Turbulent structures, particularly coherent ones like streaks, significantly influence turbulent fluxes and the dispersion of pollutants in the surface layer. These structures can be observed directly on the horizontal scans of a Doppler lidar, performed during a 13-month campaign in Dunkirk, France, an industrial city on the North Sea shore. About forty thousand quasi-horizontal scans were recorded, capturing two main types of coherent structures: organized and disorganized streaks (a third category named “others” accounting for the absence of structures). An automated classification method was developed for classifying the large dataset of lidar images. The images were pre-processed, notably for retrieving the turbulent part of the wind 1. Each image was represented by a vector of features designed to highlight the streak patterns and computed in 3 steps: (1) Gray-level Co-occurrence Matrices, (2) texture parameters such as contrast, homogeneity, correlation, or energy 2, and (3) “curve parameters” 1. A training set consisting of four hundred scans was built to train supervised learning classification algorithms. The Quadratic Discriminant Analysis classifier has the best performance among the four classification algorithms proved; it classified the training set with a cross-validation error of only 5.2 %. All four classifiers successfully discriminated the categories, with about 61% of the scans classified as coherent structures, including about 31% of organized streaks and 30% of disorganized streaks.
Maynard et al. (Thu,) studied this question.