Generalized radiative transfer emulation for imaging spectroscopy reflectance retrievals

Derivations of surface reflectance from observed radiance are fundamentallytied to the accuracy of the radiative transfer models used to simulate the interaction of light with the atmosphere and surface. These radiative transfer models are parameterized bya wide range of quantities, ranging from observation and solar geometries to assumptions about atmospheric conditions, including vertical distributions of gasesand temperatures as well as the atmospheric composition. The number of these parameters and the highly nonlinear nature of radiative transfer means that the common retrieval approach of building a coarsely-spaced linear look up table (LUT) grid of just a few parameters must necessarily compromise accuracy for computational feasibility. Here, we propose a new method called sRTMnet that facilitates the efficient creation of dense LUT grids that fully capture radiative transfer’s inherent nonlinearity. sRTMnet uses a combination of fast, reduced-order radiative transfer modeling coupled with machine-learning-based emulation to realize a factor of $\sim$1800 reduction in computation time while maintaining the accuracy of a high-fidelity RTM. We demonstrate the accuracy of sRTMnet in multiple ways. We show high accuracies of the machine-learning emulation using robust testing sets. We then use acquisitions from the airborne visible/infrared imaging spectrometer - next generation (AVIRIS-NG) to show near identical surface reflectance estimates with sRTMnet and a high-fidelity radiative transfer model (MODTRAN). Finally, we show how distributions of mapped minerals, in the style of the upcoming earth surface mineral dust source investigation (EMIT), remain consistent when using the sRTMnet-based reflectance. In addition to speed and accuracy, the fully open-source nature of sRTMnet (including simulation and emulation), provides a pathway for high-resolution imaging spectroscopy reflectance retrievals at the global scale.