Photometry on Structured Backgrounds: Local Pixelwise Infilling by Regression

The Big Picture

Imagine trying to measure the brightness of a single candle on a table covered with a crumpled, glowing tablecloth. The candle is easy enough to spot, but how much of the light reaching you comes from the candle versus the shifting glow beneath it? This is roughly the challenge astronomers face when measuring stars sitting in cosmic clouds of gas and dust near the Galactic plane.

Traditional astronomical software pipelines, the automated tools that process raw telescope images into usable catalogs, are built for cleaner situations: stars against relatively smooth, dark skies. When a star sits atop a filament of glowing interstellar gas or a wisp of dust cloud, those tools systematically misattribute background glow to the star itself. They also fail to quantify how uncertain the measurement is. These stars aren’t a fringe concern. They’re the ones that matter most for understanding star formation and the interstellar medium, the diffuse gas and dust filling the space between stars.

Andrew Saydjari and Douglas Finkbeiner at Harvard developed a technique called Local Pixelwise Infilling (LPI) that reconstructs what the background would look like if the star weren’t there, then uses that reconstruction to produce cleaner brightness measurements and more honest error bars.

Key Insight: By statistically predicting the background hidden beneath each star using its local neighborhood of pixels, LPI produces both corrected brightness measurements and reliable uncertainty estimates, something previous background-subtraction approaches could not do.

How It Works

LPI builds on Gaussian Process Regression (GPR), a statistical method that predicts unknown values from the pattern of nearby known ones. Pixels in astronomical images follow this logic, especially in smoothly varying dust clouds: knowing the surrounding pixels lets you estimate what a hidden pixel’s value should be.

Here’s the LPI workflow in practice:

1. Mask the star. For each detected source, the algorithm identifies which pixels are contaminated by starlight, based on the telescope’s Point Spread Function (the characteristic blurred halo that starlight spreads into through any real optical system).
2. Estimate local covariance. Using surrounding unmasked pixels, LPI builds a local statistical model of how background brightness varies. Covariance captures how strongly neighboring pixels tend to rise and fall in brightness together, essentially the texture of the dust or gas in that region.
3. Infill the hidden pixels. With this covariance model, the algorithm predicts what the background beneath the star should have been, along with how uncertain that prediction is.
4. Correct the flux and error. The predicted background is subtracted from the measured brightness, yielding a cleaner stellar flux. The infill uncertainty propagates directly into the reported error bar.

The difference from classical GPR is in how LPI handles covariance. Standard GPR requires choosing a specific functional form, say an exponential decay or a Matérn kernel, then fitting its parameters. LPI skips that step entirely. It estimates covariance nonparametrically, straight from the data, using only a local patch of pixels and assuming no particular formula. Because astronomical images live on pixelized grids, covariance is sampled at fixed pixel separations, which keeps the calculation tractable.

Saydjari and Finkbeiner validated LPI on synthetic images with known ground-truth backgrounds, then on real dust fields. It holds up even in crowded fields where stars overlap and photometric pipelines notoriously break down.

Why It Matters

The scale of the application backs up the method. Saydjari and Finkbeiner ran LPI on the full second data release of the Dark Energy Camera Plane Survey (DECaPS2), a dataset containing 34 billion individual detections.

On nebulous fields, LPI cleared out large numbers of >3σ outliers, readings more than three standard deviations from expected values and a standard red flag for unreliable data. Uncertainty estimates improved by a factor of roughly 2 to 3. That’s not marginal. Error bars that were previously half the size they should have been now match reality.

The practical appeal is that LPI works as a post-processing step, running after a standard photometric pipeline has already finished. No one has to rebuild complex data reduction infrastructure. It can retroactively improve archival survey data and slot into future surveys like the Legacy Survey of Space and Time (LSST) at the Vera C. Rubin Observatory, which will observe billions of Galactic-plane stars over its decade-long run.

The method isn’t limited to optical or infrared wavelengths. Any survey with pixelized images and spatially correlated backgrounds (radio, submillimeter, X-ray) could use the same approach.

Getting uncertainty estimates right has consequences beyond any single measurement. Most of what astronomers care about, from stellar distances and ages to maps of interstellar dust, comes from statistical inferences built on brightness catalogs. Feed those inferences error bars that are too small and every downstream conclusion becomes overconfident. LPI forces the data to be honest about its own limitations.

Bottom Line: LPI is a statistically grounded, scalable, post-processing method that corrects stellar brightness measurements on complex backgrounds and produces honest uncertainty estimates. Applied to 34 billion measurements in DECaPS2, it improves catalog quality and stands ready for the next generation of sky surveys.

---