Abstract
Sea fog significantly reduces visibility and poses serious risks to maritime activities in the Arctic Ocean. Convolutional neural networks, such as fully convolutional networks (FCN), are widely used for sea fog detection due to their ability to capture textural features from satellite imagery. However, their practical performance remains constrained by the lack of objective and efficient labeling strategies capable of producing region-level labels that encode the spatial continuity of fog patches. We propose a square expansion labeling method designed for convolution-based architectures, which exploits the spatial homogeneity of sea fog to transform one-dimensional CALIPSO observations into two-dimensional labels. Using this method, we develop a comprehensive dataset (2010–2018), where labels are derived from Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) classification data, and corresponding multispectral inputs are obtained from Moderate Resolution Imaging Spectroradiometer (MODIS) imagery. Based on these objectively generated labels, the FCN-based model achieves a Probability of Detection (POD) of 0.935 and an overall Critical Success Index (CSI) of 0.801 on the test set, thereby demonstrating the feasibility and effectiveness of the proposed labeling strategy for convolutional neural network–based Arctic sea fog detection.

This publication has 14 references indexed in Scilit: