Integration of Deterministic and Statistical Algorithms for Aerosol Retrieval

Aerosol optical thickness (AOT) is typically estimated from satellite radiance observations through computationally demanding deterministic retrievals based on manually constructed physical models. A statistical alternative to this deterministic method is to train regression models for AOT prediction from radiance data. This approach provides fast retrievals albeit with somewhat reduced accuracy. In this paper, we explore an integrative approach that combines statistical and deterministic algorithms to provide both inexpensive and accurate retrievals. Given a limited set of locations with AOT produced by the deterministic algorithm, and a full set of radiance data, we retrieve AOT at the remaining locations using several distinct statistical algorithms: (1) inverse distance spatial interpolation, (2) global neural networks learned on data from the entire domain, (3) region-specific neural networks, and (4) optimally weighted averaging ensembles of the first three algorithms. The integrated retrieval algorithms are evaluated using AOT and radiances observed by the Multiangle Imaging SpectroRadiometer (MISR) instrument onboard NASA’s Terra satellite during two 16-day periods in 2002 over the continental US. Results show that integration of the deterministic and statistical algorithms provide a range of options for selection of the best trade-off between accuracy and complexity. Moreover, on the statistical side, the best tradeoff between retrieval speed and accuracy was obtained through weighted averaging of global neural networks and spatial interpolation.