Yesterday we asked our collaborator Sandor Kruk for a nice image to represent our AutoML project to detect new asteroid trails in HST images and he sent us this image. Isn’t it gorgeous?

It is an image of nearby spiral galaxy NGC 5468, situated at 140 million light years away from us, with asteroid 2002 LX55 in front, identified by AutoML (the boxes show the identification). The image was taken by the Hubble Space Telescope on 29-12-2017. The credit for the image: ESA/Hubble & NASA (Hubble observation PI: Prof. Adam Riess).

By the way, the AutoML algorithms assigns a probability of 59% to those trails in the image being from an actual asteroid flying in front of Hubble’s cameras as they observed the galaxy. Isn’t it a bit too conservative here? Maybe it was because of the galaxy in the background that it gave it such a low score, because to my human eyes, it is a beautifully clear asteroid trail! What do you think?

I recently gave a summary talk at the EIROForum workshop on Big Data showing fresh from the oven results from our Deep learning project to identify asteroid trails in Hubble Space Telescope images from the European HST archive by using Google’s AutoML Vision API. This work has been done mostly in collaboration with my colleague Sandor Kruk, a Research Fellow at ESTEC, and Pablo García Martín, an engineer lead at Safran currently doing a Ph.D. in Astronomy with us applying these machine learning algorithms to data at the ESAC Science Data Centre plus other colleagues.

In short, we have managed to train an AutoML classifier by feeding training samples of between 1000 and 2000 images with asteroid trails as manually selected from our Zooniverse project Hubble Asteroid Hunter, where more than 10,000 volunteers have found over 1200 asteroid trails in all the HST ACS/WFC and WFC3/UVIS images at the ESA Hubble Science Archive. We have used both the on-cloud trainer and the edge trainer, and had to split the large images in four quadrants to maximise the ability of the network to identify trails given their relative sizes compared to the those of the images. The result has been a staggering 2,466 asteroid trails identified in 2,200 images (approximately 5% of all the images inspected), out of which 1,222 asteroid trails had already been identified by the Zooniverse volunteers while another 1,244 are new asteroid trails found. We have used 3 elapsed hours (122 computer node hours) for training the network and 7 elapsed hours (38 computer node hours) to scan the whole archive.

You can watch a full video of the presentation here:

And download the presentation in pdf here.