Development of a mathematical model for the determination of the atmospheric boundary layer height using artificial intelligence

The Atmospheric Boundary Layer Height (ABLH) is a fundamental parameter for many meteorological applications and climate change assessment and evaluation. A large number of methods for ABLH determination have been proposed; however, there is no sufficiently reliable and feasible method for this purpose. The rise of intelligence-based mathematical models for feature determination in data space has allowed their application to solve problems similar to ABLH determination. This article describes the development of a mathematical model based on artificial intelligence for ABLH determination, in which an introductory analysis of the data space from the point of view of machine learning, unsupervised, and supervised methods is presented. The methods explored are the mountain method, subtractive clustering, and the classic K-means and its soft counterpart, Fuzzy C-means. Furthermore, an analysis was conducted to determine which similarity function—whether Euclidean, Manhattan, Mahalanobis, or Cosine—best fits for ABLH estimation in each unsupervised method. For classification in a supervised fashion, the best suitable models, among others, are support vector machines and decision trees. Different internal metrics (Silhouette Index, Calinski-Harabasz score) and external metrics (root mean square error and adjusted Rand score), with a reference made by means of visual inspection by an expert, were used to evaluate the methods. The unsupervised mountain method with the Manhattan similarity function proved to be the most feasible, as it is a non-stochastic method, its computation time is reduced, and it does not require an ABLH reference. The data used was extracted from several sources: 83 days of quasi-continuous LIDAR data with 23,000 data points located at Brest, France, measured with a MiniMPL from the Meteo France LIDAR network, were used. An ABLH reference from a radiosonde adjacent to the site of the LIDAR system was used. The references range from October to December 2018. The root mean square error achieved for the whole dataset was 600 m for the mountain method. The presented method is shown to be effective for various atmospheric situations, regardless of their complexity.