Preliminary Report on Deep Learning-based Daytime Clear-Sky Radiance for VIIRS

Abstract

A fully connected “deep” neural network algorithm with the Community Radiative Transfer Model (FCDN_CRTM) is proposed 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. The model was originally trained and tested in global ocean clear-sky domain for nighttime, and were modified and applied for daytime data in this study. CRTM-simulated brightness temperatures (BTs) were defined as model labels and the clear-sky pixels were identified by an FCDN-trained clear-sky mask (FCDN_CSM) model. The preliminary result showed that the FCDN_CRTM prediction minus CRTM simulation (F-C) mean biases are only up to several tens mK for all bands. However, the corresponding standard deviation (STDs) were 4–5 times worse than training and testing data, due to the effect of the daytime solar reflection. The evaluation result suggested that further fine-tune model is needed to improve the daytime prediction accuracies, by improving input data uniformity, adjusting the model architecture, and selecting possible important features.