Neural network temperature and moisture retrieval technique for real-time processing of FengYun-4B/GIIRS hyperspectral radiances

Abstract

A fast neural network technique for retrieving vertical profiles of atmospheric temperature and water vapor from the hyperspectral infrared instrument in all-sky conditions is proposed in this study. This technique inherits from the piecewise-defined neural network (PDNN) algorithm that is presently employed operationally for the FengYun-3E Vertical Atmospheric Sounding System (VASS). A major difference from the VASS sounding is the absence of microwave observation. Thus, a new cloud-impact classification method independent of microwave radiance is developed. Additionally, the numerical weather prediction (NWP) forecast temperature can be used as the input to help obtain profile information under cloud. Validation results demonstrate that this new methodology yields higher retrieval accuracy compared to the dual-regression (DR) method currently utilized in the Geostationary Interferometric Infrared Sounder/FengYun-4B (GIIRS/FY-4B) sounding system. Improvement in retrieval accuracy can be primarily attributed to three factors: (1) the cloud-impact classification process effectively mitigates the nonlinear dependence of spectral radiance on atmospheric variables; (2) the potential influence of spectral and radiometric calibration errors of GIIRS on retrieval is minimized by employing actual GIIRS observations for network training; and (3) the incorporation of prior temperature information from forecast models. This novel approach will be used to produce the operational temperature and humidity profile products from FY4B/GIIRS.