Issue 
A&A
Volume 683, March 2024



Article Number  A122  
Number of page(s)  10  
Section  Planets and planetary systems  
DOI  https://doi.org/10.1051/00046361/202346771  
Published online  15 March 2024 
Hubble Asteroid Hunter
III. Physical properties of newly found asteroids
^{1}
Department of Theoretical Physics, Autonomous University of Madrid (UAM),
Madrid
28049,
Spain
email: pablo.garciamartin@estudiante.uam.es
^{2}
European Space Agency (ESA), European Space Astronomy Centre (ESAC), Camino Bajo del Castillo s/n, 28692 Villanueva de la Cañada,
Madrid,
Spain
^{3}
Astronomical Institute of the Romanian Academy,
5 Cutitul de Argint,
040557
Bucharest,
Romania
^{4}
University of Craiova,
Str. A. I. Cuza nr. 13,
200585
Craiova,
Romania
^{5}
Jet Propulsion Laboratory, California Institute of Technology,
Mail Stop 321100, 4800 Oak Grove Drive,
Pasadena,
CA
91109,
USA
^{6}
Bastion Technologies,
17625 El Camino Real #330,
Houston,
TX
77058,
USA
^{7}
Université Côte d’Azur, Observatoire de la Côte d'Azur, CNRS, Laboratoire Lagrange,
06304
Nice,
France
^{8}
Google Cloud,
6425 Penn Ave,
Pittsburgh,
PA
15206,
USA
Received:
28
April
2023
Accepted:
22
December
2023
Context. Determining the size distribution of asteroids is key to understanding the collisional history and evolution of the inner Solar System.
Aims. We aim to improve our knowledge of the size distribution of small asteroids in the main belt by determining the parallaxes of newly detected asteroids in the Hubble Space Telescope (HST) archive and subsequently their absolute magnitudes and sizes.
Methods. Asteroids appear as curved trails in HST images because of the parallax induced by the fast orbital motion of the spacecraft. Taking into account the trajectory of this latter, the parallax effect can be computed to obtain the distance to the asteroids by fitting simulated trajectories to the observed trails. Using distance, we can obtain the absolute magnitude of an object and an estimation of its size assuming an albedo value, along with some boundaries for its orbital parameters.
Results. In this work, we analyse a set of 632 serendipitously imaged asteroids found in the ESA HST archive. Images were captured with the ACS/WFC and WFC3/UVIS instruments. A machine learning algorithm (trained with the results of a citizen science project) was used to detect objects in these images as part of a previous study. Our raw data consist of 1031 asteroid trails from unknown objects, not matching any entries in the Minor Planet Center (MPC) database using their coordinates and imaging time. We also found 670 trails from known objects (objects featuring matching entries in the MPC). After an accuracy assessment and filtering process, our analysed HST asteroid set consists of 454 unknown objects and 178 known objects. We obtain a sample dominated by potential main belt objects featuring absolute magnitudes (H) mostly between 15 and 22 mag. The absolute magnitude cumulative distribution log N(H > H_{0}) ∝ α log(H_{0}) confirms the previously reported slope change for 15 < H < 18, from α ≈ 0.56 to α ≈ 0.26, maintained in our case down to absolute magnitudes of around H ≈ 20, and therefore expanding the previous result by approximately two magnitudes.
Conclusions. HST archival observations can be used as an asteroid survey because the telescope pointings are statistically randomly oriented in the sky and cover long periods of time. They allow us to expand the current best samples of astronomical objects at no extra cost in regard to telescope time.
Key words: methods: data analysis / astronomical databases: miscellaneous / minor planets, asteroids: general
© The Authors 2024
Open Access article, published by EDP Sciences, under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
This article is published in open access under the Subscribe to Open model. Subscribe to A&A to support open access publication.
1 Introduction
The Solar System objects (SSOs) can be studied by observing their movement over time, which requires multiple exposures separated by a few days to produce a sufficiently large orbital arc. If the SSOs have been reported to the Minor Planet Center (MPC), which is the main worldwide repository for the receipt and distribution of positional measurements of SSOs^{1}, precovery in archival data can aid in orbit determination. In this paper, we use a different method to characterise asteroids that relies on a set of observations taken from low Earth orbit in large astronomical archives.
Observatories in low Earth orbit confer a distinct advantage in that their fast motion produces a parallax effect on the asteroid’s trail shape, allowing distance determination with fewer observations. This parallax method was first devised by Evans et al. (1998) and Evans & Stapelfeldt (2002) for use with asteroids identified in Hubble Space Telescope (HST) observations. Figure 1 shows an example of an asteroid trail imaged by HST, where the parallax effect is evident in the curvature of the trail.
In the Hubble Asteroid Hunter project^{2} (Kruk et al. 2022), we identified 1,701 new asteroid trails in 19 yr of HST observations with the Advanced Camera for Surveys (ACS/WFC) and Wide Field Camera 3 (WFC3/UVIS) observations, using a deep learning algorithm on Google Cloud, trained on volunteer classifications from the Hubble Asteroid Hunter citizen science project on the Zooniverse platform. 1031 (61%) of the trails correspond to SSOs that we could not identify in the MPC database. Considering how faint these asteroids are, many of them are likely previously unidentified objects. The parallax fitting method (Evans et al. 1998) is ideal for studying this new sample of SSOs. By fitting the parallax for these trails, we can determine the approximate orbital parameters of these asteroids and derive their physical properties. Once we have determined the distance to the asteroids, we can use their apparent magnitudes to determine their intrinsic brightnesses and sizes, assuming an average albedo value.
Studying faint asteroids is essential in order to better understand their size distribution, which remains poorly understood for small objects. The shape of the magnitude distribution curve, as determined by Gladman et al. (2009), shows a shallower slope for asteroids of smaller sizes down to H = 18 mag. This result is crucial to our understanding of the evolution of the main belt along with its collisional and dynamic depletion models (Bottke et al. 2015).
Hubble Space Telescope observations taken over a long time span can be effectively used as a randomly observed asteroid survey because the telescope’s pointing selections are statistically random in the sky, and many observations are across the ecliptic where the main asteroid belt is located.
This paper is structured as follows. In Sect. 2, we describe the data that were used in this study. In Sect. 3, we outline the parallax fitting method and examine its efficacy through tests conducted on both simulated data and a sample of known asteroids in the MPC database also identified by the Hubble Asteroid Hunter citizen science project volunteers. Section 4 reports the results of the application of this method to previously unidentified asteroids, including the determination of their orbital parameters and size distributions. In Sect. 5, we compare our findings to those of previous studies. Finally, in Sect. 6 we present our conclusions.
Fig. 1 Example of asteroid trail imaged by HST in observation idj554030 from WFC3/UVIS, taken on 5 Feb 2018. The trail spans roughly 10 arcsec in RA and 8 arcsec in Dec. 
2 Data
The trails analysed in this work were obtained from the ESA Hubble Space Telescope Archive for instruments ACS/WFC and WFC3/UVIS. An objectdetection machine learning algorithm (trained with the results of a citizen science project) was used to find asteroid trails (Kruk et al. 2022) and other objects of interest such as gravitational lenses (Garvin et al. 2022) and satellite trails (Kruk et al. 2023) in Hubble images from the archive for both instruments. We used a total of 24731 HST composite images from ACS/WFC and 12 592 from WFC3/UVIS, each one of them split into four quadrant cutouts to improve asteroid trail detection efficiency. HST images from the archive were selected based on the following criteria: no calibration images or ‘darks’; exposure time greater than 100 seconds; FoV greater than 0.044° (avoiding subframes); and no grism spectral images, as their special features could be confused with trails. We also removed any image purposely targeting Solar System objects.
The volunteers of our citizen science project and also the machine learning algorithm were trained to identify cosmic rays to correctly distinguish their streaks from real asteroids. Nevertheless, the whole set of images featuring potential detections was visually checked by the authors during the first part of this project to avoid any artifact or spurious detection misidentified as an asteroid.
Our data consist of 1031 asteroid trails from unknown objects, not matching any entries in the MPC database based on their coordinates and imaging time. We also found 670 trails from known objects (objects featuring matching entries in the MPC database), which proved very useful in testing and validating the method we propose in this work. For every case, a list of right ascension (RA) and declination (Dec) coordinates forming a trail is used along with the exposure starting and ending times.
Checking the identified trails for known objects was conducted using a twostep process. The first step involved using the average position and exposure time of each trail. We fed this data to the SkyBoT service (Berthier et al. 2006, 2016), performing a cone search with a 900 arcsec radius. This preliminary step returned 22 748 possible candidates for our whole set. During the second step, we filtered these candidates for precise matches with the observed trails. For every image, ephemeris for all its possible matches was retrieved from JPL Horizons service (Giorgini et al. 1996) using the exposure time. We then analysed the approximate averaged distance between the observed trail and those obtained from JPL Horizons (as seen from HST), using a 3 arsec threshold for a positive match. The complete trail processing pipeline applied during the first part of this project is explained in full detail in Kruk et al. (2022).
3 Analysis methods
3.1 Parallax method
To obtain the distance to Earth of the unknown objects, we present a method using the parallaxinduced curvature of asteroid trails in HST images, following the work by Evans et al. (1998) and Evans & Stapelfeldt (2002). Asteroid trails imaged by HST present a very particular parallax induced by the spacecraft orbital motion around Earth while tracking a fixed target. Being in a low Earth orbit, HST’s fast pace compared to its typical exposure times enables this unique effect for relatively close objects in our Solar System. HST travels approximately onethird of its orbit during a 30 min exposure (the average time for our dataset); a stationary object situated at 2 AU from Earth (around the main belt) could feature a parallax effect of up to ~8 arcsec.
For every trail, HST orbital motion for the exposure time is obtained from Jet Propulsion Laboratory (JPL) SPICE routines (Acton 1996; Acton et al. 2018) along with Earth position with respect to the Solar System Barycenter (SSB). Our code iterates through different object distance solutions between 0.2 and 6.7 AU, obtaining predicted trails that are then compared to the observed trail, looking for a bestfit solution. An upper distance limit of 6.7 AU was chosen to potentially cover the Jupiter Trojan asteroid population and also based on the noncurvature of trails for objects located further away.
In addition to the parallax induced by the object’s distance and HST–Earth trajectories, we also need to consider the object’s motion around the Sun, which we refer to as the object’s intrinsic rate of motion as observed in our images. The observed trail is the linear combination of the parallax effect modelled by vector P(Δ, t) as a function of both observer distance (Δ) and time (t), and the object’s intrinsic rate of motion in RA and , respectively.
From Evans et al. (1998), every point of the trail can then be described as: (1)
where α_{0} and δ_{0} are the initial coordinates of the object; P_{α} and P_{δ} are the components of vector and are the intrinsic rates of motion of the object, and t is the time of the exposure being considered.
To obtain the observed object’s intrinsic rate of motion, we use a rather straightforward method. The total span of the trail in both RA and Dec coordinates must be equal to the asteroid motion rate plus the ‘deformation’ generated by the parallax effect. Knowing the starting and ending points of the trail and the standard deformation generated by the parallax at any given distance, just one unique value of is possible for each distance iteration.
The selected trail solution is the one featuring the lowest error value χ^{2} as defined in Eq. (2), comparing the predicted trails for 1300 different observer distance values (Δ) and the observed trail. We normalise χ^{2} using a 0.05 arcsec value, which corresponds to the ACS/WFC pixel resolution (Ryon 2022). The resolution achieved by WFC3/UVIS is 0.04 arcsec/pixel (Dressel 2022), but we decided to use the worst ACS/WFC case for our analysis to present a unified method for both instruments.
The χ^{2} expression is as follows: (2)
where n is the total number of existing points for the trail, x_{calc} are the sky position values from the simulated trail being fitted, and x_{obs} are the observed trail sky position values.
During the first stage of this project (Kruk et al. 2022), the trails were discretised by our extraction algorithm, finding the maxima along each pixel column of the image cutout file. Taking the resolution of HST into account, this is approximately one point every 0.04 arcsec for WCS3/UVIS and 0.05 arcsec for ACS/WFC. The average number of points for a given trail is approximately 300.
We cannot distinguish between distance solutions with a mean fittingresidual difference of lower than the instrument resolution (considered to be one pixel, or 0.05"). Therefore, the uncertainty U_{r} of our method is set by the two distance values corresponding to χ^{2} + 1 (as shown in Fig. 2).
Once we have a distance solution for a trail as seen from HST, we obtain the distance to Earth and SSB using SPICE routines. Given that all the analysed asteroids were serendipitously imaged during a single HST observation, we cannot apply any standard orbit approximation method. The object’s velocity component in the asteroid–Earth direction will always be missing and we can only obtain the limits of its possible orbital parameters for semimajor axis (a), eccentricity (e), and inclination (i).
Nevertheless, we can calculate some possible combinations of these parameters supposing a bounded orbit. The object’s estimated velocity and escape velocity are calculated from the observation (the object’s distance from the Sun can be obtained using SPICE combined with the asteroid–Earth position obtained from the parallax method) and 20 000 different velocity values for the missing asteroid–Earth axis velocity component are iterated using the escape velocity as a limit. We obtain a set of 20 000 possible orbital parameter combinations for each object, which can be plotted as a function of semimajor axis as seen in Fig. 3. We can also obtain minimum semimajor axis, eccentricity, and inclination values for the object. Objects featuring estimated velocities greater than their escape velocity are not considered in our analysis and are expected to be cases where the method will not be able to obtain a good result (see Sect. 3.2 for further details on the filtering process).
As an example of the parallax method, we show the results for object (511908) 2015 HU67 identified in HST observation idq27j020. This object is a relatively small asteroid from the middle main belt and features an apparent magnitude of 21.5 in the V band. This trail was processed using our full code pipeline, from trail coordinates to orbital parameter results. The trail fit and χ^{2} minimum are shown in Figs. 4 and 2 respectively. The graph featuring all possible inclination and eccentricity values as a function of semimajor axis is shown in Fig. 3.
As a benchmark, we use the JPL Horizons ephemeris service (Giorgini et al. 1996) to validate the parallaxobtained Earth distance and orbital parameters for this known object. We show this comparison in Table 1. Our parallax results show good agreement with the values obtained from the database.
In some cases, the trail does not present enough curvature to converge to a welldefined distance solution or to obtain a reduced uncertainty interval allowing us to calculate a meaningful absolute magnitude value. If the trail is not sufficiently curved, our χ^{2} function will be too broad, implying a large uncertainty U_{r} in the distance solutions. Trails lacking significant curvature correspond to objects seen close to HST’s orbital plane; that is, the induced parallax is close to the direction of the asteroid’s intrinsic motion. The most curved trails are those where the asteroid is seen at a large angle from HST’s orbital plane, such that the induced parallax is becoming perpendicular to the asteroid’s intrinsic motion.
An example of a trail not presenting a clear solution is object (58250) 1993 QU1 featured in the HST image with observation id j93y94010. Despite its high signaltonoise ratio and good fit (Fig. 5), the trail is too straight to yield an accurate distance solution and our method does not converge (Fig. 6). We can also see in Fig. 6 that the trail fitting yields a large number of possible solutions inside the χ^{2} + 1 threshold, as expected for straight trails. The obtained Earth distance value is 6.695 AU (the actual JPL Horizons Earth distance equals 2.382 AU). These cases were filtered out from our sample as discussed in Sect. 3.2.
Fig. 2 Asteroid (511908) 2015 HU67. χ^{2} values showing a minimum at the bestfit distance solution from Earth. The vertical line represents the actual distance for this object obtained from the JPL Horizons ephemeris. 
Fig. 3 Possible values of eccentricity (e, red) and inclination (i, blue) as a function of semimajor axis (a) for asteroid (511908) 2015 HU67 calculated using the parallax method. For every semimajor axis value, the corresponding possible eccentricity and inclination solutions are shown graphically. 
Fig. 4 Asteroid (511908) 2015 HU67. Observed trail from HST (blue) vs bestfit distance solution using the parallax method (red). χ^{2} value for this fitting is 0.13 (see Eq. (2)). 
3.2 Assessing the accuracy of the method
To obtain a preliminary measure of the accuracy of our method, we used 21 283 available knownasteroid ephemerides from the JPL Horizons service (Giorgini et al. 1996) as a proof of concept (POC) set. For every object, we input only the point coordinates (RA and Dec) and time, as a real trail observed by HST, and applied our method to calculate a distance solution. The obtained distance from Earth was then compared with the distance value provided by the JPL Horizons service. The global relative distance error (distance error divided by real object distance) results for this POC set are centred around zero and can be seen in Fig. 7. This confirms that our parallax method produces distance results that are compatible with JPL Horizons database values. No filter or goodnessoffit criteria were applied to the results of this analysis or to the input data, which explains the long tails at both ends of the error distribution. This result validates the accuracy of our Python implementation of the method from Evans et al. (1998), which was originally implemented in Fortran.
The next step is to evaluate our method in more detail and precisely define its accuracy for distance predictions. To do so, we used the 670 trails from known objects identified in the first part of this project (Kruk et al. 2022). Of these 670 asteroids, we selected the 452 featuring a permanent MPC designation to reduce potential errors in their reference orbital parameters. We ran the code using this group, which we name the knownobjectsvalidation set, and after examination we decided to apply the following criteria to filter spurious distance results that could contaminate our final population distribution. We excluded the following objects:
Objects presenting observed estimated velocities above their escape velocity. No extrasolar objects were included in the known objects set and the probability of finding one among our results was deemed extremely unlikely.
Objects above a 6.5 AU distance result or not featuring a clear χ^{2} absolute minimum (U_{r} uncertainty equal to 0), for which the method was not converging.
Objects featuring an upper or lower distance uncertainty U_{r} value equal to or greater than 35% of their actual distance. We chose this value by analysing the relative error distribution for the whole raw set (distance error divided by actual distance), which represents one standard deviation (σ) for this distribution. As mentioned above, these last two criteria are meant to filter out trails that are too straight to yield a reliable parallax solution.
Trails as nonmonotonic functions (RA or Dec acting as X–Y; see an example in Fig. 17 of Kruk et al. 2022). The current version of our interpolation code is not adapted to handling this kind of trail. We found less than five of these cases for the considered analysis set, and so the impact versus effort ratio was deemed too low to add this feature to our code. This option should be considered for further work if larger data sets should be analysed.
Trails presenting background or close sources modifying the extracted coordinates of the trail (see Kruk et al. (2022) for the full trail coordinate extraction details) or other imaging artifacts such as trails slipping out of the cutouts. As mentioned above, our code uses both the starting and ending points as boundary conditions to obtain the object’s rates of motion along with the image exposure time; it is especially sensitive to errors originating from the incorrect definition of the two ends of a trail.
Lowsignaltonoiseratio trails, making coordinates difficult to obtain and trail limits difficult to measure.
We applied these criteria, keeping a total of 178 cases as our final knownobjectsvalidation set. We consider these cases as a representative set to evaluate the accuracy of our parallaxbased method, having removed any external influence from HST imaging or from our processing pipeline. This filtering process features a very conservative approach for our selection criteria, which led us to keeping about 40% of the initial processed asteroids, optimising for detection quality over sample size.
The distancetoEarth relative error distribution for this knownobjectsvalidation set is presented in Fig. 8 (distance error divided by actual distance). This error was calculated by comparing JPL Horizons Earth distance with the computed distance using the parallax method. We obtain a mean relative error value centred at −1.1% and a standard deviation of σ = 10.5%. This error distribution is presented to the reader as a global performance indicator of our novel parallax method. It will not be applied as an average error value to the results of this work. The specific distance uncertainty is calculated for each object as explained in Sect. 3.1.
To conduct a final verification of our method, we created a synthetic validation set. Starting from the known objects validation set, we obtained ‘synthetic’ trails from JPL Horizon ephemerides for the same objects and exposure time as seen from HST. We fed these trails to our code as real HST trails to obtain a distance solution. In this case, the obtained mean relative error value is −1.5% and σ = 7.5%. The comparison of both error distributions is shown in Fig. 8. Both error distributions are closely overlapping and centred around zero, and therefore we consider that any major influence from HST imaging or from our image processing pipeline was correctly removed from our analysis.
As expected, we find a slight correlation between the actual distance from the object to Earth, or its apparent magnitude, and the distance error using the parallax method (see Fig. 9). In general, distant (and probably faint) objects yield trails that are more difficult to analyse because of the lower signaltonoise ratios and reduced parallaxinduced curvature in the trails.
Another important aspect to assess when determining the accuracy of our method is to make sure the filtering process does not present any bias, removing fainter or smaller objects and potentially reducing the number of outer main belt asteroids in our observations. Figure 10 shows the initial sample of known objects (452 asteroids) as a function of distance to Earth and absolute magnitude (both parameters obtained from JPL Horizons, considered as ground truth). Two different populations are displayed in this figure: our final known object validation set in red and the objects that were excluded during the filtering process in grey. We see a fairly constant and random removal process following both axes, without any systematic or selection effects. As intended, our filtering process is mainly driven by the characteristics of the trails, which are random, and not the physical properties of the asteroids themselves.
Orbital parameters of (511908) 2015 HU67 from parallax method vs JPL Horizons ephemeris data.
Fig. 5 Object (58250) 1993 QU1. Example of a straight trail with no parallax solution obtained. 
Fig. 6 Object (58250) 1993 QU1. The method does not converge because of the lack of curvature from the trail. The χ^{2} minimum is obtained at the upper distance limit set in our code (≈6.7 AU). The red line represents the actual Earth distance for this object from the JPL Horizons ephemeris. 
Fig. 7 Relative distance error for JPL Horizons ephemeris distance vs parallaxobtained distance for the proofofconcept set (21 283 known objects) using coordinates from JPL Horizons as simulated trails (RA, Dec, and time). xaxis is distance error divided by actual distance. No filtering or goodness of fit criteria were applied to this proofofconcept set, which explains the long tails at both ends of the distribution. 
Fig. 8 Distance to Earth relative error distribution (AU) for our knownobjectsvalidation set using HST trails (red) compared to the relative error distribution for the synthetic trails validation set created using JPL Horizons ephemerides for these same objects and time (green). xaxis is distance error divided by actual distance. 
Fig. 9 Scatter plot comparing the relative error on the distance to Earth obtained from the parallax method vs the object’s actual Earth distance and apparent magnitude. 
Fig. 10 Absolute magnitude vs distance to Earth plot, comparing our final known object validation set (red) with those removed during the filtering process (grey). We can see a fairly random removal process around both axes, without any effect depending on distance or magnitude. 
3.3 Magnitude analysis
The apparent magnitudes of the trails were obtained in the previous part of this project (Kruk et al. 2022). We start from these values to obtain the absolute magnitudes of the objects using the distance solutions obtained with the parallax method.
Most of the unknown objects were serendipitously imaged by HST in one single band, and no further information is available about their colour or asteroid type. Although some of our detections come from the same HST observation set, we cannot link the objects from one image to another, as we have no reliable information about their orbital parameters. Before calculating their absolute magnitude values, we transform the apparent magnitude of each object in any given HST filter to a Johnson V filter in order to apply a unified method to compare all of them. We assume the asteroid spectra to be mostly the reflected spectrum from the Sun, except for their particular absorption features. Therefore, the filter transformation value applied is the same used for the spectrum of the Sun. Using the known objects set as proof once again, we obtain a 0.18 mag average error for this approximation.
Using the parallaxcalculated distance and phase angle for each object, we obtain its absolute magnitude using the equation from Bowell et al. (1989): (3)
where H is the object’s absolute magnitude, V is its apparent magnitude in V, G is the photometric slope parameter, Δ is the object’s distance to Earth, and r is the object’s distance to the Sun (both in AU). We used a G = 0.15 constant value for our analysis. ϕ_{1} and ϕ_{2} parameters are functions of the phase angle α of the object.
3.4 Sizeestimation method
We use the absolute magnitude values (H) calculated in the previous section to estimate the size of the analysed objects, assuming them to be spherical and to feature a uniform surface and therefore constant albedo. To do so, we applied the equation from Harris & Harris (1997). Even if albedo is known to change from Sdominated to Cdominated taxonomy classes from the inner to the outer belt (DeMeo & Carry 2014), an average 0.15 albedo value was chosen for estimating the sizes of the objects. This is an intermediate value representative of the typical range (0.03 – 0.4) of asteroid compositional types (e.g. Mainzer et al. 2011; Popescu et al. 2018; Mahlke et al. 2022). To avoid any error induced by this assumption, we used absolute magnitude (H) to create the population size distribution graph discussed in Sect. 5 (Fig. 16). As an example, a H =19.5 object (around the median value for our unknown objects set) will feature a diameter of ≈700 m if supposing a 0.05 albedo and ≈330m if we suppose a 0.25 albedo.
4 Results
We applied the methods described to the set of 1031 unknown Solar System Objects from Kruk et al. (2022). The strict filtering criteria we present in Sect. 3.2 yield a total of 454 objects. This filtered set is ≈44% of the original number, matching the yield obtained for the knownobjectsvalidation set after cleansing. As stated in the previous sections, we prioritised high precision over a large population of objects. The majority of the input objects are purposely excluded during these filtering processes to maximise accuracy.
For the histograms of this section, data for known objects are extracted directly from the JPL Horizons database, which is considered as a baseline and therefore do not feature any uncertainty bars in any of the presented graphs. The uncertainty bars for the histograms of unknown objects were obtained using bootstrapping (Efron 1982), randomly resampling the existing population 1000 times and calculating the standard deviation for each bin.
Fig. 11 Distance to SSB for known (red) and unknown (blue) objects. Unknown object values were obtained from the parallax method, known object values from the JPL Horizons database. 
Fig. 12 Distance to Earth for known (red) and unknown (blue) objects. Unknown object values were obtained from the parallax method, known object values from the JPL Horizons database. 
4.1 Object distance
We obtained the distance to Earth and the distance to the Solar System Barycenter (SSB) for the 454 unknown objects. Results can be seen in Figs. 11 and 12. Both distance distributions feature a close overlap for known and unknown objects, centred around the main belt in both cases. There is a small bump for unknown objects at less than 1 AU from Earth, representing close objects spotted by HST. Given the limited nature of our calculated orbital parameters, we cannot obtain a real perihelion value and therefore cannot assess whether or not they could be nearEarth objects (NEOs), as we argue in Sect. 5.
Fig. 13 Absolute magnitude (H) for known (red) and unknown objects (blue) from JPL Horizons database and calculated using the distance obtained with the parallax method, respectively. The upper axis shows the approximate object size using a 0.15 albedo value. 
Fig. 14 Size distribution for unknown objects (using a 0.15 albedo value). 
4.2 Magnitude and size distributions
Despite having obtained a very close distance distribution for known and unknown objects (either to Earth or SSB), if we take a look at the absolute magnitude results displayed in Fig. 13 we can see that the unknown objects population is dominated by fainter objects centred around H = 19–20 magnitude. Both results indicate that we are mostly imaging small (<1 km) main belt objects not easily accessible using groundbased asteroid surveys.
The size distribution for the unknown objects (assuming an average albedo of 0.15 for all objects) is presented in Fig. 14. A significant portion of them are below 1 km in size.
5 Discussion
5.1 Size and absolute magnitude population distributions
A scatter plot presenting size as a function of distance from Earth is shown in Fig. 15. As we can see, the lower size limit of the detected unknown objects increases with distance from Earth, as expected for a fluxlimited survey. The unique capabilities of HST allow us to detect unknown objects up to distances of around 3 AU from Earth (with the main population centred around 2 AU) and below a 1 km estimated diameter. We find that our unknown objects population is mainly composed of small main belt asteroids (< 1 km size).
Assuming that the size distribution is the same everywhere in the asteroid belt and despite the fact that our total dataset (unknown + known objects) does not represent a dedicated or observationally unbiased asteroid survey, we can use our HST set as a random asteroid population sample given the large time span covered by our HST archival analysis (19 yr) and the statistically random HST pointing locations (Kruk et al. 2022). As shown in Figs. 11 and 15, most of our objects are found in the main belt. If we take a closer look at Fig. 15, we can clearly see the cutoff generated by the lower part of the scatter plot, which represents the observational limit of HST for a given distance. For a distance of 2.5 AU from Earth, which is roughly the end of the outer main belt, our size detection limit is 0.5 km, equivalent to H ≈ 19.5 using a 0.15 albedo. We assume our dataset to be complete up to this magnitude value regarding the main belt asteroid population.
The new population model we present is based on absolute magnitude (H) and is depicted in Fig. 16. We present the entire HST set (unknown and known asteroids), featuring 632 objects. We identify two different approximate values for the slope of our cumulative distribution function: log N(H > H_{0}) ∝ (0.56 ± 0.12) log(H_{0}) for H values between 12 and 15, and log N(H > H_{0}) ∝ (0.16 ± 0.06) log(H_{0}) for H values between 15 and 20. Above H ≈ 19–20, our dataset reaches its detection limit.
The values we obtained can be compared with the results from the SubKilometer Asteroid Diameter Survey (SKADS; Gladman et al. 2009). This dedicated and debiased survey for main belt small objects yields a slope of α = 0.3 ± 0.01 for H values between 15 and 18, in agreement with the value obtained in our analysis. The slight difference could be explained by our HST dataset bias, its incompleteness, or the different filters used in the two analyses. The steeper slope change for the H < 15 regime is also described in this work, with a reported value of α = 0.5, which is very close to the value found for the HST asteroids set (α = 0.56). Our work shows that the α = 0.3 slope value between 15 and 18 H remains consistent up to H ≈ 19–20 mag. For reference, H ≈ 15 corresponds to a size of 3 km for an albedo of 0.15.
Other previous works found results compatible with our survey. Using observations from WISE/NEOWISE in the thermal infrared, Masiero et al. (2011) found a size distribution slope for main belt objects in good agreement with that found by Gladman et al. (2009), which is considered as the benchmark for our work. Still in the infrared and using data from Spitzer, Ryan et al. (2015) found a slightly shallower population distribution slope with a ‘kink’ or inflection point around a diameter of 8 km, corresponding roughly to H ≈ 13 using a 0.15 albedo. In our case, we do not have enough observations around this magnitude value to confirm this feature. Heinze et al. (2019) performed main belt observations from the Blanco Telescope at Cerro Tololo InterAmerican Observatory, reporting a nonconstant population powerlaw slope. In this latter work, the apparent Rband population distribution shows a transition around R = 19–20, with a shallower slope for fainter magnitudes, which is in good agreement with our benchmark (Gladman et al. 2009).This shallower slope is maintained up to two fainter magnitudes than Gladman et al. (2009), which reinforces the results from our work showing a constant slope maintained up to H ≈ 19–20 mag. Maeda et al. (2021) found a similar shape for both ‘Clike’ and ‘Slike’ size population distributions featuring main belt asteroids between 0.4 and 5 km in size (roughly H ≈ 10 and H ≈ 14 using a 0.15 albedo). This similarity between the two types of asteroids supports the average nature of our population distribution, not being able to reliably distinguish asteroid classes in our unknown objects HST set. These latter authors also report an inflection point in the slope of the absolute magnitude population distribution that is located at around 1.7 mag fainter than the inflection point found in the present study. Slope values are α = 0. 55 and α = 0.13; that is, slightly shallower for the fainter end.
Fig. 15 Size vs Earth distance for the knownobjectsvalidation set (grey) and the unknown objects set (green, using a 0.15 albedo value to estimate size). For clarity, the average size and distance uncertainties are displayed next to the main population area as a black cross. 
Fig. 16 Absolute magnitude cumulative size distribution for our HST dataset. The black lines and numerical values represent the calculated approximate equivalent slopes α for log N(H > H_{0}) ∝ α log(H_{0}). 
5.2 Potential comets
An interesting possibility suggested by our findings is the existence of objects with orbits consistent to those of comets. To help us find possible candidates, we calculated the Tisserand’s parameter (Tisserand 1889) with respect to Jupiter (T_{J}) as seen in Eq. (4), using the object’s minimum inclination, eccentricity, and semimajor axis values. (4)
where a_{J} is Jupiter’s semimajor axis, a is the object’s minimum semimajor axis, e is the object’s minimum eccentricity, and i is the object’s minimum inclination.
We find 45 objects of interest featuring T_{J} < 3. This represents ≈10% of the total 454 analysed objects. Two examples are displayed in Table 2 and Fig. 17. All the 45 candidates present apparent magnitudes of between 22.2 and 25.1 mag in the given filters used by HST for each observation, making them realistic candidate new objects featuring comet orbits.
We note that T_{J} parameter provides a first indication of a cometary orbit. The Jupiter family comets are characterised by 2 ≤ T_{J} ≥ 3, while for the longperiod comets, T_{J} < 2. An enhanced criterion to identify cometary orbits is provided by Tancredi (2014); it takes into account the orbital parameters and the minimum orbital intersection distance, and discards the objects in mean motion resonances.
Two examples of potential comets found in our dataset and their calculated approximate orbital parameters.
Fig. 17 Two examples of potential comets found in our dataset. Left images correspond to j8n218010 and right images to j90i02010. The upper row represents the original trail cutouts and the lower row the identified trail centerline and trail extraction lateral limits (Kruk et al. 2022). 
5.3 Potential nearEarth objects
NearEarth objects are also a relevant class of objects, and merit a more detailed analysis to find potential candidates in our dataset. We define NEOs as objects featuring a perihelion of < l.3 AU (Bottke et al. 2002). Considering that we only have lower limits for eccentricity and semimajor axis, the calculated perihelion and aphelion values are also approximations and should only be used as an indicator of potential objects belonging to this family (shown in the lowerleft corner of Fig. 15). We find 74 potential NEO candidates. The apparent magnitudes in this case span from 20.4 to 25.1 in the given filter used by HST to take the image.
Among our NEO set, we find ~34% Amors, ~41% Apollos, and ~12% Atens. These values are close to but not coincident with those from Granvik et al. (2018), who present a population distribution for NEOs in the range of 17 < H < 25, featuring a relative share of 39.4% Amors, 54.4% Apollos, and 3.5% Atens. In our case, our analysis is limited by the approximate nature of the orbital parameters and HST observational bias for NEOs. Our dataset does not represent a complete population model for this kind of object (which is beyond the scope of this work) but a random sample of observations from HST archival data.
5.4 Further work
Absolute magnitude distributions can be used to estimate the size distribution after a careful estimation of the albedo values, such as in the study conducted by Ivezic et al. (2001) using commissioning data from SDSS and including colour information.
We are also currently analysing the light curves from all the detected asteroids (known and unknown objects). This analysis should lead to a better characterisation of the properties of small objects, which are important in order to fully understand the collisional and dynamical main belt evolution (Bottke et al. 2015). The raw light curves from this project require specialised processing in order to take into account the parallax effect and correctly project time versus brightness variation.
Beyond the discovery and physical characterisation of small asteroids in the HST archive, all three of the papers from the Hubble Asteroid Hunter project illustrate how the training of deep learning neural networks to identify objects in images – or other features in science data archives – using labels produced by large citizen science projects could be used to automatically tag objects or phenomena in scientific datasets and assist data analysts in the subsequent identification of such types of objects or phenomena, greatly enlarging the scientific potential of such datasets. Encouraged by these prolific citizen science projects, ESA has launched another two similar citizen science projects (Rosetta Zoo^{3} and GaiaVari^{4}) where volunteers are asked to classify moving features on the surface of the comet 67P as observed by the Rosetta mission or classify different types of variable stars in the Gaia catalogue. The results of these projects will also be used to train deep neural networks that will allow the identification and classification of such features in future datasets (from similar or the same missions) reliably and automatically. In the particular case of the asteroids, serendipitous observations made by JWST will probably become a substantial source of new asteroid detections, as demonstrated recently by Müller et al. (2023). The ESA Euclid mission will also enable many new object discoveries in the years to come thanks to its large field of view combined with its spatial resolution (Carry 2018; Pöntinen et al. 2020). Unfortunately, in both JWST and Euclid cases, their orbits around L2 are not suited to induce parallax in the same way as HST.
6 Conclusions
We analysed the physical properties of a set of 632 serendipitously imaged asteroids from the HST archive. These asteroids were found using a combination of a citizen science project and a machine learning algorithm as part of a previous study (Kruk et al. 2022). To obtain asteroid distances from just one observation, we applied a proven method based on the parallax effect generated by HST’s low Earth orbit (Evans et al. 1998). Approximately 40% of the raw objects found in the archive yield meaningful distance results using this method. We used these distance results to compute the absolute magnitude and approximate size of the asteroids.
Our results indicate a set dominated by main belt objects featuring sizes of below 10 km and absolute magnitudes (H) mostly between 15 and 22. The unknown objects set (454 objects) is bounded below 1 km size and between 17 and 22 for H. The absolute magnitude cumulative distribution confirms the previously reported slope change at H = 15 for main belt objects (Gladman et al. 2009), from α ≈ 0.56 to α ≈ 0.16, with this latter slope being maintained up to absolute magnitudes of around H ≈ l0 in our case.
One of the advantages of applying machine learning to find Solar System objects in complete astronomical archives is the large number of potential results obtained. This allows us to apply purposely strict filtering conditions to improve accuracy and still keep a large enough sample to obtain statistically meaningful results.
Adding automatic SSO detection pipelines to deep but FoVlimited telescopes could generate important contributions to the population of small SSOs and to our understanding of the early stages and evolution of the Solar System. In our case, the analysis of archival data spanning 19 yr limits the possibilities of any object follow up for interesting cases such as potential NEOs or comets. This drawback can be avoided using realtime processing pipelines.
We have contacted the IAU Minor Planet Center to submit the full identifications set (known and unknown objects) for the whole Hubble Asteroid Hunter Project (Kruk et al. 2022).
Acknowledgements
We thank the anonymous referee for the helpful comments and suggestions that led to the improvement of this manuscript. We acknowledge support from ESA through the Science Faculty – Funding reference ESASCISCLE154. This work was possible thanks to the Google Cloud Research Credits Program. The work of M.P. was supported by a grant of the Romanian National Authority for Scientific Research  UEFISCDI, project number PNIIIP22.1PED20213625. We want to specially thank the STScI helpdesk for their quick responses and useful advice. This publication uses data generated via the https://www.zooniverse.org/platform, development of which is funded by generous support, including a Global Impact Award from Google, and by a grant from the Alfred P. Sloan Foundation. This work has made extensive use of data from HST mission, hosted by the European Space Agency at the eHST archive, at ESAC (https://www.cosmos.esa.int/hst), thanks to a partnership with the Space Telescope Science Institute, in Baltimore, USA (https://www.stsci.edu/) and with the Canadian Astronomical Data Centre, in Victoria, Canada (https://www.cadcccda.hiaiha.nrccnrc.gc.ca/).
References
 Acton, C. H. 1996, Planet. Space Sci., 44, 65 [Google Scholar]
 Acton, C., Bachman, N., Semenov, B., & Wright, E. 2018, Planet. Space Sci., 150, 9 [Google Scholar]
 Berthier, J., Vachier, F., Thuillot, W., et al. 2006, ASP Conf. Ser., 351, 367 [NASA ADS] [Google Scholar]
 Berthier, J., Carry, B., Vachier, F., Eggl, S., & Santerne, A. 2016, MNRAS, 458, 3394 [NASA ADS] [CrossRef] [Google Scholar]
 Bottke, W. F., Brož, M., O'Brien, D. P., et al. 2015, in Asteroids IV, 701 [Google Scholar]
 Bottke, W. F., Morbidelli, A., Jedicke, R., et al. 2002, Icarus, 156, 399 [NASA ADS] [CrossRef] [Google Scholar]
 Bowell, E., Hapke, B., Domingue, D., et al. 1989, in Asteroids II, 524 [Google Scholar]
 Carry, B. 2018, A&A, 609, A113 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
 DeMeo, F. E., & Carry, B. 2014, Nature, 505, 629 [NASA ADS] [CrossRef] [Google Scholar]
 Dressel, L. 2022, in WFC3 Instrument Handbook for Cycle 30 v. 14 (Baltimore: STScI) [Google Scholar]
 Efron, B. 1982, CBMSNSF Regional Conference Series in Applied Mathematics (Philadelphia: Society for Industrial and Applied Mathematics (SIAM)) [Google Scholar]
 Evans, R. W., & Stapelfeldt, K. R. 2002, in ESA SP, 500, 509 [NASA ADS] [Google Scholar]
 Evans, R. W., Stapelfeldt, K. R., Peters, D. P., et al. 1998, Icarus, 131, 261 [NASA ADS] [CrossRef] [Google Scholar]
 Garvin, E. O., Kruk, S., Cornen, C., et al. 2022, A&A, 667, A141 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
 Giorgini, J. D., Yeomans, D. K., Chamberlin, A. B., et al. 1996, in AAS/Division for Planetary Sciences Meeting Abstracts, #28, 25.04 [Google Scholar]
 Gladman, B. J., Davis, D. R., Neese, C., et al. 2009, Icarus, 202, 104 [NASA ADS] [CrossRef] [Google Scholar]
 Granvik, M., Morbidelli, A., Jedicke, R., et al. 2018, Icarus, 312, 181 [CrossRef] [Google Scholar]
 Harris, A. W., & Harris, A. W. 1997, Icarus, 126, 450 [Google Scholar]
 Heinze, A. N., Trollo, J., & Metchev, S. 2019, AJ, 158, 232 [NASA ADS] [Google Scholar]
 Ivezic, Ž., Tabachnik, S., Rafikov, R., et al. 2001, AJ, 122, 2749 [NASA ADS] [CrossRef] [Google Scholar]
 Kruk, S., García Martín, P., Popescu, M., et al. 2022, A&A, 661, A85 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
 Kruk, S., GarcíaMartín, P., Popescu, M., et al. 2023, Nat. Astron., 7, 262 [CrossRef] [Google Scholar]
 Maeda, N., Terai, T., Ohtsuki, K., et al. 2021, AJ, 162, 280 [NASA ADS] [CrossRef] [Google Scholar]
 Mahlke, M., Carry, B., & Mattei, P. A. 2022, A&A, 665, A26 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
 Mainzer, A., Grav, T., Masiero, J., et al. 2011, ApJ, 741, 90 [NASA ADS] [CrossRef] [Google Scholar]
 Masiero, J. R., Mainzer, A. K., Grav, T., et al. 2011, ApJ, 741, 68 [Google Scholar]
 Müller, T. G., Micheli, M., SantanaRos, T., et al. 2023, A&A, 670, A53 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
 Pöntinen, M., Granvik, M., Nucita, A. A., et al. 2020, A&A, 644, A35 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
 Popescu, M., Licandro, J., Carvano, J. M., et al. 2018, A&A, 617, A12 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
 Ryan, E. L., Mizuno, D. R., Shenoy, S. S., et al. 2015, A&A, 578, A42 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
 Ryon, J. E. 2022, in ACS Instrument Handbook for Cycle 30 v. 21.0 (Baltimore: STScI) [Google Scholar]
 Tancredi, G. 2014, Icarus, 234, 66 [NASA ADS] [CrossRef] [Google Scholar]
 Tisserand, F. 1889, Bull. Astron. Ser. I, 6, 241 [Google Scholar]
All Tables
Orbital parameters of (511908) 2015 HU67 from parallax method vs JPL Horizons ephemeris data.
Two examples of potential comets found in our dataset and their calculated approximate orbital parameters.
All Figures
Fig. 1 Example of asteroid trail imaged by HST in observation idj554030 from WFC3/UVIS, taken on 5 Feb 2018. The trail spans roughly 10 arcsec in RA and 8 arcsec in Dec. 

In the text 
Fig. 2 Asteroid (511908) 2015 HU67. χ^{2} values showing a minimum at the bestfit distance solution from Earth. The vertical line represents the actual distance for this object obtained from the JPL Horizons ephemeris. 

In the text 
Fig. 3 Possible values of eccentricity (e, red) and inclination (i, blue) as a function of semimajor axis (a) for asteroid (511908) 2015 HU67 calculated using the parallax method. For every semimajor axis value, the corresponding possible eccentricity and inclination solutions are shown graphically. 

In the text 
Fig. 4 Asteroid (511908) 2015 HU67. Observed trail from HST (blue) vs bestfit distance solution using the parallax method (red). χ^{2} value for this fitting is 0.13 (see Eq. (2)). 

In the text 
Fig. 5 Object (58250) 1993 QU1. Example of a straight trail with no parallax solution obtained. 

In the text 
Fig. 6 Object (58250) 1993 QU1. The method does not converge because of the lack of curvature from the trail. The χ^{2} minimum is obtained at the upper distance limit set in our code (≈6.7 AU). The red line represents the actual Earth distance for this object from the JPL Horizons ephemeris. 

In the text 
Fig. 7 Relative distance error for JPL Horizons ephemeris distance vs parallaxobtained distance for the proofofconcept set (21 283 known objects) using coordinates from JPL Horizons as simulated trails (RA, Dec, and time). xaxis is distance error divided by actual distance. No filtering or goodness of fit criteria were applied to this proofofconcept set, which explains the long tails at both ends of the distribution. 

In the text 
Fig. 8 Distance to Earth relative error distribution (AU) for our knownobjectsvalidation set using HST trails (red) compared to the relative error distribution for the synthetic trails validation set created using JPL Horizons ephemerides for these same objects and time (green). xaxis is distance error divided by actual distance. 

In the text 
Fig. 9 Scatter plot comparing the relative error on the distance to Earth obtained from the parallax method vs the object’s actual Earth distance and apparent magnitude. 

In the text 
Fig. 10 Absolute magnitude vs distance to Earth plot, comparing our final known object validation set (red) with those removed during the filtering process (grey). We can see a fairly random removal process around both axes, without any effect depending on distance or magnitude. 

In the text 
Fig. 11 Distance to SSB for known (red) and unknown (blue) objects. Unknown object values were obtained from the parallax method, known object values from the JPL Horizons database. 

In the text 
Fig. 12 Distance to Earth for known (red) and unknown (blue) objects. Unknown object values were obtained from the parallax method, known object values from the JPL Horizons database. 

In the text 
Fig. 13 Absolute magnitude (H) for known (red) and unknown objects (blue) from JPL Horizons database and calculated using the distance obtained with the parallax method, respectively. The upper axis shows the approximate object size using a 0.15 albedo value. 

In the text 
Fig. 14 Size distribution for unknown objects (using a 0.15 albedo value). 

In the text 
Fig. 15 Size vs Earth distance for the knownobjectsvalidation set (grey) and the unknown objects set (green, using a 0.15 albedo value to estimate size). For clarity, the average size and distance uncertainties are displayed next to the main population area as a black cross. 

In the text 
Fig. 16 Absolute magnitude cumulative size distribution for our HST dataset. The black lines and numerical values represent the calculated approximate equivalent slopes α for log N(H > H_{0}) ∝ α log(H_{0}). 

In the text 
Fig. 17 Two examples of potential comets found in our dataset. Left images correspond to j8n218010 and right images to j90i02010. The upper row represents the original trail cutouts and the lower row the identified trail centerline and trail extraction lateral limits (Kruk et al. 2022). 

In the text 
Current usage metrics show cumulative count of Article Views (fulltext article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 4896 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.