A New Approach for Nighttime Sea Fog Detection Based on VIIRS/DNB Data

Nighttime sea fog poses significant challenges to ship navigation and carrier-based aircraft takeoff and landing. The day/night band (DNB) of the visible infrared imager radiometer suite (VIIRS) offers high spatial resolution and excellent night imaging capabilities, providing a new idea for nighttime sea fog detection. In this work, a nighttime sea fog detection dataset (NSFDD) was constructed using VIIRS/DNB data, which spans the period from 2017 to 2023 and covers the offshore waters of China. Considering the characteristics of the NSFDD, we proposed a new method (DA-TransUNet-CRF) by integrating the DA-TransUNet model and the fully connected conditional random fields (dense CRFs). Specifically, the hybrid feature extraction architecture of the DA-TransUNet model helps increase detection probability. Dense CRFs were incorporated into the DA-TransUNet model to address the issue of boundary ambiguity in the outputs. Experiments indicated that the DA-TransUNet-CRF method outperformed the existing detection method (auto-DCD) and three widely employed deep learning models: U-Net, Attention U-Net, and TransUNet. Specifically, the DA-TransUNet-CRF method achieved a probability of detection of 85.59% and a false alarm rate of 1.31%. This method has proven effective in enhancing detection accuracy when missed detections occur in the auto-DCD method due to inappropriate threshold selection. Furthermore, it can increase the detection performance in cases with thin cirrus cloud coverage, which often poses challenges for traditional methods. These findings provide a new reference for nighttime sea fog detection research.