Probabilistic segmentation of overlapping galaxies for large cosmological surveys

Abstract

Encoder-Decoder networks such as U-Nets have been applied successfully in a wide range of computer vision tasks, especially for image segmentation of different flavours across different fields. Nevertheless, most applications lack of a satisfying quantification of the uncertainty of the prediction. Yet, a well calibrated segmentation uncertainty can be a key element for scientific applications such as precision cosmology. In this on-going work, we explore the use of the probabilistic version of the U-Net, recently proposed by Kohl et al. (2018), and adapt it to automate the segmentation of galaxies for large photometric surveys. We focus especially on the probabilistic segmentation of overlapping galaxies, also known as blending. We show that, even when training with a single ground truth per input sample, the model manages to properly capture a pixel-wise uncertainty on the segmentation map. Such uncertainty can then be propagated further down the analysis of the galaxy properties. To our knowledge, this is the first time such an experiment is applied for galaxy deblending in astrophysics.

Realistic galaxy images and improved robustness in machine learning tasks from generative modelling

“…For example, Arcelin et al (2021) employ VAEs to learn probabilistic models of galaxies for deblending. Bretonnière et al (2021) propose to train VAEs for generating synthetic data for the upcoming Euclid survey, which can help with the preparation and calibration of algorithms. Smith et al (2022) train denoising diffusion proba-bilistic models on galaxy images and demonstrate applications for in-painting of occluded data and domain transfers.…”

“…This data set consists of individual galaxies from the contiguous COSMOS field (Mandelbaum et al 2012;Bretonnière et al 2021). The field was observed using the F814W filter with a drizzle pixel scale of 0.03 arcsec pixel −1 and limiting point source depth at 5𝜎 of 27.2 mag.…”

“…The number of filters of this last Conv2D layer depends on the number of channels of the data set the VAE is trained on. We also tested an implementation of the VAE described by Bretonnière et al (2021), however we had difficulty training this architecture successfully for our data sets, possibly due to limited training time. When training the VAE, we also found that we have to modify the training objective (4), because the models converged to either having good reconstruction capabilities but poor generation capabilities, or the model showed signs of mode collapse.…”