A New Application of Unsupervised Learning to Nighttime Sea Fog Detection

Area of Science:

• Atmospheric Science
• Remote Sensing
• Artificial Intelligence

Background:

• Traditional sea fog detection algorithms struggle to differentiate fog from stratus clouds due to similar characteristics.
• Existing methods rely on predetermined thresholds derived from climatology or simulations, limiting adaptability.

Purpose of the Study:

• To develop a novel nighttime sea fog detection algorithm using unsupervised learning.
• To improve the accuracy and adaptability of sea fog detection by optimizing thresholds for specific domains.

Main Methods:

• Utilized brightness temperatures from COMS satellite (3.7 μm and 10.8 μm channels) and OSTIA sea surface temperature data.
• Implemented the expectation-maximization (EM) algorithm, an unsupervised learning technique, for data optimization and feature identification.
• Evaluated algorithm performance using Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) vertical profile products.

Main Results:

• The unsupervised learning approach enabled the application of optimal, domain-specific thresholds for fog detection.
• The algorithm demonstrated promising results in a local domain.
• Achieved a Probability of Detection (POD) of 0.753 and a Critical Success Index (CSI) of 0.477.

Conclusions:

• The developed algorithm effectively detects nighttime sea fog, offering improved differentiation from stratus clouds.
• Unsupervised learning provides a more robust method for interpreting limited data in remote sensing applications.
• This approach enhances the reliability of sea fog detection for meteorological applications.