMSU), scan bias results from the changing local zenith angle, and other factors as well (indicated by the asymmetric scan bias pattern observed in most AMSU-A channels on both the NOAA15 and NOAA16 satellites). Air-mass bias derives from the fast (i.e. approximate) radiative transfer code of the forecast model, and depends on the temperature and moisture characteristics of the underlying atmosphere, and the surface emissivity for those channels that sense the surface. Our goal is to find and implement the best possible bias correction scheme for satellite radiance assimilation.

Most major weather centers use linear regression to correct bias. The scheme previously used at NRL was adapted from Eyre (1992), and used a simple, global scan-bias correction, along with a global linear regression against microwave brightness temperatures, to correct for bias in the difference between observed and model brightness temperatures. A different scheme due to Harris & Kelly, 1999, uses four model forecast fields (1000-300 and 200-50 hPa thickness, surface skin temperature, and total column precipitable water) to predict air-mass bias. The latitudinal dependence of air-mass bias is accounted for by performing 18 independent regressions, one for each 10-degree latitude band. The use of model fields as predictors is quite sensible and produced good results; however, the linear artifacts along the latitude band boundaries are a cause for some concern. A global air-mass regression, using the Harris and Kelly predictors plus cloud liquid water, along with a subset of observed brightness temperatures as in Eyre’s scheme, should result in superior bias correction. Preliminary results using a subset of these predictors are encouraging.