DISK DETECTIVE: DISCOVERY OF NEW CIRCUMSTELLAR DISK CANDIDATES THROUGH CITIZEN SCIENCE

To choose sources to upload to the website, we performed some initial computer-based cuts on the WISE data, informed by published debris disk searches (see above) but not limited by the Hipparcos or Tycho catalogs, for example. We utilized signal-to-noise cuts (w4snr, w1sigmpro, w4rchi2, and w4sigmpro) and some of the AllWISE catalog flags (cc_flags, xscprox, na, nb, n_2mass, and ext_flg) to remove sources that were noisy or close to known extended sources. Though many searches have used sophisticated color cuts to focus on particular kinds of disks (e.g., Koenig et al. 2012), we kept our color cuts minimal to cast as broad a net as possible. To preselect sources with infrared excesses, we merely removed all sources with [W4] > [W1]-0.25. ([W1] is magnitude in the WISE 3.4 μm band, and [W4] is magnitude in the WISE 22 μm band.) This criterion corresponds to a 26% excess over a Rayleigh–Jeans slope between those two bands. We also required that a WISE 4 excess be significant at the 5-σ level. Table 1 summarizes all these initial cuts.

We launched the site on 2014 January 28 with a first batch of subjects covering only the Galactic latitudes +30 to +40, +50 to +90, and −40 to −90, with no additional magnitude limit. At first, about 20% of the initial upload was made available on the site. We soon realized that most of the volunteer effort was being spent classifying faint, extragalactic sources, so we decided to impose a magnitude limit on the search. We chose a criterion of J < 14.5 because the subjects brighter than this magnitude were clearly concentrated in the Galactic plane, while fainter subjects appeared to be isotropically distributed on the sky. We uploaded a second batch of 272,022 subjects on 2014 May 30. This second batch covered the rest of the sky, but was limited to J < 14.5. We also deactivated all the previously uploaded and active J > 14.5 subjects on this date, so presently the J > 14.5 subjects currently make up a small subset of the data for which we have classifications.

We examine the classification data weekly to chart our progress. As of 2015 August 25, the selection of classification buttons had the distribution shown in Table 2. Clearly, the dominant false-positive rejected by the classification process is "Multiple objects in the Red Circle."

Note that the "Empty Circle in WISE images" button exists mainly to allow users a reasonable response option if, for example, there were a network glitch. None of the data actually had empty circles in all the WISE images, though some subjects did have empty red circles in a single band, and some users chose this "Empty Circle in WISE images" classification for such subjects. The rarity of this situation is reflected in the very low (0.6%) classification rate for this button.

So far, we have investigated two basic algorithms for sorting the raw data into classifications: a plurality algorithm and majority algorithm. By the plurality algorithm, the classification with the most votes becomes the official classification. By the majority algorithm, the official classification is one with ≥50% of the votes, if one exists.

To help test these algorithms, we prepared a "gold standard" data set of 500 subjects classified by the members of the science team. All of the subjects in this "gold standard" data set were classified by two or more science team members, and all tie votes were discarded. This process showed that science team members agree with one another at roughly the 82% level as to whether an object is "good" or not. A challenge in creating this "gold standard" set was that sometimes science team members ruled out candidates based on their expertise/outside information rather than strictly doing what the website asked for. There was also a range in terms of how conservative the science team members were in terms of crowded fields and ISM background. For example, the YSO experts (used to looking in the Galactic plane) on the science team tolerated background objects and "extended" objects more readily than the debris disk experts on the team.

The majority algorithm performed well at selecting objects deemed to be classified as "None of the Above/Good candidate." On a randomly selected group of about 500 subjects, 94% of those classified as "None of the Above/Good candidate" by the majority algorithm agreed with the gold standard set rankings. For the other five classification buttons, the agreement between classifications selected via the plurality algorithm and the gold standard set varied: 52% for "Extended," 54% for "Not round," 24% for "Object Moves Off the Crosshairs," and 7% for "Empty circle." This lack of agreement suggests that users may sometimes neglect to click all of the relevant buttons when a subject is "bad" for more than one reason. However, since our desire is simply to rule out false positives, this lack of agreement on the precise nature of the false positive does not hamper our study.

Deciding on classifications via the plurality algorithm generally leads to larger disagreements with the gold standard set and the classification data. The exception to this rule seems to be the case of "multiple objects in the red circle"; subjects classified as such by the plurality vote agree with the gold standard set 96% of the time.

Table 2 shows the total selection distribution between the categories on the main site from the first ∼1 million classifications. Users selected the "good" button nearly 20% of the time. Using the majority algorithm to determine a final classification, we get a 16% yield of "good" objects based on those classifications.

Some of the brighter Disk Detective subjects show diffraction spikes and noise from detector saturation, especially in DSS, Sloan, and WISE 1 images. When we launched Disk Detective we did not explicitly explain these phenomena anywhere on the site, though we readily answered questions about them in the chat forum TALK. Nonetheless, many of our first users interpreted bright stars with spikes as "oval" or in some cases "extended." On 2014 March 31, we edited the spotters guide and the tutorial at DiskDetective.org to include examples with diffraction spikes, which greatly reduced the problem, but some of these mistakes persist in our data. We expect this confusion over diffraction spikes to get better over time, as most of our classifications now come from participants who are well aware of the problem.

With nearly 300,000 total objects from AllWISE to classify and 18%–25% yield described above, we estimate that the online classification scheme at DiskDetective.org will produce up to 75,000 good objects, still a daunting number to investigate. However, the automated online classification stage is just the beginning of our vetting process. The next stage of the process aims to harvest DDOIs, subjects that we consider deserving of additional follow-up observations.

For this next stage of vetting, we created a collection of Google spreadsheets that both the science team and the superusers could edit. We first populated the spreadsheets with subjects chosen as "None of the Above/Good Candidate" using the latest classification data. We chose subjects in the right portion of the sky for any upcoming follow-up observations and also populated the lists in order of agreement fractions and brightness (in J band), ensuring that the bright subjects, and those with high agreement percentages were looked at first. Then we invited the superusers to add their own favorite objects, which they collect using tools on the TALK social network. We also invited the entire Disk Detective user community to submit subjects automatically via a Google form (but not directly edit the spreadsheet).

We coached the superusers on how to research each source in SIMBAD and VizieR (and sometimes NED) to fill in information about spectral type, proper motion, variability, parallax, prior observations, and make comments on the SED, etc. Then we checked the superusers's comments and selected the follow-up targets (DDOIs) from the list based on the following criteria.

• 1.  
  SIMBAD object descriptions excluding post-AGB stars, carbon stars, novae, Cepheids, cataclysmic variables, high-mass X-ray binaries, eclipsing binaries, galaxies, Active Galactic Nuclei, planetary nebulae, reflection nebulae, rotational variables, symbiotic stars, or Wolf–Rayet stars. Note that we did keep sources with SIMBAD object descriptions Shell Star, Orion Variable, and White Dwarf.
• 2.  
  No Long Period Variables, SR+L, Slow Irregular Variables, Miras, Semi-regular Variables, Semi-regular Pulsating Variables, or Carbon stars based on literature searches.
• 3.  
  Including only spectral types B through M according to SIMBAD, when a type is available.

Only about half of the subjects have entries in SIMBAD, so we often relied on VizieR to help us search the relevant literature. Many of the subjects had unknown spectral types, and many were severely reddened. So at this stage of the vetting process, we often simply labeled sources as "late-type based on color."

• 1.  
  No sign that the WISE 1 photometry drops out due to saturation, based on visual inspection of the SED.
• 2.  
  No known companions within 16'', except for spectroscopic binaries.
• 3.  
  For M stars and other subjects with $V-J\gt 1$, we require [W1] − [W4] > 0.9. The peak of the thermal emission from cool star photospheres may lie at long enough wavelengths that the Rayleigh–Jeans limit no longer accurately describes the photosphere's [W1] − [W4] color even in the absence of circumstellar dust. Hence, we impose this more stringent requirement of 0.9 mag of 22 μm excess for these cool subjects. For some red subjects, however, we find that the SED clearly curves upwards at W4; we do not exclude these subjects.

In the process, we naturally rediscovered many known disks, so we added the following additional criterion:

• 1.  
  No sources that have already been imaged by a pointed space mission (i.e., Spitzer, Herschel, HST) or 8 m class telescope (i.e., Keck, Gemini, VLT, Magellan; based on the published literature), except those without good quality spectral types (SIMBAD quality C or higher).

Our search includes many sources in and near the Galactic plane and many sources with no parallax measurements from Hipparcos. But since these sources require extra care, we have added two additional criteria for the purpose of this paper:

• 1.  
  No sources within 5° of the Galactic plane.
• 2.  
  Only sources with parallax measurements from Hipparcos.

Most of the subjects on the vetting spreadsheets do not meet these additional eight criteria. But sources that do meet all these criteria (plus the criteria in Table 1, of course) we label as DDOIs and place in our our queue for follow-up observations. So far we have collected 770 DDOIs in total that have survived the above vetting by the science team and/or by multiple superusers. Of these, 517 have classification histories via the main Disk Detective online classification tool. The remaining DDOIs have minimal classification data, since they were submitted directly by volunteers, and some volunteers choose to flip through images on TALK rather than the main site.

The yield at this stage varies greatly depending on how we rank the spreadsheets. Higher galactic latitudes provide higher yields, as do sources that are brighter in WISE 4. Also, as our users have become more educated, they have become better at selecting subjects to place on the spreadsheets, which raises the yield. But as of now, the typical yield at this stage of vetting (DDOIs per source on the vetting spreadsheet) is about 12%.

Multiplying the size of the input catalog by the automated vetting yield (24.4%) and then by the DDOI vetting yield (about 12%) gives a final yield of about 3% and an estimate of the total number of DDOIs we ultimately expect to discover of ∼8000. This estimate suggests that our search is presently about 10% complete. However, it is worth noting that many of these criteria, while catching a lot of the aforementioned AGB stars that pollute our "good" candidate list, also eliminate a population of YSO disks and M stars (for example, YSOs can be irregular variables).

The DDOIs have all been carefully vetted by hand and researched in the literature by multiple scientists and/or well-trained superusers. So this list of sources serves as a reference set of subjects that we can use to make decisions about how to run the vetting.

One major consideration of any citizen science project is when to remove, or "retire," subjects from classifications on the site. To make that decision, we need to know roughly how long it takes the user population to converge on an answer. Figure 2 shows how the standard deviation of the agreement for majority-algorithm-ruled "good" candidates in our DDOI set varies with number of total classifications. For these subjects, "agreement" is defined simply as the number of good classifications divided by the total number of classifications the subject has received. Between 10 and 15 classifications, the standard deviation of the agreement for this subset of "good" DDOI candidates levels off, suggesting that there would be minimal marginal benefit from requiring additional classifications beyond this point. Thus the current iteration of the site uses a conservative 15 total classifications as our benchmark for when to retire subjects.

Zoom In Zoom Out Reset image size

Table 3 lists our first batch of 50 DDOIs. Although some of the sources in Table 3 are Be stars or disk candidates that have been previously reported, 37 of these sources are new disk candidates from the Disk Detective project.

We have decided to present only stars in with Hipparcos parallaxes as a test of our spectral classification pipeline. Though Wu et al. (2013) performed a large search for stars in the Hipparcos catalog that have excess emission in W4, they used a $[{K}_{s}]-[{\text{}}W4]$ cutoff, and the All-Sky catalog. Our selection criteria $[{\text{}}W1]-[{\text{}}W4]$ and input catalogs (AllWISE) are different, so we find objects they missed, even in the Hipparcos catalog.

Table 3 lists the photometry of these disk candidates from 2MASS and SIMBAD (V band). It also lists the spectral types we derived from our FAST spectra, informed by photometry and Hipparcos distances, and WISE photometry. Eleven of the sources in Table 3 have previously been reported as disk candidates by Zuckerman et al. (2012), Kennedy & Wyatt (2012, 2013), Wu et al. (2013), Wahhaj et al. (2013), and Chen et al. (2014). We include these objects here as a check on our consistency with other searches. However, 37 of these sources have not been previously reported as disk candidates (and are not Be stars).

The WISE photometry in Table 3 is taken from the All-Sky data release and corrected for saturation according to the formulae in Patel et al. (2014). This corrected All-Sky WISE photometry is more accurate for sources brighter than about the 8th magnitude in W1. No color correction is applied to the WISE photometry, since the sources have nearly Rayleigh–Jeans spectra. If we assumed instead that the emission were entirely from a 200 K blackbody disk, the correction to WISE 4 would be −0.016 mag (Wright et al. 2010), negligible compared to our 0.25 mag WISE 1–WISE 4 excess criterion for inclusion in Disk Detective.

The vast majority of our DDOIs did not have previously identified luminosity classes. We assigned them using the approach of Patel et al. (2014), who separated dwarfs from giants in their catalog via a simple cut on an HR diagram, retaining only stars with ${M}_{V}\gt 6.0(B-V)-1.5$. Figure 3 shows an HR diagram for our candidates. The stars on Figure 3 are color coded according to spectral type: blue = B, green = A, and red = F. Most of our disk candidates are probably dwarfs since they fall below the Patel et al. (2014) cut, which is shown by the dashed line.

Zoom In Zoom Out Reset image size

If we conservatively assume that the quality of our spectra corresponds to the "C" quality flag in SIMBAD ("A" is the highest quality), then of the 50 targets with new spectral types that we present here, 48 show clear improvements in quality over the published literature, while the remaining 9 targets have spectral types of the same quality as the literature. The root-mean-square change in spectral types versus SIMBAD for all 50 targets was 2.64, with 27 objects shifted at least one subtype toward lower temperature and 12 objects shifted at least one subtype toward higher temperature.

Figures 4 and 5 summarize some more properties of the stars in Table 3: their distribution on the sky and their distance distribution. Their distance distribution peaks at roughly 110 pc; at this distance, a large telescope with adaptive optics like the Gemini Planet Imager could image an analog to the HR 8799 planetary system and comfortably detect at least two of the planets. For comparison, Patel et al. (2014) limited their study to sources with distances <120 pc. But the distance distribution of the DDOIs is grossly similar to that of the disk candidates in Wu et al. (2013), which peaks at about 90 pc. The distribution of our DDOIs also includes a long tail of objects beyond 200 pc; roughly one-fourth of our DDOIs are in this long tail, which consists mainly of B and early A dwarfs. Figure 5 also shows the distances to some well-known disks for references. The distribution of the DDOIs on the sky is shaped mainly by the range of declinations accessible to FAST and our decision to add sources to the Web site grouped by Galactic latitude.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Figure 6 compares our disk candidates with the Wu et al. (2013) color selection criteria. We did not impose such a criterion, but all of our disk candidates fall to the right of the dotted line in this figure, showing that they meet the Wu et al. (2013) criterion anyway. On average, the disk candidates in this paper are somewhat redder than those in Wu et al. (2013).

Zoom In Zoom Out Reset image size

We fit simple models to the SEDs consisting of a stellar photosphere plus a blackbody dust component using a Levenberg–Marquardt algorithm. These fits yield constraints on the dust temperature and the fractional infrared luminosity, f, the total bolometric power emitted by our single-temperature blackbody disk model divided by the total bolometric power emitted by our blackbody stellar model for each source. For stars with excess at only WISE 4, these fits yield only an upper limit on the dust temperature and a lower limit on f.

Figure 7 shows SEDs for six of the new disk candidates, together with a simple model for the flux. Our SEDs employ WISE 1 and WISE 2 fluxes from the All-Sky survey catalog corrected for saturation using the formulae in Patel et al. (2014). For these models, we supplemented the WISE and 2MASS photometry with UBVRI photometry from SIMBAD when it was available. Dashed lines show blackbody models fit to the star and to the disk component; the solid line shows the total model flux.

Zoom In Zoom Out Reset image size

Table 4 summarizes our SED modeling of all the 39 new disk candidates reported in Table 3. The temperature uncertainties are 1–σ uncertainties from the shape of the two-parameter χ² surface near the minimum. The fractional infrared excesses, f, listed in this table should all be considered lower limits. When the temperature listed is an upper limit, the listed fractional infrared excess corresponds to a blackbody disk with the listed upper limit temperature.

All of the DiskDetective.org subjects were pre-selected to have [W1]–[W4] excesses significant at the 5-σ level or better, based on the AllWISE catalog. We also checked the All-Sky catalog of WISE photometry for these sources, correcting this photometry via the correction factors in Patel et al. (2014). All of the sources in Table 4 still have [W1]–[W4] excesses significant at the 5-σ level or better using the corrected All-Sky photometry.

Additionally, eight of our DDOIs also turned out to have significant (>3σ) W3 excess based on corrected All-Sky WISE photometry, as you can read in Table 5: HD 4670, HD 7406, HD 14685, HD 19257, HD 20994, HD 164137, HD 173056, and HD 218546.

Five of the stars with newly reported excesses have known companions or secondaries. None of these is apparent in the WISE images or likely to be cool enough to create the observed WISE infrared excesses.

• HD 7406. The separation of this binary is ∼1' (Schmeidler 1987), wide enough that the companion probably does not contaminate the WISE photometry.
• HD 74389. HD 74389 has several nearby companion stars that could conceivably contaminate the SED. This star has a white dwarf companion discovered by Sanduleak & Pesch (1990), type DA1.3, V = 14.62 located 2011 (2230 au projected separation) to the west (Holberg et al. 2013), which is readily visible in the DSS Blue, Red, and IR images. Holberg et al. (2013) also report a nearby M dwarf companion to the west and a nearby sdB companion to the north. Indeed, the same DSS Blue, Red, and IR image bands also show a second background object ∼13'' to the north of the star, probably the sdB.

However, none of these objects is visible in the WISE or even the 2MASS images, which look like unsaturated clean point sources, probing a resolution at K band of <4''. Moreover, we only see a significant excess at W4, but not at W3. The W3–W4 color of an M dwarf or sdB photosphere is ∼0 since the SEDs of these objects are in the Rayleigh–Jeans limit beyond 10 μm, so the lack of a resolved companion at the K band rules out a significant contribution to the W4 excess from the companion's photospheric emission. An M dwarf or sdB unresolved by 2MASS might still be cause for concern if it were itself dusty. But since the A star is the most luminous object in the system, it seems most likely that we are observing reprocessed light from dust around the A star.

See below for further discussion of this interesting object.

• HD 23873 is a member of the Pleiades, and listed as a binary in the Hipparcos Input Catalog (Turon et al. 1993). The separation of the binary is 11000 au according to Makarov & Robichon (2001), corresponding to roughly 15.
• HD 207888 was identified as a visual binary by Scardia (1988). The separation is wide enough (∼1') that the companion probably does not contaminate the WISE photometry.
• HD 22128 is a double-lined spectroscopic binary system with a semimajor axis of roughly nine solar radii. We found a spectral type of F0IV for the combined light, but in fact, the two components are A stars; see Folsom et al. (2013) for a detailed spectral analaysis. This star was first identified as a debris disk host star by Silverstone (2000) based on Infrared Space Observatory photometry (60–100 μm). We are the first to report excess emission from this source at WISE 4, excess at the 8-σ level. Wyatt et al. (2007) considered the object to be an "anomalous system" based on the high luminosity of its inferred disk given the star's age. We note that the system appears slightly extended in 2MASS K, but not in other bands. No 18 μm flux measurement for this object appears in the AKARI catalog (Ishihara et al. 2010).
• HD 224098. This B8V star is listed as having faint optical secondary, type F0V, in Lindroos (1985). No data on the separation of the secondary is provided.

Two of our new disk candidates have known stellar variability of <=0.04 magnitudes in V band. This degree of variability is too small to affect our selection process.

• HD 152308 is classified as an α2 CVn variable star by Dubath et al. (2011) and as type A0 Cr Eu according to Renson & Manfroid (2009). The variability is 0.04 mag in the Hipparcos system, with a period of 0.94 days. Our classification yielded B9V.
• HD 218155. This star varies over a range of ∼0.03 magnitudes in the V band (Watson 2006).

Aside from normal main-sequence stars, our program reveals four new Be stars: HD 6612, HD 7406, HD 164137, and HD 218546. None of these stars appears in the Be Star Spectra (BeSS) database (Neiner et al. 2011) or in the Kohoutek & Wehmeyer (1999) catalog of Hα emission-line stars, which cataloged stars with galactic latitude $| b| \lt 10^\circ$. All show prominent Hα emission and the strong upper level Balmer absorption lines characteristic of classical Be stars (e.g., Porter & Rivinius 2003; Rivinius et al. 2013). Two of these stars, HD 7406 and HD 218546, show significant W3 excess, in addition to their W4 excess. The B1e star HD 7406 has additional strong He i absorption lines; we detect weak He i absorption in the other stars. Our high S/N spectra reveal no trace of other permitted or forbidden emission lines, confirming our classification of them as classical Be stars rather than pre-main-sequence Herbig Ae/Be stars (Raddi et al. 2013; Rivinius et al. 2013).

These stars illustrate the promise of Disk Detective for identifying new classical Be stars. Although Be stars have large infrared excesses due to free–free emission (e.g., Gehrz et al. 1974; Dougherty et al. 1994; Bragg & Kenyon 2002), most have been identified from optical spectroscopic surveys (e.g., Slettebak 1985; Bragg & Kenyon 2002; Porter & Rivinius 2003; Raddi et al. 2013, 2015; Rivinius et al. 2013) optical photometric surveys (e.g., Wisniewski & Bjorkman 2006), and IR spectroscopic surveys (e.g., Chojnowski et al. 2015). Uniform selection based on infrared excess from WISE data provides a flux-limited survey fairly independent of interstellar reddening.

With only four Be stars, it is premature to analyze spectroscopic properties and statistics for the WISE sample of classical Be stars. However, it is worth noting that the frequency of late-type Be stars (three B8–B9 stars and one B1 star) is somewhat larger than expected from current samples where B4 and earlier stars are much more common (e.g., Bragg & Kenyon 2002; Rivinius et al. 2013). We plan a detailed analysis of a larger sample in a future paper (A. S. Bans et al. 2016, in preparation).

One of our new candidate debris disks appears to orbit HD 74389 A, an A0V star with a white dwarf companion discovered by Sanduleak & Pesch (1990). It also has a possible M dwarf companion and sdB companion reported by Holberg et al. (2013). There are three known planetary systems with white dwarfs as distant companions: Gl 86 (Queloz et al. 2000; Mugrauer & Neuhäuser 2005), HD 27442 (Butler et al. 2001; Chauvin et al. 2006; Raghavan et al. 2006; Mugrauer et al. 2007), and HD 147513 (Porto de Mello & da Silva 1997; Mayor et al. 2004). However, the HD 74389 system appears to contain the first debris disk around a star with a white dwarf companion.

The white dwarf companion, HD 74389 B, has V mag 14.62, and is located 20.11 arcsec to the west (Holberg et al. 2013). It is readily visible in the DSS Blue, Red, and IR images. These bands also show a background object ∼13'' to the north of the star, possibly the M dwarf companion reported by Holberg et al. (2013).

It is interesting to ponder the origin of the disk around HD 74389 A and how the post-main-sequence evolution of HD 74389 B may have affected it. Though the white dwarf may presently have a projected separation of 2230 au, it could have been two to four times closer in when it was on the main sequence, thanks to stellar mass loss. The disk may have merely survived the evolution of the higher-mass star, or it may represent a signature of dynamical changes to the system, like planet exchange (e.g., Kratter & Perets 2012). It could even have been built from mass loss by the companion, via the process described by Perets & Kenyon (2013).