P2.2
Using regression models to enhance signals in a dispersive radiative field: Reducing stray light corruption in the limb profiler of the Ozone Mapper Profiler Suite (OMPS)
John W. Bergman, Computational Physics, Inc., Boulder, CO; and L. E. Flynn, J. Hornstein, and J. Lumpe
As the technology of remote sensing using radiometric observations advances and we are confronted with weaker and weaker signals, stray light corruption will become an increasingly important challenge. Stray light is particularly important for instruments that sample a large dynamic range such as the OMPS limb profiler. In that case, only a very small fraction of photons straying from the high-intensity region of spectral/viewing angle space can dominate measurements in the low intensity region. For the OMPS limb profiler, stray light represents a small linear perturbation to the overall observed energy, even though it dominates observations for some wavelengths and viewing angles. By exploiting those characteristics, we have found that both iterative techniques based on a Taylor expansion of the inverse stray light operator and linear regression models can effectively reduce stray light corruption from the OMPS measurements provided there is sufficient sampling of the measured radiant energy.
Regression models are extremely efficient in operational application because their cost is incurred during offline training. However, they can perform no better than the data used to train them and only work well for systems whose dynamical operators are largely linear. Regression models can be problematic in an operational application if the instrument undergoes changes (e.g., pixel failure in the detector); with the instrument in orbit, retraining can be difficult. Thus, it is important to both characterize stray light with instrument tests before it is launched and to find stray light removal techniques that are flexible and can be altered to accommodate instrument changes. We examine three variations of an ozone retrieval algorithm that utilizes regression models to characterize photon dispersion and other instrument effects. These three methods are compared in terms of their sensitivity to model error, their sensitivity to errors in the assumed background atmospheric conditions, and complications resulting from dead pixels.
Poster Session 2, Applications and Exploitation of NPOESS and GOES-R Data Products II
Wednesday, 17 January 2007, 2:30 PM-4:00 PM, 217D
Previous paper Next paper