Denoising in astronomy and astrophysics

How it is used

Almost every astronomical measurement arrives buried in unwanted signal. Detectors add electronic noise, the sky itself glows, the atmosphere blurs and radio transmitters on the ground swamp faint cosmic sources. Denoising means recovering the underlying signal from such degraded data. In this field the methods are applied to time series (the changing brightness of a star, quasar or supernova), to images of galaxies, and to spectra, which record how much light arrives at each wavelength. In the examples here, models are fitted to brightness measurements taken at scattered times and in several colour bands, to radio data spread across time and frequency, and to the continuous output of instruments listening for gravitational waves.

Why researchers use it

Researchers reach for these methods because the interesting part of the signal is often weaker than the noise around it, and because the noise is rarely simple. Classical filtering assumes noise is random and unchanging, but instrument noise drifts, and disturbances in a detector's surroundings leak into the measurement in ways no formula describes. A model learned from data can stand in for that missing description. The examples show two broad habits. One is to fit a flexible statistical model, such as a Gaussian process, which treats a light curve as a smooth function with an estimated uncertainty, so sparse and irregular measurements can be read at any moment. The other is to train a network on simulated data, where the true signal is known, and then apply it to real observations.

In the record

Limits

The weakness is circularity. Models are usually trained on simulations, so they learn the noise the simulation contains, and real instruments misbehave in ways no simulation anticipates. A denoiser can then invent structure that looks plausible but is not there, smoothing over a genuine short-lived event or sharpening a glitch into a signal. Removing noise also discards information about how the measurement was made, which matters when the result feeds into a claim about uncertainty. In this field such errors are hard to catch, because there is often no independent view of the same object to check against.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Related

• Denoising: the idea, across every field
• AI in astronomy and astrophysics

Every article