Deriving boundary layer height from aerosol lidar using machine learning: KABL and ADABL algorithms

Abstract

The atmospheric boundary layer height (BLH) is a key parameter for many meteorological applications, including air quality forecasts. Several algorithms have been proposed to automatically estimate BLH from lidar backscatter profiles. However recent advances in computing have enabled new approaches using machine learning that are seemingly well suited to this problem. Machine learning can handle complex classification problems and can be trained by a human expert. This paper describes and compares two machine-learning methods, the K-means unsupervised algorithm and the AdaBoost supervised algorithm, to derive BLH from lidar backscatter profiles. The K-means for Atmospheric Boundary Layer (KABL) and AdaBoost for Atmospheric Boundary Layer (ADABL) algorithm codes used in this study are free and open source. Both methods were compared to reference BLHs derived from colocated radiosonde data over a 2-year period (2017–2018) at two Météo-France operational network sites (Trappes and Brest). A large discrepancy between the root-mean-square error (RMSE) and correlation with radiosondes was observed between the two sites. At the Trappes site, KABL and ADABL outperformed the manufacturer's algorithm, while the performance was clearly reversed at the Brest site. We conclude that ADABL is a promising algorithm (RMSE of 550 m at Trappes, 800 m for manufacturer) but has training issues that need to be resolved; KABL has a lower performance (RMSE of 800 m at Trappes) than ADABL but is much more versatile.