Using deep learning to identify asteroid trails on ESA's Hubble data archive
2020-11-08, 08:40–08:40, Posters

The ESAC Science Data Centre is conducting Machine Learning experiments to provide AI-supported added value services to our users in the mid-term future. This presentation will describe some of the experiments conducted and our analysis of the applicability of Machine Learning in different steps of our value chain (data production, curation, archiving, valorisation, dissemination and support to users in their exploitation).

In particular, in the area of data valorisation, I will present results from one such projects, were we have used the Google AutoML Vision API to identify asteroid trails on archival HST images. For this project, we have used labels of asteroid trails from volunteer markings from the Zooniverse project to train a Deep Learning Convolutional Neural Network embedded into Google's AutoML Vision API and then have used it to scan the whole archive to look for new asteroid trails that the volunteers had not identified. I will describe how the Precision and Recall of the machine learning model depend on the pre-processing of the images and show the very promising results from this analysis. I will end up by elaborating on how these type of initiatives will help enriching the search for data-driven patterns on archival data in the mid-term future.

Theme – Cloud Computing at Different Scales, Machine Learning, Statistics, and Algorithms, Citizen Science Projects in Astronomy