The Edge-On Galaxies in the DESI Survey (EGIDE): Sample Building and Photometry

Abstract

We present the EGIDE (Edge-on Galaxies in the DESI survey) project—a catalog of 149,215 edge-on galaxy candidates created using the data of the DESI Legacy Imaging Survey DR10 images. The catalog size is ten times greater than its predecessor and covers more than half of the sky. It is constructed in an automatic way, utilizing the full power of manual annotations from the GalaxyZoo volunteers, implemented in the Zoobot neural model, which was fine-tuned to search for edge-on galaxies specifically. To ensure the credibility of the dataset, subsequent manual supervision was performed. The EGIDE catalog provides homogeneous SExtractor photometry in the $griz$ bands, total stellar mass estimates, redshifts for 98% of the sample, star formation rates, and other information. All of this is publicly available at The Edge-on Galaxy Database site. The preliminary analysis focused on differences between edge-on galaxies in the so-called blue sequence and red cloud populations. These galaxies demonstrate distinct properties: the number of redder galaxies decreases with increasing $a/b$ ratio faster than that of the bluer galaxies; galaxy thickness varies with galaxy color: red sequence galaxies are thicker than blue cloud galaxies; the flattening ratio $q=b/a$ increases significantly with total stellar mass $M_{★}$ only among redder cloud galaxies. It is an intriguing result that the same trend of q increasing at the high-mass end is detected by both the statistical models of figures of revolution and direct observations of edge-on galaxies in EGIDE independently. The full extent of this relationship’s validity can only be determined after properly accounting for the contributions of the bulge and the PSF.

1. Introduction

Edge-on disk galaxies, which are by definition inclined at nearly 90° to the line of sight, are a unique laboratory for studying the physical processes governing the internal and external evolution of galaxies and larger structures in the Universe. This orientation allows one to obtain, from observations, a lot of information about galaxies that is not available at other inclination angles. Consequently, there is a rich body of results and ongoing interest in these objects dating back to early pioneering works [1,2,3]. We list some of them below (see also a recent review in [4]).

First of all, for edge-on galaxies, we can directly obtain information on the thickness of the stellar disk and the vertical distribution of stellar matter within it. In the same way, we can estimate the thickness and properties of a gaseous disk [5]. The presence of so-called thick and thin disks, which are characterized by different scale heights of the stellar population $h_z$, is clearly established in our Galaxy using a variety of methods (for example, five methods applied to APOGEE DR17 in [6]). In other galaxies, these two disks were first discovered in the edge-on galaxy NGC 4565 by [3]. The exponential scale height of the thick disk is larger than that of the thin disk (∼1 kpc and ∼0.3 kpc in the Milky Way [7]), and the latter comprises younger populations. These observations also hold true for other galaxies [8], which raises the question of whether thick disks are heated up during evolution or are born thick. Different possible causes of vertical heating have been proposed, such as bars, tidal interactions, major or minor mergers, spiral arms, radial migration, adiabatic heating, and scattering by clumps or giant molecular clouds [9,10].

This discussion has been reignited with the beginning of observations with the James Webb Space Telescope (JWST). With the ability to observe edge-on regular disk galaxies up to $z\approx 5$ [11] and even beyond, JWST can now potentially directly observe all epochs during which disks formed. Data from the Hubble Space Telescope [12] and early JWST-based work suggest that disk thickness is almost constant across early epochs. The separation into thin and thick disks in Ref. [13] reveals that the thick disk forms first, and the transition epoch to two disks varies in time depending on the total stellar mass $M_{★}$. The results, however, are not yet settled, as various JWST samples lead to different numbers: Ref. [14] suggests that disks have intermediate thickness at birth and then thicken due to the combined effects of decreasing surface mass density and violent gravitational instabilities, but they are not born as thin as reported in Ref. [15].

Quantitatively, the thickness of disks may be characterized by the intrinsic flattening $q_{int}$ (also called the oblateness or aspect ratio), defined as the ratio of disk height to disk radius. For observed galaxies, we see the apparent flattening $q_{vis}=b/a$, where b and a are the disk minor and major axes. Due to projection effects, as well as the influence of the bulge and dust, we find $q_{vis}\ge q_{int}$, but the true intrinsic galaxy shapes can still be inferred from these measurements in a statistical sense by rotating various shapes and measuring the resulting projection’s b and a. This study was initiated by Edwin Hubble in the 1930s, and since Ref. [16] we have known that $q_{int}\approx 0.25$, although individual galaxies have a fairly wide range of thicknesses, including ultra-thin galaxies, for which $q\approx 0.1$ [17,18]. Statistical inference of $q_{int}$ for large samples such as GAMA DR3 [19] and SDSS [20,21], along with comparisons with cosmological simulations (Illustris, IllustrisTNG, EAGLE) in [22], reveal that the high fraction of thin disks observed in real galaxies does not match the simulations (see also [23,24]). There is an ongoing debate as to whether this effect is real or relates to an incorrect accounting of the spherical component [25] or perhaps to the fact that the apparent $q_{vis}$ does not properly reflect the true three-dimensional shape [24]. Recent estimates of oblateness over a wide $M_{★}$ range in Ref. [26] show that, for 40% of galaxies $q_{int}<0.2$, resulting in an extremely large number of “flat” dwarfs and an even greater inconsistency with simulations. Since edge-on galaxies have the advantage that we can measure flattening directly rather than statistically, larger samples of galaxies with the inclination $i\approx 90°$ are needed to resolve this discrepancy in $q_{vis}$ versus $q_{int}$.

To further complicate the situation, there are indications from JWST observations that high-redshift galaxies may not exhibit oblate (disky) geometry but instead contain a significant number of prolate objects [27,28], which could be misidentified as edge-on systems. The edge-on orientation of these prolate galaxies may also be a possible explanation for the nature of Little Red Dots [29]. Potentially, such objects can also help to distinguish between different models of dark matter origin [30]. Moreover, knowing the disk thickness in edge-on galaxies allows us to understand the properties of the dark halo. For example, the vertical force gradient can be measured as in UGC 7321, and the full halo shape can be recovered [31]. Another approach is that, under plausible assumptions about the Toomre parameter Q, we can derive the proportionality $h_z/h_R\propto M_{disk}/M_{tot}$, where $h_R$ is the radial disk scale and $M_{tot}$ is the total mass within a certain radius [32,33,34,35,36,37]. In other words, this imposes an independent constraint on the dark halo mass. Finally, the existence of superthin galaxies with observed $q<0.1$ (e.g., IC 5249) and with weak or no galactic bulges (cf. [38]) may imply the presence of an unusually massive dark halo with a compact core [36,38,39,40]. An alternative explanation derived from the TNG100 simulation in Ref. [41] links superthin galaxies to disk radial growth driven by specific mergers. In their picture, the existence of superthin galaxies is explained by subhalos that fall in at small angles and in the direction of rotation, i.e., when angular momentum is continuously pumped into the system since $z\approx 1$.

Many disk features are best seen when the disk is viewed edge-on and can be reliably studied only in this orientation. These include disk warps [42,43,44], flaring [8,45], central X-shaped structures from boxy/peanut-shaped (B/PS) bulges [46,47,48], polar rings [49,50,51,52], and some low-surface-brightness structures surrounding galaxies [53]. Photometric breaks in radial disk profiles and truncations are also more visible in edge-on galaxies [8,54], as are structures with lower surface brightnesses (Savchenko et al., in prep.). Additionally, the significantly lower uncertainty in the measured orientation parameters (inclination i and position angle) of edge-on galaxies is important when conducting research where precise knowledge of these parameters is crucial. This applies to studies related to the rotation velocity of the galaxy, such as Tully–Fisher measurements [55], or research that depends on the mutual orientation with cosmic filaments [56,57] or satellites [58,59].

Last but not least, an important advantage of galaxies viewed edge-on is the ability to study processes for which knowledge of the height above the disk plane is essential. For example, extraplanar diffuse ionized gas (DIG) contaminates photometry in the H$\alpha$ filter when viewed face-on and must be separated from the disk emission [60,61]. Another example is the presence of extraplanar cold gas [62,63], which may be essential for the formation of lenticular S0 galaxies [64]. The disk vertical scale height $h_z$ is also crucial when volumetric information is needed, as in studies of large-scale star formation [65,66], disk gravitational stability [67,68,69,70,71], secular bar formation [72], and many others.

For all the reasons listed above, edge-on disk galaxies are therefore of particular interest for various areas of extragalactic astrophysics. Although disk thickness can now be easily studied in many available simulations, real observations are still preferred due to the possible presence of unphysical conditions in simulations (e.g., [73]). The importance of edge-on galaxies is emphasized by the efforts spent on compiling catalogs (and surveys, such as GECKOS [74]) devoted to them. The classic Flat Galaxies Catalog (FGC) [75] and the Revised Flat Galaxy Catalog (RFGC) [76] together contain around 4000 large ($a\ge 40$ arcsec) galaxies. A sample one and a half times larger was found in SDSS and presented in the $gri$ EGIS (Edge-on Galaxies In SDSS) catalog [77]. Subsequent efforts allowed the construction of an even larger sample of 16,551 edge-on galaxies in the Pan-STARRS survey (EGIPS (https://www.sao.ru/edgeon/catalogs.php?cat=EGIPS (accessed on 10 June 2026)), [78,79]), where edge-on galaxies were visually inspected after automatic selection from the Pan-STARRS DR2 survey. There are also specialized samples, such as the 2MASS-selected Flat Galaxy Catalog (2MFGC) [80] or [81,82] from the same datasource, the B/PS-related subsample in [83], a catalog of low-surface-brightness edge-on galaxies [84], and others.

The number of disk galaxies with inclination between $i_1$ and $i_2$ varies as $cos{i}_1-cos{i}_2$, which gives ≈1.75% of the total population for $i=89°$–$90°$. Therefore, in the forthcoming era of contemporary wide-field surveys, such as Rubin, Roman, Euclid, 4MOST, and others, it is anticipated that samples of approximately ${10}^5$ highly inclined galaxies will be identified. The manual detection of such sizable samples is well beyond the reasonable capabilities of any individual or group. For example, the largest relevant project, Galaxy Zoo, has already collected 104 million annotations for a million galaxies [85], but only around 15,000 edge-on galaxies with reliable annotations were found (see figure 3). The detection of the axis ratio q from photometric data, which is another method for finding edge-on galaxies, is susceptible to various influences, including projection effects, significant errors, and challenges in automatic decomposition. At the same time, the simplicity of form is a distinguishing characteristic of edge-on galaxies, which exhibit geometry and features (e.g., dust lanes) that are relatively straightforward to comprehend. In a previous study, an approach was proposed for EGIPS [78] that involved the implementation of an automatic detection system utilizing a neural network, which yielded satisfactory results.

In this study, we propose an expanded methodology, whereby a pre-trained model from an external dataset is utilized and subsequently fine-tuned through the application of transfer learning. This approach is employed for the identification of edge-on galaxies within the wide-field DESI Legacy Imaging DR10 survey. The resulting sample is referred to as EGIDE (The Edge-on Galaxies in the DESI Survey) (https://www.sao.ru/edgeon/catalogs.php?cat=EGIDE (accessed on 10 June 2026)). The EGIDE catalog, which we present here, contains more than 145,000 edge-on galaxies. The methodology employed here can be successfully applied to forthcoming Rubin, Roman, and Euclid data, and the EGIDE sample size is analogous to the expected sizes of future samples from these telescopes.

The present work is structured as follows. In Section 2, we describe the DESI DR10 data and outline the procedure used to detect edge-on galaxies. Section 3 presents an overview of the photometry extraction, its precision and the obtained results. In Section 4, we list information about distances to the galaxies in EGIDE. Section 5 discusses the completeness of our sample. In Section 6, we discuss our results, specifically the galaxy color-magnitude relation in Section 6.1, the color-flattening relation in Section 6.2, the mass-disk thickness relation in Section 6.3, and the number of galaxies versus flattening in Section 6.4. Finally, Section 7 summarizes our conclusions. In the Appendix A, Appendix B, Appendix C and Appendix D we provide additional details about neural network training, data cross-validation, and the methods used.

Throughout this paper, we assume a standard flat ΛCDM cosmology with ${\Omega }_m=0.3$, ${\Omega }_{\Lambda }=0.7$ and ${\mathrm{H}}_0$ = 70 km/s/Mpc, and use the AB magnitude system.

2. EGIDE Sample Building

As a source of observational images, we use the DESI Legacy Imaging Surveys in this work. The DESI Legacy Imaging Surveys (DESI Legacy) represent a comprehensive astrophysical campaign designed to provide the essential imaging dataset for target selection of the Dark Energy Spectroscopic Instrument (DESI; [86]). Initially, DESI Legacy delivers a uniform imaging coverage of approximately 14,000 square degrees of the sky in the optical $grz$ bands, combining data from three individual telescopes: the Dark Energy Camera Legacy Survey (DECaLS; [87]), the Beijing–Arizona Sky Survey (BASS; [88]), and the Mayall z-band Legacy Survey (MzLS; [86]). These observations are complemented by photometry at infrared wavelengths using data from the Wide-field Infrared Survey Explorer (WISE) mission, thereby providing spectral energy distribution coverage from 0.4 μm to 4.5 μm. Additional data are provided in catalogs obtained using the TRACTOR probabilistic astronomical source detection algorithm [89]. Another significant source of information is the recently released DESI Data Release 1 (DESI DR1), which presents spectroscopic redshift measurements for millions of galaxies [90].

Throughout this work, we use images from the latest version of DESI Legacy available, which is DR10 (For details about the release, see https://www.legacysurvey.org/dr10/description/ (accessed on 10 June 2026).). The two most notable changes in this release are, first, the inclusion of i-band observations for the first time, and, second, the extension of the southern ($\delta \le$ 32.375°) part of the survey [91]. The survey footprint is well represented in Figure 1. It is easy to see that the full DESI DR10 covers significant fractions of both the northern and southern celestial hemispheres and covers more than 20,000 square degrees of the sky. In addition to its extensive coverage, DESI Legacy DR10 has another advantage: the significant depth of the images, which reaches on average 28.5–29 mag arcsec⁻² within a 10 arcsec square box at the $3\sigma$ level in the g-band [53].

All coadded observations in DESI Legacy are grouped into so-called “bricks”. Each brick has an approximate size of 0.25° × 0.25° and is defined in terms of RA/DEC coordinates. There are 366,898 unique bricks in DR10, which have some small border overlap, and this is the reason why we see duplicate detections later. Each image has an identical pixel scale of 0.262 arcsec/pix. The median PSF FWHM is approximately 1.29 arcsec, 1.18 arcsec, and 1.11 arcsec for $grz$, respectively [86]. Note that starting from DESI DR9, the PSF model includes an inner and an outer component, where the latter represents large extended wings by a power law for g and r bands or a Moffat profile for z-band (https://www.legacysurvey.org/dr10/psf/ (accessed on 10 June 2026)).

To construct a sample of edge-on galaxies in the DESI survey, we used an approach similar to the one used in [79] to make the EGIPS database [78]. We downloaded all available images (“bricks”) in all available bands, and then performed an initial object search using the SExtractor package [92]. During this search, we excluded objects that were too small by allowing the SExtractor parameter A_IMAGE to be greater than 10 pixels (which roughly corresponds to a lower limit of the Kron radius of 9 arcsec). We also excluded objects that were too round (ELLIPTICITY < 0.4), as they usually cannot represent edge-on galaxies, and objects that were too elliptical (ELLIPTICITY > 0.95), because no real galaxies showed such ellipticities, but only image artifacts such as leaked signal from oversaturated bright stars. This preliminary stage resulted in $1.4\times {10}^6$ objects. To find edge-on galaxies in this catalog, we utilized an artificial neural network classifier.

All necessary technical details about the neural network training are given in Appendix A. Here, for convenience, we provide only some of them. The main difference from our previous work [79] is that we trained a neural network classifier from scratch using our training set of edge-on galaxies solely, and not edge-on galaxies. In this work, we use another approach called fine-tuning [93], where we start with a neural network already pretrained on a broad generic task and then adapt it to our task of edge-on galaxies search. Starting from a pre-trained model instead of a random initial condition allows one to achieve a better convergence of the training process, especially for limited training samples [94]. The contemporary and upcoming wide-field surveys, such as Rubin, Roman, and Euclid, provide a plethora of data to build heavy foundation models that can be further fine-tuned to more specific tasks.

We use Zoobot as the base for our search. Zoobot [95] is a Python package that leverages deep learning to characterize the detailed appearance of galaxies. It is trained on around 92 million labels collected from Galaxy Zoo volunteers, who answered a series of questions about each galaxy’s morphology (see the example question tree in figure 4 in Ref. [96]). Zoobot learns to answer all of these diverse tasks simultaneously, acquiring a rich, transferable representation of galaxy structure that can be efficiently adapted to new related tasks. We trained nine independent models by repeatedly re-running the Zoobot fine-tuning process in order to reduce possible misclassifications of individual models. We used edge-on RFGC and EGIPS galaxies from the previous works as positive examples [79]; the training size is several thousand examples (see Appendix A for exact numbers). The performance of individual classifiers of the ensemble is significant, achieving 97% accuracy (see Figure A1). We select a galaxy as a candidate to be located edge-on if the majority of the ensemble models (at least five out of nine) vote for it.

Automated selection of galaxies works as a “black box” and, in principle, can be susceptible to errors of the first and second kind. Bright bars, stellar spikes due to the PSF, or other elongated structures can easily mimic edge-on galaxies; therefore, additional efforts are always needed to ensure the validity of the sample. The application of our neural ensemble resulted in a considerable number of false detections caused by horizontal lines due to CCD leakage around bright stars. This is a known problem in DESI Legacy, and examples of such stars can be easily found in the DESI Viewer (https://www.legacysurvey.org/viewer (accessed on 10 June 2026)). The number of such lines greatly exceeds the number of edge-on galaxies; thus, even a small fraction of false positive detections of them significantly contaminates the classification results. To solve this issue, we created a sample of 5000 such artifacts and trained another Zoobot classifier to separate these artifacts from edge-on disks. This genuinely fixed the problem.

To ensure the reliability of our results, as the last step of the sample building pipeline, we manually check all pre-selected candidates. From the 192,593 net-selected objects, we remove 43,378 or roughly 25% of the initial sample. These removals include not only clearly false detections but also cases where there is an edge-on galaxy within the cropped image that probably activates the network weights, as well as genuinely elongated objects that are nevertheless not edge-on (see examples in Figure A2). We also remove relatively few duplications caused by brick overlap (see above). We want to emphasize that a quarter of false detections constitutes only 3% of the initial sample pre-selected by SExtractor, and it is a satisfactory result considering the difficulty of some cases and the small training dataset. This rate can be clearly lowered with additional training, and future missions such as the wide-field surveys Rubin, Roman, and Euclid will clearly benefit from the EGIDE size, if it is used for updating weights.

The final Edge-on Galaxies in the DESI Survey (EGIDE) sample consists of 149,215 galaxies. Several examples of edge-on or nearly edge-on galaxies are presented in Figure 2 and in Appendix D. The sky distribution of all selected candidates is presented in Figure 1. It is clear that galaxies of our sample uniformly cover the DESI DR10 footprint without significant gaps. Overall, the distribution of EGIDE galaxies covers more than half of the sky. Note, however, that not all bands are available for every galaxy, with i-band images being the most deficient, as they cover only 72% of the entire sample.

In

Table 1. The size of the intersection of the EGIDE sample with the surveys we are comparing with. The first column shows the acronym for the survey; the next columns show the reference, the intersection size, and the median Kron radius $a_r$ of the galaxies in the intersection. The lower part of the table, below the horizontal line, is dedicated to the surveys included in the RCSEDv2 project.

Survey	Ref.	Size, N	${\mathit{a}}_{\mathit{r}}$ [Arcsec]	Survey	Ref.	Size, N	${\mathit{a}}_{\mathit{r}}$ [Arcsec]
RFGC	[76]	135	32.1	EGIS	[77]	2726	22.9
EGIPS	[78]	8294	22.6	HyperLEDA	[97]	60,838	16.0
DESI DR1	[97]	30,833	13.6	REGALADE	[98]	145,876	13.6
RCSEDv2	[97]	30,163	16.0
SDSS	[21]	17,027	16.0	FAST	[99]	406	26.3
2dFGRS	[100]	3869	14.3	2dFLenS	[101]	381	13.8
6dFGS	[102]	3067	22.1	Hectospec	[103]	279	13.3
eBOSS	[21]	2392	14.8	CfA	[104]	125	19.2
LAMOST	[105]	1621	16.8	UZC	[106]	55	49.3

, we provide the intersection of the EGIDE sample with various extragalactic databases, including previously constructed catalogs of edge-on systems such as RFGC, EGIS, and EGIPS. We provide this information in case the reader is interested in edge-on galaxies within a specific survey and may be planning future research using this information.

3. Photometry

3.1. Photometry Making

To obtain photometric parameters of the galaxies in our sample, we used the SExtractor software package (version 2.24.2) [92]. To perform photometry, we downloaded images of each galaxy in all available bands, using preliminary estimates of its coordinates and sizes that were obtained during the initial sample construction. To reduce the contamination of galaxy photometry by foreground stars, we decided to mask them out (even though the deblending algorithm of SExtractor effectively separates objects from each other, in some cases, it fails to do so when a star is projected onto a galactic body). To make a stellar mask, we queried the Gaia database [107] to obtain coordinates and magnitudes of stars in the image and then covered each star with a circular mask whose radius depends on the star’s magnitude in the G band. We empirically found that a mask with radius

$$r\left(G\right)\left[\mathrm{arcsec}\right]=9.33exp\left(-0.3(G-12.7)\right)+2.10$$

effectively covers stars over the entire range of magnitudes. For consistency, we used the same mask for all bands.

Once the mask was ready, we automatically ran SExtractor for all bands and obtained a set of photometric parameters, including coordinates, sizes, fluxes, and position angles. To illustrate the photometry results, we used the estimates of geometric parameters (central coordinates, position angles, and major- and minor-axes sizes) to draw ellipses that contain the Kron fluxes. The resulting ellipses produced by SExtractor were checked (Images with them are available at the https://www.sao.ru/edgeon/catalogs.php?cat=EGIDE site (accessed on 10 June 2026).). There are some problematic situations where a close or foreground object distorts the ellipse boundary, or rare situations where the segmentation process splits objects into pieces. These often result in severe variation in ellipse properties between bands, such as position angle, and most suspicious cases can thus be easily identified; see examples of such cases in Figure A2. We provide an additional test, where we check the difference in the position angle of ellipses between bands and find that a difference of more than 2° occurs for ∼9000 cases, which is a small fraction. We validate all results without such problematic cases. Overall, the quality of our SExtractor aperture photometry was decent, and the comparison with other surveys in Appendix B strengthens this point. It is important to mention that, as was noted in [108], z-band images in the northern portion of the DESI footprint are significantly affected by pattern noise subtraction. It is not an issue in this work, since z-band photometry was not used in the analysis; however, the z-band magnitudes presented in EGIDE should be used with caution.

In total, for most of the galaxies in EGIDE, we measured the semi-major a and semi-minor b axes of the ellipse, its position angle and ellipticity, as well as the Petrosian [109] and Kron [110] magnitudes in each available image and band. We compare the results with other available measurements in Appendix B and find overall a reasonable match (see Figure A5). All magnitudes were corrected for Galactic extinction using the GMS25 map from Ref. [111]. This map gives similar values as the widely used recalibration of SFD data in Ref. [112], but was constructed with a hundred times more stars than the one used for SFD recalibration. We additionally check that usage of [112] extinctions, accessed via dustmaps [113], will not alter the results. The exact corrections were calculated from $E(B-V)$ reddening for $R_V=3.1$ using table 6 coefficients in Ref. [112]. It is important to note that edge-on galaxies obviously have significant internal extinction, but it is difficult to measure (cf. [114]). We derived the K-correction in $griz$ bands using the kcorrect software (Available at https://github.com/blanton144/kcorrect (accessed on 10 June 2026), we use version 5.1.9.) [115]. The estimated median K-correction is around 0.2 mag in the g-band, 0.1 mag in the r-band, and 0.05 mag in the i-band. Finally, corrections for SDSS stellar magnitudes to account for differences in bands with DESI Legacy were estimated using the B4–B6 equations from Ref. [86].

3.2. Oblateness Estimation: PSF Correction and Bias Sources

Oblateness estimates $q=b/a$ made by SExtractor are affected by various effects: image smearing by the PSF, finite resolution, and bulge impact all make small galaxies appear “rounder”. Since we use oblateness estimates for analysis throughout this work, in this subsection, we discuss the scales of these effects and how they can be reduced in our measurements.

Image smearing by the PSF alters the observed brightness distribution of astronomical objects, resulting in systematic biases of measured structural parameters, including major and minor axis lengths. As a consequence, the observed axis ratio is shifted toward rounder values and needs to be corrected. The values a and b of the major and minor axes presented in this work are based on the A_IMAGE and B_IMAGE output parameters of SExtractor, which are derived from the square roots of eigenvalues of the covariance matrix ${\widehat{C}}_{\mathrm{object}}$ of the object brightness distribution. Convolution of a brightness distribution with a kernel results in the addition of their covariance matrices:

$${\widehat{C}}_{\mathrm{observed}}={\widehat{C}}_{\mathrm{object}}+{\widehat{C}}_{\mathrm{PSF}},$$

therefore, the intrinsic covariance matrix can be estimated as

$${\widehat{C}}_{\mathrm{object}}={\widehat{C}}_{\mathrm{observed}}-{\widehat{C}}_{\mathrm{PSF}}.$$

Assuming that the PSF can be approximated by a circular Gaussian with dispersion $\sigma$, the intrinsic (deconvolved) object sizes can be estimated as

$$\mathrm{A}\_{\mathrm{IMAGE}}_{\mathrm{object}}^2=\mathrm{A}\_{\mathrm{IMAGE}}_{\mathrm{observed}}^2-{\sigma }^2,$$

$$\mathrm{B}\_{\mathrm{IMAGE}}_{\mathrm{object}}^2=\mathrm{B}\_{\mathrm{IMAGE}}_{\mathrm{observed}}^2-{\sigma }^2.$$

And thus:

$$q=\left({\displaystyle \frac{b}{a}}\right)=\sqrt{{\displaystyle \frac{\mathrm{B}\_{\mathrm{IMAGE}}_{\mathrm{observed}}^2-{\sigma }^2}{\mathrm{A}\_{\mathrm{IMAGE}}_{\mathrm{observed}}^2-{\sigma }^2}}},\mathrm{where}\sigma ={\displaystyle \frac{\mathrm{FWHM}}{2\sqrt{2ln2}}}.$$

Therefore, we correct all a and b values using the appropriate PSF $\sigma$ of the DESI Legacy survey in each band. We note that this correction is approximate because real galaxy profiles and the PSF are generally non-Gaussian, while SExtractor moments are computed from weighted pixel distributions.

The effect of this correction is presented in Figure 3, where panel (a) shows the distribution before correction and panel (b) shows it after correction. We can clearly see how, after correction, the q values become smaller, especially for galaxies with small sizes, which appear rounder due to the PSF smearing effect. Now both small and large galaxies demonstrate thin objects, and the isolines appear straighter, as expected. We additionally filter out galaxies with q smaller than $0.1$, a value interpolated from the lower boundary of larger galaxies, to remove those for which the correction results in unphysically small values.

After correction, we can obtain unrealistically small b values due to PSF variation and resolution effects, which are limited by the signal-to-noise level and other factors. To estimate the number of affected objects, we performed a numerical experiment. In this experiment, we created images of artificial galaxies with given values of a and b and then performed SExtractor photometry in the same way we did for real galaxies in our sample. We then compared the true (input) values of galactic parameters with the measured ones. Since the rounding effect should depend strongly on both the true values of galaxy size and its axis ratio (small and thin galaxies should be more affected by rounding), we performed this experiment for a range of combinations of a and $q=b/a$.

We used the Imfit MakeImage [116] package to create 5000 model images of a galaxy with a pure exponential brightness distribution, normalized to magnitudes of 16 and 17 in the r-band (which are typical for galaxies in our sample; see the next section), convolved them with the r-band PSF image of the Legacy Survey, and then co-added them to a random position in a real DESI Legacy field in order to simulate the impact of the noise from real images. We then used SExtractor to perform photometry and obtain observed values of the a and b parameters.

The results of this experiment are presented in Figure 3, panel (c), where for each combination of observed values of a and $b/a$, we show the relative error of the q measurement:

$${\displaystyle \frac{q_{\mathrm{measured}}-q_{\mathrm{true}}}{q_{\mathrm{true}}}}.$$

As expected, the error in the measured q value is positive, i.e., galaxies are observed to be rounder than they are in reality, and the smaller and thinner the galaxy, the larger the error. The errors for the 16th and 17th magnitude models are similar and can be used interchangeably. Using the isolines as presented, we can arguably select only galaxies with a relatively small error in q estimation. It is also clear from this figure that the majority of simulated galaxies in this experiment have $\delta q<15\%$ after PSF correction. Note also the strong agreement between the density distribution of the estimated $b/a$ after correction and the presented isolines.

Last but not least, it is important to discuss here the bulge influence, which can increase the measured thickness if not taken into account. The bulge component light can be properly measured only after photometric decomposition, which is a tedious task if performed well. We will try to glimpse its effect indirectly using several approaches below.

In principle, the knowledge of morphological Hubble types can provide the needed $B/T$ (see, for example, figure 8 in [77]). For some edge-on galaxies in the intersection with the HyperLEDA database [117], see Table 1, we have estimated Hubble type t. There are only several thousand objects in total with available t, which are not enough for a full analysis, but in principle, this information can be used for the estimation of the bulge influence on q. According to the HyperLEDA database, in the intersection, we have 176 Sa, 289 Sb, and 1863 Sc galaxies. For them, we can estimate oblateness q and its standard deviation equal $0.26\pm 0.10$, $0.24\pm 0.08$, and $0.24\pm 0.07$, respectively. If we assume $B/T$ values similar to those found from photometric decomposition in EGIS [77] (see figure 8), then the bulge-to-total ratio will be approximately 0.6/0.3/0.2 for Sa/Sb/Sc bins accordingly. Given that there is no drastic difference in linear sizes between galactic types under consideration, we can assume only a moderate influence of bulge on q.

Bulge influence can potentially be approximated by metrics of galactic bulge strength, such as the concentration index C, the Sérsic index n, or the so-called Gini-$M_{20}$ bulge parameter (GMB), which is less sensitive to dust and mergers [118]. All of these, however, require calculating additional nontrivial parameters, whose estimation may also be unstable. We utilize the advantage of the already calculated CAS [119] statistics for EGIPS data and estimate the GMB parameter as $GMB=-0.693\times M_{20}+4.95\times \mathrm{Gini}-3.96$ (see [118]). A GMB value greater than zero indicates greater bulge domination, while a lower value indicates greater disk domination. We see a statistically significant anticorrelation between GMB and $a/b$ in EGIPS galaxies, with Spearman’s coefficient $\rho =-0.47$, and $\rho =0.43$ for GMB versus $(g-i)$ color. This could provide a clue about bulge influence, which may reasonably be expected to be of the same order in the EGIDE sample.

Finally, we carried out the same experiment as above with a model image of a galaxy, but with an additional bulge component to estimate the relative error of the q measurement. All parameters were similar, and we added a Sersic bulge with $B/T=0.2$, which is expected to be a typical value in the sample (see figure 6 in [77]). The results are presented in Figure 3, lower right. As one might expect, for smaller galaxies the effect is significant; after PSF correction, it reaches a relative error of 30–50%. The same limit that we impose for reliable q estimation gives a conservative 25% error in q, which is smaller than the variance within individual Hubble types, as we showed previously.

To ensure the validity of the obtained results, in all places where it matters (Section 6.2, Section 6.3 and Section 6.4), we use only the subsample with $b/a$ estimated above the threshold line presented in Figure 3, consisting of 61,775 galaxies.

3.3. Photometry Results

A summary of some of the obtained photometry results and a comparison with the analogous EGIPS measurements are presented in Figure 4. With the green line in Figure 4, we also show properties for a $cz$-limited subsample of EGIDE galaxies between 5000 km/s and 15,000 km/s (see Section 5 for justification). In the top left panel, we show the angular Kron radius in arcseconds. The EGIDE distribution is shifted toward smaller angular sizes than EGIPS, with the average size almost twice as small (13.5 arcsec versus 22 arcsec). The reason is that we detect more distant galaxies in the DESI Legacy (see Section 4). Despite this, there are still many extended galaxies in our data, including 12,599 objects larger than the EGIPS median of $a>22$ arcsec in the r-band. Note that the size parameter presented here differs from figure 6 in [78], where the semi-major axis $a_r$ was shown before multiplication by the Kron coefficient.

The top right panel in Figure 4 shows the distribution of the apparent Petrosian magnitude in the r-band. The EGIDE and EGIPS distributions are drastically different: there are almost no galaxies with $m_r\le 16$ mag in the new sample, and the distribution has a limit at $m_r\approx 20$ mag. The absence of galaxies with bright apparent $m_r$ does not mean that EGIDE contains fainter objects than EGIPS. Quite the contrary, we actually have more massive and more luminous galaxies in EGIDE (Compare also with Figure 7 in the Section 5.); they are simply much more distant (see Figure 5).

In the bottom left panel of Figure 4, we show the $(g-r)$ color distribution from EGIPS and EGIDE in the same photometric system. Here, it is evident that the majority of galaxies in our sample exhibit rather red colors. These optical colors are similar to those from EGIPS: its median value $(g-r)=0.60$ mag divides the EGIDE sample in half, i.e., close to its median. At the same time, for the volume-limited sample from EGIDE (green line), we see two clearly separated peaks, which are absent in the other distributions. Thus, it is clear that from the colors alone, red and blue galaxies are well separated only in a small $cz$ range.

In the bottom right panel in Figure 4, we show the $b/a$ distribution in the r-band. We can see that indeed all galaxies in EGIDE are highly flattened objects, with a median value of $q\approx 0.25$. We want to emphasize again that, as Section 2 describes, we do not use q information directly during selection. Since many of the obtained $b/a$ values are small, these are indeed edge-on galaxies. Compared to the previous EGIPS sample [78], the new distribution is shifted by $+0.05$–$+0.10$ in terms of quartiles. Another difference is a larger tail of galaxies with rather round isophotes of the ellipses produced by SExtractor in EGIDE. As we discuss in Section 3.2, these differences may be caused by various factors, which may relate to the deeper surface brightness limit in DESI images compared to Pan-STARRS, PSF, and bulge influence, or to the fact that some candidates have orientations other than edge-on. Nevertheless, we still have many thin galaxies in EGIDE. For instance, in [41], the condition $a/b>5$ is used as the boundary for “superthin” galaxies. We prefer to traditionally define superthin galaxies as those with major-to-minor axial ratios $a/b>10$ [17,18,38], and refer to objects with $a/b>5$ as “thin.” We have 8,278 thin galaxies in EGIPS, while the same criterion in the newly formed EGIDE yields an almost three times larger sample consisting of 22,958 objects. For “superthin” galaxies with $q<0.1$, the proportion is reversed: 63 and 32 in EGIPS and EGIDE, respectively.

For all bands under consideration, we show the measured q and the standard deviations in

Table 2. Number of galaxies and the median $q=b/a$ value for the EGIPS and EGIDE samples. Values are presented for each band under consideration for the whole sample, and additionally for the large ($a_{band}>20$ arcsec) galaxies.

	EGIPS		EGIDE		EGIPS: $\mathit{a}>{20}^{\prime \prime }$		EGIDE: $\mathit{a}>{20}^{\prime \prime }$
Band	N	q	N	q	N	q	N	q
g	16,550	0.205 ± 0.070	148,064	0.239 ± 0.079	10745	0.180 ± 0.050	25,351	0.203 ± 0.076
r	16,557	0.210 ± 0.057	142,076	0.249 ± 0.079	10226	0.186 ± 0.046	17,897	0.212 ± 0.080
i	16,558	0.215 ± 0.055	103,210	0.241 ± 0.075	9690	0.191 ± 0.046	9834	0.206 ± 0.075
z	16,556	0.214 ± 0.058	145,070	0.240 ± 0.073	8524	0.187 ± 0.045	11,625	0.199 ± 0.069

. Besides the already mentioned difference between EGIPS and EGIDE, we also note that the flattening in the g-band is, on average, smaller than in the r-band in all cases, and that the dispersion is largely similar between bands. Galaxies with a semi-major axis larger than 20 arcsec show smaller relative thickness, probably due to the lower impact of the PSF roundening, as Table 2 suggests, and have lower uncertainties in the EGIPS case.

Last but not least important is a comparison with independent oblateness q measurements available in other surveys. We provide such a comparison in Appendix C. In total, we compare with ellipticity components from Dark Energy Spectroscopic Instrument Data Release 1 (DESI DR1, [90]), $R2/R1$ (minor-to-major axis ratio) measurements in REGALADE [120], the same axis ratio logr25 from HyperLEDA data [117], and also statistical q modeling from [26]. As Figure A6 demonstrates, all of these measurements have distributions very similar to what we measure here, thus validating the small q values found in this work. One visible peculiarity is the significantly smaller values obtained from DESI DR1, which may likely relate to the methodology they used: in all examined cases, probabilistic TRACTOR models result in much greater a than observed. Another interesting feature of Figure A6 comes from the comparison with the [26] model for galaxies in the $log{M}_{★}$ mass bin from 9.0 $M_{\odot }$ to 9.5 $M_{\odot }$. We can see that $b/a$ for subsamples obtained according to different methods and data actually fit the left and right wings of the reconstructed model in this mass bin relatively well. This may imply that the oblate model in Ref. [26] fits the actual thickness of galaxies we obtain directly for edge-on galaxies very well. On the other hand, this may be a coincidence, and that requires further testing (see also discussion in Section 6.3).

4. Redshift Distribution

For a better understanding of the objects we work with, for example, to derive stellar masses, we need information about their distances. To obtain this, we use a recently compiled database called REGALADE (the Revised Galaxy List for the Advanced Detector Era, [120]). This new all-sky catalog combines data from various surveys (DESI, Pan-STARRS, Cosmicflows, SDSS, etc.) for more than 80 million galaxies out to a distance of 2000 Mpc ($z\le 0.37$ or $cz\le$ 111,000 km/s).

For 98% of the galaxies in EGIDE (145,876 objects), we find z information within 5 arcsec of REGALADE sources, and for the majority of our data (137,678 galaxies or 92%), the Euclidean distance between coordinates is less than 1 arcsec. Since the authors of Ref. [120] mix all spectroscopic, photometric, and redshift-independent distances in the same table, this could potentially bias the results. To ensure reliability, we additionally collect z measurements from three other sources. The first is the HyperLEDA database [117], from which we take data in the CMB frame of reference. The second source is the DESI DR1 [90], which contains spectroscopic redshift measurements for $z<4$. However, since DESI DR1 is included in REGALADE, we use the specific Stellar Mass and Emission Line Catalog Value-Added Catalog (EMLines VAC), which contains information for galaxies with reliable redshift measurements. In both cases, we match galaxies within 1 arcsec distance. The third source we consult is the second version of the Reference Catalog of Spectral Energy Distributions of galaxies (RCSEDv2 (https://rcsed2.voxastro.org/ (accessed on 10 June 2026)), [121]). We find significant agreement between the redshifts z from the different databases; see the comparison in Appendix B. Throughout the paper, we verify that all dependent results remain the same when using different z sources.

The resulting redshift distribution is presented in Figure 5. It is clear that the effective depth of the EGIDE catalog is significantly increased compared to EGIPS: the median velocity is three times larger, equal to ∼33,750 km/s, which corresponds to a luminosity distance of about 520 Mpc for the adopted cosmology. Note also that the $cz$ distribution of EGIDE galaxies has a long tail that extends to the REGALADE boundary of $D=2000$ Mpc. Compared to EGIPS, we also see slightly fewer large galaxies in the $cz<5000$ km/s covered local volume, which is due to the selection criteria (see also Table 1). These galaxies are too large to be detected by the neural network and were manually added in EGIPS. Thus, we want to emphasize that EGIPS and RFGC can be used as good supplements to complete EGIDE in the $cz<5000$ km/s velocity range.

5. Completeness

To assess the completeness of the constructed EGIDE sample, we use several approaches. The first is to perform the classical $V/V_m$ test, also known as the luminosity-volume test, which indicates whether a sample of objects is uniformly distributed in space [122], as expected from their fluxes. For the 40% most luminous galaxies in EGIDE (58,302 objects), we obtain $V/V_m=0.4636\pm 0.0012$ in the g-band, indicating a high degree of completeness for this subsample. Inclusion of galaxies with lower flux results in a more incomplete subsample ($V/V_m=0.4398\pm 0.0010$ for the 50% quantile). From the comparison with EGIPS in Section 3, we also know that the size-selected sample is also incomplete: $V/V_m=0.5781\pm 0.0016$ for galaxies with a Kron r-band radius larger than 15 arcsec.

Two additional tests are implemented and shown in Figure 6. Here, we plot the $logN$-$log{a}_r$ cumulative distribution (completely analogous to figure 3 in [78]), which should have a slope of $-3$ for a uniform distribution of galaxies if we neglect cosmological effects. We can see that indeed the average slope is close to $-3$, but there is a knee in the distribution around $a_r\approx 17$ arcsec. Before the knee, the slope is closer to $-4$, and after it, it is $-2.2$. For the populations of blue $0.2<(g-r)<0.6$ mag and red $0.7<(g-r)<1.2$ mag galaxies (see motivation in Section 6.1), the situation is the same. On the right side of Figure 6, we plot the $logN$-apparent Petrosian magnitude cumulative distribution for all $griz$ bands. For a uniform distribution, the slope of such a distribution should be $0.6$. For the part between 12 mag and 17 mag, where the increase is steady, we see slopes slightly larger, around $0.65$–$0.7$, which may indicate an excess of faint galaxies in EGIDE. As in Figure 4, we see the limiting magnitude to be $m=19$ mag.

As a final test, we plot stellar mass versus redshift in Figure 7. Total stellar mass $M_{★}$ estimates were made using the calibrations from [123] with DESI g and r absolute stellar magnitudes. These estimates are in good agreement with SED-estimated masses from CIGALE [124] from DESI DR1 [90] and with $M_{★}$ estimates in REGALADE; see details in Appendix B. The distribution shows a traditional shape with a visible selection bias, where only more massive (i.e., more luminous) galaxies are visible at larger distances. Comparing with the EGIPS sample, we can clearly see that EGIDE lacks some galaxies in the nearby volume but has significantly improved depth in terms of $cz$ and also contains more massive galaxies (median value $log{M}_{★}/M_{\odot }=10.56$ in EGIDE versus $log{M}_{★}/M_{\odot }=10.34$ in EGIPS). We schematically mark the observational limitations for both surveys with lines in Figure 7. It is reasonable to analyze two distinct subsamples: (1) galaxies with 5000 km/s $<cz<$ 15,000 km/s, which include both faint and luminous objects and correctly represent the flux dynamic range; and (2) galaxies in the $log{M}_{★}/M_{\odot }=10.7–11.4$ mass bin, which correctly represent the distance range. These subsamples can be considered volume- or mass-complete, and we test all the aforementioned results with both to see if there is any difference in the conclusions.

6. Discussion

In this Section, we present some derived results about the EGIDE sample. Specifically, we discuss the galaxy color–magnitude diagram, the color–flattening relation, and the disk thickness dependence on stellar mass.

6.1. Galaxy Color-Magnitude Diagram

In Figure 8, we present the $(g-r)$ color versus absolute magnitude $M_g$ diagram (CMD) for galaxies in EGIDE. We use the r-band instead of i because the latter is available for only 70% of the sample. For easier comparison with previous results, we show three separate redshift bins: the nearby bin $cz$ < 10,000 km/s, the intermediate bin 10,000 < $cz$ < 30,000 km/s, and the bin for 85% of all galaxies in the sample with $cz$ < 60,000 km/s. For comparison, we also overlay SDSS DR12 [125] data as contours for the same $cz$ bins, corrected to fit DESI magnitudes (see Section 3). In all three subpanels of the SDSS data, we can clearly see the red cloud, preferentially populated by early-type disk galaxies, which are also brighter, and the dimmer blue cloud, formed by spiral galaxies [126]. For $cz$ < 10,000 km/s, we see good agreement between EGIDE and SDSS DR12 data, with edge-on galaxies equally populating the red and blue clouds. For galaxies with larger redshifts (second and third panels), as well as for the entire EGIDE sample, we see, first, that galaxies tend primarily to occupy the red sequence with no clear presence of the so-called green valley, and second, that the color $(g-r)$ is shifted to redder and, to a lesser extent, to bluer values compared to SDSS photometry. This fact—that the SDSS contours are shifted toward bluer galaxies—may be explained by higher internal extinction in edge-on galaxies relative to galaxies with arbitrary inclinations, as well as by the deeper limit of the DESI Legacy observations. Note, however, that this effect should be more important for star-forming galaxies with high dust influence. The high prominence of red-sequence galaxies in Figure 8 may be due to a significant fraction of lenticular galaxies in the constructed EGIDE sample. Unlike late-type galaxies, edge-on lenticular galaxies may be harder to detect due to the absence of spiral arms or prominent dust lanes. Potentially, there can also be some elongated elliptical galaxies of type E7 with $q\ge 0.3$, but this type is rare, and the detection of S0 galaxies is much more probable.

Optical colors alone are not optimal to clearly separate red and blue clouds in CMD, with $NUV-r$ often implemented as a better choice [127]. However, when available, it is even better to use for this purpose the physically motivated star formation rate (SFR)-$M_{★}$ plane. In Figure 7, we show these values for 30,986 EGIDE galaxies with available SED models with estimated SFR from DESI DR1. Overlaid on the density distribution in Figure 7 are contours for the general sample from the DESI Early Data Release (EDR) (https://data.desi.lbl.gov/doc/releases/edr/ (accessed on 10 June 2026)). We choose to use EDR over DR1 data for practical reasons because it produces essentially the same contours but faster due to being 16 times smaller. To plot the contours, we select only morphological EDR data with exponential, Sérsic, and de Vaucouleurs models (EXP, SER, and DEV codes) and with $z<0.37$ to match the EGIDE distance volume. We can clearly see the star-forming main sequence (SFMS), which is essentially the equivalent of the blue cloud in the galactic CMD. A slightly less populated “red and dead” area with more massive early-type galaxies (ETGs) is also evident in Figure 7. We see that the EGIDE galaxies coincide well with the general population contours from DESI EDR. For later use, we mark the separation line between ETG and SFMS galaxies as $log\mathrm{SFR}=0.85log{M}_{★}-8.74$, which is similar to H$\alpha$-related local SFMS in [128].

6.2. Color–Flatness Relation

In Figure 9, we show $(g-i)$ color versus $a/b$ for the EGIDE galaxies. Despite the ten times larger sample size and other principal differences between the samples, we observe a similar picture to what was found for EGIPS [78]. The obvious cluster or overdensity of galaxies represents thick, $a/b\sim 3$–5, and red, $(g-i)\sim 1.2$–$1.3$ mag, objects. The thinner galaxies with $a/b>5$ are, on average, bluer by $0.2$–$0.4$ mag. Overall, the distribution is visually shifted toward a smaller $a/b$ compared to the analogous figure 10 for EGIPS [78], with a heavily visible “tail” around $a/b\sim 5$. The running medians in both samples are very close to each other, confirming the same trend. The result in Figure 9 holds if we use subsamples with only large galaxies ($a_r>15$ arcsec), as we show in the upper right panel.

Similar conclusions were already made previously for smaller samples of galaxies in $(g-r)$ color as well (e.g., figure 4 in [129] or figure 7 in [77]). The internal extinction variation due to a small inclination difference from the edge-on orientation cannot be the explanation, because when $a/b$ increases, we expect to find galaxies closer to edge-on, and thus they must be redder. Existing statistical models from [114] that connect extinction and inclination predict an effect that is similar in magnitude to what we observe, but reversed. Thus, there must be another explanation, which may potentially be the effect of the bulge, because light from a bulge would geometrically lead to a greater axis ratio (see Section 3.2). Indeed, as Figure 7 and Figure 8 clearly show, we have many galaxies with a significant spherical component. While whether disk thickness depends on morphology is a debated question [130,131,132], such a dependency could potentially have a similar effect on color (see figure 8 for EGIS in [77]). Note that in the SAMI survey [133], the authors measure that the $(g-i)$ color of bulges is only slightly redder by $0.12\pm 0.02$ mag than the color of disks, which they also attribute to higher metallicities rather than older stellar populations. Such a difference is rather small to be important for the measured effect if $B/T\sim 0.5$ or less.

We also plot in Figure 9 the star-forming galaxy subsample, selected as in Figure 7, to check whether such separation affects the $(g-i)$ versus $a/b$ relation. This results in an almost constant shift between the two subpopulations of about $0.15$–$0.2$ mag, even for thin galaxies with $a/b>7$, where galaxies around the SFMS are bluer (see lines in the upper right panel of Figure 9). This experiment shows that there is indeed a difference between these subpopulations, but the trend is almost the same for both. Note that even for star-forming galaxies, those with larger bulges should show redder colors; therefore, selecting a subsample of SFMS galaxies cannot fully remove the bulge effect.

On the bottom row of Figure 9, we show the same q versus $(g-i)$ dependence for the mass-limited ($log{M}_{★}/M_{\odot }=10.7–11.4$, left) and $cz$-limited (5000 km/s $<cz<$ 15,000 km/s, right) subsamples. These subplots demonstrate interesting results. The most massive galaxies are indeed redder and lie above the median line for the whole EGIDE sample, and demonstrate an almost constant color of $(g-i)\approx 1.15\pm 0.1$ mag with flattening. On the other side, the volume-limited subsample shows clear bimodality with only a few galaxies in the transition zone between two distinct clouds. These two clouds are equally well populated; the redder one demonstrates rounder galaxies with $a/b\sim 3.5\pm 1.0$ and $(g-i)\approx 1.0\pm 0.14$ mag, while the blue galaxies are thinner with $a/b\sim 5.0\pm 0.9$ and a color of $(g-i)\approx 0.5\pm 0.2$ mag. Further investigation of this dependence is beyond the scope of this paper, but such a clear separation, combined with the fact that massive $log{M}_{★}/M_{\odot }>10.7$ galaxies have almost the same color regardless of axis ratio, clearly needs to be explained. One of the possible directions to explain such dependence is the well-known relation of dispersion versus stellar age population [134].

6.3. Mass–Thickness Relation

As another scientific application of the newly built EGIDE sample, we compare how the flattening q depends on the total stellar mass $M_{★}$. To ensure consistency of the methodology, i.e., that the same photometry is used for mass derivation, we use the [123] calibrations for mass derivation (see also Appendix B for details and comparison with other mass measurements). In Figure 10, we show results in the r-band across five orders of magnitude in mass bins. Despite the small $q\sim 0.2$–$0.3$ across the whole $M_{★}$ range, we can see in the Figure that q is not constant: it increases toward the high-mass end of the distribution. We also see a slight increase in $b/a$ for dwarf galaxies, but it is far less significant.

To validate this result, we compare it with other studies. Recently, in Ref. [26], a broad range of surveys (GAMA, DESI, ALFALFA) with hundreds of thousands of galaxies at arbitrary inclinations was used to infer the abundance of intrinsically flat galaxies needed to reproduce the observed abundance of highly elongated systems in projection. To do so, the authors in Ref. [26] used two statistical models for r-band images: an oblate model, where $a=b>c$, and a triaxial model, where $a>b>c$ (in this notation, $q=c/a$). In Figure 10, we show the peaks of the q distribution for the oblate model from their work (for mass bin $log{M}_{★}/M_{\odot }=9–9.5$, see their Figure 6; data for other mass bins were obtained via private communication with José Benavides). It is evident that while the statistically derived q values from [26] are smaller, which is expected by definition, we can see agreement between the trends in the two datasets. Their sample is limited to ${10}^{11}{M}_{\odot }$, but the steepness of the trend, the location of the minimum, and the slightly larger q values for dwarfs are all consistent (Spearman correlation $\rho =0.61$). This similarity between two explicitly different samples, when one represents direct measurements from observations and the other a statistical modeling approach, cross-validates both studies. Note, however, that measurements of the axis ratio q in both samples may suffer from PSF smearing, resolution, and bulge effects, as discussed in Section 3.2.

Several other findings support this result. For example, the lower envelope curve of $b/a$ versus $M_{★}$ for galaxies in SDSS DR7 from figure 1 in Ref. [135] should clearly describe only edge-on systems, and it is consistent with Figure 10. In Ref. [132], the authors perform photometric decomposition for a subsample of EGIS galaxies and find that $h_z/h_R$ increases for brighter galaxies in the SDSS i band (see figure 12, left). In Ref. [136], the authors note that dwarf galaxies, which are located on the left side of the mass distribution, are systematically thicker than the general population. Finally, for a subsample of our galaxies from DESI DR1, we plot $q_{\mathrm{DESI}}$, derived using the two ellipticity components ${ϵ}_1$ and ${ϵ}_2$ from the TRACTOR models, versus the same masses we use. Although these are automatic measurements and can be susceptible to errors, $q_{\mathrm{DESI}}$ broadly shows the same trend. However, the $q_{\mathrm{DESI}}$ values are lower, and the minimum is shifted toward larger galaxies.

The observed trend in Figure 10 can be justified at least qualitatively. If disks are in a submarginal stability regime with Toomre Q assumed constant and slightly above unity, then we can show, as in Ref. [137], that the vertical-to-radial disk scale is proportional to the ratio of disk mass to total mass: $z_0/h\propto M_d/M_{tot}\propto M_{★}/M_h,$ where $M_h$ is the mass of the dark halo, for example, $M_{200}$. If we take into account any stellar mass–halo mass (SMHM) relationship, for example, as presented in figure 14 in Ref. [138], we can broadly assume $M_{★}\propto M_h^{\alpha }$, where $\alpha \approx 1.4$ for $log{M}_h/M_{\odot }\le 12$. Substituting into the previous relationship, we can derive that

$$z_0/h\propto M_{★}/M_h\propto M_{★}^{1-1/\alpha }.$$

This can, in principle, justify the observed trend in Figure 10, if we assume that $b/a$ and $z_0/h$ are close and well-correlated values. Note, however, that the derived simple power law predicts an increase in flattening larger than observed for realistic $\alpha$; thus, it is probably also governed by additional processes.

Interestingly, it was shown in Ref. [139] for 49 edge-on bulgeless disk galaxies that $z_0/h$ is higher when rotation speed $V_c$ is lower, with a clear transition region around $V_c=120$ km/s. This is a reverse trend to the picture presented here and in Ref. [26], if we assume the influence of the bulge on $b/a$ to be small. Note, however, that the sample in Ref. [139] is rather small and that the decision to use bulgeless galaxies can, in principle, bias the results, since we can expect that processes that form or grow a massive bulge can also heat the disk [140]. On the other hand, if we roughly estimate $M_{200}$ using the approximation $M_{200}/M_{\odot }=1.074\times {10}^5\times V_c^{3.115}$ from Ref. [141] and the transition value $V_c=120$ km/s from [139], we obtain $M_{200}/M_{\odot }\approx 3\times {10}^{11}$. This value is close to the beginning of the SMHM peak in figure 14 of [138], showing that indeed such rotation speed $V_c$ can be in the transition region and affect $z_0/h$ due to different halo contributions, as described in the previous paragraph.

Of course, the trend presented in Figure 10 could be related to factors other than the disk, such as the spheroidal component. Thus, for galaxies at the high-mass end, we can expect to find lenticular galaxies, which may have a large bulge-to-total ratio $B/T$, which in turn will increase the b value. In Ref. [25], the authors found the fraction of stellar mass in the spheroidal component in TNG50 simulated galaxies and evaluated concentration versus stellar mass for galaxies from the HSC-SSP survey. They indeed found (figures 5 and 6) that the spherical component fraction shows a similar U-shaped trend with $M_{★}$. Another piece of evidence is Ref. [132], where figure 13 exactly shows that flattening $h_z/h_R$ increases as $B/T$ increases.

As another independent test of the spheroidal component influence, we can proceed as follows. Based on the SFR–stellar mass relation presented in Figure 7, we separate the star-forming main-sequence galaxies from the red cloud using the separation line shown in that figure. For each of these two classes, we present the mass-q relation separately on the right side of Figure 10. Note that these are not exactly the same galaxies as in the left subplot, because we have SFR information only for a subsample of 30833 objects, which is roughly 20% of the EGIDE sample size. In Figure 10, we see that galaxies from the so-called red cloud, which potentially have a larger spherical component, follow exactly the same previously demonstrated trend, as do galaxies from the main sequence. Similarly to that, if we instead split galaxies into blue and red according to the CMD presented in Figure 8 with colors $0.2\le (g-r)\le 0.6$ and $0.6\le (g-r)\le 1.2$, then for both types of galaxies we will see an increase similar to that presented in Figure 10.

6.4. Count Drop with Color

Previously, for the EGIPS sample, we detected a difference in the number distribution of galaxies with $a/b$ for galaxies of different colors (see figure 11 in Ref. [78]). Thus, the axis ratio functions of red ($1.0\le (g-i)\le 1.4$) and blue ($0.4\le (g-i)\le 0.8$) galaxies are noticeably different; i.e., the number of redder galaxies drops with increasing $a/b$ faster than for the bluer galaxies. Overall, the slope coefficient k in the relation $logN\propto -k\times a/b$ was found to form a U-shape with color (see figure 2 in Ref. [78] and also Ref. [142]).

Here in Figure 11, we present a similar analysis but for the $(g-r)$ color. Motivated by the CMD distributions in Figure 8, we present fits for $logN$ versus $a/b$ in the g-band for blue and red galaxies with conditions 0.2 mag $<(g-r)<0.6$ mag and 0.7 mag $<(g-r)<1.2$ mag, respectively. We indeed see the same picture as before: the decline of red galaxies is faster than that of the bluer ones. The effect is smaller than that presented in [78], probably because of the smaller and more distant galaxies, but it is still noticeable. The slope values k are larger because the number of galaxies in the $a/b<5$ bins is much larger than in EGIPS, while the right-side end of the distribution is similar to that in EGIPS (see also Figure 9 in Section 6.2). On the right panel of Figure 11, we present the slope k in various color bins. The rate of decline becomes gradually flatter for bluer colors and remains nearly the same for galaxies with $(g-r)<0.6$ mag and $(g-r)>0.7$ mag. This can be explained by the fact that younger and bluer stellar populations form thinner disks, as the well-known vertical dispersion versus age relation dictates [143,144]; thus, such galaxies have higher counts in the $a/b>7–8$ bins. Note that the same conclusions about k in $N\propto exp(-k\times a/b)$ can, in principle, be derived from Figure 9.

We also plot in Figure 11 the same dependency for the mass-limited and $cz$-volume-limited subsamples, defined in Section 5. For the latter, there are not enough galaxies with $(g-r)>0.9$ to fit the linear regression properly, so we limit the k estimation for it to this value. We can clearly see that for the volume-limited sample, the galaxy count N drop is almost the same: $k\approx 0.9$. For the mass-limited data, the decline coefficient k increases across the entire presented color range. From the comparison with the full EGIDE, it is clear that the peak value of $k\approx 1.1$, as well as the whole dependence for redder galaxies, is indeed defined more by massive galaxies.

7. Conclusions

We have presented the EGIDE (Edge-on Galaxies in the DESI Survey) catalog, a new large-scale sample of 149,215 edge-on galaxy candidates selected from the DESI Legacy Imaging Surveys DR10. The catalog is publicly available through the Edge-on Galaxy Database (https://www.sao.ru/edgeon/catalogs.php?cat=EGIDE (accessed on 10 June 2026)). EGIDE is approximately ten times larger than its predecessor, Edge-on Galaxies in the Pan-STARRS survey (EGIPS), and covers more than half of the sky, providing a homogeneous dataset for studying the vertical structure of disk galaxies and statistical properties of edge-on systems. The edge-on selection methodology combines a neural network search using a fine-tuned Zoobot model with manual visual inspection, resulting in a reliable sample.

For each galaxy in EGIDE, the catalog provides homogeneous SExtractor photometry in the $griz$ bands (Kron and Petrosian magnitudes, ellipse parameters), total stellar masses estimated using color-based calibrations, and redshifts for 98% of the sample, which were collected primarily from REGALADE. Cross-identifications with HyperLEDA, RCSEDv2, DESI DR1 and other surveys (see Table 1) are also included.

The EGIDE dataset has the following key properties:

• Galaxies in EGIDE have a median angular half-size of the major axis equal to 13.5 arcsec and are highly flattened objects with $q=b/a\approx 0.24$ (Figure 4, Table 2 and Figure A6). Measurements of the axis ratio q may suffer from PSF smearing, resolution, and bulge effects, especially for smaller galaxies. We correct for PSF smearing, impose a constraint on the relative error of q, and discuss the bulge influence in Section 3.2.
• Due to methodological reasons, the sample lacks very close and extended edge-on galaxies; thus, previous surveys of edge-on galaxies such as RFGC, EGIS, and EGIPS should be seen as complementary to the EGIDE dataset.
• The $V/V_m$ test indicates that the sample is mostly complete for the 40% most luminous galaxies ($V/V_m=0.4636\pm 0.0012$), with a depth limit of $m\approx 19$ mag in the r-band (Figure 6).
• The median redshift of EGIDE is $cz\approx \mathrm{33,750}$ km s⁻¹ and is three times larger than that of EGIPS (see Figure 5). The galaxies are distributed up to a distance of 2000 Mpc and are slightly more massive than those from EGIPS (median value $log{M}_{★}/M_{\odot }=10.56$ in EGIDE versus $log{M}_{★}/M_{\odot }=10.34$ in EGIPS, see Figure 7, left).
• Edge-on galaxies in EGIDE populate both the red and blue sequences (see Figure 7 and Figure 8, right). Compared to the SDSS DR7 reference sample, EGIDE galaxies are shifted to fainter absolute magnitudes and exhibit a broader range of colors, which is potentially explained by the greater depth of the DESI Legacy images and by internal extinction.

The analysis presented in this paper leads to the following results:

• The color–inverse flattening diagram confirms the previously found bimodality: a dense cloud of thicker ($a/b\sim 3$–5) and redder ($(g-i)\sim 1.1$–$1.3$ mag) galaxies is followed by a population of thinner ($a/b>5$) galaxies that are bluer by $0.2$–$0.4$ mag (Figure 9). This bimodality appears to be a fundamental property of disc galaxies, and is better visible for the volume-limited subsample, while massive galaxies form a single cluster with near-constant color. Despite the tenfold increase in sample size, the running medians of EGIDE and EGIPS are nearly identical on these plots, but there is also a noticeable difference between galaxies from the star-forming main sequence (SFMS) and non-SFMS areas.
• The flattening $q=b/a$ increases with total stellar mass $M_{★}$ at the high-mass end (Figure 10, left). The same trend is independently recovered from the statistical oblate models of Ref. [26], and from DESI DR1 ellipticity measurements, cross-validating our approach. We found that the increase in $b/a$ is driven by both red cloud early-type galaxies, as well as SFMS galaxies (Figure 10, right).
• The number of red galaxies drops with increasing $a/b$ faster than that of blue galaxies (Figure 11). The decline coefficient k in $N\propto exp(-k\times a/b)$ is $k\approx 1.10$ for red galaxies and $k\approx 0.97$ for blue galaxies in the full sample. For the volume-limited subsample, $k\approx 0.9$ for all $(g-r)$ colors, while for the mass-limited subsample, k increases across the entire color range, indicating that the change in k is primarily driven by massive systems.
• Comparisons with HyperLEDA, DESI DR1, RCSEDv2 and REGALADE show good agreement in redshift (NMAD ≈ 20–170 km/s, Figure A3), total stellar mass $M_{★}$ (NMAD ≈ 0.15 dex, Figure A4), and apparent magnitude (NMAD ≈ 0.10 mag, Figure A5). Flatness measurements from these independent sources also confirm the small $b/a$ values found in EGIDE (see Figure A6).

In summary, EGIDE provides a large, reliable, and publicly accessible catalog of edge-on galaxies. The first scientific results strengthen, confirm, and extend earlier findings on color-flattening and mass-flattening relations, and highlight the hypothetical importance of the spheroidal component in shaping the observed flattening of massive galaxies. The methodology developed here can be directly applied to upcoming wide-field surveys such as Euclid, Roman, and LSST, where samples of similarly large size will become routinely available. The EGIDE catalog is well-suited for the search for specific objects (e.g., superthin galaxies, polar rings, boxy bulges, galactic fountains, etc.), as well as for future studies of disk vertical structure and comparisons with cosmological simulations with robust statistical analyses across mass, color, and environment.