Simulation-Based Inference for Probabilistic Galaxy Detection and Deblending

TL;DR

This work introduces BLISS, a probabilistic, simulation-based inference framework for simultaneous detection, deblending, and measurement of galaxies and stars in LSST-like images. BLISS combines forward amortized variational inference with tiling and a denoising autoencoder to produce per-tile posteriors over source counts, centroids, and types, plus noiseless reconstructions for deblending. On LSST-like simulations, BLISS improves aperture flux posteriors for blended and faint objects and demonstrates how detection uncertainty can be propagated to flux measurements, potentially mitigating blending-induced systematics in next-generation cosmological analyses. The approach shows promise for scalable, uncertainty-aware cataloging in crowded fields and provides a pathway to extend to multi-band data and realistic morphology, with ongoing work to integrate these uncertainties into downstream cosmology.

Abstract

Stage-IV dark energy wide-field surveys, such as the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST), will observe an unprecedented number density of galaxies. As a result, the majority of imaged galaxies will visually overlap, a phenomenon known as blending. Blending is expected to be a leading source of systematic error in astronomical measurements. To mitigate this systematic, we propose a new probabilistic method for detecting, deblending, and measuring the properties of galaxies, called the Bayesian Light Source Separator (BLISS). Given an astronomical survey image, BLISS uses convolutional neural networks to produce a probabilistic astronomical catalog by approximating the posterior distribution over the number of light sources, their centroids' locations, and their types (galaxy vs. star). BLISS additionally includes a denoising autoencoder to reconstruct unblended galaxy profiles. As a first step towards demonstrating the feasibility of BLISS for cosmological applications, we apply our method to simulated single-band images whose properties are representative of year-10 LSST coadds. First, we study each BLISS component independently and examine its probabilistic output as a function of SNR and degree of blending. Then, by propagating the probabilistic detections from BLISS to its deblender, we produce per-object flux posteriors. Using these posteriors yields a substantial improvement in aperture flux residuals relative to deterministic detections alone, particularly for highly blended and faint objects. These results highlight the potential of BLISS as a scalable, uncertainty-aware tool for mitigating blending-induced systematics in next-generation cosmological surveys.

Figures (11)

• Figure 1: BLISS inference pipeline outline. Schematic demonstrating the procedure for detecting, deblending, and reconstructing galaxies with BLISS. The procedure begins with an astronomical image of a galaxy blend, which is split into padded tiles (orange border) and passed through the detection encoder. The detection encoder outputs centroid locations for each detected galaxy that can be used to center each padded tile on its corresponding source. The centered padded tiles are then passed to the classification and deblending encoder. Finally, the outputs of each encoder can be used to construct a posterior catalog of the galaxy blend, and a corresponding noiseless reconstruction of the original image. For more details on each of the steps (blue circles) in this outline, see Section \ref{['sec:method']}.
• Figure 2: Detection performance as a function of SNR. In this figure we present the performance of the trained BLISS detection encoder using different probability thresholds (color gradient, blue to red) and SEP (dashed, black) on our testing set of galaxy blends. We show the precision, recall, and $F_{1}$ score in the same equally-log-spaced SNR bins. In the bottom-right plot, we additionally show the true SNR distribution of our dataset-blend dataset using the same SNR bins. For more details on this figure, see Section \ref{['sec:detection-results']}.
• Figure 3: Completeness as a function of blendedness. This figure shows the fraction of detected true objects (i.e., recall) as a function of their blendedness for SEP and the BLISS detection encoder. The BLISS detection encoder's recall for different probability thresholds is shown in the blue to red gradient curves, while SEP's is shown with a black dashed curve. The blendedness bins shown are chosen so that there are the same number of sources in each of them. See Section \ref{['sec:detection-results']} for details.
• Figure 4: 2D Histogram of classification probability and SNR. In these plots we show a 2D histogram of the classification probability of a given source being a galaxy and its true SNR, for every galaxy (left) and star (right) in the testing dataset-blend. This probability is the output from the binary encoder conditioned on true counts and centroids of every source. For more details on this figure, see Section \ref{['sec:binary-results']}.
• Figure 5: Classification performance as a function of SNR for galaxies and stars. This figure shows the precision (green), recall (purple), and $F_{1}$ score (pink) as a function of true SNR obtained by applying the binary encoder to our testing dataset-blend of galaxies and star blends. The SNR bins in each case are chosen so that there are an equal number of light sources on each bin. We condition the binary encoder on the true counts and centroids of every source in this dataset. For more details on this figure, see Section \ref{['sec:binary-results']}.