EGIPS: Edge-on Galaxies in Pan-STARRS

• EGIPS is a homogeneous catalog of nearly edge-on disc galaxies from Pan-STARRS1 DR2, selected using deep convolutional neural networks and expert visual vetting.
• The methodology integrates SExtractor photometry with non-parametric morphological measures, achieving high completeness (>96%) and purity (~99.2%) in galaxy classification.
• EGIPS data enables detailed studies of galactic disks, bulge properties, boxy/peanut structures, and satellite kinematics, offering refined scaling relations for galaxy evolution research.

The Edge-on Galaxies in the Pan-STARRS Survey (EGIPS) refers to a suite of catalogues constructed from the Pan-STARRS1 (PS1) survey data, providing a homogeneously selected, statistically robust sample of nearly or purely edge-on disc galaxies covering approximately three quarters of the sky north of declination $-30^\circ$. EGIPS leverages deep convolutional neural networks and expert visual validation to assemble samples ranging in size from $1.7 \times 10^4$ to $1.3 \times 10^5$ galaxies, enabling precise photometric, structural, and dynamical studies of galactic disks, bulges, and their satellites on a scale previously unattainable.

1. Survey Construction and Catalogue Definition

The principal EGIPS catalogue comprises $16\,551$ edge-on galaxies identified from Pan-STARRS1 DR2 imaging north of declination $-30^\circ$, corresponding to roughly $3/4$ of the celestial sphere. Reliable SExtractor photometry is achieved for objects with $13.8 < r_\mathrm{Petro} < 17.4$ mag, and $63\%$ of catalogue members have spectroscopic redshifts tabulated in HyperLEDA ($\mathrm{cz}_\mathrm{CMB} \approx 11\,600$ km s$^{-1}$, $1.7 \times 10^4$0 Mpc) (Makarov et al., 2022). In an expanded release using refined neural-network methodology and more permissive magnitude and size limits, $1.7 \times 10^4$1 edge-on disk galaxies are catalogued, with a major axis threshold of $1.7 \times 10^4$2, $1.7 \times 10^4$3 mag, and central surface brightness $1.7 \times 10^4$4--$1.7 \times 10^4$5 mag arcsec$1.7 \times 10^4$6 (Savchenko et al., 2023).

Each entry in EGIPS includes astrometry, homogeneous multi-band photometry (g, r, i, z, y), calibrated non-parametric morphology indices, and classification metadata. The catalogue infrastructure supports programmatic queries and download in ASCII/FITS via the web interface at https://www.sao.ru/edgeon/.

2. Machine Learning Classification Pipeline

Galaxy candidate selection employs a deep convolutional neural network (CNN) ensemble trained initially on positive examples from the EGIS (SDSS-based) catalogue and negative examples from HyperLeda (disk galaxies with $1.7 \times 10^4$7), stars, and empty fields. The typical CNN architecture comprises three convolutional blocks (two convolutional layers + batch normalization per block), max-pooling, $1.7 \times 10^4$8 dropout, a single fully connected layer, and a two-neuron output layer for edge-on/other discrimination (Makarov et al., 2022, Savchenko et al., 2023). The model encompasses $1.7 \times 10^4$9 trainable parameters per network.

Training utilizes aggressive data augmentation (random flips, rotations, zoom up to $1.3 \times 10^5$0, noise injection) yielding $1.3 \times 10^5$1 images. Ensemble learning (five or eleven independent CNNs) further suppresses biases and increases purity: a candidate is accepted if at least the majority of networks exceed $1.3 \times 10^5$2.

Detection rate (completeness) and reliability (purity) are rigorously benchmarked:

• On independent RFGC samples: $1.3 \times 10^5$3, $1.3 \times 10^5$4, and $1.3 \times 10^5$5.
• For disks with $1.3 \times 10^5$6, $1.3 \times 10^5$7 mag, and $1.3 \times 10^5$8 mag arcsec$1.3 \times 10^5$9, EGIS completeness is $16\,551$0. Further completeness maps produced via injection-recovery of synthetic galaxies yield $16\,551$1 completeness under similar surface-brightness constraints (Savchenko et al., 2023).

A critical final step is visual vetting: all candidates are inspected by at least three astronomers using the Zooniverse platform, eliminating artifacts, asterisms, and mis-inclined disks. Only $16\,551$2 of final objects have inclination $16\,551$3.

3. Photometric and Morphological Parameterization

Photometry is conducted using SExtractor (g, r, i, z, y) with custom deblending and artifact-rejection schemes. Each galaxy's shape (semi-major, semi-minor axis, ellipticity, position angle) and flux (MAG_AUTO/Kron, MAG_PETRO/Petrosian) parameters are supplied with photometric errors, band quality flags, and extinction corrections (using Schlegel et al. 1998; Schlafly & Finkbeiner 2011 maps). Agreement with EGIS SDSS-based photometry yields typical dispersions $16\,551$4 mag, $16\,551$5 mag, $16\,551$6 mag; the photometric range is empirically valid for $16\,551$7 mag (Makarov et al., 2022).

Non-parametric morphological statistics are provided using the statmorph package:

• Gini ($16\,551$8): quantifies pixel flux inequality.
• $16\,551$9: second-order moment of brightest 20% of the flux.
• Concentration ($-30^\circ$0): $-30^\circ$1, ratio of radii enclosing 80% and 20% of flux.
• Asymmetry ($-30^\circ$2): normalized residual after 180° rotation.
• Smoothness ($-30^\circ$3): normalized residual after image smoothing.

These metrics enable objective comparisons of bulge/disk prominence, bar strength, interaction signatures, and star-forming clumpiness.

4. Scientific Results and Structural Scaling Relations

Stellar Populations, Internal Extinction, and Colour–Thickness Trends

In the $-30^\circ$4 vs. $-30^\circ$5 colour–magnitude plane, EGIPS edge-on galaxies span both the red sequence and blue cloud, closely tracking SDSS DR12 populations. Notably, edge-on red-sequence galaxies are redder than randomly oriented controls by $-30^\circ$6 mag, consistent with enhanced internal extinction (Makarov et al., 2022). In the blue cloud, median colour trends linearly with disk flatness (inverse thickness $-30^\circ$7):

$-30^\circ$8

with $-30^\circ$9. Thinner blue-cloud disks are bluer, suggesting a coupling between vertical structure and stellar populations or dust.

Disk/Bulge Structural Parameters and Scaling

Multilinear regressions leveraging SExtractor sizes and EGIS decompositions yield:

$3/4$0

where $3/4$1 is the observed semi-major axis, $3/4$2 and $3/4$3 are the radial and vertical scale-lengths, $3/4$4 is the edge-on central surface brightness, and $3/4$5 is the bulge-to-total ratio. Incorporating all predictors, the scatter reduces to $3/4$6 dex. Predictive relations for $3/4$7 given observed quantities are tabulated (Makarov et al., 2022).

Boxy/Peanut (B/PS) Bulge Analysis

A dedicated study of $3/4$8 EGIPS galaxies exhibiting clear boxy/peanut-shaped (X-structure) bulges reveals:

• Median X-structure angle $3/4$9
• Median B/PS bulge scale-length $13.8 < r_\mathrm{Petro} < 17.4$0 kpc
• B/PS-to-total intensity ratio $13.8 < r_\mathrm{Petro} < 17.4$1
• Strong anticorrelation between $13.8 < r_\mathrm{Petro} < 17.4$2 and $13.8 < r_\mathrm{Petro} < 17.4$3 and correlation of $13.8 < r_\mathrm{Petro} < 17.4$4 with stellar mass; i.e., more massive galaxies harbor larger, flatter B/PS bulges (Smirnov et al., 20 Jan 2026).

A comparison with TNG50 cosmological simulations, using identical photometric decomposition techniques, demonstrates that simulated B/PS bulges are systematically smaller and fainter, with lower $13.8 < r_\mathrm{Petro} < 17.4$5 and lower side-on X-angles ($13.8 < r_\mathrm{Petro} < 17.4$6 in TNG50 vs. $13.8 < r_\mathrm{Petro} < 17.4$7–$13.8 < r_\mathrm{Petro} < 17.4$8 in EGIPS). This suggests limitations in current bar/bulge formation scenarios in the simulations.

5. Satellite Populations and Dynamical Inference

A systematic search for physical satellites using the EGIPS main sample identifies $13.8 < r_\mathrm{Petro} < 17.4$9 satellites around $63\%$0 primaries under strict photometric and kinematic ($63\%$1 km s$63\%$2, $63\%$3 kpc) criteria. Focusing on $63\%$4 satellites with $63\%$5 km s$63\%$6 redshift precision and strong gravitational boundness, the study yields

• Mean projected satellite–host separation $63\%$7 kpc
• Mean 1D velocity dispersion $63\%$8 km s$63\%$9

Projected masses determined via

$\mathrm{cz}_\mathrm{CMB} \approx 11\,600$0

adopt a nearly linear scaling with total $\mathrm{cz}_\mathrm{CMB} \approx 11\,600$1-band luminosity:

$\mathrm{cz}_\mathrm{CMB} \approx 11\,600$2

The resulting average $\mathrm{cz}_\mathrm{CMB} \approx 11\,600$3 ratio $\mathrm{cz}_\mathrm{CMB} \approx 11\,600$4 is typical of massive local spirals and consistent with Milky Way, M31, and M81 kinematics. The slope $\mathrm{cz}_\mathrm{CMB} \approx 11\,600$5 reflects a mild decrease of $\mathrm{cz}_\mathrm{CMB} \approx 11\,600$6 with increasing $\mathrm{cz}_\mathrm{CMB} \approx 11\,600$7 (Smirnov et al., 2023).

6. Data Products and Access

EGIPS data products include:

• General properties: HyperLEDA PGC ID, EGIPS IAU-style names, coordinates, extinction estimates, spectroscopic redshifts, and visual classification statistics.
• Photometric and morphological tables per band: Pan-STARRS skycell indices, geometric parameters with errors, fluxes, magnitude corrections, bad pixel fractions, morphological indices.

A summary table structure, as provided in the catalogues, is as follows:

Parameter	Meaning	Typical Range
ra, dec	Equatorial coordinates (deg, J2000)	Sky region $\mathrm{cz}_\mathrm{CMB} \approx 11\,600$8
a, b	Semi-major/minor axes (arcsec)	$\mathrm{cz}_\mathrm{CMB} \approx 11\,600$9
magpetro, magauto	Petrosian/Kron magnitudes (g,r,i,z,y)	$^{-1}$0
gini, m20, conc.	Non-parametric morphology metrics (per statmorph)	Survey-specific
votes, pctgood	Visual classification tally, fraction “good” views	$^{-1}$1
cz, cz3k	Redshifts (heliocentric / CMB frame, km s$^{-1}$2)	$^{-1}$3 km s$^{-1}$4

Public access is maintained at https://www.sao.ru/edgeon/ with advanced query options.

7. Implications and Future Directions

EGIPS enables analyses that transcend the inherent orientation biases in local galaxy samples and provides a strong empirical foundation for:

• Disentangling vertical vs. radial stellar populations and dust via inclination-invariant edge-on selection.
• Systematic mapping of dynamical scaling relations in disk, bulge, and B/PS components, facilitating direct confrontation with cosmological hydrodynamical simulations.
• Probing the environmental and satellite population demographics of massive spirals, with reliable dynamical masses by satellite kinematics.
• Statistical studies of disk heating, extinction, and secular evolution across barred and unbarred systems, free from projection ambiguities.

Future applications foreseen in the literature include detailed 3D stellar density reconstructions, systematic dust extinction mapping in disk planes, and improved calibration of scaling relations employing the uniform structural and photometric framework established by EGIPS (Makarov et al., 2022).

A plausible implication is the need for further refinement of cosmological simulations (e.g., TNG50), which currently underpredict the size, structure, and vertical strengths of bar- and B/PS-dominated bulges as observed in EGIPS (Smirnov et al., 20 Jan 2026).