Ocean Fog Detection using Himawari-8 data over the Yellow sea with Machine Learning Approaches

Abstract

Ocean fog (OF) is a phenomenon in which the visibility distance is less than 1km over the ocean due to the droplets. OF has a role not only as a moisture source, for the plants when it enters the land, also as an obstacle to maritime traffic. Many harbors set up fog detectors on the land to monitor OF occurrence near their port, but it covers a limited area. Recently, satellite remote sensing which covers wider area was usually applied on this criterion, but it is hard to identify the OF condition because of the complexity of generation condition and optical-thermal properties. Thus, in this study, machine learning approaches (e.g., random forest, support vector machine, logistic regression) were used to observe OF occurrence. As spatial coverage, temporal coverage is also important for maritime traffic, so the geostationary satellite (i.e., Himawari-8) data were used. The study area is the Yellow sea, which is suffering from OFs frequently. The Cloud-Aerosol Lidar with Orthogonal Polarization data were used to get OF location as follows Wu et. al. (2015). For the temporal seamless monitoring, infrared channels 0of Himawari-8 were used as the input images. From the input images, not only thermal feature such as mean, maximum and minimum, but also spatio-temporal feature such as roughness of buffer area, temporal anomaly. Additional post processing was applied to check the reliability of each OF pixels.