Multi-source Data Fusion Network for Sea Fog Detection

Sea fog seriously affects the safety of maritime activities. This paper presents the development of a dataset for sea fog detection research and a multi-source data fusion sea fog detection network called MDF-SFNet. We collected all sea fog events observed in the Yellow Sea and Bohai Sea regions (118.1° E-128.1° E, 29.5° N-43.8° N) from March 2016 to December 2022. The Himawari-8 satellite data and sea surface temperature data are from the National Meteorological Center of China. This portion of the dataset is then accurately labeled to produce the final dataset. In addition, this paper proposes a multi-source data fusion sea fog detection model. This model not only takes advantage of Transformers to extract global input information, but also considers the traditional method of sea fog detection by meteorological experts based on physical characteristics and threshold methods. In the model, we introduce a gate module named Gate for adaptive threshold selection, an FFN module for multi-source data fusion, and an encoder module to facilitate dense prediction. MDF-SFNet extracts features from visual and meteorological physical properties, and continuously fuses the two features within the model. Experimental results show that our model achieves superior sea fog detection performance, with an F1 score of 0.742, a POD (probability of detection) of 0.753, and a CSI (critical success index) of 0.576. MDF-SFNet shows improved detection performance compared to existing state-of-the-art deep learning networks for sea fog detection.