Many eyes on the archive

On 21 February 2017, six days after the launch of the Backyard Worlds: Planet 9 project, a volunteer posting on the project’s discussion board pointed out a “small dipole” in one of the animated image sets: a faint black-and-white pair of blobs that flickered as the frames cycled. Two days later another volunteer submitted the same source through the project’s reporting form. It was the 117th submission the team had received. A month after that, on 19 March, the SpeX spectrograph on NASA’s 3 m Infrared Telescope Facility showed that the dipole was a T5.5 brown dwarf about 34 parsecs away, moving across the sky at about 0.7 arcseconds per year (Kuchner et al. 2017).

Nothing about this object was hidden in the usual sense. It already had an entry in the AllWISE catalogue. What it lacked was an identity: nobody had noticed that the entry moved, and the earlier motion searches of WISE data, which worked on single exposures, had not reached its brightness. The volunteers saw it in coadded images about 0.9 magnitudes deeper than those searches. The gain came from the coadds and the four-and-a-half-year baseline, and from people willing to look at more than a million small animated images.

This chapter is about that kind of discovery. Its central argument can be put in one sentence: human attention often recovers signals that a pipeline’s assumptions threw away, but that is not the same as humans out-detecting an optimal algorithm at equal photon counts. Separating those two claims, and measuring what the human stage really costs, is where the algorithmic opportunity lies.

Three ways the public finds things

Amateurs have contributed to astronomy for as long as there has been astronomy, and in organised form at least since the American Association of Variable Star Observers was founded in 1911 (Kuchner et al. 2017). For planetary bodies the contributions now fall into three groups, which often meet in the same discovery.

The first is distributed classification. A professional team prepares images from a large archive (WISE, Hubble, the Dark Energy Camera, SOHO), cuts them into small “subjects”, and asks volunteers on a web platform, usually Zooniverse, a simple question about each. The volunteers do not need a telescope or software, and each sees only a small part of the data.

The second is skilled independent search. Some volunteers grow into experts who write their own catalogue queries, build software, and design searches the original project never planned. Backyard Worlds produced several such people, and some now co-author papers and attend the science team’s calls (Faherty et al. 2021).

The third is independent observation. Amateur astronomers with their own telescopes, or with access to public spacecraft data, still discover comets, measure asteroid positions to Minor Planet Center standards, and confirm other people’s discoveries. Amateur occultation timing is covered in Chapter 25, and volunteer transit searches in Chapter 30.

A discovery chain typically passes through all three: someone notices an anomaly, a catalogue lookup or orbit fit tests it, and a telescope secures the evidence.

A short history

Comet hunters with their own telescopes

Independent comet hunting survived the arrival of the large surveys. On 23 July 1995 Alan Hale, in Cloudcroft, New Mexico, and Thomas Bopp, in Stanfield, Arizona, independently found the comet now called C/1995 O1 (Hale–Bopp), at that time 7.15 au from the Sun (NASA, Hale–Bopp). Its diffuse appearance, not its point-source brightness, was what caught the eye. The earlier visual and photographic history of comet hunting is told in Chapter 3 and Chapter 4.

The most consequential recent amateur discovery is the second known interstellar object. On 30 August 2019 Gennady Borisov reported a comet he had found at his own observatory, MARGO, in Nauchnij, with a 0.65 m f/1.5 astrograph and a CCD. The Minor Planet Center’s circular records a condensed coma about 7 arcseconds across, confirmations by other observers including Hidetaka Sato and Peter Birtwhistle, early hyperbolic-orbit calculations by Davide Farnocchia, Bill Gray and David Tholen, and a caution that all the observations were made low in the sky close to the Sun (MPEC 2019-R106). Separately, Piotr Guzik and colleagues’ software flagged the object to them on 8 September; their early fit to 447 positions over 21.1 days gave an eccentricity of 3.379 ± 0.020, far from anything bound to the Sun (Guzik et al. 2019). That number was an early solution, not the current orbit. Discovery by an amateur, confirmation by a network, and computational recognition of interstellar origin were three separate contributions. Chapter 14 covers interstellar objects in general.

SOHO: a spacecraft archive as a public telescope

The Solar and Heliospheric Observatory, launched in 1995 and in routine operation from April 1996, carries the LASCO coronagraphs, which block the solar disc and image the corona out to many solar radii. Their images are public, and they show a region of sky that is hard to search from the ground: the immediate neighbourhood of the Sun, where sungrazing comets brighten and die. Amateurs began searching the images and reporting comets that the mission’s own analysis had not picked out.

Karl Battams and Matthew Knight trace the organised effort, which later became NASA’s Sungrazer Project, to 2000. Doug Biesecker curated it from the start through 2003, Doug Hammer for part of 2003, and Battams from October 2003 onward (Battams & Knight 2017). NASA’s current project page gives the launch as 2006, which seems to refer to a later programme identity and should not erase the earlier documented work (NASA, Sungrazer Project). By 31 May 2016 the count stood at 3,138 new comets, 2,692 of them members of the Kreutz family of sungrazers, in a search dominated by amateurs. On 25 March 2024 Hanjie Tan, originally from Guangzhou, a participant since childhood and by then a doctoral student in Prague, identified SOHO’s 5,000th comet. NASA put his personal total at over 200 comets since 2009 (NASA 2024). His biography is a reminder that “amateur”, “volunteer” and “professional” are stages in a life, not fixed categories.

The tradition continues on other instruments. In late March 2025 Vladimir Bezugly, Michael Mattiazzo and Rob Matson independently found comet C/2025 F2 (SWAN) in images from SOHO’s SWAN instrument, which maps ultraviolet Lyman-alpha light, not the visible light of the coronagraphs (APOD 2025 April 18). In 2022 NASA ran an open challenge to automate the LASCO search. It drew 596 registrants from 73 countries and 212 submissions, selected seven winners, and turned up two previously unknown comets (NASA 2022). The announcement gives no completeness curve, false-positive rate or timing, so it cannot support any claim about how a machine search compares with the human one.

Precovery by click

Archives also invite a different task: not finding unknown objects but finding known ones in old images. In July 2011 Enrique Solano, Carlos Rodrigo, R. Pulido and Benoît Carry opened a Spanish Virtual Observatory project that predicted where known near-Earth asteroids should appear in Sloan Digital Sky Survey images and asked volunteers to find and measure them. In its first fifteen months, 3,226 registered users made more than 167,000 measurements. After checking, these became 2,351 accepted new positions of 551 near-Earth asteroids. For 130 of those asteroids the images predated the earliest known observation, so they were true precoveries, and 24 of the 130 were potentially hazardous (Solano et al. 2014). Those are four different denominators (clicks, accepted positions, improved objects, precoveries), and a careless reader can turn 167,000 into a discovery count. Precovery as a computational problem is the subject of Chapter 13.

In 2013 Zooniverse announced Asteroid Zoo, with the company Planetary Resources, proposing to search more than three million Catalina Sky Survey images and to use the volunteers’ labels to train better automatic detectors (Zooniverse 2013). The announcement documents a design, not a yield, but the idea that volunteer labels should become training data was there from the start.

Backyard Worlds

Marc Kuchner, Jacqueline Faherty, Aaron Meisner, Adam Schneider, Laura Trouille and colleagues launched Backyard Worlds: Planet 9 on 15 February 2017. Its stated aim was to find objects with significant motion in WISE infrared images, with a focus on brown dwarfs and on possible distant planets in our own Solar System (Schneider et al. 2026). The name points to the Planet Nine hypothesis of Chapter 8, but the project’s yield has been brown dwarfs, not planets, and its catalogue is not a calibrated exclusion map for a distant planet (Chapter 9 discusses what such a map requires). Early on, roughly 76,000 unique IP addresses had made more than 3.4 million classifications (Kuchner et al. 2017). An IP address is not a person, so the participant number is only indicative.

The project’s history is also a history of volunteers turning into instrument builders. Dan Caselden wrote WiseView, an interactive blinker for WISE epochs; Frank Kiwy wrote AstroToolBox, a Java suite for catalogue overlays, epoch blinking and photometric typing; the collaboration built WRAP to tie image inspection to cross-survey catalogue identification (Faherty et al. 2021; Brooks et al. 2023). By 2026 the project had published 3,006 motion-confirmed L and T dwarf candidates (Schneider et al. 2026) and a candidate Y dwarf found by a neural network and a volunteer working together (Meisner et al. 2026). Both are described below.

Hubble, active asteroids and Rubin

Sandor Kruk, Bruno Merín, Pablo García Martín, Marcel Popescu and colleagues launched Hubble Asteroid Hunter on 21 June 2019, just ahead of International Asteroid Day. It ran until August 2020. Volunteers looked for the curved trails that asteroids leave across Hubble images, and their labels trained a commercial neural detector that then searched the whole archive (Kruk et al. 2022). A 2024 paper used the curvature of the trails to estimate distances (García-Martín et al. 2024).

Colin Orion Chandler and colleagues launched Active Asteroids on 31 August 2021. It shows volunteers cutouts of known minor planets from Dark Energy Camera (DECam) images and asks whether each shows a tail or coma. The project’s 2024 overview is the best-documented account of crowd classification for planetary science that exists, including an honest accounting of its throughput shortfall (Chandler et al. 2024). Follow-up continues in the team’s own papers. A 2025 study by Kennedy Farrell, Chandler and colleagues found a second activity episode on the known quasi-Hilda comet 362P using new and archival images (Farrell et al. 2025). In 2025 NASA listed a related project, Rubin Comet Catchers, which shows volunteers animated sequences from the Vera C. Rubin Observatory rather than single images (NASA, Rubin Comet Catchers). Its page describes a workflow and gives no discovery count yet. The physics of cometary activity belongs to Chapter 27.

What a volunteer actually sees

A volunteer never sees photons. They see an image that has been coadded, differenced, clipped, stretched and coloured, and each step decides which signals are visible.

The image model

A common model for the $k$-th epoch image is

where $B_k$ is the diffuse sky and detector structure, $P_k$ is the point-spread function (PSF, the image of a point source), $F_{s,k}$ is the flux of source $s$, $\mathbf x_s(t_k)$ is its position at time $t_k$, and $\epsilon_k$ is noise. The noise is Poisson in the electron counts plus read noise, and resampling, stacking and subtraction correlate it between pixels. More important for a human search is the structured contamination: cosmic rays, optical ghosts, latent images, saturated stars, background galaxies, and residuals from imperfect registration. These are not extra Gaussian variance. They are a population of impostors, and the tutorials of every project discussed here are largely catalogues of impostors.

For a nearby star or brown dwarf the position follows a tangent-plane model,

where $\boldsymbol\mu$ is the proper motion, $\varpi$ the parallax and $\mathbf p(t)$ the known parallax factor set by Earth’s (or the spacecraft’s) orbit. A Solar System body needs a projected orbit and the observer’s actual ephemeris instead. Either way, a person watching a flipbook is testing temporal coherence: something moves on a plausible path while everything else stays put.

The Backyard Worlds flipbook

The original Backyard Worlds display is specified exactly (Kuchner et al. 2017). It starts from four time-resolved unWISE coadds $E_1,\dots,E_4$ in each of the W1 (3.4 µm) and W2 (4.6 µm) bands, and forms differences against the median of the other pair:

The differences are clipped to between $-0.5\sigma$ and $9.5\sigma$, a smoothed background is removed with a 12 × 12 pixel top-hat, an arcsinh stretch is applied, and W1, W2 and their mean are mapped to blue, red and green. Each frame is a 256 × 256 pixel subtile, one sixty-fourth of an unWISE tile of 1.56° × 1.56°. With 18,240 tiles, that is 1,167,360 nominal subtiles, before overlaps and gaps in epoch coverage. Stationary sources largely cancel; a moving source survives as a pair of positive and negative residuals.

The display is a good one for eyes. It is not a sufficient statistic: it throws away information that a full motion fit would use, and it has a sensitivity set by its own clipping and stretch.

Why a slow mover looks like a dipole

Take a single source of flux $F$ that moves by a small displacement $\boldsymbol\delta$ between two epochs. The difference image is

by a first-order Taylor expansion of $P$. The gradient of a peaked PSF is positive on one side and negative on the other, so the residual is a dipole, a bright lobe in the direction of motion and a dark lobe behind. Three consequences follow from this one line.

First, a slow mover can be recognised before its two images have separated, because the dipole appears at sub-PSF displacements. Second, the dipole’s amplitude scales with $|\boldsymbol\delta|$, so it fades to nothing as the displacement shrinks. Third, a small registration error between epochs produces exactly the same leading-order pattern around a stationary star. A dipole alone cannot tell motion from misalignment. The discriminant is consistency: a real mover’s dipoles all point along one track across epochs, while registration residuals point the same way for every star in the field. More votes cannot repair a temporal signal that is not there; a longer baseline can.

Coadds, depth and what volunteers gain

Why did volunteers find a brown dwarf that earlier WISE searches had missed? Mostly because they were looking at coadds. For independent, background-limited exposures, stacking $N$ of them raises signal-to-noise by $\sqrt N$ and the limiting magnitude by

With $N=100$ exposures the gain is 2.5 magnitudes. The original Backyard Worlds coadds reached roughly W1 = 16.5 and W2 = 15.5 (Vega), spanned about 4.5 years from early 2010 to mid-2014, and kept sensitivity for objects moving slower than about 6 arcseconds per day within each epoch’s combination. The first discovery, at W2 = 15.37 ± 0.09, sat about 0.9 magnitudes below the single-exposure limit of the earlier searches (Kuchner et al. 2017). The volunteers were not seeing below the noise. They were seeing deeper images than anyone had searched for motion. The physics of stacking is the subject of Chapter 17 and Chapter 18, and the WISE mission itself is discussed in Chapter 24.

A matched filter makes the point sharper. For a template $s_\theta$ with parameters $\theta$ (position, motion) and data $d$ with Gaussian covariance $C$, the optimal score is

as derived in Chapter 1. A person’s judgement combines colour, shape, motion and context in a way that is not calibrated in units of $Z$. Nothing in the published record shows a volunteer beating a matched filter applied to the same coadds. What the record shows is volunteers finding things in data nobody had filtered that way, and in places where the automated filters had been told not to look.

Trails from a moving telescope

Hubble presents a different picture. During an exposure from $t_{k,0}$ to $t_{k,1}$, a moving object paints

a trail convolved with the PSF. From the ground the trail is usually straight. Hubble is in low Earth orbit, and its own motion during an exposure bends the tracks of nearby objects into visible curves. In a tangent-plane approximation,

where $\boldsymbol\omega$ is the object’s intrinsic angular motion, $\mathbf R_\perp$ the component of the telescope’s position perpendicular to the line of sight, and $\Delta$ the object’s distance. The curvature term scales as $1/\Delta$, so the shape of the trail encodes distance. García-Martín and colleagues used full SPICE ephemerides rather than this schematic form. For each trail they scanned 1,300 trial distances between 0.2 and 6.7 au, fitted the intrinsic motion, and kept the distance with the smallest residual, using a 0.05 arcsecond positional scale (García-Martín et al. 2024). A single trail leaves the radial velocity unconstrained, so they sampled 20,000 line-of-sight velocities subject to the orbit being bound. The result is a family of orbits, not an orbit.

Turning the trail’s brightness into a size takes the usual reflected-light relations,

where $H$ is the absolute magnitude, $m$ the apparent magnitude, $r$ and $\Delta$ the heliocentric and observer distances in au, $\Phi(\alpha)$ the phase correction at phase angle $\alpha$, and $p_V$ the geometric albedo (Chapter 2 derives them). Since $p_V$ is assumed, the diameter is an estimate. A human-verified curved trail is good evidence that an asteroid was there, and a weak constraint on what it was. The general problem of detecting streaks is covered in Chapter 11.

Activity as a residual

Active Asteroids asks a third kind of question. The object is known, its position predicted; the question is whether its image is wider than an inactive body’s would be. Written as a model,

with $F_n P_{\rm trail}$ the trailed nucleus and $F_c[P*C_\eta]$ a coma or tail template $C_\eta$ convolved with the PSF, depending on orientation and profile parameters $\eta$. The project does not fit this likelihood; volunteers judge the images by eye. But the model makes the confusions explicit. Poor seeing, an unresolved blend, trailing, or a galaxy behind the target can each broaden an image, and only an object-centred feature that persists across epochs is evidence of activity. Even then, a tail does not by itself say why the body is active: sublimating ice, an impact, or fast rotation can all release dust (Chandler et al. 2024).

The statistics of many votes

A vote fraction is a score, not a probability

Let $z_i\in\{0,1\}$ say whether subject $i$ contains the target, and let $y_{iu}\in\{0,1\}$ be volunteer $u$‘s answer. The simplest aggregate is the vote fraction

over the $n_i$ volunteers who saw subject $i$. Hubble Asteroid Hunter showed each cutout to ten volunteers and called $f_i$ the “probability” of a trail; cutouts with $f_i>0.5$ became 1,488 trail labels (Kruk et al. 2022). Calling $f_i$ a probability does not make it one. It is a score, and its relation to the posterior probability of a real object depends on how good the volunteers are, how rare the target is, and how independent the votes are.

The log-odds of a crowd

A simple generative model makes the dependence explicit. Give each volunteer a sensitivity $a_u=P(y=1\mid z=1)$ and a false-positive rate $b_u=P(y=1\mid z=0)$, and let $\pi_i$ be the prior probability that subject $i$ is a target. If votes are independent given $z_i$, Bayes’ theorem gives the posterior log-odds as a sum:

Each “yes” adds $\log(a_u/b_u)$ and each “no” adds $\log[(1-a_u)/(1-b_u)]$, which is negative. Good volunteers (high $a_u$, low $b_u$) contribute large increments; a volunteer who says yes to everything ($a_u=b_u=1$) contributes nothing, as they should. Weighted voting is an approximation to this sum.

Votes are not independent

The log-odds sum treats fifteen votes as fifteen independent pieces of evidence. They are not. Volunteers see the same display, read the same tutorial, and are fooled by the same impostors. If every pair of votes on a subject has correlation $\rho$, the effective number of independent votes among $n$ is

With $n=15$ and $\rho=0.1$, $n_{\rm eff}=6.25$; with $\rho=0.2$, it is about 3.9. An aggregation rule calibrated under independence will be overconfident by roughly the ratio $n/n_{\rm eff}$ in its log-likelihood. This is a model, not a measured correlation for any project; none of the papers discussed here reports one. It is a warning about redundancy: the fifteenth look at an image shown the same way adds much less than the first.

Weighting volunteers in practice

Active Asteroids built a weighting scheme from three measured quantities per volunteer (Chandler et al. 2024). Training accuracy is $M_{1,u}=Y_{\rm training}/T_{\rm training}$, the fraction of training images (all of which show activity) that the volunteer marked correctly. Experience is $M_{2,u}=\tfrac14\log_{10}\min(T_{\rm total},10^4)$, a logarithm of the number of images classified, capped at 10,000 so that it runs from 0 to 1. The optimism correction is $M_{3,u}=1-Y_{\rm sample}/T_{\rm sample}$, one minus the fraction of real (non-training) images marked active. The combined weight and subject score are

The optimism term exists because of a measured problem. About 35% of users, some 2,400 people, said “yes” to more than 20% of images, when the team expected no more than about 1% of images to deserve a flag. Since every training image contains activity, a volunteer who always answers “yes” scores perfectly on training accuracy. $M_3$ penalises that strategy. In the language of the generative model, $M_1$ estimates $a_u$ and $M_3$ is a proxy for $1-b_u$; the paper does not frame it this way, but the correspondence explains why both terms are needed.

The team tuned the threshold so that every image on a “control list” of vetted active objects was retained, and counted how many images the team would then have to inspect. Across about 170,000 development images, the unweighted vote fraction needed a threshold of 46.66% and sent 2,513 images to the experts. The combined weighting, at threshold $L_i\ge0.473$, sent 1,718, which is 795 fewer, a 32% reduction. Each metric on its own did worse than the combination. The team also found that the number of extraneous flagged images kept falling as the number of votes per image grew up to fourteen, and kept fifteen votes in the live project.

That is a real workload saving, but read what it is: retention of a constructed control list. It is not a blind measurement of completeness over all tail shapes. The same paper shows a real tail, on asteroid (62412) 2000 SY178, that scored 0.35 and fell below the threshold, and reports that in a later comparison experts had already flagged 138 thumbnails that the volunteer selection missed (Chandler et al. 2024). Many weighted votes still miss faint activity.

Aggregating clicks

When volunteers also mark positions, their clicks must be merged. Hubble Asteroid Hunter asked volunteers who saw a trail to click its two ends, and clustered the endpoints with HDBSCAN, a density-based clustering algorithm, with minimum cluster size and minimum samples both set to five (Kruk et al. 2022). This turns ten noisy annotations into one consensus geometry. It is cheap: for $m$ clicks per subject, low-dimensional density clustering costs roughly $O(m\log m)$, trivial next to neural inference or human time. It does not produce a photon-noise likelihood, and the endpoints of faint trails are uncertain enough that later measurement needed a 3 arcsecond tolerance when matching to known objects.

From candidates to populations

Discovering one object establishes that it exists. Using a sample of discoveries to say how many such objects there are requires the selection function

the probability that an object with properties $\theta$ survives every stage: archive coverage, preprocessing, the display, the human response, any machine ranking, submission, expert review and follow-up. Chapter 16 shows how surveys measure $S$ by injecting synthetic objects. No citizen-science project discussed here has published a full $S(\theta)$, and the Backyard Worlds authors say directly that their volunteers’ mixture of flipbook searches, custom catalogue queries and image analyses makes completeness “difficult to establish” (Schneider et al. 2026).

The projects in detail

Backyard Worlds: from flipbook to spectrum

The Backyard Worlds pipeline, as published in 2017, is unusually easy to reproduce (Kuchner et al. 2017):

1. Build time-resolved unWISE coadds in W1 and W2.
2. Split each of the 18,240 tiles into 64 subtiles of 256 × 256 pixels.
3. Form the four differences $D_1,\dots,D_4$ defined above.
4. Clip to $-0.5\sigma$ to $9.5\sigma$, with representative noise of 2.1 and 6.9 Vega nanomaggies in W1 and W2.
5. Subtract a 12 × 12 pixel top-hat-smoothed background and apply an arcsinh stretch.
6. Map W1 to blue, W2 to red and their mean to green, and animate the four frames.
7. Ask volunteers to mark movers and dipoles; let them discuss on the forum and submit strong candidates through the “Think You’ve Got One” (TYGO) form.
8. Check each submission against the images, catalogues and astrometry, and request follow-up.

Image preparation costs $O(KN)$ for $N$ pixels and $K$ epochs, since every operation is local and fixed-size. The human cost is $O(rS)$ classifications for $S$ subjects at redundancy $r$, times a variable inspection time. Neither number translates into GPU-hours.

The candidate stream grew large. By November 2025 the TYGO form had received over 102,000 submissions, every one inspected by the science team (Schneider et al. 2026). Those include duplicates, known objects, rejected artefacts and real movers that are not L or T dwarfs, so dividing the final catalogue size by 102,000 gives no meaningful false-positive rate. The inclusion criteria for the 2026 catalogue were discernible motion and a photometric spectral type of L0 or later. Motions came from existing proper-motion catalogues or from weighted least-squares fits to individual detections,

where $\mathbf x_k$ is the measured position at epoch $k$ with covariance $\Sigma_k$. The authors avoid mixing heterogeneous astrometric systems. The lesson for anyone cross-matching moving objects is simple: a nearest-neighbour match without propagating the motion to each catalogue’s epoch can attach another star’s photometry to a real mover. WRAP was built partly to prevent exactly this, by showing the user the image alongside each candidate catalogue row (Brooks et al. 2023).

The result is 3,006 new motion-confirmed L and T dwarf candidates, 2,357 with L-type and 649 with T-type photometric spectral types, plus 80 more candidates without significant motion measurements (Schneider et al. 2026). Photometric types are typically good to about ±1 subtype for field dwarfs, but these are candidates awaiting spectra. The 2020 Spitzer programme shows what the next step looks like. It observed 95 candidates selected as especially cold, confirmed the motion of 75, and found five motion-confirmed sources with Y-dwarf-like colours (Meisner et al. 2020). Because the sample was chosen to be cold, its success fraction says nothing about the project’s overall reliability.

Hubble Asteroid Hunter: humans train the machine

Hubble’s archive is a hostile place for a simple streak detector. Trails curve by different amounts, exposures vary, backgrounds are complex, and the standard cosmic-ray rejection in DrizzlePac partly erases bright moving sources, removing the centre of a trail and leaving its edges. The project prepared 37,323 composite images, which became 149,292 quadrants for automated inspection. Volunteers saw a subset: 144,559 uploaded cutouts, of which 132,878 were unique and 11,681 uploaded twice (Kruk et al. 2022). By July 2020, 11,482 volunteers had made 1,783,873 classifications.

The volunteers did more than vote on asteroids. On the discussion board they tagged other things they noticed, and the team turned those tags into labels: 1,343 cosmic rays, 698 gravitational-lens arcs and 1,673 satellite trails, alongside the 1,488 asteroid trails. These four classes trained a Google Cloud AutoML Vision object detector, with a random 70/15/15 split into training, validation and test sets. At a confidence threshold of 0.5 and an intersection-over-union of 0.5, asteroid detection achieved 73.6% precision and 58.2% recall, an F1 score of 65.0%. Modelling the confusing classes explicitly, rather than treating every elongated feature as a possible asteroid, was a sensible design that came directly from the volunteers.

Run over all 149,292 quadrants, including images added to the archive up to March 2021, AutoML proposed 2,041 trails. Of these, 1,044 matched volunteer trails within 4 to 5 arcseconds, so AutoML found 997 that volunteers had not, and volunteers found 444 that AutoML had not. Three of the authors inspected every candidate in the union. The flow from there is shown below.

There is a small inconsistency in the source. $1{,}488+2{,}041-1{,}044=2{,}485$, but the paper reports a union of 2,487, and its expert-review categories sum to 2,487. The downstream numbers are kept as published.

After detection came measurement. The team reran DrizzlePac on the original exposures without the trail-erasing step and extracted each centreline with a spline fit and a nine-pixel-wide aperture. About a third of trails needed manually supplied positions because the per-point signal-to-noise was too low, which is human help in measurement rather than discovery. Known objects were matched with a wide SkyBoT search (22,748 candidates within 900 arcseconds), Hubble-specific JPL Horizons ephemerides and a 3 arcsecond tolerance (Kruk et al. 2022).

The unmatched trails are fainter. The known objects have a median magnitude of 21.5; the unmatched ones a median of 23.1, 1.6 magnitudes fainter, and the catalogue reaches about 24.5. Near the ecliptic the trails imply about 80 asteroids per square degree brighter than 24.5. But 1,031 unmatched trails are not 1,031 new asteroids with orbits. They are observations of unidentified objects, some of which may be the same body, and most of which are too old and too short to recover. The 2024 parallax analysis, after quality filtering, kept 632 objects, 178 known and 454 unknown (García-Martín et al. 2024). The number 454 counts known objects in the first paper and unknown objects in the second; they are different populations.

Active Asteroids: ephemeris-conditioned cutouts

Active Asteroids makes the human task tractable by conditioning on known objects. Its pipeline, HARVEST, is written in Python with compiled utilities, Bash and Slurm, and a MySQL database (Chandler et al. 2024). In outline:

for each newly available calibrated DECam exposure:
    ingest metadata and estimate the depth (SNR 10)
    query which known minor planets are predicted in the field
    reject objects < 1 mag brighter than the depth, bad geometry, crowding
    reject predicted trails longer than 15 pixels (since 15 Aug 2023)
    cut object-centred FITS and PNG thumbnails
select a batch across dynamical classes and orbital phases
interleave known-active training images, with immediate feedback
collect 15 independent volunteer answers per image, then retire it
compute time-dependent volunteer weights and scores L_i
send L_i >= 0.473, plus forum escalations, to the science team
check other epochs and archives; request telescope time if warranted

Training images appear with probability 50% for a volunteer’s first ten classifications, 20% up to fifty, 10% up to a hundred and 5% thereafter. The 15-pixel trail cut was added because volunteers often mistook trailing for activity. The depth cuts, typically demanding one and later two magnitudes of margin above the estimated limit, reduce the workload. All of these are selection cuts that remove faint or fast objects, not improvements in sensitivity.

By July 2023 roughly 8,300 volunteers had made 6.7 million classifications of about 430,000 images, drawn from some 141,000 DECam exposures. The project’s yield, at the time of the 2024 paper, was previously unknown activity on fifteen asteroids and one Centaur (Chandler et al. 2024). Of the fifteen, four were active quasi-Hilda asteroids, seven Jupiter-family comets and four active asteroids (one main-belt comet and three candidates). The one Centaur, C/2014 OG392, was in fact found by the team while testing the workflow before launch; volunteers later flagged it independently, as they did other known active Centaurs.

The paper is most valuable for its arithmetic about throughput. HARVEST produces roughly 5,000 vetted thumbnails per day as the DECam archive grows. Volunteers retired a mean of 805 ± 678 images per day, with peaks around 8,000. The archive already held about 18 million vetted thumbnails. At the peak rate, classifying them would take 2,250 days, about six years; at the mean rate, about 60 years. And the deficit is growing: the mean daily shortfall is 5,000 − 805 = 4,195 images.

For Rubin the paper’s forecast is starker. Its authors estimate about 4 million minor-planet detections per night in LSST, about 1.2 million after HARVEST’s existing vetting cuts. One night of LSST would then take 150 days of volunteer effort at the peak rate. Tightening the depth margin to two or three magnitudes would reduce the load by about 30% or 60%, but would remove every thumbnail of an object fainter than about magnitude 20 at the stricter cut, discarding exactly the faint Centaurs the project most wants (Chandler et al. 2024). These are the authors’ forecasts, not measured Rubin performance. They make the point that the problem is a queueing problem before it is a classification problem. Chapter 15 describes the Rubin data flow these numbers would sit on.

Sungrazers: a learned trajectory prior

A LASCO search starts with time-ordered coronagraph images, background corrections suited to the instrument, and visual comparison against stars, cosmic rays and moving coronal structure. LASCO C2 has 11.9 arcsecond pixels and C3 56 arcsecond pixels, both on 1,024 × 1,024 detectors; their limiting magnitudes are roughly 9 and 8, depending on passband and the comet’s colour (Battams & Knight 2017). Those bright limits reflect the glare of the corona, not a weak telescope, and cannot be compared with an asteroid survey’s point-source depth.

Kreutz comets follow characteristic tracks with seasonal visibility, and experienced searchers carry those tracks in their heads. In effect they run a matched filter over a learned bank of trajectories. Battams and Knight require at least five images to confirm a Kreutz candidate. When the nominal cadence rose from three to five full-resolution images per hour after 2010, short-lived faint comets became eligible that would previously have appeared in too few frames. The raw annual totals jumped. Normalised by the number of images, the Kreutz discovery rate actually fell, and Battams and Knight argue that the underlying arrival rate of Kreutz comets stayed roughly constant.

Here a discovery count reflects the observing system rather than the sky. A track-before-detect search over LASCO images, summing PSF-weighted residuals along plausible paths, would cost $O(NKV)$ for $N$ pixels, $K$ frames and $V$ trajectory hypotheses. A family-restricted bank cuts $V$ but builds in the same bias as an experienced human: comets outside the known families are harder to find. Chapter 37 treats track-before-detect in general.

Amateur astrometry and campaigns

Amateurs with telescopes need calibrated images, accurate times, an astrometric solution against a star catalogue, a measurement of the moving object, and a clean report to the Minor Planet Center. Astrometrica, Herbert Raab’s long-lived Windows program, combines image alignment, blinking, PSF fitting, automatic moving-object identification, a link to the MPC orbit file, and track-and-stack (Astrometrica). Tycho Tracker advertises GPU-accelerated synthetic tracking on AMD, NVIDIA and Apple Silicon hardware, which brings motion-compensated stacking (Chapter 18) to small observatories (Tycho Tracker). No benchmark is published there, so no speed claim is adopted here.

The International Astronomical Search Collaboration (IASC) gives student and amateur teams image sets in month-long campaigns (IASC). Teams overlay known objects, solve astrometry, measure moving sources and submit reports. The instructions say plainly that if a source is hidden in one image, the team should report the images in which it is visible, rather than invent the missing position; three measurements can still be reported (IASC FAQ). IASC defines a “faint detection” as one not first found by the survey’s own processing, and distinguishes preliminary detections, which are unverified, from provisional ones, which the MPC has verified. Later observations can still change who gets discovery credit (IASC Faints). A campaign certificate is not the same as a numbered asteroid.

The SVO precovery project shows the difference between a targeted search and a blind one. It processed metadata for 938,046 SDSS DR8 images, kept predicted positions brighter than V = 22, and achieved detections around V = 21 to 21.5. Each position was measured by at least 15 users; if their consensus scatter was below 1 arcsecond in both coordinates, staff inspected it, and cases still unresolved after 45 users were examined separately (Solano et al. 2014). Why had the SDSS moving-object pipeline missed these asteroids? Its algorithm failed for motions faster than 0.5° per day and needed detections in enough filters, and some images fell outside its catalogue release. Of the accepted measurements, 1,696 came from images taken before May 2007 that were absent from the published moving-object catalogue, so the gain was not just newer data. Because the search was conditioned on a known orbit, it avoided almost all blind trajectory trials, at the cost of depending on that orbit being good enough. Chapter 13 develops this computational trade.

Neural ranking and the human eye: SMDET-1

The most recent result in this chapter is also the clearest example of a hybrid search. SMDET is a recurrent convolutional network, with 3D convolutional layers and 2D convolutional long short-term memory (LSTM) layers, trained on synthetic fast-moving sources injected into unWISE images (Meisner et al. 2026). It divides the sky into image sequences 2.9 arcminutes square, each with 16 time-resolved coadds per band in W1 and W2 spanning about 10.5 years, segments any moving source it sees, and ranks the sequences by the probability that they contain one. The network was run over the whole sky. Dan Caselden, who began as a Backyard Worlds volunteer, then inspected the 10,000 highest-ranked sequences by eye.

One of them was SMDET-1, moving at about 1.3 arcseconds per year at Galactic latitude +2.25°, in the crowded plane of the Milky Way. It crosses a brighter background source and stays blended with it in WISE. Spitzer resolves it better, with a 4.5 µm magnitude of 14.574 ± 0.089 and a colour limit $[3.6]-[4.5]>2.81$. That colour implies a Y-type object with an effective temperature below 391 K and a photometric distance less than 7.4 parsecs (Meisner et al. 2026). These are limits for a candidate. There is no measured parallax or spectrum yet.

SMDET-1 shows what a clean-source catalogue can lose. The object is not exceptionally faint; it was absent from useful motion catalogues because it was blended. A sequence model working on pixels, trained on injections, recovered it, and a person confirmed it. It also exposes an unsolved problem. A “top 10,000” list is a review budget, not a false-alarm probability, and nobody yet knows what the network ranked below position 10,000. Auditing that is the same problem as auditing a volunteer threshold.

Landmark results

Limitations

Selection is heterogeneous and largely unmeasured. No project here has measured $S(\theta)$, and the control-list thresholds of Active Asteroids are not population completeness: a real tail scored 0.35 against a threshold of 0.473. Motion confirmation shows that something moves. A photometric type is not a spectral type, and a spectral type is not a mass.

Positive-only training leaves the false-positive rate unmeasured. All Active Asteroids training images contain activity, so a volunteer who always says yes looks perfect on them, which is why the optimism term exists. That term could in turn penalise a genuine specialist who happens to see an enriched subset. This is an assessment, not a demonstrated failure. Training sets with both positives and negatives, chosen across a range of difficulty, would let projects estimate $a_u$ and $b_u$ directly.

Hubble detection depends on morphology. The method works poorly for trails shorter than about 20 pixels, which makes it inefficient for distant, slow objects such as trans-Neptunian objects (Kruk et al. 2022). Hubble’s pointings were chosen for other science, so its asteroid sample is serendipitous, not uniformly complete, and selecting trails with measurable curvature also selects distance and geometry.

Participation is extremely unequal. In the SVO project, 15% of registered users measured any asteroid and 1% contributed half the measurements (Solano et al. 2014). In Hubble Asteroid Hunter the mean was 155 classifications per volunteer and the maximum 85,333 (Kruk et al. 2022). Participant counts are a poor measure of capacity; retention and mentoring decide how many people become experts.

The computational evidence is thin. Neither the Hubble nor the SMDET papers give enough information about hardware, training cost or inference cost to compare human time with GPU time at equal completeness. Neural precision and recall at an intersection-over-union threshold are not astronomical false-alarm probabilities. The Hubble split was random over an archive that contains repeated cutouts and related exposures, so a split grouped by observation would be a stronger test of generalisation. That is a methodological recommendation, not an allegation of leakage.

Some limits are physical and some are not. Missing coverage, too few photons, unresolved confusion and orbital degeneracy are physical. Registration errors, trail-erasing cosmic-ray rejection, catalogue flags near bright stars, display choices and slow cross-identification are engineering. Follow-up capacity and data access are organisational. A faster classifier helps only with the second kind.

Where the room is

The citizen-science literature demonstrates very well that humans and machines are complementary, and demonstrates very little about how to spend modern compute on the combination. That gap is the opportunity. What follows is assessment and proposal, not demonstrated result; no order-of-magnitude speedup at equal sensitivity has yet been shown in this field.

Optimise the queue, not the classifier

The wall-clock cost of a hybrid search is roughly

where $N_{\rm inspect}$ items each take human time $\tau_h$ and $N_{\rm follow}$ candidates each take follow-up time $\tau_f$. In every project here the last two terms dominate. A faster tensor-core kernel for $T_{\rm detect}$ achieves nothing if it sends more ambiguous candidates to people. A more precise proposal stage can speed up the whole search while leaving the image processing unchanged.

Replace fixed redundancy with sequential decisions

Hubble used ten votes per image and Active Asteroids fifteen, for every image. The log-odds sum above suggests a sequential rule instead: keep asking until the posterior crosses an upper or lower threshold, retiring clear negatives early, sending disagreements to experienced volunteers, and continuing to collect evidence on rare promising cases. This is the logic of Wald’s sequential probability ratio test, discussed in Chapter 38. If easy negatives could be retired after two votes instead of fifteen, the saving on that subset alone would be 7.5-fold; this is an estimate from assumed vote counts, not a measured overall speedup. The available evidence urges care, since Active Asteroids found extraneous flags still falling up to fourteen votes. The opportunity lies in routing different subjects differently, not in lowering every retirement count. The thresholds must be calibrated on held-out real objects and injections, including confusing negatives, and the correlation between votes must be estimated rather than assumed zero.

Put physical models before neural ones

A plausible GPU hierarchy, in the spirit of Chapter 40, is:

batch calibrated pixels with variance and mask channels
compute cheap background, PSF, motion and morphology features
rank with a compact sequence model or a physical residual score
fit richer trajectory, PSF and contaminant models to the survivors
show ambiguous and anomalous cases to people, with multi-epoch context
randomly audit rejected cases; measure injected-source recovery
update the routing, keeping an independent evaluation set

Each project’s documented failures suggest a specific physical model. For Hubble, fit curved templates parametrised by distance and observer motion instead of an unconstrained curve bank. For WISE, fit a moving source jointly with its stationary neighbours, so that a blend like SMDET-1 or a companion like BD+60 1417B does not depend on a clean catalogue detection. For activity, compare each image with the expected trailed PSF and test object-centred residuals against other epochs. Batched convolution, local PSF fitting, resampling and fixed-shape sequence inference all suit GPUs well. An FFT can amortise a shared convolution over a whole image at $O(N\log N)$, and tree summations can share partial paths between trajectories (Chapter 21, Chapter 39). Neither removes the cost of a high-dimensional trajectory bank by itself.

The upstream costs are not small. Building unTimely, the time-domain unWISE catalogue that later Backyard Worlds work draws on, took 226,620 CPU-hours on Intel Haswell processors to produce 43,459,534,445 detections in 616,806 catalogues, a 4.3 TB release (Meisner et al. 2023). That is shared catalogue production, not a search bill. Any redesign should state which stage it accelerates: preparation, ranking, human review or follow-up.

Keep an exploration channel

The most interesting objects may lie outside any model’s training set. Volunteers on Hubble Asteroid Hunter tagged lens arcs and satellite trails without being asked, and those tags became classes. A good hybrid system reserves a random audit channel and a human anomaly channel alongside the high-ranked candidates, and records real objects that had low consensus. In active learning, the image that teaches the network most (an odd detector defect) is often not the one with most scientific value (a very cold candidate), and the two need separate budgets.

Benchmark the whole chain

The missing piece is an end-to-end benchmark. Inject moving sources and realistic tails into raw or minimally processed images, before the steps that can erase them, varying magnitude, motion, curvature, crowding and distance from bright stars. Run the same subjects through machine-only, human-only and hybrid routes, and report completeness at a fixed review budget, expert-confirmed precision, hardware, runtime and follow-up latency. Put catalogue-rejected cases such as BD+60 1417B in a protected challenge set. Define the target population first, so that a magnitude cut that drops faint Centaurs is never credited as a GPU speedup.

The most credible route to a large gain is to reuse image preparation, coalesce archive reads, batch physical scoring, and cut redundant human inspection while keeping audited exceptions. A large network run on every cutout may cost more than a physical prefilter followed by a small model. Chapter 42 gathers this and the book’s other open problems. Here the summary is short: the binding constraint is people’s attention, and the best algorithms will be the ones that spend it most carefully.

References

1. Battams, K., Knight, M. M. (2017). SOHO comets: 20 years and 3000 objects later. Philosophical Transactions of the Royal Society A 375, 20160257. arXiv:1611.02279
2. Brooks, H., Kirkpatrick, J. D., Caselden, D. et al. (2023). WRAP: A Tool for Efficient Cross-Identification of Proper Motion Objects Spanning Multiple Surveys. arXiv:2312.09294
3. Chandler, C. O., Trujillo, C. A., Oldroyd, W. J. et al. (2024). The Active Asteroids Citizen Science Program: Overview and First Results. The Astronomical Journal 167, 156. arXiv:2403.09768
4. Faherty, J. K., Gagné, J., Popinchalk, M. et al. (2021). A Wide Planetary Mass Companion Discovered Through the Citizen Science Project Backyard Worlds: Planet 9. The Astrophysical Journal. arXiv:2112.04678
5. Farrell, K. A., Chandler, C. O., Trujillo, C. A. et al. (2025). Recurrent Cometary Activity Discovered on Quasi-Hilda Jupiter Family Comet 362P/(457175) 2008 GO98. The Astrophysical Journal Letters 987, L35. arXiv:2507.17808
6. García-Martín, P., Kruk, S., Popescu, M. et al. (2024). Hubble Asteroid Hunter III. Physical properties of newly found asteroids. arXiv:2401.02605
7. Guzik, P., Drahus, M., Rusek, K. et al. (2019). Initial characterization of interstellar comet 2I/Borisov. Nature Astronomy. arXiv:1909.05851
8. Hill, D. (2022). Winners Selected for the NASA SOHO Comet Search with Artificial Intelligence Open-Science Challenge. NASA Science. science.nasa.gov
9. IASC (2026). International Astronomical Search Collaboration: home page, FAQ and faint-detection pages. iasc.cosmosearch.org; FAQ; Faints
10. Kruk, S., García Martín, P., Popescu, M. et al. (2022). Hubble Asteroid Hunter: I. Identifying asteroid trails in Hubble Space Telescope images. Astronomy & Astrophysics 661, A85. arXiv:2202.00246
11. Kuchner, M. J., Faherty, J. K., Schneider, A. C. et al. (2017). The First Brown Dwarf Discovered by the Backyard Worlds: Planet 9 Citizen Science Project. The Astrophysical Journal Letters. arXiv:1705.02919
12. Meisner, A. M., Faherty, J. K., Kirkpatrick, J. D. et al. (2020). Spitzer Follow-up of Extremely Cold Brown Dwarfs Discovered by the Backyard Worlds: Planet 9 Citizen Science Project. The Astrophysical Journal. arXiv:2008.06396
13. Meisner, A. M., Caselden, D., Schlafly, E. F. et al. (2023). unTimely: a Full-sky, Time-Domain unWISE Catalog. The Astronomical Journal. arXiv:2209.14327
14. Meisner, A. M., Caselden, D., Marocco, F. et al. (2026). SMDET-1: a Nearby Y Dwarf Candidate. arXiv:2608.00046
15. Minor Planet Center (2019). MPEC 2019-R106: Comet C/2019 Q4 (Borisov). minorplanetcenter.net
16. NASA Science (2026). C/1995 O1 (Hale-Bopp). science.nasa.gov
17. NASA Science (2026). The Sungrazer Project. science.nasa.gov
18. NASA Science (2025). Rubin Comet Catchers. science.nasa.gov
19. Nemiroff, R., Bonnell, J. (2025). APOD: 2025 April 18, Comet C/2025 F2 SWAN. science.nasa.gov
20. Raab, H. (2026). Astrometrica: research grade CCD astrometry. astrometrica.at
21. Schneider, A. C., Kuchner, M. J., Faherty, J. K. et al. (2026). Three Thousand Motion-Confirmed L and T Dwarf Candidates from the Backyard Worlds: Planet 9 Citizen Science Project. arXiv:2604.01323
22. Solano, E., Rodrigo, C., Pulido, R., Carry, B. (2014). Precovery of near-Earth asteroids by a citizen-science project of the Spanish Virtual Observatory. Astronomische Nachrichten 335, 142–149. arXiv:1302.5375
23. Thomas, V. (2024). ESA, NASA Solar Observatory Discovers Its 5,000th Comet. NASA Science. science.nasa.gov
24. Tycho Tracker (2026). Tycho Tracker. tycho-tracker.com
25. Zooniverse (2013). Anyone fancy an asteroid? Zooniverse blog. blog.zooniverse.org