Remote Measurement of Active Whitecaps Using Deep Learning Journal of Atmospheric and Oceanic Technology

Whitecaps generated by wave breaking and air entrainment can be classified as active (stage A) or residual (stage B). Measurement of each stage individually is essential for accurate parameterization of air–sea interaction processes, but conventional methods used for separation in visible images are subjective. In this study, this problem is solved using a pipeline for active whitecap fraction measurement. In this pipeline, a new horizon detection method is developed to stabilize and rectify images, and a deep learning model based on U-Net is trained and validated to identify and extract active whitecaps. The model demonstrates robust prediction accuracy even when images are contaminated by sun glint. The model is applied to 48 h of video footage collected during a cruise in Gulf of Mexico. It is determined that, as a function of wind speed, the active whitecap fraction has significant variability and disparity compared to previous research. This finding indicates that secondary factors should be considered for accurate whitecap parameterization. This is explored using a random forest, which indicates that sea surface temperature, swell, and wave age are important to the active whitecap fraction. The precise impact of sea surface temperature is further explored using analyses of variance (ANOVA), which suggest it has a positive correlation with the active whitecap fraction.