Detection of dawn sea fog/low stratus using geostationary satellite imagery

Traditional satellite-based detection of dawn sea fog/low stratus (SFLS) is difficult because of the weak reflec-tivity in the visible at low solar elevation angles and the contamination of the reflected sunlight in the mid-infrared. Here, based on single geostationary satellite measurements acquired by China's Fengyun 4A (FY-4A), we propose a dawn SFLS detection algorithm using the joint Fully Convolutional Network and Conditional Random Field (FCN-CRF), which are well known for image semantic segmentation under low contrast conditions. We train the FCN-CRF detection algorithm using FY-4A measurements over the Yellow Sea, where some dawn SFLS events are long-lived, providing relatively time-invariant dawn SFLS samples for training. We design a SFLS labelling technique using the satellite observations before and after dawn to train the FCN-CRF detection for dawn SFLS. A test against buoy visibility observations shows that the FCN-CRF detection is able to detect dawn SFLS with satisfactory accuracy, with a probability of detection (POD) of 84.9%, a false alarm ratio (FAR) of 8.7%, a critical success index (CSI) of 78.5% and a hit rate score (HR) of 87.4%.