Feature selection for CIE standard sky classification

Abstract There are several compilations of sky classifications that refer to Meteorological Indices (MIs) (variables usually recorded at meteorological ground stations), due to the scarcity of sky scanner devices that can supply the experimental data needed to apply the CIE standard sky classification. The use of one rather than another MI is never justified, because there is no standardized criterion for their selection. In this study, forty-three MIs, traditionally used to define different sky conditions, are reviewed. Feature Selection (FS) is a key step in the design of a sky-classification algorithm using MIs as an alternative to data from sky scanners. Four procedural methods for FS -Pearson, Permutation Importance, Recursive Feature Elimination, and Boruta- are applied to an extensive data set of MIs that includes CIE standard sky classification data, which was used as a reference. The use of FS procedures significatively reduced the original set of MIs, permitting the construction of different classification trees with high performance for the sky classification. In the case of the Pearson FS method, the classification tree only used two MIs. The advantage of the Pearson FS method is that it functions independently from the machine-learning algorithm used latter for the sky classification.