Galaxy Zoo 2: detailed morphological classifications for 304,122 galaxies from the Sloan Digital Sky Survey

Introduction

The Galaxy Zoo project (Lintott et al., 2008) was launched in 2007 to provide morphological classifications for nearly one million galaxies drawn from the Sloan Digital Sky Survey (SDSS; York et al., 2000) Main Galaxy Sample (Strauss et al., 2002). This scale of effort was made possible by combining classifications from hundreds of thousands of volunteers via a web-based interface. In order to keep the task at a manageable level of complexity, only the most basic morphological distinctions were requested, enabling the separation of systems into categories of elliptical (early-type), spiral (late-type) and mergers.Galaxy Zoo is archived at http://zoo1.galaxyzoo.org. Following the success of this project (Lintott et al., 2008, 2011), the same methodology of asking for volunteer classifications was launched in 2009 with a more complex classification system. This paper presents data and results from this second incarnation of Galaxy Zoo, called Galaxy Zoo 2 (GZ2). These data comprise detailed morphologies for more than 300,000 of the largest and brightest SDSS galaxies.The Galaxy Zoo 2 site is archived at http://zoo2.galaxyzoo.org.

While the morphological distinction used in the original Galaxy Zoo (GZ1) – that which divides spiral and elliptical systems – is the most fundamental, the motivation for GZ2 was that galaxies demonstrate a much wider variety of morphological features. There is a long history of enhanced classifications (see Buta 2013 for a historical review), but the most well-known approach (Hubble, 1926) included a division between barred and unbarred spirals, resulting in the famous ‘tuning fork’ diagram. Further distinctions ordered ellipticals based on their apparent roundness and spirals on a combination of tightness and distinction of the arms and size of the central bulge. Along the late-type sequence, these traits are often correlated with physical parameters of the systems being studied (Roberts & Haynes, 1994), with spirals becoming (on average) redder, more massive, and less gas-rich for “earlier” locations in the sequence.

Morphological features can clearly provide insights into the physical processes that shape the evolution of galaxies. Most obviously, merger features reveal ongoing gravitational interactions, but even the presence of a central bulge in a disk galaxy is likely to indicate a history of mass assembly through significant mergers (Martig et al. 2012 and references therein). On the other hand, galactic bars and rings reveal details of slower, secular evolution and stellar orbital resonances. For example, bars, are known to drive gas inwards and are related to the growth of a central bulge (reviews are given in Kormendy & Kennicutt 2004; Masters et al. 2011). Careful classifications of morphological features are thus essential if the assembly and evolution history of galaxies is to be fully understood.

Traditional morphological classification relied on visual inspection of small numbers of images by experts (e.g., Sandage, 1961; Sandage & Bedke, 1994; de Vaucouleurs et al., 1991; Buta, 1995; Buta, Corwin & Odewahn, 2002). However, the sheer size of modern data sets (such as the SDSS Main Galaxy Sample) make this approach impractical. Detailed classifications of limited subsets of SDSS images have been made through huge efforts of a small number of experts. Fukugita et al. (2007) and Baillard et al. (2011) determined modified Hubble types for samples of 2253 and 4458 galaxies, respectively; the largest such effort to date is Nair & Abraham (2010), who provide detailed classifications of 14,034 galaxies. Galaxy Zoo 2 includes more than an order of magnitude more systems than any of these. Furthermore, each galaxy has a large number of independent inspections, which permits estimates of the classification likelihood (and in some cases the strength of the feature in question). The size of GZ2 allows for a more complete study of small-scale morphological features and their correlation with many other galaxy properties (e.g., mass, stellar and gas content, environment), while providing better statistics for the rarest objects.

The use of proxies for morphology — such as colour, concentration index, spectral features, surface brightness profile, structural features, spectral energy distribution or some combination of these — is a common practice in astronomy. However, proxies are not an adequate substitute for full morphological classification, as each has an unknown and likely biased relation with the features being studied. For example, most ellipticals are red and most spirals are blue; however, interesting subsets of both types have been found with the opposite colour (Schawinski et al., 2009; Masters et al., 2010). With a sufficiently large set of galaxies, the diversity of the local population can be fully sampled and the relationship between morphology and proxies can be quantified.

Automated morphological classification is becoming much more sophisticated, driven in part by the availability of large training sets from the original Galaxy Zoo (Banerji et al., 2010; Huertas-Company et al., 2011; Davis & Hayes, 2013). However, these methods do not yet provide an adequate substitute for classification by eye. In particular, as Lintott et al. (2011) note, such efforts typically use proxies for morphology as their input (especially colour), meaning they suffer from the objections raised above. The release of the dataset Galaxy Zoo 2 will be of interest to those developing such machine learning and computer vision systems.

The GZ2 results were made possible by the participation of hundreds of thousands of volunteer ‘citizen scientists’. The original Galaxy Zoo demonstrated the utility of this method in producing both large-scale catalogues as well as serendipitous discoveries of individual objects (see Lintott et al. 2011; Fortson et al. 2012 for reviews of Galaxy Zoo 1 results). Since then, this method has been expanded beyond galaxy morphologies to include supernova identification (Smith et al., 2011), exoplanet discovery (Fischer et al., 2012; Schwamb et al., 2012) and a census of bubbles associated with star formation in the Milky Way (Simpson et al., 2012; Kendrew et al., 2012), as well as a variety of “big data” problems outside of astronomy.See http://www.zooniverse.org/ for the full collection.

Several results based on early Galaxy Zoo 2 data have already been published. Masters et al. (2011, 2012) use GZ2 bar classifications to measure a clear increase in bar fraction for galaxies with redder colours, lower gas fractions, and more prominent bulges. Hoyle et al. (2011) showed that the bars themselves are both redder and longer in redder disk galaxies. Skibba et al. (2012) demonstrated that a significant correlation exists between barred and bulge-dominated galaxies at separations from $0.15$–$3$ Mpc. Kaviraj et al. (2012) used GZ2 to study early-type galaxies with visible dust lanes, while Simmons et al. (2013) discovered a population of AGN host galaxies with no bulge, illustrating how black holes can grow and accrete via secular processes. Finally, Casteels et al. (2013) quantify morphological signatures of interaction (including mergers, spiral arms, and bars) for galaxy pairs in the SDSS. This paper describes the data used in these studies, and goes further by quantifying and adjusting for classification biases and in comparing GZ2 classifications with other results.

This paper is organised as follows. Section 2 describes the sample selection and method for collecting morphological classifications. Section 3 outlines the data reduction and debiasing process, and Section 4 describes the tables that comprise the public data release. Section 5 is a detailed comparison of GZ2 to four additional morphological catalogues that were created with SDSS imaging. We summarise our results in Section 6.

This paper uses the WMAP9 cosmological parameters of $H_{0}=71.8$ km/s/Mpc, $\Omega_{m}=0.273$, and $\Omega_{\Lambda}=0.727$ (Hinshaw et al., 2012).

Project description

The primary sample of objects used in Galaxy Zoo 2 comprise approximately the brightest 25% of the resolved galaxies in the SDSS North Galactic Cap region. The sample is generated from the SDSS Data Release 7 (DR7) ‘Legacy’ catalogue (Abazajian et al., 2009), and therefore excludes observations made by SDSS for other purposes, such as the SEGUE survey. Spectroscopic targets come from the SDSS Main Galaxy Sample (Strauss et al., 2002).

Several cuts on the data were applied to the DR7 Legacy sample for selection in GZ2. The goal was to include only the nearest, brightest, and largest systems for which fine morphological features can be resolved and classified. GZ2 required a Petrosian half-light magnitude brighter than 17.0 in the $r$-band (after Galactic extinction correction was applied), along with a size limit of petroR90_r$>$3 arcsec (petroR90_r is the radius containing 90% of the $r$-band Petrosian aperture flux). Galaxies which had a spectroscopic redshift in the DR7 catalogue outside the range $0.0005<z<0.25$ were removed; however, galaxies without reported redshifts were kept. Finally, objects which are flagged by the SDSS pipeline as saturated, bright or blended without an accompanying nodeblend flag were also excluded. The 245,609 galaxies satisfying all these criteria are referred to as the ‘original’ sample.

An error in the selection query meant that the ‘original’ sample initially missed objects to which the SDSS photometric pipeline (Stoughton et al., 2002) assigned both blended and child flags. These are objects that have been deblended from a larger blend (hence child), and have been identified as blended themselves (hence blended; due to containing multiple peaks). However, these are ‘final’ objects, as the SDSS deblender doesn’t attempt to further deblend already deblended objects. These galaxies, which are typically slightly brighter, larger and bluer than the general population, were added to the GZ2 site on 2009-09-02. These additional 28,174 galaxies are referred to as the ‘extra’ sample.

In addition to galaxies from the DR7 Legacy, GZ2 also classified images from Stripe 82, a multiply-imaged section along the celestial equator in the Southern Galactic Cap. The selection criteria were the same as for the Legacy galaxies, with the exception of a fainter magnitude limit of $m_{r}<17.77$. For the Stripe 82 sample only, GZ2 includes multiple images of individual galaxies: one set of images at single exposure depth, plus two sets of co-added images from multiple exposures. The coadded images combined 47 (south) or 55 (north) individual scans of the region, resulting in an object detection limit approximately two magnitudes lower than in normal imaging (Annis et al., 2011).

The primary sample for GZ2 analysis consists of the combined ‘original’, ‘extra’, and Stripe 82 normal-depth images with $m_{r}\leq 17.0$. We have verified that there are no significant differences in the classifications between these sub-samples (i.e., no significant bias is introduced by the fact that they were classified at different times) and thus can be reliably used as a single data set. This is hereafter referred to as the GZ2 main sample (Table 1), and is used for the bulk of the analysis in this paper. Data from both the Stripe 82 normal-depth images with $m_{r}>17.0$ and the two sets of coadded images are separately included in this data release.

2 Image creation

Images of galaxies for classification were generated from the SDSS ImgCutout web service (Nieto-Santisteban, Szalay & Gray, 2004) from the Legacy and Stripe 82 normal depth surveys. Each image is a $gri$ colour composite $424\times 424$ pixels in size, scaled to $(0.02\times$petroR90_r) arcsec per pixel.

Coadded images from Stripe 82 were generated from the corrected SDSS FITS frames. Frames were combined using Montage (Jacob et al., 2010) and converted to a colour image using a slightly modified version of the SkyServer asinh stretch code (Lupton et al., 2004), with parameters adjusted to replicate the normal SDSS colour balance. The parameterisation of the stretch function used is:

where $Q=3.5$ and $\alpha=0.06$. The colour scaling is [1.000,1.176,1.818] in $g$, $r$ and $i$, respectively.

The first set of Stripe 82 coadded images were visually very different from the single-depth images. Changing the colour balance to maximise the visibility of faint features, however, resulted in more prominent background sky noise; since each pixel is typically dominated by a single band, the background is often brightly coloured by the Lupton et al. (2004) algorithm. Due to concerns that this noise would be an obvious sign that the images were from deeper data (potentially biasing the classifications), we created a second set of coadd images in which the colour of background pixels was removed. This was achieved by reducing the colour saturation of pixels outside of a soft-edged object mask.

The original and desaturated coadd image sets are labeled ‘stripe82_coadd_1’ and ‘stripe82_coadd_2’, respectively (Table 1). Subsequent analysis revealed very few differences between the classifications for the images using the two coadd methods (see §4.2).

3 Decision tree

Morphological data for Galaxy Zoo 2 were collected via a web-based interface. Volunteers needed to register with a username for their classifications to be recorded. Like Galaxy Zoo 1, classification begins with the user being shown an SDSS colour composite image of a galaxy alongside a question and set of possible responses. More detailed data is then collected via a multi-step decision tree. In this paper, a classification is defined as the total amount of information collected about one galaxy by a single user completing the decision tree. Each individual step in the tree is a task, which consists of a question and a finite set of possible responses. The selection of a particular response is referred to as the user’s vote.

The first GZ2 task is a slightly modified version of GZ1, identifying whether the galaxy is either “smooth”, has “features or a disk”, or is a “star or artifact”. The appearance of subsequent tasks in the interface depends on the user’s previous responses. For example, if the user clicks on the “smooth” button, they are subsequently asked to classify the roundness of the galaxy; this task would not be shown if they had selected either of the other two responses.

The GZ2 tree has 11 classification tasks with a total of 37 possible responses (Figure 1 and Table 2). A classifier selects only one response for each task, after which they are immediately taken to the next task in the tree. Tasks 01 and 06 are the only questions that are always answered for each and every classification. Once a classification is complete, an image of the next galaxy is automatically displayed and the user can begin classification of a new object. Importantly, in no case could a volunteer choose which galaxy to classify.

Data from the classifications were stored in a live Structured Query Language (SQL) database. In addition to the morphology classifications, the database also recorded a timestamp, user identifier, and image identifier for each classification.

4 Site history

Galaxy Zoo 2 launched on 2009-02-16 with the ‘original’ sample of 245,609 images. The ‘extra’ galaxies from the Legacy survey were added on 2009-09-02. The normal-depth and the first set of coadded Stripe 82 images were mostly added on 2009-09-02, with an additional $\sim 7700$ of coadded images added on 2010-09-24. Finally, the second version of the coadded images were added to the site on 2009-11-04.

For most of the duration of GZ2, images shown to classifiers were randomly selected from the database. To ensure that each galaxy ultimately had enough responses to accurately characterize the likelihood of the classification, images with low numbers of classifications were shown at a higher rate toward the end of the project. The main sample galaxies finished with a median of 44 classifications; the minimum was 16, and $>99.9\%$ of the sample had at least 28 classifications. The ‘stripe82_coadd_2’ galaxies had a median of 21 classifications and $>99.9\%$ had at least 10 (Figure 2).

The last GZ2 classifications were collected on 2010-04-29, with the project spanning just over 14 months. The archived site continued to be maintained, but classifications were no longer recorded. The final dataset contains 16,340,298 classifications (comprising a total of 58,719,719 tasks) by 83,943 volunteers.

Data reduction

In a small percentage of cases, individuals classified the same image more than once. In order to treat each vote as an independent measurement, classifications repeated by the same user were removed from the data, keeping only their votes from the last submission. Repeat classifications occurred for only $\sim 1\%$ of all galaxies. The removal of the repeats only changed the morphological classifications for $\lesssim 0.01\%$ of the sample.

2 Individual user weighting and combining classifications

For example, if a question has three possible responses, and the galaxy corresponds best to response $a$, then the vote fractions for responses $(a,b,c)$ might be $(0.7,0.2,0.1)$.

If an individual votes for response $a$, then $\kappa=(0.7+(1-0.2)+(1-0.1))/3=0.8$

If an individual votes for response $b$, then $\kappa=((1-0.7)+0.2+(1-0.1))/3=0.467$

If an individual votes for response $c$, then $\kappa=((1-0.7)+(1-0.2)+0.1)/3=0.4$

Votes which agree with the majority thus have high values of consistency, whereas votes which disagree have low values.

Each user was assigned an overall consistency ($\bar{\kappa}$) by taking the mean consistency of every response. From the distribution of results for the initial iteration (Figure 3), a weighting function is applied that down-weights classifiers in the tail of low consistency.

For this function, $w=1$ for $\sim 95\%$ of classifiers and $w<0.01$ for only $\sim 1\%$ of classifiers. The vast majority of classifiers are thus treated equally; there is no up-weighting of the most consistent classifiers. The top panel of Figure 3 also shows that the lowest-weighted classifiers completed only a handful ($<10$) of objects on average. This may demonstrate either that the volunteers are becoming more accurate as they classify more galaxies, or that inconsistent people are less likely to remain engaged with the project; further work on user behaviour is needed to distinguish between the two possibilities.

After computing $\kappa$, vote fractions were recalculated using the new user weights, and then repeated a third time to ensure convergence. For each task, individual responses are combined to produce the total vote count and a vote fraction for each task. The weighted votes and vote fractions generated by Equation 4 are used exclusively hereafter when discussing GZ2 votes and vote fractions; for brevity, we typically drop the term “weighted”.

3 Classification bias

The vote fractions are adjusted for what is termed classification bias. The overall effect of this bias is a change in observed morphology fractions as a function of redshift independent of any true evolution in galaxy properties, a trend also seen in the Galaxy Zoo 1 data (Bamford et al., 2009). The SDSS survey is expected to be shallow enough to justify an assumption of no evolution, and so the presumed cause is that more distant galaxies, on average, are both smaller and dimmer in the cutout images. As a result, finer morphological features are more difficult to identify. We note that this effect is not limited to crowd-sourced classifications; expert and automatic classifications must also suffer from bias to some degree, although smaller sample sizes make this difficult to quantify.

Figure 4 demonstrates the effect of classification bias for the GZ2 tasks. The mean vote fraction for each response is shown as a function of redshift; the fraction of votes for finer morphological features (such as identification of disk galaxies, spiral structure, or galactic bars) decreases at higher redshift. The trend is strongest for the initial task of separating smooth and feature/disk galaxies, but almost all tasks exhibit some level of change.

Part of the observed trends in type fractions at high redshifts is due to the nature of a magnitude-limited sample; high-redshift galaxies must be more luminous to be detected in the SDSS and are thus most likely to be giant red ellipticals. However, there is clear evidence of classification bias in GZ2 even in luminosity-limited samples. Since this bias contaminates any potential studies of galaxy demographics over the sample volume, it must be corrected to the fullest possible extent.

Bamford et al. (2009) corrected for classification bias in the GZ1 data for the elliptical and combined spiral classes. Their approach was to bin the galaxies as a function of absolute magnitude ($M_{r}$), the physical Petrosian half-light radius ($R_{50}$), and redshift. They then computed the average elliptical-to-spiral ratio for each ($M_{r}$,$R_{50}$) bin in the lowest redshift slice with significant numbers of galaxies; this yields a local baseline relation which gives the (presumably) unbiased morphology as a function of the galaxies’ physical, rather than observed parameters. From the local relation, they derived a correction for each ($M_{r}$,$R_{50}$,$z$) bin and then adjusted the vote fractions for the individual galaxies in each bin. The validity of this approach is justified in part since debiased vote fractions result in a consistent morphology-density relation over a range of redshifts (Bamford et al., 2009). We modify and extend this technique for the GZ2 classifications.

There are two major differences between the GZ1 and GZ2 data. First, GZ2 has a decision tree, rather than a single question and response for each vote. This means that all tasks, with the exception of the first, depend on responses to previous tasks in the decision tree. For example, the bar question is only asked if the user classifies a galaxy as having “features or disk” and as “not edge-on”. Thus, the value of the vote fraction for this example only addresses the total bar vote fraction among galaxies that a user has classified as disks and are not edge-on, and not as a function of the total galaxy population (see Casteels et al. 2013 for further discussion).

Applying the thresholds to galaxies for deriving the bias correction does increase the number of bins with large variances; however, it is critical for reproducing accurate baseline measurements of individual morphologies. The correction derived from well-classified galaxies is then applied to the vote fractions for all galaxies in the sample.

The second major difference is that the adjustment of the GZ1 vote fractions assumed that the single task was essentially binary. Since almost every vote in GZ1 was for a response of either “elliptical” or “spiral” (either anticlockwise, clockwise, or edge-on), this ratio was employed as the sole metric. No systematic debiasing was done for the other GZ1 response options (“star/don’t know” or “merger”), and the method of adjusting the vote fractions assumes that these other options do not significantly affect the classification bias for the most popular responses. This is not possible for GZ2: many tasks have more than two possible responses and represent a continuum of relative feature strength, rather than a binary choice.

The debiasing method relies on the assumption that for a galaxy of a given physical brightness and size, a sample of other galaxies with similar brightnesses and sizes will (statistically) share the same average mix of morphologies. This is quantified using the ratio of vote fractions $(f_{i}/f_{j})$ for some pair of responses $i$ and $j$. We assume that the true (that is, unbiased) ratio of likelihoods for each task $(p_{i}/p_{j})$ is related to the measured ratio via a single multiplicative constant $K_{j,i}$:

The unbiased likelihood for a single task can trivially be written as:

with the requirement that the sum of all likelihoods for a given task must be unity:

Multiplying (6) by the inverse of (7) yields:

The corrections for each pair of tasks can be directly determined from the data. At the lowest redshift bin, $\frac{p_{i}}{p_{j}}=\frac{f_{i}}{f_{j}}$ and $K_{j,i}=1$. From Equation 5:

This can be simplified by defining $C_{j,i}\equiv\text{log}_{10}(K_{j,i})$ and substituting into (13):

The correction $C_{j,i}$ for any bin is thus the difference between $f_{i}/f_{j}$ at the desired redshift and that of a local baseline, if the ratios between vote fractions are expressed as logarithms.

Bins for $M_{r}$ range from $-24$ to $-16$ in steps of $0.25$ mag, for $R_{50}$ from to $15$ kpc in steps of $0.5$ kpc, and for $z$ from $0.01$ to $0.26$ in steps of $0.01$. These bin ranges and step sizes are chosen to maximize the parameter space covered by the bias correction. Only bins with at least 20 galaxies are used in deriving a correction.

Since each unique pair of responses to a question will have a different local baseline, there are $\binom{n}{2}$ correction terms for a task with $n$ responses. For $n=2$, this method is identical to that described in Bamford et al. (2009).

The baseline morphology ratios for the GZ2 tasks are shown in Figure 5 for the first two responses in each task. To derive a correction for bins not covered at low redshift, we attempted to fit each baseline ratio with an analytic, smoothly-varying function. The baseline ratio for the responses to Tasks 01 and 07 is functionally very similar to the GZ1 relation (Figure A5 in Bamford et al., 2009). This ratio can be fit with an analytic function:

where $\{s_{1},s_{2},s_{3},s_{4},s_{5},s_{6},s_{7},s_{8},s_{9}\}$ are minimised to fit the data.

None of the other tasks are well-fit by a function of the form in Equation 15. For these, a simpler function is used where both $M_{r}$ and $R_{50}$ vary linearly:

where $\{t_{1},t_{2},t_{3},t_{4},t_{5}\}$ are the parameters to be minimized. Equation 18 is fit to all other tasks where enough non-zero bins exist to get a good fit. Finally, for pairs of responses with only a few sampled bins, we instead directly measured the difference bin-by-bin between the local ratio and the measured ratio at higher redshift. Galaxies falling in bins that are not well-sampled are assigned a correction of $C_{i,j}=0$ for that term; this is necessary to avoid overfitting based on only a few noisy bins.

This method succeeds for most GZ2 tasks and responses. Figure 4 illustrates the comparison between the mean raw and debiased vote fractions as a function of redshift. The debiased results (thick lines) are flat over $0.01<z<0.085$, where $L^{\star}$ galaxies ($M_{r}\sim-20.44$; Blanton et al., 2003) are within the detection limit of the survey and there are fewer empty bins. The debiased early- and late-type fractions of 0.45 and 0.55 agree with the GZ1 type fractions derived by Bamford et al. (2009) for the same selection criteria. The bar fraction in disk galaxies is approximately 0.35, slightly higher than the value found by using thresholded GZ2 data in Masters et al. (2011).

4 Angular separation bias

The vote fractions also suffer from a bias which depends on the angular separation between galaxies. For some classifications, participants perceive a galaxy’s morphology differently when it has a close apparent companion. Casteels et al. (2013) found that this bias is particularly strong for Task 08 (“odd features”) and its “merger” classification. The mean merger vote fractions of both physically close galaxies with similar redshifts and projected pairs with very different redshifts increase strongly as a function of decreasing angular separation. This results in projected pairs of non-interacting galaxies being classified as mergers. To determine an unbiased estimate of the mean probability for a given classification, Casteels et al. (2013) subtracted the mean probabilities of projected pairs (with very different redshifts) from physically close pairs (with similar redshifts) for each projected separation bin. This results in a residual probability which is considered to represent the true change in morphology due to strong tidal interaction. While such a correction can be applied in a statistical way to the mean vote fractions for a given classification, applying such a correction to the individual vote fractions is not as straightforward.

The vote fractions presented in this data release have not been corrected for angular separation bias and readers using GZ2 data to study very close pairs are advised to keep this in mind, particularly for Task 08. Fortunately, the angular separation bias is minimal for the rest of the classifications and can usually be ignored. A detailed discussion of the angular separation bias (and how it affects individual classifications) is given in Casteels et al. (2013).

The catalogue

The data release for GZ2 includes the vote counts and fractions (raw, weighted, and debiased) for each task in the classification tree for each galaxy. Data for the five subsamples described below can be accessed at http://data.galaxyzoo.org, and are available on CasJobshttp://skyserver.sdss3.org/casjobs/ in SDSS Data Release 10 (Ahn et al., 2013). Abridged portions of each data table are included in this paper (Tables 5–9).

Table 5 also includes an abbreviated version of the classification designated as $gz2\_class$. It is intended to serve as a quick reference for the consensus GZ2 classification; any quantitative analyses, however, should use the vote fractions instead. A description of how the string is generated is given in Appendix A.

Table 6 gives the GZ2 classifications for the 42,462 main sample galaxies without spectroscopic redshifts. To compute the debiased likelihoods, we used the morphology corrections obtained for galaxies in the spectroscopic main sample. SDSS photometric redshifts (Csabai et al., 2003) are used to derive $M_{r}$ and $R_{50}$ for each galaxy in the photometric sample and select the appropriate correction bin. The mean error in the redshift of the photometric sample (from the SDSS photo-$z$) is $\Delta~{}z=0.021$ (a fractional uncertainty of 27%), compared to the spectroscopic accuracy of $\Delta~{}z=0.00016$ (0.3%). Since the size of the redshift bins in $C_{j,i}$ is 0.01, a shift of several bins can potentially produce a very large change in the debiased vote fractions.

Since the redshift can have a strong effect on classification bias, galaxies with spectroscopic and photometric redshifts from the SDSS are separated; we do not recommend that their debiased data be combined for analysis. For science cases where the main driver is the number of galaxies, however, it may be possible to combine the raw vote fractions for the two samples.

2 Stripe 82

Data for Stripe 82 is reduced separately from the GZ2 main sample. This is due to the deeper magnitude limit of the samples (both normal and coadded) as well as the improved seeing in the latter. Since different image qualities potentially affect the debiasing, all three Stripe 82 samples are individually adjusted for classification bias. The method is the same as that used for the spectroscopic main sample galaxies – the only difference is that the threshold for classification in the coadded sample is lowered from 20 to 10 votes.

Table 7 gives classifications for the Stripe 82 normal-depth images with spectroscopic redshifts. Galaxies in this table with $m_{r}<17.0$ also appear in Table 5; however, the corrections for classification bias here are derived based only on Stripe 82 data, and so debiased likelihoods and flags are slightly different. Classifications for galaxies with photometric redshifts only are not included.

Tables 8 and 9 contain classifications for the first and second sets of coadded Stripe 82 galaxies with spectroscopic redshifts. Since both the number of galaxies and the average number of classifications per galaxy are a small fraction of that in the main sample, though, the corrections encompass a smaller range of tasks and phase space in ($M_{r}$,$R_{50}$,$z$). The increased exposure time and improved seeing, however, means that the effect of classification bias is lessened at lower redshifts; the raw vote fractions may thus be more suitable for some science cases that require deeper imaging.

Figure 6 compares the results of the Task 01 classifications for the GZ2 main and Stripe 82 samples. The distributions of the responses for both the main sample and Stripe 82 normal-depth show similar behavior as a function of redshift. This applies both when using thresholded vote fractions and the raw likelihoods. The type fractions for the coadded data, however, are significantly different – there is a significant increase at all redshifts in the fraction of responses for “features or disk”. This increases the fraction of unclassified galaxies (and subsequently decreases the fraction of smooth galaxies) when using thresholds, and a similar shift of vote fractions from smooth to feature/disk when using the raw likelihoods.

This difference demonstrates why the main sample corrections cannot be applied to the coadded images. The likely cause is that the coadded data allows classifiers to better distinguish faint features and/or disks, due to both improved seeing (from $1.4\arcsec$ to $1.1\arcsec$; Annis et al., 2011) and higher signal-to-noise ratio.

The biggest systematic difference is for the response to Task 05 (bulge prominence) of the bulge being “just noticeable”. The mean fraction in coadd2 is 35% higher than that in coadd1. This effect is opposite (but not equal) to that for an “obvious” bulge, for which coadd1 is 13% higher; this may indicate a general shift in votes toward a more prominent bulge. A similar but smaller effect is seen in classification of bulge shapes for edge-on disks (Task 09), where votes for “no bulge” in coadd1 data go to “rounded bulge” in coadd2. The specific cause for these effects as it relates to the image quality is not investigated further in this paper.

For most morphological questions, the two versions of coadded images showed no significant difference. While either set of coadded data can likely be used for science, we recommend using coadd2 if choosing between them. The overall consistency indicates that the votes for both could potentially be combined and treated as a single data set; this could be useful for increasing the classification accuracy for deeper responses (such as spiral arm properties) within the GZ2 tree.

3 Using the classifications

Since GZ2 is intended to be a public catalogue for use by the community, we present two examples of how classifications can be selected. Actual use will depend on the individual science case, and additional cuts (e.g., making a mass or volume-limited sample) may be required to define the parameters more appropriately.

The first use case suggested for the GZ2 data is the selection of pure samples matching a specific morphology category. This is appropriate for when some finite number of objects with clear morphological classifications is required (perhaps for individual study or an observing proposal), but there is no requirement to have a complete sample. An example would be the selection of three-armed spirals. The simplest way is to search for galaxies in the GZ2 spectroscopic main sample (Table 5) with t11_arms_number_a33_3_flag$=1$, which returns 308 galaxies. Inspection of the flagged images shows that they are all in fact disk galaxies with three spiral arms, with no object that is a clear false positive.

Comparison of GZ2 to other classification methods

To assess both the scope and potential accuracy of the GZ2 classifications, we have compared our results to four morphological galaxy catalogues (including the previous version of Galaxy Zoo). All four catalogues contain classifications based on optical SDSS images and have significant overlaps with the galaxies in GZ2.

Galaxy Zoo 1 (Lintott et al., 2011): Citizen science

Nair & Abraham (2010) : Expert visual classification

EFIGI (Baillard et al., 2011) : Expert visual classification

Huertas-Company et al. (2011) : Automatic classification

A summary of the agreement between GZ2 and other catalogues is given in Table 4; the remainder of this section discusses the results in more detail.

The galaxies in GZ2 are a subset of GZ1, with $248,883$ in both catalogues. The similarities between GZ1 and Task 01 in GZ2 allow their results to be compared in detail. We analyzed vote fractions for the GZ1 “elliptical” category as compared to GZ2 “smooth” galaxies, and combined responses for all three GZ1 spiral categories to the GZ2 “features or disk” response.

Galaxies are slightly more likely to be identified as a spiral in GZ2 than in GZ1. Figure 9 shows the distribution of the difference between spiral classifications, using the debiased likelihoods for combined spirals for GZ1 and “features or disk” galaxies in GZ2. The slight leftward skew indicates that a galaxy is more likely to be identified as a spiral in GZ2 compared to GZ1. When restricted only to galaxies in the joint clean samples ($p>0.8$), the spread is greatly reduced and the distribution is centered around a difference of zero, indicating that the two agree very well for classifications with high levels of confidence.

Based on classifications from galaxies in both projects, GZ2 is more conservative than GZ1 at identifying spiral structure. A possible explanation is that this is a bias from classifiers who are anticipating subsequent questions about the details of any visible structures. An experienced classifier, for example, would know that selecting “features or disk” is followed by additional questions, none of which offer options for an uncertain classification. If the classifier is less confident in identifying a feature, it is possible they would avoid this by clicking “smooth” instead.

In addition to the angular separation bias discussed in §3.4, GZ2 responses to Task 08 (“odd feature”) also suffer from crosstalk. This is the result of more than one response being applicable for some galaxies, which forces the participant to choose the one they consider most relevant. For example, a merging galaxy may display a strong dust lane, be highly irregular in shape, and have a disturbed appearance. While “merger”, “dust lane”, “irregular” and “disturbed” are all possible classifications, the participant will usually choose the “merger” classification and information about the other morphological features is lost. For close pairs, this crosstalk is a function of angular separation – the fraction of galaxies classified as mergers increases with decreasing separation, while the other “odd feature” classifications lose votes correspondingly (Casteels et al., 2013). We note that in later incarnations of Galaxy Zoowww.galaxyzoo.org it is possible to select multiple classifications from the “odd feature” task.

To summarize, GZ1 and GZ2 share nearly 250,000 galaxies that have been classified in both samples. The separation of early and late-type galaxies from the two projects is mostly consistent, especially for high-confidence ($p>0.8$) galaxies. GZ2 classifications are more conservative than GZ1 at identifying spiral structure for intermediate vote fractions. Mergers identified in GZ1 appear at a very high rate in GZ2 as “odd” galaxies, although classification as a merger is complicated by cross-talk between other GZ2 responses to Task 08.

2 Expert visual classifications

The standard for detailed morphological classifications for many years has come from visual identifications by individual expert astronomers. We compare the GZ2 classifications to two SDSS morphological catalogues generated by small groups of professional astronomers: Nair & Abraham (2010, hereafter NA10) and EFIGI (Baillard et al., 2011). The fact that GZ2 and both expert catalogues used data from the same survey allows for direct comparison of the results.

The catalogue of NA10 is based on images of 14,034 galaxies from SDSS DR4. Galaxies were selected from a redshift range of $0.01<z<0.1$, with an extinction-corrected apparent magnitude limit of $g<16$. In comparison, the GZ2 sample is deeper, spans a larger redshift range, and contains a more recent data release. 12,480 galaxies were classified in both GZ2 and NA10 – this comprises nearly all (89.9%) of the NA10 catalogue, but only 4.5% of GZ2.

NA10 is based on visual classifications of monochrome $g$-band images by a single astronomer (P. Nair). The data include RC3 T-types (a numerical index of a galaxy’s stage along the Hubble sequence; de Vaucouleurs et al., 1991) as well as measurements of bars, rings, lenses, pairs, interactions, and tails. The NA10 data does not contain information on the likelihood or uncertainty associated with morphological features, although it does measure some features by their relative strengths (dividing barred galaxies into strong, medium, and weak classes, for example).

EFIGI consists of classifications of 4,458 galaxies, which are a subset of the RC3 catalogue with 5-colour imaging in SDSS DR4. Almost all galaxies in EFIGI are at $0.0001<z<0.08$. Classifications on composite $gri$ images were performed by a group of 11 professional astronomers, each of whom classified a subset of 445 galaxies. A training set of 100 galaxies was also completed by all 11 astronomers to adjust for biases among individual classifiers. 3,411 galaxies are in both EFIGI and GZ2. This constitutes 77% of EFIGI and 1.2% of the GZ2 sample.

T-types in EFIGI were assigned using a slightly modified version of the RC3 Hubble classifications. Peculiar galaxies were not considered a separate type, and ellipticals were subdivided into various types: compact, elongated (standard elliptical), cD (giant elliptical), and dwarf spheroidals. The remaining morphological information, dubbed “attributes”, is divided between six groups:

spiral arms: arm strength, arm curvature, rotation

texture: visible dust, dust dispersion, flocculence, hot spots

dynamics: bar length, inner ring, outer ring, pseudo-ring, perturbation

EFIGI attributes were measured on a five-step scale from 0 to 1 (0, 0.25, 0.50, 0.75, 1). For some attributes (e.g., arm strength, rings), the scale is set by the fraction of the flux contribution of the feature relative to that of the entire galaxy. For others (e.g., inclination or multiplicity), it ranges between the extrema of possible values.

The EFIGI and NA10 catalogues were compared in detail by Baillard et al. (2011). T-type classifications for the two catalogues strongly agree; EFIGI lenticular and early spirals have slightly later average classifications in NA10, while later EFIGI galaxies have slightly earlier NA10 T-types. EFIGI has a major fraction of galaxies with slight-to-moderate perturbations with no interaction flags set in the NA10 catalogue, indicating that NA10 is less sensitive toward more benign features (e.g., spiral arm asymmetry). The bar length scale is consistent between the two samples; good agreement is also found for ring classifications.

The lack of sensitivity to weak bars from NA10 may also be related to the design of the GZ2 interface. When asked if a bar is present, the image shown in the web interface is an icon with two examples of a barred galaxy (Figure 1). The example image has the bar extending across the disk’s full diameter, fitting the typical definition of a strong bar. With this as the only example (and no continuum of options between the two choices), GZ2 participants may not have looked for bars shorter than the disk diameter, or have been less confident in voting for “yes” if they were identified. Results from Hoyle et al. (2011) show that classifiers are fully capable of identifying weak bars in other contexts.

The EFIGI bar length attribute is measured with respect to $D_{25}$, the decimal logarithm of the mean isophote diameter at a surface brightness of $\mu_{B}=25$ mag arcsec-2. A value of 1.0 (the strongest bar) extends more than half the length of $D_{25}$, while the median value of 0.5 would be about one-third the length of $D_{25}$. The overall fraction of barred galaxies in EFIGI is 42% (1439/3354); this is essentially unchanged if only oblique galaxies are considered (915/2099 = 44%). This is significantly higher than the mean bar fraction of Masters et al. (2011), at 29.5%, but consistent with results using automated ellipse-fitting techniques (Barazza, Jogee & Marinova, 2008; Aguerri, Méndez-Abreu & Corsini, 2009).

The higher fraction in EFIGI is due to the contributions of galaxies with bar length attributes of 0.25, the majority of which have GZ2 vote fractions below 0.5. If only EFIGI galaxies at 0.5 and above are considered to be barred, then the bar fraction falls to 17%. Only some of the galaxies in the 0.25 EFIGI bin are being classified by GZ2 as barred, however, Baillard et al. (2011) defines these as “barely visible” bars.

2.2 Rings

NA10 classify three types of ringed galaxies based on criteria from Buta & Combes (1996): inner rings (between the bulge and disk), outer rings (external to the spiral arms), and nuclear rings (lying in the bulge region). In GZ2, rings can be identified only if the user selects “yes” for the question “Anything odd?” Since the “odd feature” task has seven responses, of which only one can be selected, any galaxies with multiple “odd” features will have votes split among the features, with only one option achieving a plurality (see §5.1). While this means that some galaxies with rings may have low vote fractions in the GZ2 classifications, those with high vote fractions are typically strong and distinct.

Figure 13 shows a moderate correlation between the EFIGI ring attributes and the GZ2 ring vote fractions. The relationship is dominated by galaxies for which the methods agree strongly on either no ring or a ring with high contributions to the total galaxy flux. For intermediate (between 0.25 and 0.75) values of the EFIGI ring attribute, the GZ2 vote fraction has relatively little predictive power.

2.3 Mergers and interacting galaxies

Galaxies in GZ2 are classified as mergers in Task 08 “anything odd?” NA10 classify possible mergers in two ways: both as pairs of objects and as galaxies with visible interaction signatures. The paired objects are sorted by relative separation (close, projected, apparent, or overlapping pairs), and interacting galaxies by morphology (disturbed, warp, shells, tails, or bridges).

Results from both expert catalogues are consistent with Casteels et al. (2013), who found that the mean vote fraction for mergers increases with decreasing projected separations ($r_{p}$), but then drops off significantly for the closest pairs at $r_{p}<10$ kpc. At these separations, the GZ2 votes for Task 08 go instead to the “irregular” and “disturbed” responses.

2.4 T-types

One of the primary challenges for morphological classification in GZ2 is matching the classification tree to T-types, which are not a category in the decision tree. The classifications from expert catalogues are thus extremely valuable as a calibration sample.

Figure 16 shows the percentage of galaxies identified as having either a disk or features from the first question in the GZ2 tree, colour-coded by their NA10 T-types. There is a clear separation in the GZ2 fractions for galaxies classified as E versus Sa–Sd. Disk galaxies, including S0, have a median fraction for the GZ2 “features or disk” question of 0.80, with a standard deviation of 0.29. Disks with few GZ2 votes for “feature” are found to be primarily lenticular (S0) galaxies. If only galaxies with T-types Sa or later are considered, the peak at lower GZ2 vote fractions disappears. The median GZ2 vote fraction for these galaxies is 0.88, with a standard deviation of 0.23. The highest GZ2 vote fraction for an elliptical galaxy in NA10 is 0.741; therefore, any cut above this includes galaxies exclusively identified by NA10 as late-type.

For disk galaxies with spiral structure, Task 10 in GZ2 asked users to classify the “tightness” of the arms. This had three options: tight, medium, or loose, accompanied with icons illustrating example pitch angles (Figure 1). This allows investigation of the parameters which contribute to the Hubble classification of late-type galaxies which depends on both spiral arm and bulge morphology; tight spirals are presumed to be Sa/Sb, medium spirals Sb/Sc, and loose spirals Sc/Sd.

There are less than 30 galaxies classified by GZ2 as smooth and as Sa or later-type by NA10. Individual inspection reveals that these galaxies show no evidence of a disk, and so their NA10 classification is purely bulge-related. There also exist $\sim 700$ galaxies classified by GZ2 as smooth but as S0 or S0/a by NA10; these are mostly smooth, face-on galaxies with prominent bulges.

Overall, a clear trend is demonstrated for looser GZ2 spiral arms to correspond with later spiral T-types from expert classifications. High vote fractions are mostly Sa/Sb galaxies for tight winding, Sb/Sc galaxies for medium winding, and Sc/Sd galaxies for loose winding. Individual galaxies, however, can show significant scatter in their GZ2 vote fractions and do not always separate the morphologies on the level of the Hubble T-types. Classifications of spiral galaxies into subcategories (Sa, Sb and Sc) by experts have been shown to be dominated by bulge classification, and to pay little attention to the arm pitch angle, despite the original definition of the late-type categories.

Having considered the effect of spiral arm tightness, we examine the relationship between bulge morphology and T-type. Disk galaxies in GZ2 are also classified by the visible level of bulge dominance (Task 05), irrespective of whether spiral structure is also identified. This task has four options: “no bulge”, “just noticeable”, “obvious”, and “dominant” (Figure 1).

The left side of Figure 18 shows the distribution of NA10 T-types for galaxies based on their GZ2 vote fractions for bulge prominence, including only galaxies with at least 10 votes for Task 05. Vote fractions for both the “no bulge” and “dominant” responses peak strongly near zero and tail off as the vote fraction increases. The responses to the middle options (“just noticeable” and “obvious”) are both symmetrically distributed around a peak near 0.5.

The link to T-type is more sharply defined for GZ2 bulge prominence than for spiral tightness, according to expert classifications. Very clean samples of late-type (Sb–Sd) spirals can be selected using only the “no bulge” parameter; additional samples with $\sim 10$% contamination can be selected with the “just noticeable” and “obvious” distributions. Elliptical galaxies that have bulge prominence classified in GZ2 are most often “dominant”, but there is no obvious separation of ellipticals from disk galaxies based on this task alone.

Since Hubble types are based on both the relative size of the bulge and the extent to which arms are unwound (Hubble, 1936), we explored whether the combination of Tasks 05 and 10 from GZ2 can be mapped directly to T-types. The numerical T-types from NA10 were fit with a linear combination of the GZ2 vote fractions for the bulge dominance and arms winding tasks. The best-fit result using symbolic regression (Schmidt & Lipson, 2009), however, depends only on parameters relating to bulge dominance:

This technique assumes that the difference in morphology is well-defined by mapping T-types to a linear scale, which is far from being justified. Figure 19 shows the distribution of the GZ2-derived T-type from Equation 19 compared to the NA10 values. The large amounts of overlap between adjoining T-types show that this clearly does not serve as a clean discriminator. One could make a cut between the earliest (Sa) and latest (Sd) spiral types based only on the vote fractions. Alternatively, the relative numbers of galaxies could be used as the weights to construct the probability of a given T-type. This has yet to be conclusively tested.

The distributions in Figure 19 also show that S0 galaxies in particular would typically be mistakenly judged as later types (overlapping strongly with Sa) on average using only this metric. This is consistent with the “parallel-sequence” model of van den Bergh (1976) and later revised by several groups (including Cappellari et al., 2011; Laurikainen et al., 2011; Kormendy & Bender, 2012).

Finally, we note that Simmons et al. (2013) identified a significant effect in which nuclear point sources, such as AGN, can mimic bulges in the GZ2 classifications. This has not been accounted for in this analysis, but could potentially be addressed by separating the sample into AGN and quiescent galaxies (via BPT line ratios) and looking for systematic differences between the two samples.

2.5 Bulge prominence

EFIGI measures the bulge/total light ratio $(B/T)$ in each galaxy, with the attribute strength corresponding to the relative contribution of the bulge. Elliptical galaxies have $B/T=1$ and irregular galaxies $B/T=0$. Baillard et al. (2011) show that $B/T$ is correlated with arm curvature and anti-correlated with the presence of flocculent structure and hot spots, consistent with movement along the Hubble sequence.

2.6 Arm curvature

EFIGI also measures the arm curvature of each galaxy, with classifications very similar to the “tightness of spiral arms” question (Task 10) in GZ2. If both expert and citizen science classifiers agree, one would expect galaxies with high GZ2 vote fractions for tight spirals to have EFIGI classifications at 0.75–1.0; GZ2 galaxies classified as medium spirals to be centered around 0.5; and loose spirals to have arm curvatures of 0.0–0.25.

3 Automated classifications

Huertas-Company et al. (2011, HC11) have generated a large set of morphological classifications for the SDSS spectroscopic sample using an automated Bayesian approach. The broad nature of their probabilities (four broad morphological categories), do not directly relate to the majority of the GZ2 fine structure questions, such as bar or spiral arm structure. Comparison between the two samples, however, is useful to demonstrate the effect that smaller-scale features (as classified by GZ2) may have on automatically-assigned morphologies.

The sample classified by HC11 is limited to galaxies with $z<0.25$ that have both good photometric data and clean spectra. Their total of 698,420 galaxies is approximately twice the size of GZ2. The HC11 sample goes to fainter magnitudes, with more than 400,000 galaxies below the GZ2 limit of $m_{r}>17$. Their morphological classification algorithm is implemented with support vector machine (SVM) software that tries to find boundaries between regions in $N$-dimensional space, where $N$ is determined by criteria including morphology, luminosity, colour, and redshift (Huertas-Company et al., 2008). The training set is the 2,253 galaxies in Fukugita et al. (2007), which are already classified by T-type. Each galaxy is assigned a probability of being in one of four subclasses: E, S0, Sab, and Scd (the latter two combining their respective late-type categories).

We note that the inclusion of colour means that HC11 classifications are not purely morphological, but include information about present-day star formation as well as the dynamical history which determines morphology. Studies of red spiral (Masters et al., 2010) and blue elliptical galaxies (Schawinski et al., 2009), for example, demonstrate the advantages of keeping these criteria separate.

Huertas-Company et al. (2011) directly compared their results to the GZ1 sample from Lintott et al. (2011). They found that robust classifications in GZ1 (flagged as either confirmed ellipticals or spirals) have median probabilities of 0.92 according to their algorithm, indicating that sure GZ1 classifications are also sure in their catalogue. They also showed a near-linear relationship between the GZ1 debiased vote fraction and the HC11 probabilities. This is one of the first independent confirmations that the vote fractions may be related to the actual probability of a galaxy displaying a morphological feature.

Figure 21 shows the distributions of the HC11 early- and late-type probabilities for GZ2 galaxies robustly identified ($p>0.8$) as either smooth or having features/disks. The median HC11 early-type probability for GZ2 ellipticals is 0.85, and the late-type probability for GZ2 spirals is 0.95. This confirms the result that robust classifications in Galaxy Zoo agree with the automated algorithm for broad morphological categories.

An exception to this is a population of galaxies classified as “smooth” by GZ2, but which have very low early-type probabilities from HC11 (Figure 21). The mean GZ2 vote fraction for these galaxies is consistent with those with high early-type probabilities – these galaxies are not marginally classified as ellipticals in GZ2. The roundness of the galaxy (Task 07 in GZ2) seems to play some role, as the low-HC11 smooth galaxies have fewer round galaxies and many more “cigar-shaped” galaxies in this sample. A high axial ratio might train the HC11 algorithm to infer the existence of a disk; the absence of any obvious spiral features or bulge/disk separation (verified by eye in a small subsample of the images) lead GZ2 to categorise these as “smooth”. There is a clear dependence on apparent magnitude; the lower peak disappears if only galaxies with $r<16$ are included. Early-type galaxies that disagree with the HC11 classification are also significantly bluer, with respective colours of $(g-r)=0.67$ and $(g-r)=0.97$. Since the SVM method does include SDSS colours as a parameter, we conjecture that the low HC11 early-type probability is in part due to the fact that they are blue, in addition to morphological features such as shape and concentration.

Figure 21 also shows the distribution of “uncertain” galaxies, for which none of the responses for Task 01 had a vote fraction $>0.8$. The HC11 probability for these galaxies is bimodal, with the larger fraction classified as HC11 late-type and a smaller fraction as HC11 early-type.

Conclusions

We present the data release for the Galaxy Zoo 2 (GZ2) project. GZ2 uses crowd-sourced votes from citizen scientist classifiers to characterize morphology of more than 300,000 galaxies from the SDSS DR7. GZ2 classified $gri$ colour composite images selected on the basis of magnitude ($m_{r}<17$), angular size ($r_{90}>3\arcsec$), and redshift ($0.0005<z<0.25$) criteria. Deeper images from Stripe 82 are also included at both normal and coadded image depths.

GZ2 expands on the original Galaxy Zoo results by classifying a large array of fine morphological structures. In addition to previous distinctions between elliptical and spiral galaxies, GZ2 identifies the presence of bars, spiral structure, dust lanes, mergers, disturbed/interacting morphologies, and gravitational lenses. It also quantifies the relative strengths of galactic bulges (both edge-on and face-on), the tightness and multiplicity of spiral arms, and the relative roundness of elliptical galaxies. Classification was done via a multi-step decision tree presented to users in a web-based interface. The final catalogue is the result of nearly 60 million individual classifications of images.

Vote fractions for each response are also adjusted for classification bias, the effect of fine morphological features being more difficult to detect in smaller and fainter galaxies. Corrections to determine the debiased vote fractions are derived directly from the GZ2 data itself.

The final catalogue consists of five tables, comprising morphological classifications for the GZ2 main sample (separated into galaxies with spectroscopic and photometric redshifts) and galaxies from Stripe 82 (for normal-depth and two sets of coadded images with spectroscopic redshifts). Data for each galaxy includes (for each response) the raw and weighted number of votes, the raw and weighted vote fractions, the debiased vote fraction, and an optional flag which indicates if a feature has been robustly identified. Portions of the data are presented in Tables 5–9; full machine-readable tables are available at http://data.galaxyzoo.org and in SDSS Data Release 10.

We have compared the GZ2 classifications in detail to several other morphological catalogues. Early and late-type classifications are consistent with results from the original Galaxy Zoo, especially for galaxies in the clean samples. Expert catalogues (Nair & Abraham, 2010; Baillard et al., 2011) show good agreement for galaxies with medium to strong bars; GZ2 is less confident in identifying expert-classified weak and/or nuclear bars. In ringed galaxies, GZ2 recovers the majority of outer rings, but relatively few inner or nuclear rings due to the design of the GZ2 question. Pairs and interacting galaxies are more difficult to reliably cross-match in a clean sample, although Casteels et al. (2013) have already shown that the GZ2 “loose winding arms” parameter is a reliable proxy for interaction. The GZ2 bulge dominance parameter strongly correlates with the Hubble T-type from both expert catalogues. Adding GZ2 measurements of the spiral arm tightness, though, does not increase the T-type classification accuracy. Automated classifications from Huertas-Company et al. (2011) agree well with GZ2 in separating elliptical and late-type spirals, although identification of S0 galaxies still represents a challenge.

GZ2 contains more than an order of magnitude more galaxies than the largest comparable expert-classified catalogues (NA10, EFIGI) while still classifying detailed morphological features not replicable by automated classifications. GZ2 data have already been used to demonstrate a relationship between bar fraction and the colour, gas fractions, and bulge size of disk galaxies (Masters et al., 2011, 2012), as well as studies of the bar colour and length itself (Hoyle et al., 2011). The size of the catalogues has allowed for the discovery and study of comparatively rare objects, such as early-type dust lane galaxies (Kaviraj et al., 2012) and bulgeless AGN hosts (Simmons et al., 2013). Direct use of the GZ2 likelihoods has also been used to quantify the environmental dependence on morphology, showing a correlation for barred and bulge-dominated galaxies (Skibba et al., 2012) and identifying reliable signatures of interaction from GZ2 data (Casteels et al., 2013).

The scientific productivity of the Galaxy Zoo project has already shown that the use of multiple independent volunteer classifications is a robust method for the analysis of large datasets of galaxy images. This public release of the Galaxy Zoo 2 catalogue intends to build on this success, by demonstrating the reliability and benefit of these classifications over both expert and automated classifications. We publicly release these Galaxy Zoo 2 classifications both as a rich dataset that can be used to study galaxy evolution, and as training sets for refining future automated classification techniques.

Acknowledgments

The data in this paper are the result of the efforts of the Galaxy Zoo 2 volunteers, without whom none of this work would be possible. Their efforts are individually acknowledged at http://authors.galaxyzoo.org.

The development of Galaxy Zoo 2 was supported by The Leverhulme Trust. KWW and LFF would like to acknowledge support from the US National Science Foundation under grant DRL-0941610. CJL acknowledges support from the Science and Technology Facilities Council (STFC) Science in Society program. KS gratefully acknowledges support from Swiss National Science Foundation Grant PP00P2_138979/1. TM acknowledges funding from the STFC ST/J500665/1. RCN was partially supported by STFC grant ST/K00090X/1. BDS acknowledges support from Worcester College, Oxford, and from the Oxford Martin School program on computational cosmology. Please contact the author(s) to request access to research materials discussed in this paper.

This research made use of Montage, funded by the National Aeronautics and Space Administration’s Earth Science Technology Office, Computation Technologies Project, under Cooperative Agreement Number NCC5-626 between NASA and the California Institute of Technology. Montage is maintained by the NASA/IPAC Infrared Science Archive. It also made extensive use of the Tool for OPerations on Catalogues And Tables (TOPCAT), which can be found at http://www.starlink.ac.uk/topcat/ (Taylor, 2005, 2011).

We thank the referee for useful comments that improved the content and structure of this paper.

Funding for the SDSS and SDSS-II has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, the U.S. Department of Energy, the National Aeronautics and Space Administration, the Japanese Monbukagakusho, the Max Planck Society, and the Higher Education Funding Council for England. The SDSS website is http://www.sdss.org/.

The SDSS is managed by the Astrophysical Research Consortium for the Participating Institutions. The Participating Institutions are the American Museum of Natural History, Astrophysical Institute Potsdam, University of Basel, University of Cambridge, Case Western Reserve University, University of Chicago, Drexel University, Fermilab, the Institute for Advanced Study, the Japan Participation Group, Johns Hopkins University, the Joint Institute for Nuclear Astrophysics, the Kavli Institute for Particle Astrophysics and Cosmology, the Korean Scientist Group, the Chinese Academy of Sciences (LAMOST), Los Alamos National Laboratory, the Max-Planck-Institute for Astronomy (MPIA), the Max-Planck-Institute for Astrophysics (MPA), New Mexico State University, Ohio State University, University of Pittsburgh, University of Portsmouth, Princeton University, the United States Naval Observatory, and the University of Washington.

References

Appendix A Generating the abbreviation for a GZ2 morphological classification

As part of the GZ2 data release (Tables 5–9), we provide a short abbreviation ($gz2\_class$) that indicates the most common consensus classification for the galaxy. We emphasise that the intent is not to create a new classification system; rather, this is only a convenient shorthand for interpreting portions of the GZ2 results.

The $gz2\_class$ string is generated for each galaxy by taking the largest debiased vote fraction (beginning with Task 01) and selecting the most common response for each subsequent task in the decision tree.

Galaxies that are smooth (from Task 01) have $gz2\_class$ strings beginning with ‘E’. Their degree of roundness (completely round, in-between, and cigar-shaped) is represented by ‘r’,‘i’, and ‘c’, respectively.

Galaxies with features/disks have $gz2\_class$ strings beginning with ‘S’. Edge-on disks follow this with ‘er’, ‘eb’, or ‘en’ (with the second letter classifying the bulge shape as round, boxy, or none). For oblique disks, the letter following ‘S’ is an upper-case ‘B’ if the galaxies have a bar. The bulge prominence (‘d’ = none, ‘c’ = just noticeable, ‘b’ = obvious, ‘a’ = dominant). Both bars and bulges follow the same general trends as the Hubble sequence, although the correspondence is not exact. If spiral structure was identified, then the string includes two characters indicating the number ($1,2,3,4,+,?$) and relative winding (‘t’=tight, ‘m’=medium, ‘l’=loose) of the spiral arms.

Finally, any feature in the galaxy the users identified as “odd” appears at the end of the string in parentheses: ‘(r)’=ring, ‘(l)’=lens/arc, ‘(d)’=disturbed, ‘(i)’=irregular, ‘(o)’=other, ‘(m)’=merger, ‘(u)’=dust lane.

Objects that are stars or artifacts have the $gz2\_class$ string ‘A’. For example:

SBc2m = barred disk galaxy with a just noticeable bulge and two medium-wound spiral arms

Seb = edge-on disk galaxy with a boxy bulge

Sc(I) = disk galaxy with a just noticeable bulge, no spiral structure, and irregular morphology.

Sample images of the twelve most common $gz2\_class$ strings are shown in Figure 24.