DeepDISC-Roman: Detection, Instance Segmentation, and Classification for Roman with Deep Learning

Wide-Field Science – Regular
Xin Liu / University of Illinois – Urbana-Champaign, PI

The Nancy Grace Roman Space Telescope will deliver deep high-quality images in the near infrared for precision cosmology and beyond. As both the sensitivity and depth increase, larger numbers of blended (overlapping) sources will occur. If left unaccounted for, blending would result in biased measurements of sources that are assumed isolated, contaminating key inferences such as photometry, photometric redshift, galaxy morphology, and weak gravitational lensing. In the Roman era, efficient deblending techniques are a necessity and thus have been recognized a high priority. However, an efficient and robust deblending method to meet the demand of next-generation deep-wide surveys is still lacking.

Leveraging the rapidly-developing field of computer vision, the open-source DeepDISC-Roman will provide a new versatile deep learning framework for the Roman research community. It makes it easy to efficiently process Roman images and accurately identify blended galaxies with the lowest latency to maximize science returns. The approach is interdisciplinary and fundamentally different from traditional methods, combining state-of-the-art survey data with the latest deep learning tools. A unique feature of DeepDISC-Roman is the robust quantification of the uncertainty of the prediction, which can be then propagated into the final error budget for precision cosmology. Because of limitations with the previous deep learning applications, a new framework is under development, leveraging Detectron2 – Facebook AI Research’s next-generation open-source platform for object detection and segmentation. This program will support transforming DeepDISC-Roman from a proof-of-concept pilot study to a fully featured, developed and science-ready platform for astronomical object detection, instance segmentation, classification, and beyond. Leveraging existing software infrastructure and production-ready packages, DeepDISC-Roman will be trained and validated using a hybrid of real data and more realistic simulations by combining traditional image simulations with deep generative models. DeepDISC-Roman can be applied to many other higher-level downstream science applications such as photometric redshift estimation and galaxy morphology inferences. It will combine Roman with Rubin and Euclid to leverage the wider optical-to-near-infrared coverage to improve the reliability of photometric redshifts and to facilitate the deblending of Rubin ground-based images. The program has strong implications for a wide range of subjects, from efficiently detecting transients and solar system objects to the nature of dark matter and dark energy.

DeepDISC-Roman will deliver several key scientific, software, and data products to maximize Roman science. It will produce a versatile deep-learning framework which can be integrated into the analysis software provided by the Roman Science Centers. It addresses the Roman research and support participation program through “Development of Roman analysis software beyond that provided by the Science Centers” and “Development of algorithms for joint processing with data from other space- or ground-based observatories such as deblending algorithms, photometric redshift training and calibration, or forced photometry”. DeepDISC-Roman will complement and augment activities of the Roman Science Centers but does not overlap/duplicate them. The proposed work should be performed now, rather than closer to launch or post-lauch, because deblending is fundamental to the Roman data analysis infrastructure and critical for enabling many downstream applications. As part of the proposed work, the program will train undergraduate students through the Students Pushing Innovation at the National Center for Supercomputing Applications. Finally, the program will develop a diverse and inclusive scientific workforce and clearly defines roles and responsibilities for all team members toward pursuing those goals.