Up-to-date taxonomies and spectra are the most important tools to constrain the composition and surface mineralogy of asteroids with various methods. Those typically rely on various modeling techniques and underlying comparison with laboratory spectral measurements of meteorite samples and/or samples returned from space missions (Reddy et al., 2015). So far, this allowed for the identification of several meteorite parent bodies, for example, the howardite–eucrite–diogenite (HED) meteorites and asteroid (4) Vesta (or more broadly V-types) (McCord et al., 1970), LL chondrites, the S-type asteroids (Itokawa or Eros) (Wetherill et al., 1988; Nakamura et al., 2011) or the F-type asteroids, and the ureilite meteorites (thanks to the observed fall of the 2008 TC3 asteroid, which impacted the Earth in 2008 (Jenniskens et al., 2010). However, there are still multiple challenges linking asteroids types with meteorites and estimating their geochemical composition from remote observations. For a full review, we refer the reader to Reddy et al. (2015).

Asteroid taxonomic schemes are tied up to the type and wavelength coverage of data used for their classification as well as the number of objects studied. Early classifications were based on color, albedo, and spectral shape (Chapman et al., 1975). Those divided asteroids into three basic categories: C-carbonaceous, S-siliceous, and the U-unknown class, that is, objects that did not fit the two previous categories. Later taxonomies included an extended number of objects and features used in the classification. Several taxonomies became popular over the years; those include the Tholen taxonomy based on the Eight-Color Asteroid Survey (ECAS) survey and albedos (Tholen, 1989), the S3OS2 classification which used visible spectra obtained at the 1.52-m telescope at ESO (La Silla), or the Buss taxonomy based on the Small Main-Belt Asteroid Spectroscopic Survey (SMASS) observations in visible wavelengths (Bus and Binzel, 2002). The current Bus–DeMeo asteroid taxonomy has used the principal component analysis (PCA) of 371 asteroid spectra obtained in the visible (VIS) and near-infrared (NIR) ranges, that is, from 0.45 to 2.45 μm (DeMeo et al., 2009). The system presented 24 classes and provided an extension to an earlier taxonomy based on VIS spectra only (Bus, 1999). An additional Xn class was later added by Binzel et al. (2019). That system helped explain the compositional distribution of the asteroid main-belt and the delivery efficiency of various taxonomic types to the near-Earth asteroid population (DeMeo and Carry, 2014; Carry et al., 2016; Barucci et al., 2017; Binzel et al., 2019; Devogèle et al., 2019). Laboratory work and asteroid observations later identified space weathering processes that cause the transformation from S though Sq to Q-types in that taxonomic system (Strazzulla et al., 2005; Vernazza et al., 2009a).

Recently, neural networks have been used in the literature to perform asteroid classification. The study by Penttilä et al. (2021) presents the results of training an artificial neural network for the needs of the Gaia mission. The data used for this task were obtained from the spectra of 586 objects from DeMeo et al. (2009) and Binzel et al. (2019) in the range of 0.45 and 1.05 μm, spanned across 11 taxonomic types (A, B, C, D, K, L, Q, S, T, V, and X). Original types assigned by DeMeo et al. (2009) were processed, such that a few subclasses were generalized into their main equivalents (i.e., assigning S to types such as Sa and Sq). The data were enhanced with additional synthetic samples formed by using the principal component analysis. This study reports 86% unbalanced prediction accuracy for the task of taxonomic type prediction but cannot be directly compared to this study as it operates on slightly different data (more objects with different types assigned), a different feature set, and a different evaluation procedure. Furthermore, this study assesses the performance of the model on the selected wavelength range but does not quantify the importance of individual features.

The Bus–DeMeo taxonomy relies on spectroscopic measurements which are challenging to obtain for a large number of objects in a reasonable amount of time. Therefore, multi-filter photometry obtained in large sky surveys is often used to assign the taxonomic type to asteroids and perform compositional and evolutionary studies (e.g., Zellner et al. (1985); Carvano et al. (2010); Gil-Hutton and Licandro (2010); Sykes et al. (2000) and others).

Furthermore, surveys typically produce enormous amounts of data but can only cover a limited number of bands. For example, the Gaia mission was expected to provide spectra for about 100 000 objects in the spectral range from 0.325 to 1.1 μm (Mignard et al., 2007). The Large Synoptic Survey of the Vera Rubin Observatory is expected to provide observations in the ugrizy bandpasses for hundredths of thousands of near-Earth objects and millions of main-belt asteroids (Jones et al., 2015), and the Euclid mission will provide (visible (a broad g-r-i), Y, J, and H photometry for about 150 000 asteroids (Carry, 2018). Sergeyev et al. (2021) extracted visible colors for about ∼ 10⁵ asteroids. Those vast amounts of data require the use of fast and reliable machine learning approaches to classify those objects and link them to the current taxonomies. Furthermore, the robustness of different classification methods may have implications to the derived taxonomic distributions across the solar system, and thus its formation and evolution theories. Various studies have already performed the scientific analysis of different “big data” sets.

For example, Sykes et al. (2000) studied the distribution of color indices of asteroids and comets observed in the course of the Two Micron All Sky Survey (2MASS) and derived an approximate compositional map of the asteroid belt. Later, based on Sloan Digital Sky Survey (SDSS) data, Carvano et al. (2010) created a taxonomic classification compatible with Bus taxonomy and studied the distribution of different asteroid types in the solar system. A more detailed in-depth analysis of the SDSS data was later performed by DeMeo and Carry (2014) who also discussed the constraints on the formation and evolution of the solar system arising from the derived compositional distribution of asteroids. Popescu et al. (2016, 2018) assigned taxonomic classifications to objects observed in the course of VISTA-VHS survey and found multiple V- and A-type candidates, which are crucial in the context of the missing mantle problem (i.e., lack of basaltic mantle material as compared to iron core material present among meteorites, e.g., Burbine et al. (1996); Scott et al. (2010). Other studies focused on searching for specific taxonomic types using multi-filter data. Licandro et al. (2017) analyzed the MOVIS-C catalog to identify the V-type asteroids using infrared colors (Y-J) and (J-Ks). Similar studies were also performed based on the SDSS magnitudes, such as Solontoi et al. (2012); Roig and Gil-Hutton (2006); Oszkiewicz et al. (2014). DeMeo et al. (2019) investigated A-type candidates extracted from the SDSS catalog. Spectrophotometric characterization is also often used to constrain the composition of near-Earth objects Mommert et al. (2016); Navarro-Meza et al. (2019); Erasmus et al. (2017); Harris and Davies (1999); Bolin et al. (2020); De León et al. (2010). DeMeo et al. (2014) discussed the distribution of very red D-type asteroids located in the main belt (typically found among Jupiter trojans and in the outer main belt) and their potential origins. Thus, classification tools are essential in understanding asteroid composition and revealing multiple processes and more generally the evolution of the solar system.

These studies could benefit from data and observations made in filters that correspond to spectral features that best predict common taxonomic types. Current large surveys are an important source of information on asteroid composition and thus help understand various mechanisms and the overall formation and evolution of the solar system. However, most of those large surveys are dedicated to other astronomical objects and asteroids are only serendipitous objects. There has not yet been a large survey strictly dedicated to studying asteroids. Therefore, the optical setups of surveys are not optimized to taxonomically characterize asteroids. Furthermore, the large surveys are typically limited in the number of bands covered; thus, selecting those that improve the link to asteroid taxonomy can maximize the science output for solar system science. In this study, we investigate which spectral features provide the most information in the context of the Bus–DeMeo taxonomy, that is, which features should be observed to optimize future multi-filter surveys toward asteroid compositional studies. We also investigated various classification methods to verify their predictability of complexes and individual classes.

In section 2, we discuss the data, while section 3 describes parameterization and methodology used. In section 4, we present our results, and in section 5, we provide extended discussion. Conclusions are in section 6.

The dataset of asteroids used in the study consists of 371 objects included in Bus–DeMeo taxonomy (DeMeo et al., 2009) as well as 195 additional objects from Binzel et al. (2019). Most of the near-infrared data were collected at the NASA IRTF with the Spex instrument in its low-resolution (R ∼150) prism mode, and a majority of the visible wavelength data is from the Small Main-Belt Asteroid Spectroscopic Survey (R ∼100) Bus and Binzel (2002). For detailed information on that data, we refer the reader to the original studies. Each object is described by a set of visible and near-infrared spectral measurements (in the range of 0.45–2.45 μm), from which a set of spectral features is calculated, which is then used as an input for classification. Each object is also assigned to one of the 24 classes (taxonomic types) as defined in the study by DeMeo et al. (2009). For the purpose of this study, the dataset was limited to only contain the taxonomic types represented by more than 10 objects, reducing the data size from 566 to 504 objects split into 12 taxonomic types (S, Sq, Sr, C, Ch, X, Xk, L, V, D, K, and Q), as presented in Table 1. This was done in order to remove severely underrepresented taxonomic types with very few objects, for which machine learning methods, based on the statistical properties of the data, are not expected to work.

There were two prediction tasks considered in the scope of this study: the prediction of twelve taxonomic types and simplified version of the classification, with four complexes as target (C-complex, S-complex, X-complex, and end members). The labels for the prediction of complexes were assigned based on what the taxonomic type of each asteroid belonged to, as described in DeMeo et al. (2009).

The distribution of objects in two-dimensional space (resulting from taking the two main principal components of the spectra) with respect to taxonomic types is presented in Figure 1. The preparation of feature sets obtained from the spectral data is described in detail in the following sections.

For all methods mentioned in this section, except for gradient boosting, the implementations from scikit-learn (Pedregosa et al., 2011) package were used.

The top features that we found (Tables 6, 7, 9, 11; Figure 4) correspond to the most prominent spectral characteristics. For classification of complexes, spectral slope and reflectance covering the 1-μm absorption band contributed to the majority of the classification with reflectances at other wavelengths providing much less information. This is not surprising, given that the slope is the most prominent feature of spectra and the 1-μm band is the main feature separating the complexes containing the 1-μm olivine/pyroxene absorption (S-complex and some types in the end-member population) from the featureless ones (C- and X-complexes). In the balanced accuracy experiment, the 0.65 μm hydration feature also ranked high. That hydrated minerals feature is present in many C-complex objects and the analog meteorites (Rivkin et al., 2002) and is uncommon in the more thermally transformed types (e.g., V-types from the end members group), which aids the separation of the end members. Furthermore, that parameter splits the C- and Ch-types and C from S/X complex objects, which are redder and typically have higher reflectance values at those wavelengths. Reflectances in the near-infrared part of the spectra did not appear among the top features for complex prediction. This is consistent with the fact that Bus–DeMeo retained the complex classification from the earlier Bus taxonomy, which was based on the visible part of the spectra only.

For the type classification based on spectra, the top features contributing to the classifications were slope and reflectances relating to 1-μm absorption band and interestingly spectra in the region 1.35–1.4 μm. That spectral region characterizes multiple types. For example, the Q- and Sq-types show a shallow feature around 1.3 μm; C- and Ch-types have a small positive slope that starts around 1.3 μm; D-types show a gentle kick around 1.35 μm; and the V-types show maximum reflectance in that range. That region does not have a clear mineralogical interpretation. Surprisingly, the 2-μm region ranked below the top 20 important features. In the DeMeo et al. (2009) taxonomy, the 2-μm pyroxene band was the most prominent feature clearly separating the space in the PCA component space. In the classification task, this feature seems redundant as other parameters cover spectral features of multiple other classes simultaneously and already provide a plenitude of information. The 2 μm feature may have also ranked higher if more types containing the feature (e.g., R, O, and Sv) were taken under the analysis, thus creating the need for characterizing the distinction in the reflectance values. Those types were skipped due to low number statistics.

Altogether the low number of features needed to classify objects is not surprising, and analogous to PCA in DeMeo et al. (2009), where the first 4–5 dimensions provide the most information and then the variance in the subsequent dimensions drops to essentially zero. Generally, the top features separating the types and complexes found by the machine learning algorithms appear to focus on reasonable traits of the spectra relating to physical characteristics of asteroids and analogous meteorites.

The 1-μm olivine/pyroxene absorption band appears in both the classification of types and complexes. It is one of two the most striking absorption bands in asteroid spectra tied to olivine and pyroxene content (Gaffey et al., 1993; Sanchez et al., 2014). For the classification of complexes, it is a clear distinction between olivine-containing asteroids (S-complex, V-types, and Q-types) and the rest. For the classification of types, it additionally helps assessing the depth of the absorption band.

It is worth noting that the S-complex asteroids were found broadly compatible with ordinary chondrites (OC) based on the 1-μm absorption band (Chapman and Salisbury, 1973) and after understanding that the space weathering processes affect the spectral slopes, causing the asteroid spectra to appear redder than that of meteorites (Sasaki et al., 2001; Chapman, 2004; Strazzulla et al., 2005). Furthermore, the Hayabusa sample return space missions provided mineralogical argument connecting the OC meteorites with S-complex asteroids (Nakamura et al., 2011; Yurimoto et al., 2011). However, as indicated by Vernazza et al. (2016) and others, the S-complex asteroids are a diverse group, and other meteorite groups may originate from this complex as well.

On the other hand, the hydrated Ch/Cgh-type asteroids (that do not contain the 1-μm band) are partly linked with CM chondrites (Rivkin, 2012; Bland and Travis, 2017; Vernazza et al., 2017; Carry et al., 2019). Hiroi et al. (1993) also suggested a link with the CI/CM meteorites. Some C-types that do not show hydration have also been linked to IDPs Vernazza et al. (2015); Carry et al. (2021). Part of the C-complex is not sampled by the meteorite collection Vernazza et al. (2015), which highlights some of the challenges in linking asteroid spectral types with their meteorite analogs.

The featureless types are mostly distinguished by their spectral slopes. For the X-complex, we only included the X and Xk types, which tend to have a higher spectral slope than that of the C-complex objects (C, Ch) considered in this study. The Xk types are considered parent bodies of mesosiderates (Vernazza et al., 2009b). The X-type in the Bus–DeMeo taxonomy does not distinguish the E, M, and P Tholen types (known as X-complex degeneracy) and thus may correspond to multiple meteorite groups.

During our analysis, we excluded several spectra types for which there were not enough data (Sa, Sv, B, Cb, Cg, Cgh, Xc, Xe, Xn, T, A, O, and R). These rare types are not classified in our analysis. Since we are performing fivefold stratified cross-validation, we require at least five spectra in each type. Moreover including objects with low number statistics may lead to a bias-variance tradeoff phenomenon, that is, overfitting. It is most evident for complex models like multilayer perceptrons.

The exclusion means that some classes and their features are not represented in our data set. Therefore, the top features that we found do not optimize those classes. Penttilä et al. (2021) merged those subclasses with their main equivalent (Sa with S). Although this process increases sample size, it may negatively influence the inference for the merged types. Furthermore, merging several subtypes into a single class may lead the machine learning algorithms toward the conclusion that the features that discriminate between them are irrelevant. Therefore, we decided to omit those types altogether.

The spectral data were not corrected for multiple factors such as phase reddening, temperature, grain size, space weathering effects, or impact darkening. Many of those mechanisms are poorly constrained for the different taxonomic types and sometimes even object-dependent.

The spectra of S-complex asteroids show redder spectral slopes with increasing phase angle (the effect is known as phase reddening), and the depth of the one- and two-micron bands increases up to 70 and 55 of phase angle, respectively, and then decreases (Sanchez et al., 2012). However, most main-belt asteroids are observed in the phase angle range 0–30°, where the effect is minuscule. The effect plays a more important role for near-Earth asteroids (NEAs) that can be observed at higher phase angles. Binzel et al. (2019) demonstrated the phase reddening effect for 433) Eros and explored the slope correction values for a few other asteroids. Based on that Binzel et al. (2019) suggested that the effect may be object-dependent and thus not easily accounted for.

Particle size affects the depth and slope of absorption features and overall reflectance. Large particle sizes typically result in deeper absorption bands, bluer spectral slope, and lower reflectance Reddy et al. (2015); Mustard and Hays (1997). Generally, larger asteroids tend to have smaller grain sizes as revealed by thermal inertia studies (Delbo et al., 2015). However, the grain size distribution for most individual asteroids (except those visited by space missions) is unknown thus correcting for particle size is impossible.

Temperature causes changes in the band centers and depths as well as broadens or narrows down the bands (thus also affects the bar area ratios) (Singer and Roush (1985); Hinrichs and Lucey (2002); Burbine et al. (2009)). The variation in temperature is more significant for objects on eccentric orbits like NEAs and less for main-belt objects.

Shock impact darkening can suppress the absorption bands and reduce the overall reflectance, which can lead to misclassification of objects (Kohout et al., 2020, 2014) as it tends to “move” objects even between complexes. Overall, phase reddening corrections play an important role in constraining the space reddening as they mostly affect spectral slope (Reddy et al., 2015). Temperature variations and shock impact darkening have a larger impact on mineralogical characterization (Reddy et al., 2015).

All the listed mechanisms affect our understanding of the asteroid–meteorite connection. However, applying corrections for the aforementioned factors is not viable with the current state of knowledge for all the asteroids of different taxonomic types in our sample.

Furthermore, the Bus–DeMeo taxonomy is simply based on observed spectral features and was not controlled for those factors. By training to recreate the Bus–DeMeo types, we trained to classify objects based on biased data into the Bus–DeMeo biased classes. Thus, those mechanisms do not affect our results.

However, these mechanisms can significantly affect asteroid spectra and have to be considered when new taxonomies, possibly based on much larger data, are derived (e.g., that based on the Gaia space mission data (Cellino et al., 2020)). Last, as shown by the OSIRIS-REx mission, rubble-pile asteroids may contain a mixture of various materials, like the basaltic exogenic material found on a primitive asteroid (101,955) Benu (DellaGiustina et al., 2021). Another example of such mixed exogenous composition is asteroid 2008 TC3 that impacted the Earth on October 7, 2008, and the recovered fragments contained many different ureilitic and chondritic lithologies (Bischoff et al., 2010). How common those objects are and how should they be handled in future taxonomies have to be carefully considered.

Our analysis did not take into account albedo information. In contrast to the Tholen taxonomy, the Bus–DeMeo taxonomy did not take into account geometric albedo, thus introducing that the additional parameter could confuse the classification of types in the X-complex (otherwise split into E, M, and P types in the Tholen taxonomy) or at the very least be a spurious parameter that does not contribute to the classification of classes that are not defined by albedo. Furthermore, as shown by the confusion matrix, there are more misclassifications within the complexes, which have similar albedos (except for the X-complex). Since we focus on the Bus–DeMeo taxonomy, we do not include geometric albedo in our analysis as well.

In this study, we addressed the automatic classification of taxonomic types and complexes according to the DeMeo et al. (2009) taxonomy. We assessed machine learning methods’ capability of recreating the classification with the data used in DeMeo et al. (2009) to create the taxonomy. We showed that machine learning methods can be used to recreate the Bus–DeMeo taxonomy. Moreover, this taxonomic scheme is robust in the sense that most of the types and complexes remain in place, despite the increased sample size, which is not always the case for taxonomies based on PCA analysis when more data are available.

Furthermore, the experiments were carried out to record the difference in performance on the spectra, rather than five principal components. We found that the prediction accuracy improved across both tasks and all methods, which proves that the original feature set misses some important information required to distinguish different types and complexes. In our experiments, a multilayer perceptron with three layers of 32 neurons, a stochastic gradient descent solver, batch size of 32, and adaptive learning rate with the initial value of 0.1 performed best for both tasks, closely followed by the support vector machine with the RBF kernel.

Sequential feature selection was performed on the task of taxonomic type prediction to quantify the importance of individual features. In the experiments, the feature set was reduced to highlight the areas of interest that mostly contribute to making the prediction. For complex prediction, the top five features were sufficient for obtaining 93% prediction balanced accuracy, whereas six features resulted in 81% balanced accuracy in the case of taxonomic type prediction. This shows that the whole spectrum contains redundant features which are not informative for predicting the output. On the one hand, this may lead to degradation of classification, due to overfitting to noise. On the other hand, it shows spectra contain more information than necessary for taxonomic-type prediction and that validates the reduction to a few narrowband filters.

For future space and ground-based surveys that aim for taxonomic characterization of asteroids in the Bus–DeMeo scheme, we would recommend the use of six narrowband filters that cover the reflectances at 0.65, 1.0, and 1.4 μm and additional couple filters that allow for estimation of the spectral slope in the 0.45–2.5 μm. According to this study, those would be sufficient to obtain prediction balanced accuracy on the level of 93 and 81% for prediction complexes and types, respectively. Narrowband filters were already utilized in the J-PLUS survey (Morate et al., 2021); however, those could be further optimized to more efficiently link the observed asteroids to the Bus–DeMeo taxonomy.