A Deep Neural Network for Achieving Spectrally Consistent and Seamless Infrared Radiance Measurements Across Geostationary Satellite Domains

The NASA Clouds and the Earth's Radiant Energy System (CERES) project provides the scientific community with observed top-of-atmosphere (TOA) shortwave and longwave fluxes for climate monitoring and climate model validation. To achieve this goal, CERES relies on TOA broadband fluxes derived from geostationary satellite (GEO) imagery to account for the diurnal flux variations between the CERES observation intervals. Consistent global flux derivation depends on accurate and consistent cloud retrievals. Scene-dependent spectral measurement inconsistency of the instruments that make up the contiguous ring of GEO observations (GEO-Ring), as well as limb darkening effects, can cause discontinuities in derived cloud properties and radiative fluxes at the boundaries of adjacent imager domains. Although the algorithms utilize radiative transfer models to account for instrument-band-dependent atmospheric correction and viewing zenith angle (VZA) dependency, small discontinuities may persist due to uncertainties inherent to the multiple imager-specific algorithms. Furthermore, while hyperspectral-instrument-based spectral band adjustment factors may effectively account for spectrally induced bias, they are less effective at reducing variance owed to the specific composition of the viewed scene, which is challenging to robustly characterize. As such, this article highlights the use of a deep neural network (DNN) to resolve spectral-and VZA-induced biases between GEO-Ring imagers. The DNN uses available infrared (IR) channels from the GEO instruments, along with viewing and solar illumination geometry, to estimate homogenized, VIIRS-like IR radiances for use in the GEO cloud algorithm. This approach is effective at mitigating scene-dependent spectral variance and VZA dependency, resulting in consistent radiance measurements across the GEO-Ring, thereby leading toward a more seamless global cloud assessment.