Simulation-Based Inference for Probabilistic Galaxy Detection and Deblending

This paper introduces BLISS, a simulation-based inference framework that utilizes convolutional neural networks and denoising autoencoders to probabilistically detect, deblend, and measure galaxy properties in crowded astronomical images, demonstrating significant improvements in flux accuracy for blended and faint objects compared to deterministic methods.

The Gist

Imagine you are looking at a crowded night sky through a telescope. In the past, the stars and galaxies were far enough apart that you could easily point to one and say, "That's a star," or "That's a galaxy."

But now, imagine the universe has gotten so crowded that the lights are overlapping. It's like looking at a city at night from a plane: the streetlights, car headlights, and building windows all blur together into a giant, glowing mess. In astronomy, this is called blending.

For the next generation of telescopes (like the Rubin Observatory), this "mess" will be the biggest problem. If we can't tell where one galaxy ends and another begins, we can't measure them accurately. And if we can't measure them accurately, we can't understand how the universe is expanding or what dark energy is.

This paper introduces a new tool called BLISS (Bayesian Light Source Separator) to solve this problem. Here is how it works, explained simply:

1. The Old Way vs. The New Way

• The Old Way (The "Cut-and-Paste" Method): Traditional tools look at a blurry image and try to draw a box around the brightest spot. They say, "Okay, this blob is one object." If two galaxies are overlapping, the old tool might miss one entirely or measure the combined light of both as if it were just one giant, weird galaxy. It's deterministic: it gives you one answer and says, "This is it."
• The New Way (BLISS): BLISS is like a detective who doesn't just guess; it calculates probabilities. Instead of saying, "There is definitely one galaxy here," it says, "There is a 90% chance there are two galaxies here, and here is where their centers likely are, with a 10% chance they are actually three." It embraces uncertainty rather than ignoring it.

2. How BLISS Works: The Three-Step Detective Team

BLISS uses a team of three AI "detectives" (neural networks) that work together to clean up the image.

• Detective #1: The Spotter (Detection Encoder)
  Imagine you are looking at a foggy photo of a crowd. The Spotter scans the image and says, "I see a face here, and another face there." It doesn't just guess the location; it draws a fuzzy circle around where the face might be. It tells us: "I'm pretty sure there's a galaxy here, but I'm not 100% sure exactly where its center is."
• Detective #2: The ID Card Reader (Classification Encoder)
  Once the Spotter finds a "face," the ID Reader steps in. Is this a star (which is just a point of light) or a galaxy (which is a swirling cloud of stars)? It looks at the shape and says, "95% chance this is a galaxy."
• Detective #3: The Restorer (Deblending Encoder)
  This is the magic trick. Imagine you have a photo of two people hugging, and their clothes are blended into a single color. The Restorer takes the "fuzzy circles" from the Spotter and the "ID" from the Reader, then uses a special AI trick (an autoencoder) to digitally "un-hug" them. It reconstructs what each galaxy would look like if it were standing alone in a clean, noise-free room.

3. The "Tile" Strategy

The universe is huge, and the images are massive. BLISS doesn't try to solve the whole puzzle at once. Instead, it cuts the image into small, overlapping square tiles (like a mosaic).

• It solves the puzzle for one tile, then moves to the next.
• Because the tiles overlap, the AI can see the neighbors. If a galaxy is half in one tile and half in another, the AI uses information from both to figure out the whole picture.

4. Why "Probabilistic" is a Game Changer

The paper's biggest breakthrough isn't just that BLISS can separate the galaxies; it's that it keeps track of how unsure it is.

• The "Gambler" Analogy:
  • Old Method: A gambler bets $100 on a horse and says, "This horse will win." If the horse loses, the gambler is wrong, and the bet is lost.
  • BLISS: A gambler says, "I think Horse A has a 60% chance, Horse B has a 30% chance, and Horse C has a 10% chance."
  • The Result: When scientists use BLISS to measure the brightness (flux) of these galaxies, they don't just get one number. They get a range of possibilities. The paper shows that by using this "range of possibilities" (sampling many different scenarios), they get a much more accurate average measurement than if they just picked the single "most likely" answer.

5. The Bottom Line

The authors tested BLISS on simulated images that look exactly like what the Rubin Observatory will see in the future. They found that:

1. It finds more faint galaxies that other tools miss.
2. It separates overlapping galaxies much better.
3. It reduces errors in measuring how bright galaxies are, especially when they are very crowded and faint.

In summary: BLISS is a smart, probabilistic AI system that acts like a super-powered photo editor for the universe. It doesn't just try to force a blurry image into a clear one; it calculates the odds of what's really there, allowing astronomers to measure the cosmos with a level of precision that was previously impossible. This is a crucial step toward understanding the dark energy that is pushing our universe apart.

Technical Summary

Here is a detailed technical summary of the paper "Simulation-Based Inference for Probabilistic Galaxy Detection and Deblending" (BLISS).

1. Problem Statement

The next generation of Stage-IV dark energy surveys, such as the Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST), will observe an unprecedented density of galaxies. A primary challenge in these deep surveys is blending, where the light from multiple galaxies and stars overlaps within a single pixel or resolution element.

• Impact: Blending introduces systematic errors in weak gravitational lensing (cosmic shear) and photometry. It biases shape measurements, distorts flux estimates, and leads to "unrecognized blends" (where two sources are identified as one), potentially biasing cosmological parameters like $S_8$ by more than $2\sigma$.
• Limitations of Current Methods: Traditional tools like SourceExtractor (SExtractor) use deterministic thresholding and cannot handle overlapping sources effectively (assigning pixels to only one object). While modern deblenders (e.g., Scarlet, MuSCADeT) use joint modeling, they are often deterministic and fail to quantify the inherent uncertainty in source properties, which is crucial for propagating errors into cosmological analyses.

2. Methodology: The BLISS Framework

The authors propose BLISS (Bayesian Light Source Separator), a simulation-based inference (SBI) framework that approximates the posterior distribution of light source properties. BLISS combines deep learning with variational inference to perform detection, classification, and deblending while quantifying uncertainty.

Core Architecture

BLISS processes images by splitting them into overlapping padded tiles (53 $\times$ 53 pixels) to handle arbitrarily large images. It employs three neural network encoders trained via Forward Amortized Variational Inference (FAVI):

1. Detection Encoder:
   • Input: Padded tiles.
   • Output: A probabilistic catalog for the tile, specifically the posterior distribution over the number of sources ($n_t \in \{0,1\}$) and the centroid location ($\ell_t$) of the source within the tile.
   • Mechanism: Uses a Bernoulli distribution for source presence and a bivariate Gaussian for centroid location.
2. Classification Encoder:
   • Input: Centered padded tiles (cropped around the detected centroid).
   • Output: The probability that a detected source is a galaxy ($b_t=1$) or a star ($b_t=0$).
   • Mechanism: Outputs a parameter for a Bernoulli distribution.
3. Deblending Encoder (Deterministic):
   • Input: Centered padded tiles containing blended sources.
   • Output: A latent vector ($z_t \in \mathbb{R}^8$) representing the noiseless morphology and flux of the central galaxy.
   • Mechanism: Trained as an autoencoder (AE) to reconstruct isolated galaxies. When applied to blends, it learns to ignore flux from overlapping neighbors. The reconstruction is deterministic (point estimate), but uncertainty is propagated by sampling inputs from the detection encoder.

Training Strategy

• Data: Simulated single-band ($i$-band) images mimicking LSST Year-10 coadds using the GalSim package. The dataset includes parametric galaxy models (bulge + disk) and stars from the LSST DESC DC2 catalog.
• Loss Function: The detection and classification encoders minimize the Kullback-Leibler (KL) divergence between the true posterior and the variational approximation. The deblending encoder minimizes the negative log-likelihood of the reconstructed image under a Gaussian noise model.
• Amortization: A single set of neural network weights is shared across all tiles, allowing for rapid inference (seconds for megapixel images) on GPUs.

3. Key Contributions

• Probabilistic Deblending: Unlike previous deterministic deblenders, BLISS provides a full posterior distribution for source counts and positions. This allows for the propagation of detection uncertainty into downstream measurements.
• Hybrid Architecture: It uniquely combines a probabilistic detection/classification pipeline with a deterministic autoencoder-based deblender. While the deblender itself is deterministic, the authors demonstrate that sampling from the detection posterior and feeding these samples into the deblender yields a flux posterior that captures detection uncertainty.
• Scalability: The tiling strategy and amortized inference enable the processing of large survey images efficiently, overcoming the computational bottlenecks of traditional MCMC-based Bayesian methods.
• Open Source: The code and datasets are made publicly available, facilitating reproducibility and further development.

4. Key Results

The authors evaluated BLISS on simulated LSST-like images across several metrics:

• Detection Performance:
  • BLISS achieves comparable precision to SourceExtractor (SEP) across a wide range of Signal-to-Noise Ratios (SNR).
  • Crucially, BLISS demonstrates superior recall (completeness) for faint and highly blended sources compared to SEP, recovering a larger fraction of objects that traditional thresholding misses.
• Classification:
  • The classifier achieves >95% F1-score for galaxies with SNR > 10 and >70% for stars with SNR > 12.
  • At very low SNR, the model tends to default to the prior (classifying noisy stars as galaxies due to the galaxy-dominated training set), highlighting a need for better priors in low-SNR regimes.
• Deblending Accuracy:
  • When true centroids are known, the deblender significantly reduces residuals in flux, size, and ellipticity measurements compared to non-deblended measurements.
  • However, the method produces some non-physical artifacts in the tails of reconstructed galaxies, though these are buried by noise in realistic measurements.
• Probabilistic Flux Recovery (The "Samples" Approach):
  • The most significant finding is in Section 4.6. By drawing 100 samples from the detection posterior and averaging the resulting flux measurements, BLISS produces a flux posterior distribution.
  • Result: This probabilistic approach drastically reduces the median flux residual for highly blended galaxies compared to deterministic methods (MAP or SEP). For highly blended objects, the residual error dropped from ~160% (MAP) and ~400% (SEP) to significantly lower values when using the sampling approach. This confirms that accounting for detection uncertainty is vital for accurate photometry in crowded fields.

5. Significance and Future Work

• Cosmological Impact: BLISS offers a scalable pathway to mitigate blending-induced systematics in LSST and other Stage-IV surveys. By providing uncertainty-aware catalogs, it enables more robust constraints on cosmological parameters like dark energy equation of state and $S_8$.
• Limitations & Future Directions:
  • Tiling Artifacts: Detection probability drops near tile boundaries, a known artifact of the current independent tile assumption. Future work will explore overlapping tiles or dependency modeling between tiles.
  • Model Bias: The current deblender is trained on parametric galaxies. Future iterations will use Variational Autoencoders (VAEs) trained on realistic, non-parametric galaxy morphologies to capture full morphological uncertainty.
  • Real Data: The framework needs to be adapted for multi-band data, spatially variable PSFs, and real survey backgrounds (e.g., using LSST DC2 simulations).

In summary, BLISS represents a paradigm shift from deterministic cataloging to probabilistic, simulation-based inference, demonstrating that explicitly modeling detection uncertainty yields substantial improvements in the accuracy of galaxy measurements in crowded, blended fields.