An optimization approach for aerosol retrievals using simulated MISR radiances

Currently, many satellite-based aerosol retrievals make use of lookup tables (LUTs) containing precomputed solutions to the radiative transfer (RT) equation. The benefit of this strategy is the avoidance of expensive runtime calculations, but its main drawback is that the LUTs discretize what is inherently a continuous, multivariate solution space. The operational retrieval algorithm for the Multi-angle Imaging SpectroRadiometer (MISR), for example, compares the observations to a set of 74 aerosol mixtures, each composed of particle models having prescribed optical properties and size distributions. In a recent “blind” study comparing the performance of several satellite retrieval algorithms on simulated data over a black surface, the MISR algorithm performed reasonably well in recovering the “true” spectral aerosol optical depths (AODs), but because the correct aerosol model was not contained within the MISR LUT, the retrieved AODs were biased low by ~ 14%. This motivated an investigation of whether an optimization approach, in which the aerosols are modeled by a set of continuously variable parameters recovered using nonlinear least-squares, could improve the results. In this paper, we demonstrate that such an approach using Levenberg–Marquardt optimization yields superior accuracy. Advances in computer speed, development of more efficient RT codes, and algorithm innovations will be necessary for this approach to satisfy the demands of a global, production-level satellite aerosol retrieval process, especially when used in conjunction with future instruments having enhanced sensitivity to diverse aerosol properties.

Research highlights
► Satellite-based aerosol retrievals often use precomputed lookup tables (LUTs).
► This avoids computational load but limits solutions to discrete, prescribed models.
► If the correct aerosol is not present in the LUT, results may be biased.
► Using simulated satellite data, we investigated an alternative retrieval approach.
► Levenberg-Marquardt optimization yielded superior results compared to the LUT method.