Description
Identification of tidal streams in galaxy images is a difficult task as they are very faint compared to other stellar components. Additionally, the large influx of images expected from upcoming large-scale sky surveys will make it increasingly unrealistic to expect to be able to identify stellar streams by visual examination alone.
Here we will present a machine learning classifier trained to identify streams in images of simulated galaxies, with the aim to be able to use it on images from surveys. We use the Auriga suite of high-resolution cosmological simulations of Milky Way-mass galaxies to create a dataset of over 1500 images of galaxies with tidal features which can be automatically labelled, pixel-by-pixel, according to the origin of stellar particles. This labelling would be difficult and time consuming to do by hand with observations. We also avoid the issues associated with combining existing observational datasets, which would be necessary to get a large enough training sample. For example, differences in observing equipment and data processing pipelines can introduce irrelevant patterns that the classifier picks up on, reducing its effectiveness. We use this set of images to train a UNet - a type of convolutional neural network architecture which identifies structures in images through pixel-by-pixel labelling, allowing localisation of streams within an image.
We will also assess performance on observations and, if needed, fine-tune on smaller observational datasets (e.g. upcoming data from Euclid), labelled by experts. In the future, we plan to further extended this to analyse the morphology of streams.
