A comprehensive study of different feature selection methods and machine-learning techniques for SODAR structure classification

Abstract

Listen

Translate article

Sonic detection and ranging (SODAR) echogram provides atmospheric boundary layer (ABL) structure and its height. In the present research, highly accurate and reliable machine-learning methods have been derived and successfully applied on the SODAR echogram for the ABL structure. Five feature selection methods and eight classification methods have been studied in terms of their performance. 133 statistical features have been calculated from 1698 SODAR echogram images. To ensure the unbiased estimation of different structures, machine-learning methods have been used. Furthermore, ten cross-validations have been used to find accuracy. It is found that the boosted tree classifier (overall prediction performance 52.02%) has the highest prognostic presentation with 133 features. After application of the Laplacian method for feature selection, the classifier (overall prediction performance 62.19%) has the highest prognostic presentation with 20 features. The large variability analysis indicates the choice of a classification method for performance variation. Identification of optimal machine-learning methods for SODAR echogram applications is a crucial step towards the ABL structure application, providing an automatic structure classification method for atmospheric and pollutants studies.