Exploring Potential Improvement of Daytime Clear-Sky Radiance for VIIRS by Using Deep Learning Approach

Abstract

AbstractEmulating daytime top-of-atmosphere clear-sky radiances in mid-wave infrared bands (MWIR) with a radiative transfer model is more challenging than nighttime, as the solar reflection is an ineligible component in calculating radiances. A deep learning-based radiative transfer forward model (FCDN_CRTM) was developed to explore the efficiency and accuracy of reproducing the Visible Infrared Imaging Radiometer Suite (VIIRS) clear-sky radiances in five thermal emission M (TEB/M) bands for daytime data by using the community radiative transfer model (CRTM)-simulated brightness temperatures (BTs) as reference labels. Early multi-experiments in the global ocean clear-sky domain for nighttime data indicated that the prediction accuracies using FCDN_CRTM are comparable with CRTM simulation, but the speed to predict one day of VIIRS BTs was more than 50 times fast than CRTM simulation. This study extended the model to include daytime data prediction. Using the same model architecture as nighttime data to predict the TOA radiances for daytime showed that the FCDN_CRTM prediction minus CRTM simulation (F-C) mean biases are within several hundredths of a Kelvin and comparable with the nighttime for all bands. However, the standard deviations (STDs) were 5–10 times worse than training and testing data. This study explores the potential improvement of FCDN_CRTM for daytime data prediction by comparing the data internal physical characteristics and distribution between day and night, particularly for MWIRs.KeywordsFully Connected “deep” Neural Network (FCDN)Solar reflectionCRTMDeep learningMid-Infrared Window Bands (MWIR)