Automated galaxy sizes in Euclid images using the Segment Anything Model

TL;DR

This study demonstrates that the Segment Anything Model (SAM) can automatically and accurately identify galaxy disk truncations in Euclid-like images, enabling large-scale galaxy size measurements without extensive preprocessing or training.

Contribution

The paper shows that SAM can effectively detect galaxy truncations and measure sizes in large datasets, with minimal preprocessing and no need for domain-specific training.

Findings

SAM achieves ~3% deviation from manual size measurements

Performance improves to ~1% deviation when excluding problematic cases

SAM operates effectively without extensive preprocessing or additional training

Abstract

Stellar disk truncations, also referred to as galaxy edges, are key indicators of galactic size, determined by the radial location of the gas density threshold for star formation. Accurately measuring galaxy sizes for millions of galaxies is essential for understanding the physical processes driving galaxy evolution over cosmic time. In this study, we aim to explore the potential of the Segment Anything Model (SAM), a foundation model designed for image segmentation, to automatically identify disk truncations in galaxy images. With the Euclid Wide Survey poised to deliver vast datasets, our goal is to assess SAM's capability to measure galaxy sizes in a fully automated manner. SAM was applied to a labeled dataset of 1,047 disk-like galaxies with $M_* > 10^{10} M_{\odot}$ at redshifts up to $z \sim 1$, sourced from the HST CANDELS fields. We 'euclidized' the HST galaxy images by creating composite RGB images, using the F160W (H-band), F125W (J-band), and F814W + F606W (I-band + V-band) HST filters, respectively. Using these processed images as input for SAM, we retrieved various truncation masks for each galaxy image under different configurations of the input data. We find excellent agreement between the galaxy sizes identified by SAM and those measured manually (i.e., by using the radial positions of the stellar disk edges in galaxy light profiles), with an average deviation of approximately $3\%$. This error reduces to about $1\%$ when excluding problematic cases. Our results highlight the strong potential of SAM for detecting disk truncations and measuring galaxy sizes across large datasets in an automated way. SAM performs well without requiring extensive image preprocessing, labeled training datasets for truncations (used only for validation), fine-tuning, or additional domain-specific adaptations such as transfer learning.