Machine Learning Methods Applied to All-Sky Cloud Coverage Estimation

Recent developments in sensors and algorithms have allowed an automatic acquisition of atmospheric data both from satellites and ground stations. Within this context, the cloud monitoring becomes important for understanding the cloud formation process, and consequently, for modelling the weather in general. For this study we explore several machine learning (ML) models and their ability to classify the all-sky images according to the level of cloud coverage (CC). A supervised validation was employed on the daytime images of one station of the Meteorites Orbits Reconstruction by Optical Imaging (MOROI) network (Nedelcu et al, RoAJ 2018), obtained during a period of 2 years. Thus, from each image a set of photometric-based features were extracted, and the image was labelled to fit one of three categories (i.e <20% CC, 20-80% CC, >80% CC). Next, a set of ML models were selected according to their ability to work and classify tabular data. This was followed by testing the importance of each feature for improving the score. As a final step, a hyper-parameter tuning was applied, in order to obtain the best configuration of each ML method. We obtained a best score of 89% accuracy for 3 label CC classification. Among the 84 features computed, 16 were found relevant for improving the score. The feature extraction is independent of sensor specifications, hence, the process can be scaled and applied to multiple all-sky camera configurations. The results entail future studies covering the MOROI network, which are already collecting data in real-time.