Abstract Coherent structures such as streaks are essential components of the surface layer dynamics, impacting turbulent fluxes and pollutants' dispersion. These structures can be observed using horizontal scans from a Doppler lidar after extracting the turbulent component of the radial wind. Almost 40,000 scans of this type were recorded in Dunkerque (French coast on the North Sea) for 13 months, from June 2021 to July 2022. An automated classification method was developed to detect the two levels of streak coherence that were identified by eye on the lidar images, called organized and disorganized streaks. For each image, a set of textural numerical features was generated, which reduced the data set dimensionality and allowed to proceed with a reduced training set of only 400 images. Four supervised learning classification algorithms were trained, and the Quadratic discriminant analysis provided the lowest classification error, estimated at 5.2% using the block cross‐validation method. The algorithm detected streaks on 61% of the data set and successfully distinguished the two types of streaks. While the wind direction (offshore or onshore) had no direct impact, a minimum level of mechanical turbulence was required to allow streaks to form (wind speed or friction velocity ) and the streaks' level of coherence increases from disorganized to organized when the atmosphere becomes convective (upward turbulent heat flux ). The turbulence organization thus answers to the atmospheric stability in opposite ways in the surface layer and the mixed layer above and depends on seasonality.
Maynard et al. (Mon,) studied this question.