Classification of Continuous Sky Brightness Data Using Random Forest

Abstract

Sky brightness measuring and monitoring are required to mitigate the negative effect of light pollution as a byproduct of modern civilization. Good handling of a pile of sky brightness data includes evaluation and classification of the data according to its quality and characteristics such that further analysis and inference can be conducted properly. This study aims to develop a classification model based on Random Forest algorithm and to evaluate its performance. Using sky brightness data from 1250 nights with minute temporal resolution acquired at eight different stations in Indonesia, datasets consisting of 15 features were created to train and test the model. Those features were extracted from the observation time, the global statistics of nightly sky brightness, or the light curve characteristics. Among those features, 10 are considered to be the most important for the classification task. The model was trained to classify the data into six classes (1: peculiar data, 2: overcast, 3: cloudy, 4: clear, 5: moonlit-cloudy, and 6: moonlit-clear) and then tested to achieve high accuracy (92%) and scores (F-score = 84% and G-mean = 84%). Some misclassifications exist, but the classification results are considerably good as indicated by posterior distributions of the sky brightness as a function of classes. Data classified as class-4 have sharp distribution with typical full width at half maximum of 1.5 mag/arcsec2, while distributions of class-2 and -3 are left skewed with the latter having lighter tail. Due to the moonlight, distributions of class-5 and -6 data are more smeared or have larger spread. These results demonstrate that the established classification model is reasonably good and consistent.