Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Neural Posterior Estimation for Cataloging Astronomical Images from the Legacy Survey of Space and Time

Yicun Duan, Xinyue Li, Camille Avestruz, Jeffrey Regier, LSST Dark Energy Science Collaboration

TL;DR

This work tackles the challenge of constructing statistically coherent astronomical catalogs from LSST-like coadded images, a problem made more acute by blending and high source density. It introduces neural posterior estimation (NPE) with a spatially autoregressive variational family (BLISS) to infer a marginal posterior p(z|x) over tile-based source catalogs directly from coadded, multiband images. Across the LSST DC2 simulated sky survey, BLISS outperforms the traditional LSST coadd pipeline in detection, flux estimation, star/galaxy classification, and galaxy ellipticity estimation, while providing well-calibrated, albeit slightly overdispersed, posterior intervals. The results demonstrate the feasibility of scalable, uncertainty-aware cataloging for LSST-scale data and outline strategies to mitigate model misspecification and extend to nonparametric modeling and EM-based integration for future improvements.

Abstract

The Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST) will commence full-scale operations in 2026, yielding an unprecedented volume of astronomical images. Constructing an astronomical catalog, a table of imaged stars, galaxies, and their properties, is a fundamental step in most scientific workflows based on astronomical image data. Traditional deterministic cataloging methods lack statistical coherence as cataloging is an ill-posed problem, while existing probabilistic approaches suffer from computational inefficiency, inaccuracy, or the inability to perform inference with multiband coadded images, the primary output format for LSST images. In this article, we explore a recently developed Bayesian inference method called neural posterior estimation (NPE) as an approach to cataloging. NPE leverages deep learning to achieve both computational efficiency and high accuracy. When evaluated on the DC2 Simulated Sky Survey -- a highly realistic synthetic dataset designed to mimic LSST data -- NPE systematically outperforms the standard LSST pipeline in light source detection, flux measurement, star/galaxy classification, and galaxy shape measurement. Additionally, NPE provides well-calibrated posterior approximations. These promising results, obtained using simulated data, illustrate the potential of NPE in the absence of model misspecification. Although some degree of model misspecification is inevitable in the application of NPE to real LSST images, there are a variety of strategies to mitigate its effects.

Neural Posterior Estimation for Cataloging Astronomical Images from the Legacy Survey of Space and Time

TL;DR

This work tackles the challenge of constructing statistically coherent astronomical catalogs from LSST-like coadded images, a problem made more acute by blending and high source density. It introduces neural posterior estimation (NPE) with a spatially autoregressive variational family (BLISS) to infer a marginal posterior p(z|x) over tile-based source catalogs directly from coadded, multiband images. Across the LSST DC2 simulated sky survey, BLISS outperforms the traditional LSST coadd pipeline in detection, flux estimation, star/galaxy classification, and galaxy ellipticity estimation, while providing well-calibrated, albeit slightly overdispersed, posterior intervals. The results demonstrate the feasibility of scalable, uncertainty-aware cataloging for LSST-scale data and outline strategies to mitigate model misspecification and extend to nonparametric modeling and EM-based integration for future improvements.

Abstract

The Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST) will commence full-scale operations in 2026, yielding an unprecedented volume of astronomical images. Constructing an astronomical catalog, a table of imaged stars, galaxies, and their properties, is a fundamental step in most scientific workflows based on astronomical image data. Traditional deterministic cataloging methods lack statistical coherence as cataloging is an ill-posed problem, while existing probabilistic approaches suffer from computational inefficiency, inaccuracy, or the inability to perform inference with multiband coadded images, the primary output format for LSST images. In this article, we explore a recently developed Bayesian inference method called neural posterior estimation (NPE) as an approach to cataloging. NPE leverages deep learning to achieve both computational efficiency and high accuracy. When evaluated on the DC2 Simulated Sky Survey -- a highly realistic synthetic dataset designed to mimic LSST data -- NPE systematically outperforms the standard LSST pipeline in light source detection, flux measurement, star/galaxy classification, and galaxy shape measurement. Additionally, NPE provides well-calibrated posterior approximations. These promising results, obtained using simulated data, illustrate the potential of NPE in the absence of model misspecification. Although some degree of model misspecification is inevitable in the application of NPE to real LSST images, there are a variety of strategies to mitigate its effects.
Paper Structure (38 sections, 7 equations, 16 figures, 4 tables)

This paper contains 38 sections, 7 equations, 16 figures, 4 tables.

Table of Contents

  1. Introduction
  2. The LSST DESC DC2 Simulated Sky Survey
  3. Catalog simulation
  4. Simulation of raw images
  5. Image processing and coaddition
  6. DC2 as a Bayesian model
  7. Methodology: Bayesian inference
  8. Our target: a marginal of the posterior
  9. Our variational family
  10. Neural Posterior Estimation
  11. Performance evaluation strategy
  12. Our primary benchmark: the LSST coadd catalog
  13. Matching detections and true objects
  14. Bootstrap confidence intervals
  15. Frequentist coverage of credible intervals
  16. ...and 23 more sections

Figures (16)

  • Figure 1: Coadded DC2 images for two DC2 patches. Each image is 4100$\times$4100 pixel. Each patch contains tens of thousands light sources. Hundreds of them are bright enough to be visible in these renderings.
  • Figure 2: Examples of BLISS and LSST detections. BLISS detects some faint and/or overlapping sources missed by the LSST detection pipeline. The legend marks sources detected only by BLISS (circles), only by LSST (triangles), by both (stars), by neither (crosses), with BLISS localization errors (diamonds), and with LSST localization errors (squares). Several LSST predictions may appear to be correct, but they are in fact off by more than our one-pixel threshold. These images are $80 \times 80$ pixels captured in the r band.
  • Figure 3: Detection precision (left) and recall (right). Each bar in the top row of bar plots represents the denominator for computing a point directly below it (in a line plot).
  • Figure 4: Detection precision (left) and recall (right) stratified by blendedness. Each bar in the bar plot represents the denominator for computing the point directly below it (in the line plot).
  • Figure 5: Highest-density 90% credible intervals for the light source positions, formed according to the variational distribution and plotted against ground truth. Each interval represents a particular light source in a random sample of 470 light sources from our DC2 test set. The 45-degree line, which indicates perfect estimation, passes through the 94.0% and 93.8% of intervals for horizontal and vertical positions, respectively.
  • ...and 11 more figures