Improving the Uncertainty Estimation Performance of Retrieving Surface Reflectances from Remote Sensing Imaging Spectroscopy Data

The current and next generation imaging spectroscopy missions such as AVIRIS-NG, EnMAP, CHIME, and SBG measure radiances over a wide range of spectral bands. T hese measurements can be used to infer about surface properties, and t his infere n ce depends on accurate inversion of surface reflectances and their uncertainties. Optimal estimation (OE), which amounts to performing maximum a posteriori estimation using Gaussian statistics and a linearized forward model, has recently been demonstrated to perform this task well in many settings. Despite this, some issues remain: in particular, a single retrieval with OE, as currently implemented in the ISOFIT software, takes o n the order of one CPU core second, varying depending on the exact setup . While this demand is not prohibitive for a single retrieval, future instruments will be capable of obtaining even hundreds of thousands of spectra per second, with which the implied computational demands become very large. Furthermore, due to the approximations inherent in OE, the obtained uncertainties may sometimes unde r estimate true uncertainties. We explore alternative algorithms to obtain and describe those uncertainties, and demonstrate that it is possible to reliably retrieve very good surface reflectance estimates while reducing the computational load by as much as two to three orders of magnitude.