Speeding up probabilistic imaging spectroscopy retrievals with Accelerated Optimal Estimation

Several upcoming imaging spectroscopy missions will produce enormous quantities of high-dimensional radiance data later this decade. In order to use these data for inferring surface properties, reflectance spectra need to be derived from the radiances. Recent work shows that probabilistic inversion methods, such as Optimal Estimation, are capable of solving the problem and producing high-quality reflectance estimates while at the same time performing atmospheric correction. However, these methods tend to be too slow for the quantities of data that we are expecting. Accelerating these probabilistic surface reflectance retrievals is therefore important for getting the most out of these future missions. In this presentation we describe Accelerated Optimal Estimation (AOE) for retrieving surface reflectances from imaging spectroscopy data. The AOE method solves the standard Optimal Estimation (OE) problem by decomposing the optimization space into two components: a high- and a low-dimensional subspace. In the high-dimensional part we can efficiently construct an accurate Gaussian approximation of the conditional posterior probability, while the low-dimensional part retains most of the non-linearity of the problem. We demonstrate with AVIRIS-NG data, that AOE speeds up the OE retrieval significantly compared to a reference OE implementation. The AOE method also produces better-converged estimates of the maximum a posterior states when we apply it to an AVIRIS-NG scene with 160,000 pixels. We validate the posterior distributions using Markov Chain Monte Carlo methods.