Assimilation of Hyperspectral Infrared Observations with Optimal Spectral Sampling


Current data assimilation (DA) systems use only a small fraction of the thousands of channels from hyperspectral infrared sounding instruments, including IASI, AIRS, and CrIS. An alternative that retains nearly the full hyperspectral information content at greatly reduced computation cost, while filtering instrument noise, is to assimilate the optimal spectral sampling (OSS) node radiances. The nodes are a set of optimally-selected spectral points, the node radiances are monochromatically calculated at the nodes using a fast forward model, and the nodes and weights are determined by a training procedure that ensures channel radiances can be accurately computed as a weighted sum of node radiances. For DA, we calculate the node observed radiances by inverting the OSS node-to-channel relationship in a least-squares sense. DA procedures can then be applied directly to the node radiances after some practical concerns are mitigated.  The node-based approach has a quality-control advantage where emission from low clouds may contaminate channel radiances even in the stronger absorption bands because channels are wider than absorption lines, while there are uncontaminated monochromatic node radiances that have useful sensitivity to atmospheric column temperature and humidity closer to the cloud tops. Four preliminary, proof-of-concept global observing system experiments were conducted with the NOAA three-dimensional variational data assimilation system: a baseline that uses the 150 operational IASI channels; an experiment that uses a large portion (5014) of the IASI band 1 and 2 channels; and two experiments that assimilated 256 IASI-OSS node radiances with different specifications of observational errors. One of those two node-based experiments performed better than the other experiments for most metrics of forecast skill. The results of these preliminary experiments are encouraging: the node-based approach provided positive impacts on the forecast for some metrics, regions, and lead times, despite the fact that there are aspects of the node-based processing that have not yet been optimized.

Resource LinkAssimilation of Hyperspectral Infrared Observations with Optimal Spectral Sampling