Hyperspectral Infrared Sounder Cloud Detection Using Deep Neural Network Model

Abstract

Detection of cloud contaminated field of views (FOV) from satellite hyperspectral infrared sounders is essential for numerical weather prediction. A new cloud detection model is developed for the cross-track infrared sounder (CrIS) using the artificial deep neural network (DNN) technique. The truth cloud information used is from another instrument of Visible Infrared Imaging Radiometer Suite (VIIRS) deployed on the same platform of CrIS. The training data set is built from CrIS-VIIRS collocated measurements randomly selected from different months to represent different atmospheric and surface conditions. Then, we use the VIIRS cloud mask collocated within the CrIS footprint to train the CrIS spectra for cloud detection. Specifically, the CrIS spectra were transformed into their principal components (PCs), with only the top 75 PCs used as the predictors rather than the entire CrIS channels, for the purpose of better regression and convergence during the training process and faster prediction. Results were examined globally by the considered truth derived from the VIIRS cloud mask. Generally, the spatial distribution of the proposed CrIS cloud detection result agrees with that from the VIIRS, with a high model accuracy of 93%. Further analysis indicates that the proposed CrIS cloud detection result is slightly better over daytime than nighttime with the accuracy values of 94% versus 91%. The ocean areas have a higher cloud detection accuracy than continental land with accuracy values of 95% versus 88%. In addition, sometimes the DNN model would recognize the thin cloud as clear sky, as their spectra are very similar to each other. False detected pixels are also found over snow- or ice-covered and desert areas. This is possibly due to the VIIRS cloud mask that has a relatively low accuracy over these areas.