CDS&E: Detection, Instance Segmentation, and Classification for Astronomical Surveys with Deep Learning (DeepDISC)

The next generation of wide-field deep astronomical surveys will deliver unprecedented amounts of images of the sky through the 2020s and beyond. As both the sensitivity and depth increase, larger numbers of blended (overlapping) sources will occur. Blending would result in biased measurements, contaminating key astronomical inferences. Having efficient deblending techniques is thus a high priority for the future of astronomical research. However, an efficient and robust method to detect, deblend, and classify sources is still lacking for massive surveys. In this project, scientists at the University of Illinois, Urbana-Champaign will develop a versatile deep learning framework for image deblending and source detection. This work will make it easy to efficiently process wide-deep survey images and will accurately identify blended sources with the lowest possible latency to maximize science returns. Moreover, this work will provide robust uncertainties of detection inferences, which are critical for enabling precision cosmology. The proposed work has broad implications for a wide range of subjects, including detecting transients and solar system objects to probing the nature of dark matter and dark energy. As part of this project, the PI will also develop and teach a dedicated summer outreach program to engage young girls in STEM. This research program will leverage the rapidly developing field of computer vision to build a new deep learning platform for astronomical object detection, instance segmentation, classification, and beyond. It will adapt the latest open-source algorithms in computer vision for object detection and segmentation. The approach is interdisciplinary, combining state-of-the-art astronomical survey data with the latest deep learning tools. The new platform will be trained and validated using a hybrid of real data and realistic simulations that are built by combining traditional image simulations with generative models. It will be fully featured to enable higher-level downstream science applications such as photometric redshift estimation and galaxy morphology inferences. All codes generated will be open source to enable broad community usage. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.