Medium-to-large fluctuations and coherent structures (mlf-cs's) can be observed using horizontal scans from single Doppler lidar or radar systems. Despite the ability to detect the structures visually on the images, this method would be time-consuming on large datasets, thus limiting the possibilities to perform studies of the structures properties over more than a few days. In order to overcome this problem, an automated classification method was developed, based on the observations recorded by a scanning Doppler lidar (Leosphere WLS100) installed atop a 75 m tower in Paris's city centre (France) during a 2-month campaign (September–October 2014). The mlf-cs's of the radial wind speed are estimated using the velocity–azimuth display method over 4577 quasi-horizontal scans. Three structure types were identified by visual examination of the wind fields: unaligned thermals, rolls and streaks. A learning ensemble of 150 mlf-cs patterns was classified manually relying on in situ and satellite data. The differences between the three types of structures were highlighted by enhancing the contrast of the images and computing four texture parameters (correlation, contrast, homogeneity and energy) that were provided to the supervised machine-learning algorithm, namely the quadratic discriminant analysis. The algorithm was able to classify successfully about 91 % of the cases based solely on the texture analysis parameters. The algorithm performed best for the streak structures with a classification error equivalent to 3.3 %. The trained algorithm applied to the whole scan ensemble detected structures on 54 % of the scans, among which 34 % were coherent structures (rolls and streaks).