Multipixel Retrieval of Structural and Optical Parameters in a 2-D Scene With a Path-Recycling Monte Carlo Forward Model and a New Bayesian Inference Engine

Physics-based retrievals of atmosphere and/or surface properties are generally multi- or hyperspectral in nature; some use multi-angle information as well. Recently, polarization has been added to the available input from sensors and accordingly modeled with vector radiative transfer (RT). At any rate, a single pixel is processed at a time using a forward RT model predicated on 1-D transport theory. Neighboring pixels are sometimes considered but, generally, just to formulate statistical constraints on the inversion based on spatial context. Herein, we demonstrate the power to be harnessed by adding bona fide multipixel techniques to the mix. We use a forward RT model in 2-D, sufficient for this demonstration and easily extended to 3-D, for the response of a single-wavelength imaging sensor. The data, an image, is used to infer position, size, and opacity of an absorbing atmospheric plume somewhere in a deep valley in the presence of partially known/partially unknown aerosol. We first describe the necessary innovation to speed-up forward multidimensional RT. In spite of its reputation for inefficiency, we use a Monte Carlo (MC) technique. However, the adopted scheme is highly accelerated without loss of accuracy by using “recycled” MC paths. This forward model is then put to work in a novel Bayesian inversion adapted to this kind of RT model where it is straightforward to trade precision and efficiency. Retrievals target the plume properties and the specific amount of aerosol. In spite of the limited number of pixels and low signal-to-noise ratio, there is added value for certain nuclear treaty verification applications.