Neural network estimation of atmospheric profiles using AIRS/AMSU observations: Improved uncertainty assessments

Abstract

Neural networks are developed for estimating the rms accuracy profiles of individual infrared and microwave atmospheric temperature and humidity profile retrievals, thus potentially significantly improving their assimilation into numerical weather prediction models. Currently most assimilation processes compute retrieval variances or error-covariance matrices as ensemble averages over diverse profiles, or simply flag problematic retrievals. Although retrieval accuracies vary considerable from profile to profile because of clouds, even in cloud-free cases they can differ markedly. The ability to estimate accurately the variances of individual profiles is one of the benefits of hyperspectral infrared and microwave sounding. The variance-estimating neural network was trained to estimate the logarithm of variance, which was then mapped to standard deviation. Examples utilizing AIRS/AMSU/HSB soundings [1] on the NASA Aqua satellite and those from a proposed hyperspectral microwave sounder [2], [3] show that when the predicted rms errors for a single altitude are stratified, they agree with the actual rms errors within perhaps ten percent of the dynamic range of the stratifications thus significantly improving the potential for accurately weighting soundings against model parameters during assimilation. Simple quality indicators using the new variance estimates also favorably compare to AIRS Level 2 Version 5 quality flags.