CzSL: Learning from citizen science, experts and unlabelled data in astronomical image classification

Abstract Citizen science is gaining popularity as a valuable tool for labelling large collections of astronomical images by the general public. This often achieved at cost poorer quality classifications made amateur participants, which are usually verified employing smaller data sets labelled professional astronomers. Despite its success, citizen alone will not be able to handle classification current and upcoming surveys. To alleviate this issue, projects have been coupled with machine learning techniques in pursuit more robust automated classification. However, existing approaches neglected fact that, apart from amateurs, (limited) expert knowledge problem also available along vast amounts unlabelled that yet exploited within unified framework. paper presents an innovative methodology capable taking advantage expert- amateur-labelled data, featuring transfer labels between experts amateurs. The proposed approach first learns convolutional autoencoder then exploits via pre-training fine-tuning neural network, respectively. We focus on galaxy Galaxy Zoo project, we test binary, multi-class, imbalanced scenarios. results demonstrate our solution improve performance compared set baseline approaches, deploying promising different confidence levels labelling.