Background-Source Separation in astronomical images with Bayesian probability theory (I): the method

A probabilistic technique for the joint estimation of background and sources with the aim of detecting faint and extended celestial objects is described. Bayesian probability theory is applied to gain insight into the coexistence of background and sources through a probabilistic two-component mixture model, which provides consistent uncertainties of background and sources. A multi-resolution analysis is used for revealing faint and extended objects in the frame of the Bayesian mixture model. All the revealed sources are parameterized automatically providing source position, net counts, morphological parameters and their errors.

💡 Research Summary

The paper introduces a novel Bayesian Background‑Source Separation (BSS) method designed to jointly estimate the background and detect astronomical sources, especially faint and extended objects, in imaging data. The authors begin by reviewing existing source detection techniques such as SExtractor, sliding‑window methods, maximum‑likelihood PSF fitting, and wavelet‑based algorithms, highlighting their reliance on predefined source morphologies, fixed detection thresholds, and often inadequate handling of spatially varying backgrounds.

Building on Bayesian probability theory (BPT), the authors formulate a two‑component mixture model in which each image pixel is probabilistically assigned to either a “background‑only” class or a “background‑plus‑source” class. The data are assumed to follow Poisson statistics, appropriate for X‑ray and other photon‑limited regimes. The background is modeled with a flexible two‑dimensional spline (B‑spline) whose control points and amplitudes are treated as free parameters, allowing the model to capture smoothly varying backgrounds as well as sharp variations caused by exposure‑time maps, CCD gaps, or instrumental artifacts.

Two prior distributions for the source signal are explored: a non‑informative uniform prior and a power‑law prior that reflects the empirical distribution of source fluxes in many astronomical surveys. The choice of prior directly influences detection sensitivity and false‑positive rates, a relationship that is quantified through extensive simulations.

Parameter inference is performed via Markov Chain Monte Carlo (MCMC) sampling, yielding posterior probability density functions (PDFs) for all model parameters. From these posteriors, a per‑pixel “source existence probability” is derived, which replaces the conventional signal‑to‑noise threshold. Pixels with high source probability are flagged as candidate detections.

To enhance sensitivity across a range of spatial scales, the method incorporates a multi‑resolution analysis akin to a wavelet transform. At each resolution level, source probabilities are computed, allowing simultaneous detection of point‑like sources and extended structures such as galaxy clusters or supernova remnants. When multi‑band data are available, the probability maps from each band are combined in a Bayesian manner, further lowering detection limits.

The algorithm proceeds through four main stages: (1) initial estimation of spline background parameters; (2) MCMC sampling of the mixture model to obtain posterior PDFs; (3) construction of multi‑scale source probability maps and candidate selection; (4) Bayesian model comparison using Bayes factors to suppress false positives, followed by estimation of source parameters (position, net counts, morphology) and their uncertainties, fully accounting for background uncertainty.

The authors address false‑positive control by comparing the evidence (marginal likelihood) of the background‑plus‑source model against the pure background model. If the pure background model has higher evidence, the pixel is not considered a source, effectively mitigating spurious detections in complex background regions.

Performance is evaluated on three simulated datasets with varying background levels and source intensities. The BSS method is benchmarked against the wavelet‑based “wavdetect” algorithm. Results show a 20‑30 % increase in detection completeness for low‑signal‑to‑noise sources and a reduction of photometric errors by roughly 15 % on average. The impact of prior choice is demonstrated: the power‑law prior improves sensitivity to faint sources at the cost of a modest increase in false detections, whereas the uniform prior yields a more conservative catalog.

The method is further applied to real ROSAT All‑Sky Survey (RASS) data. BSS recovers faint, low‑surface‑brightness galaxy clusters and extended supernova remnants that were missed by SExtractor, while automatically correcting for exposure‑time variations and CCD gaps via the spline background model. The resulting source catalog exhibits consistent photometry with existing ROSAT catalogs and includes a substantial number of newly identified faint sources.

In conclusion, the BSS technique offers a statistically rigorous, fully Bayesian framework that simultaneously estimates background and detects sources without the need for arbitrary thresholds or predefined source shapes. Its flexibility makes it applicable to any photon‑limited imaging regime (X‑ray, γ‑ray, infrared) and especially valuable for large‑scale surveys and multi‑band analyses. Future work will focus on accelerating the MCMC sampling, integrating the method into real‑time pipelines, and applying it to specific scientific problems such as galaxy‑cluster searches in deep fields.