Abstract
A fully connected deep neural network (FCDN) clear-sky mask (CSM) algorithm (FCDN_CSM) was developed to assist the FCDN-based Community Radiative Transfer Model (FCDN_CRTM) to reproduce the Visible Infrared Imaging Radiometer Suite (VIIRS) clear-sky radiances in five thermal emission M (TEB/M) bands. The model design was referenced and enhanced from its earlier version (version 1), and was trained and tested in the global ocean clear-sky domain using six dispersion days’ data from 2019 to 2020 as inputs and a modified NOAA Advanced Clear-Sky Processor over Ocean (ACSPO) CSM product as reference labels. The improved FCDN_CSM (version 2) was further enhanced by including daytime data, which was not collected in version 1. The trained model was then employed to predict VIIRS CSM over multiple days in 2020 as an accuracy and stability check. The results were validated against the biases between the sensor observations and CRTM calculations (O-M). The objectives were to (1) enhance FCDN_CSM performance to include daytime analysis, and improve model stability, accuracy, and efficiency; and (2) further understand the model performance based on a combination of the statistics and physical interpretation. According to the analyses of the F-score, the prediction result showed ~96% and ~97% accuracy for day and night, respectively. The type Cloud was the most accurate, followed by Clear-Sky. The O-M mean biases are comparable to the ACSPO CSM for all bands, both day and night. The standard deviations (STD) were slightly degraded in long wave IRs (M14, M15, and M16), mainly due to contamination by a 3% misclassification of the type Cloud, which may require the model to be further fine-tuned to improve prediction accuracy in the future. However, the consistent O-M means and STDs persist throughout the prediction period, suggesting that FCDN_CSM version 2 is robust and does not have significant overfitting. Given its high F-scores, spatial and long-term stability for both day and night, high efficiency, and acceptable O-M means and STDs, FCDN_CSM version 2 is deemed to be ready for use in the FCDN_CRTM.
Highlights
We first summarize the FCDN_CSM v1 and discuss v2 architecture and data preprocessing in detail
The updated Advanced Clear-Sky Processor over Ocean (ACSPO) clear-sky mask (CSM) data was used as reference labels, which ACSPO version 2.4 updated to allow the Community RadiativeTransfer Model (CRTM) simulation to be conducted at the pixel level instead of in the coarse grid, significantly improving the Visible Infrared Imaging Radiometer Suite (VIIRS) observations and CRTM calculations (O-M)
Because the input data cover all seasons and more features are used for daytime, a more complex fully connected deep neural network (FCDN) model is required to achieve adequate learning of the CSM texture during the model training
Summary
Statistical, nonlinear approximation instead of a complicated physicalbased model in ANNs results in a more computationally efficient method to achieve a similar result to those of the physical-based model, without significant loss of accuracy [16,17,18] In recent years these advantages have attracted an increasing number of remote sensing scientists to explore AI-based CSM algorithms [14,15,21] by designing various machine learning architectures, such as random forest (RF), support vector machine, ANN, and convolution neural network; training the models by using selected effective sensor measurements and geophysical conditions together with atmosphere and surface ancillary data; and predicting CSM with the well-trained model.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
