Retrieval of cloud-cleared atmospheric temperature profiles from hyperspectral infrared and microwave observations

Abstract

This thesis addresses the problem of retrieving the temperature profile of the Earth's atmosphere from overhead infrared and microwave observations of spectral radiance in cloudy conditions. The contributions of the thesis are twofold: improvements in 1) microwave instrumentation and 2) hyperspectral signal processing and estimation algorithms. The NPOESS Aircraft Sounder Testbed-Microwave (NAST-M) passive spectrometer was designed, fabricated and deployed. NAST-M provides accurate brightness temperature measurements in 16 channels near the oxygen absorption lines at 50-57 GHz and 118.75 GHz, permitting the first reliably accurate retrieval images of temperature profiles and precipitation structure in cloudy areas. The correlation structure of the NPOESS Aircraft Sounder Testbed-Infrared (NAST-I) instrument noise was analyzed in the spectral and spatial domains using the Iterated Order-Noise (ION) algorithm [1] for two representative flights. Results indicate that vibration-induced noise was the dominant component, but that it could be significantly reduced by filtering in the spatial domain. Novel multi-pixel cloud clearing and temperature profile retrieval algorithms were developed for simulated Atmospheric Infrared Sounder (AIRS) and Advanced Microwave Sounding Unit (AMSU) radiances using neural networks. RMS temperature profile retrieval errors of -0.5 K were obtained for all levels of the atmosphere from 0-15 km in clear air at a horizontal resolution of 2000 km2 and a vertical resolution of 1 km. RMS radiance errors under cloudy conditions for altitudes from 0 to 10 kilometers ranged from 1.25 K to 0.1 K for radiance retrievals near 15 microns, and from 0.8 K to 0.05 K for radiance retrievals near 4 microns.

(cont.) Validation of the simulation results with NAST observations was hampered by the lack of a statistically-diverse data set accompanied by cloud truth. An upper bound on cloud-clearing performance (NEAT) was estimated to be approximately a factor of two worse than the simulation results accompanied by ground truth. An improvement of approximately 25 percent in RMS radiance cloud-clearing performance was realized by rejecting 20 percent of soundings based on a neural network-derived metric.