African Dust Property Retrieval and Cloud Detection from Hyperspectral Infrared Radiances using Optimal Estimation

Proper characterization of aerosol loading can correct systematic errors in radiance assimilation and increases the accuracy of numerical weather prediction solutions. In particular, airborne mineral dust directly influences forecast models by inducing errors in satellite retrieved temperature and water vapor profiles, with sensors incorrectly associating cooler aerosol layers as characteristic of the surface. Despite this, aerosols remain one of the most poorly constrained constituents of the environmental system. Limitations in our ability to predict and represent discrete sources and sinks necessitates daily, global observations. Polar orbiting satellites are currently the best operational platform for global observations, but have a major issue: half their the orbital tracks become useless for retrievals reliant upon the visible spectrum. We have developed an aerosol property retrieval in the infrared leveraging the current generation of hyperspectral radiometers.

Our algorithm uses spectrally resolved nadir infrared radiance data to determine optical depth, effective radius, and mineral family of dust layers and quantify information content within hyperspectral sounder signal. Successful development of which could result in the effective doubling of polar orbiter observations of dust by including the nighttime scans. This inverse modeling framework primarily relies upon three components: 1. the observed radiances from Scanning High-resolution Interferometer Sounder (S-HIS), 2. the Line-by-Line Radiative Transfer Model with Discrete Ordinances (LBL-DIS) to act as the forward model, estimating radiances for potential atmospheric states, and 3. optimal estimation (OE) to drive the state towards the observation. This is a Bayesian system, producing the output described above, as well as quantitative uncertainty estimates.

Three spectrally orthogonal mineral dust types and one liquid water habit were incorporated into the forward model, representing the most probable sources of interference within window channels. The optimal estimation process attempts to find a state space (adjusting optical thickness and effective radius) combination of these layers that most accurately produces the observation spectrum. Unfortunately, the problem is ill-posed, as there are multiple combinations that can fall within the acceptable convergence criteria. We found the potential for an incorrect convergence is reduced the most through proper constraint of layer geometry (height and thickness). Since our method relies on observations diverging from expected blackbody emissions within window channels, one would expect surface temperature error to be pivotal. However, errors in retrieved optical depths above 0.15 were insignificant within a +/- 2 K surface emission window. This works above an otherwise clear surface, but clouds create a problem. In their presence, which dust often is, it becomes impossible for the optimal estimation framework to appropriately attribute radiance changes to dust without an accurate estimation of the cloud top temperature. Instead of attempting a dust retrieval in cloud contaminated scenes, the OE process effectively screens them out when the water cloud state is sufficiently convergent.

The training dataset come from the Hurricane Storm Sentinel (HS3) mission, which provided a low altitude (in comparison to LEO; 20 km vs 700 km), hyperspectral radiances from S-HIS, as well as collocated retrievals from the Cloud Physics Lidar (CPL). We will evaluate the retrieval efficacy for data from the 2013 deployment. Initial results found that after screening cloud contaminated scenes, residuals between S-HIS and CPL retrieved aerosol optical depths were distributed around -0.07, with S-HIS retrievals having a low bias when compared to CPL retrievals greater than 0.15.