The miniJPAS and J-NEP Surveys: Machine Learning for Star-Galaxy Separation

Abstract

We present a supervised machine learning classification of sources from the Javalambre Physics of the Accelerating Universe Astrophysical Survey (J-PAS) Pathfinder datasets: miniJPAS and J-NEP. Leveraging crossmatches with spectroscopic and photometric catalogs, we construct a robust labeled dataset comprising 14,594 sources classified into extended (galaxies) and point-like (stars and quasars) objects. We assess dataset representativeness using UMAP analysis, confirming broad and consistent coverage of feature space. An XGBoost classifier, with hyperparameters tuned using automated optimization, is trained using purely photometric data (60-band J-PAS magnitudes) and combined photometric and morphological features, with performance thoroughly evaluated via ROC and purity–completeness metrics. Incorporating morphology significantly improves classification, outperforming the baseline classifications available in the catalogs. Permutation importance analysis reveals morphological parameters, particularly concentration, normalized peak surface brightness, and PSF, alongside photometric features around 4000 and 6900 Å, as crucial for accurate classifications. We release a value-added catalog with our models for star-galaxy classification, enhancing the utility of miniJPAS and J-NEP for subsequent cosmological and astrophysical analyses.

1. Introduction

The Javalambre Physics of the Accelerating Universe Astrophysical Survey (J-PAS, www.j-pas.org (accessed on 23 November 2025)) is a ground-based imaging survey that will map thousands of square degrees using a unique set of 56 filters, providing a low-resolution spectrum for every pixel of the sky (the wavelength resolution is $R_{\lambda }=\lambda /\Delta \lambda$, so that $R\sim 60$ is the approximate value in the intermediate wavelength range), without the target selection biases or fiber collision issues that typically affect spectroscopic surveys. Designed and optimized to achieve highly accurate photometric redshifts, its scientific objectives span from cosmology to galaxy evolution [1].

Here we focus on data from the Pathfinder camera, which was used to test the telescope performance and execute the first scientific operations. The camera features a 9 k × 9 k CCD, with a 0.3 deg² field of view and 0.225 arcsec pixel size. This effort led to the miniJPAS [1] and J-NEP [2] surveys, which covered $1{\mathrm{deg}}^2$ of the AEGIS field (four pointings around $(\mathrm{RA},\mathrm{Dec})=({215}^{\circ },{53}^{\circ })$) and $0.29{\mathrm{deg}}^2$ of the North Ecliptic Pole (one pointing centred at $(\mathrm{RA},\mathrm{Dec})=(260.6^{\circ },65.8^{\circ })$), respectively. Both surveys use the 56 J-PAS filters together with the u, g, r, and i broad bands, for a total of 60 bands (note that the J-PAS survey itself employs only the i broad band in addition to the 56 narrow bands). However, J-NEP adopted longer exposure times, reaching between 0.5–1.0 magnitudes deeper than miniJPAS, which itself reaches the target depth of J-PAS–namely, $m_{\mathrm{AB}}\approx 24$ in the broad-band filters (5$\sigma$ in a 3″ aperture; $m_{\mathrm{AB}}$ denotes AB magnitudes, defined such that a source with constant flux density of $3631\mathrm{Jy}$ has $m_{\mathrm{AB}}=0$).

While miniJPAS targeted the well-studied Extended Groth Strip field, an area observed by many experiments, J-NEP fully covers the time-domain field of the James Webb Space Telescope’s North Ecliptic Pole [3], a region chosen for its low galactic extinction, absence of bright stars, and continuous visibility from Webb’s northern hemisphere.

Our main goal here is to provide a machine learning-based classification of J-NEP sources as an alternative to the Bayesian stellar and galactic loci classifier (SGLC [4], in jnep.StarGalClass; throughout this work, for reproducibility, we explicitly reference the table names available at www.j-pas.org/datareleases (accessed on 23 November 2025)). This aims both to validate the quality of the SGLC, possibly outperfoming it, and to offer an independent classification that can be used to assess the robustness of subsequent scientific analyses. In particular, our approach incorporates the full photometric information, which is crucial for faint galaxies whose morphology closely resembles that of stars.

The SGLC is based on morphological information, specifically on the bimodal distribution observed in the concentration–magnitude diagram, which distinguishes compact point-like objects from extended sources. López-Sanjuan et al. [4] modeled these two populations to compute the probability that a given source is compact or extended. This model, combined with appropriate priors, was then used to estimate the Bayesian probability that a source is a star or a galaxy [1].

In Baqui et al. [5], we obtained the machine learning classification of miniJPAS sources, available in minijpas.StarGalClass. Here, taking advantage of new spectroscopic data and an improved methodology, we revise the classification of miniJPAS sources and provide a machine learning classification of J-NEP sources. This paper is organized as follows. Section 2 describes the training data, Section 3 presents the feature configuration, Section 4 details the classification methodology and examines the representativeness of the labeled set, and Section 5 reports the results. Finally, conclusions are drawn in Section 6.

2. Data

In this work, we primarily used the jnep.MagABDualObj and minijpas.MagABDualObj tables from the data releases J-NEP-PDR202107 and MINIJ-PAS-PDR201912, respectively, available at https://www.j-pas.org/datareleases (accessed on 23 November 2025). These catalogs provide photometric and morphological measurements for all sources detected in the r band, with forced photometry in the remaining filters. The r band was adopted as the detection image owing to its depth, image quality, and completeness, as detailed in Bonoli et al. [1]. All queries employed in this work are listed in Appendix A.

To ensure data quality, we required objects to satisfy mask_flags=0, which excludes detections affected by image boundaries, bright-star halos, or artifacts. For the flags field, which encodes issues encountered during photometric extraction, we retained objects with values 0 (clean, no problems detected), 1 (neighbor, the object is close to bright neighbors or contains bad pixels), 2 (blended, originally detected together with another source but successfully deblended), 4 (saturated, containing one or more pixels at or near saturation), and 6 (blended + saturated, a combination of the previous two cases). These situations can still yield reliable photometry and remain relevant for galaxy and cluster science, whereas more severe conditions (e.g., truncation at image edges, incomplete aperture measurements, or failed deblending) were excluded. This yields a total of 25,198 objects for miniJPAS and 7290 objects for J-NEP within the range $15<r<23.5$.

Since training supervised machine learning models requires ground-truth labels, we cross-match miniJPAS and J-NEP data with surveys that provide reliable photometric or spectroscopic classifications into extended (galaxies) and point-like (stars and quasars) sources. This is a crucial step, as the performance of a supervised ML model depends critically on the quality and representativeness of the adopted training set. The surveys used for crossmatching in this work are:

• SDSS DR12 (Section 2.1): the final Data Release (DR12) of the third-generation Sloan Digital Sky Survey (SDSS-III [6]), which collected imaging and spectroscopy from 2008 to 2014, providing wide-area multiband photometry and spectra of millions of sources.
• Gaia EDR3 (Section 2.2): the Early Data Release 3 of the Gaia mission [7], delivering high-precision astrometry, broad-band photometry, and radial velocities for over 1.5 billion stars in the Milky Way.
• DEEP3 DR4 (Section 2.3): spectroscopic redshift survey conducted with DEIMOS on the Keck II telescope in the Extended Groth Strip (EGS) [8]. It provides secure galaxy redshifts and spectra, forming part of the AEGIS survey program.
• Binospec (Section 2.4): spectroscopic data from Binospec [9], a high-throughput optical spectrograph on the 6.5-m MMT, covering 370–1000 nm with wide field of view, used to obtain redshifts for faint galaxies.
• DESI DR1 (Section 2.5): the first data release of the Dark Energy Spectroscopic Instrument [10], based on the first 13 months of the main survey. It provides high-confidence redshifts for 18.7 million objects (13.1 M galaxies, 1.6 M quasars, 4.0 M stars) over more than 9000 deg², making it the largest extragalactic spectroscopic sample to date and a key resource for cosmology and large-scale structure studies.
• HSC-SSP PDR2 (Section 2.6): the second public data release of the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP [11]), providing deep, wide-field imaging in five broad bands ($grizy$), with exquisite seeing from the Subaru telescope.

2.1. Crossmatch with SDSS DR12

We used the table minijpas.xmatch_sdss_dr12, which contains the crossmatch with SDSS DR12 and includes both the spectroscopic classification spCl and the photometric classification class, expected to be reliable for $15<r<20$ [5]. We verified the latter by comparing DESI, DEEP3, SDSS (spectroscopic), and Gaia sources (taken as ground truth) against the SDSS photometric classification. The photometric classification yields 802 galaxies and 1228 stars within $15<r<20$, for a total of 2030 labeled sources. The spectroscopic classification provides 355 galaxies and 220 point-like sources (94 stars and 124 quasars) for $r<23.5$, for a total of 573 labeled sources.

We also used the analogous table jnep.xmatch_sdss_dr12. The photometric classification yields 156 galaxies and 724 stars within $15<r<20$, for a total of 880 labeled sources. No spectroscopic labels were available, since J-NEP lies outside the SDSS footprint.

2.2. Crossmatch with Gaia EDR3

We used the tables minijpas.xmatch_gaia_edr3 and jnep.xmatch_gaia_edr3, which contain the crossmatches with Gaia EDR3. As we are interested in parallax, Gaia EDR3 is essentially equivalent to the newer Gaia DR3. Stars were identified as objects with a parallax measured at a significance greater than 3$\sigma$, i.e., parallax_over_error > 3. This selection yields 989 stars for miniJPAS and 570 stars for J-NEP.

2.3. Crossmatch Between miniJPAS and DEEP3 DR4

We crossmatched miniJPAS with $r<23.5$ with the DEEP3 DR4 catalog alldeep.egs.2012jun13 (sites.uci.edu/deep3/deep3zcat (accessed on 23 November 2025)) In the case of duplicates, we retained the entry with the highest ZQUALITY, and, if necessary, the one with the smallest redshift error Z_ERR. The crossmatch was performed using the Python module onexmatch v0.1.0 (github.com/valerio-marra/onexmatch (accessed on 23 November 2025)) [12], adopting a maximum separation of 1 arcsec and an ambiguity threshold of 0.5 arcsec, yielding 8279 matched objects as shown in Figure 1.

As shown by the diagnostic metrics, the positional offsets in Right Ascension and Declination are well described by a zero-mean isotropic 2D Gaussian whose standard deviation is equal to the average of the measured ${\sigma }_{\Delta \mathrm{RA}}$ and ${\sigma }_{\Delta \mathrm{DEC}}$; the resulting Kullback–Leibler divergence is ≲0.1. The adoption of an ambiguity threshold of 0.5 arcsec removed 10 ambiguous DEEP3 matches associated with duplicated miniJPAS sources, whose positional differences were below this threshold—about twice the pixel scale and comparable to the PSF size. See Appendix C for more details.

After crossmatching, we retained only sources with ZQUALITY ⩾3, yielding 6583 galaxies, 433 active galactic nuclei (AGNs), and 5 stars. Unlike quasars (QSOs), many AGNs exhibit extended morphologies, which introduces ambiguity in the classification process. To mitigate this, we restricted point-like sources to objects with morph_prob_star > $0.9$ and redshift Z > $0.1$, and stars to those with morph_prob_star > $0.9$ and Z < $0.01$. The morph_prob_star values are provided by the jnep.StarGalClass table, which estimates stellar probability from morphological information [4]. This selection resulted in 36 point-like sources (35 AGNs and 1 star). This filtering step was necessary to construct a clean labeled dataset. Our study focuses on a binary star-galaxy classification, rather than a three-class star-AGN/QSO-galaxy scheme, which would require a significantly larger and more balanced AGN sample.

2.4. Crossmatch Between J-NEP and Binospec Data

We crossmatched J-NEP sources with $r<23.5$ with data that was obtained by Christopher Willmer using Binospec at the MMT observatory. Details about the dataset are provided in Hernán-Caballero et al. [2]. The crossmatch was performed using the onexmatch code, adopting a maximum separation of 1 arcsec and an ambiguity threshold of 0.5 arcsec, and yielded 957 matched sources, as shown in Figure 2. The adoption of an ambiguity threshold of 0.5 arcsec removed 67 ambiguous Binospec matches associated with duplicated miniJPAS sources, mostly with positional differences below 0.2 arcsec (Figure A1).

After crossmatching, we retained only the 726 objects with ZQUALITY ⩾ 3 and Z_ERR < 0.01. Since the Binospec catalog does not provide spectroscopic classifications, we assigned class labels based on morphology and redshift: objects with morph_prob_star $>0.9$ and $Z<0.01$ were classified as stars, and those with morph_prob_star $<0.1$ and $Z>0.01$ as galaxies. This yielded 53 stars and 593 galaxies, for a total of 646 objects. There are 69 objects with morph_prob_star $>0.1$ and $Z>0.01$. These could correspond either to QSOs or to compact galaxies; we exclude this small set of sources from the catalog.

2.5. Crossmatch Between miniJPAS and DESI DR1

We crossmatched miniJPAS sources with $r<23.5$ with the DESI DR1 catalog (zall-pix-iron.fits at data.desi.lbl.gov/public/dr1/spectro/redux/iron/zcatalog/v1 (accessed on 23 November 2025)). The crossmatch was performed using the onexmatch code, adopting a maximum separation of 1 arcsec and an ambiguity threshold of 0.5 arcsec, and yielded 2865 matches, as shown in Figure 3. The adoption of an ambiguity threshold of 0.5 arcsec removed 1232 ambiguous DESI matches associated with duplicated miniJPAS sources, mostly with positional differences below 0.05 arcsec (Figure A1).

After crossmatching, as suggested in [10], only sources with ZCAT_PRIMARY = True, ZWARN = 0 and DELTACHI2 > 25 were retained, for a total of 2640 objects. Of these, 2307 are galaxies and 333 are point-like sources (124 QSOs and 209 stars).

2.6. Crossmatch Between miniJPAS and HSC-SSP PDR2

We used the photometric classification provided by x_extendedness_value (with x = g, r, i, z, y) from the catalog minijpas.xmatch_subaru_pdr2, which contains the crossmatch with the HSC-SSP PDR2 catalog. To ensure consistency, we retained only the 11,960 objects for which all five x_extendedness_value measurements agree.

We tested the HSC-SSP classification in the range $18.5<r<23.5$ against DESI, DEEP3, SDSS (spectroscopic) and Gaia sources, used as ground truth; see

Table 1. Confusion matrices in three r-band magnitude bins, comparing DESI, DEEP3, SDSS (spectroscopic) and Gaia sources (used as ground truth) against HSC-SSP classifications.

	18.5–20.5		20.5–22.5		22.5–23.5
	Pred. Stars	Pred. Gal.	Pred. Stars	Pred. Gal.	Pred. Stars	Pred. Gal.
true stars	79	10	92	21	6	4
true gal.	0	385	5	1523	9	2350

. We find that for $r>22.5$ a non-negligible fraction of stars, though small in absolute numbers, are misclassified as galaxies. To minimize this effect, we restrict the HSC-SSP labels to the range $18.5<r<22.5$, where the bright limit avoids saturation. This selection yields 4406 galaxies and 946 stars, for a total of 5352 objects.

2.7. Final Labeled Set

The crossmatched catalogs from the previous sections partially overlap. When dealing with duplicate J-PAS sources, we assigned class labels according to the following priority order, reflecting the reliability of each classification: DESI, Binospec, DEEP3, SDSS (spectroscopic classification), Gaia, HSC-SSP, and SDSS (photometric classification).

Table 2. Summary of external catalogs used to crossmatch and label sources in miniJPAS and J-NEP. Values in parentheses indicate the total number of matched sources before duplicate removal. When a source appears in multiple catalogs, labels are assigned based on a priority order corresponding to the table’s top-down sequence, with DESI given the highest priority. As a result, the final number of adopted labels is smaller than the per-catalog match count.

Crossmatch Source	miniJPAS (Galaxies/Stars)	J-NEP (Galaxies/Stars)
DESI DR1	2307/333 (2307/333)	–
Binospec	–	593/53 (593/53)
DEEP3 DR4	5748/16 (6583/36)	–
SDSS DR12 (spectro)	189/124 (355/218)	–
Gaia EDR3	–/826 (–/989)	–/565 (–/570)
HSC-SSP PDR2	2323/765 (4406/946)	–
SDSS DR12 (photo)	157/261 (802/1228)	105/229 (156/724)
Total	10,724/2325	698/847

summarizes the labeled dataset, comprising 14,594 unique sources. The corresponding r-band magnitude distributions of the retained sources from each survey are shown in Figure 4.

Figure 5 shows the r-band magnitude distributions of stars and galaxies for the miniJPAS and J-NEP surveys. In both surveys, galaxies dominate at $r\gtrsim 21$, while stars are more prevalent at $r\lesssim 18.5$. This results in a class imbalance in the training set, which can bias both the performance of the classifier and the interpretation of its metrics. To avoid misleading conclusions from this imbalance, we evaluate performance using purity-completeness curves, which provide a balanced view of classification quality independent of the relative proportions of stars and galaxies.

Finally, Figure 6 shows the histogram of the magnitude differences between the SDSS and J-PAS r-band total flux magnitudes for point sources. This provides an additional crosscheck of the quality of the crossmatch. We restrict the analysis to point sources because J-PAS does not provide a total-flux model magnitude for extended objects, as MAG_AUTO is measured within an aperture of two Kron radii and therefore misses a non-negligible fraction of the galaxy flux in the outskirts.

As stated earlier, we perform a binary classification into extended and point-like sources. Point-like sources include both stars and QSOs, which are morphologically similar at the J-PAS resolution. However, the number of spectroscopically confirmed QSOs in our sample is insufficient to support a reliable three-class classification [13,14,15,16,17] that would exploit the low-resolution J-PAS spectra. As is customary, we refer to extended sources as galaxies and to point-like sources as stars throughout this paper.

3. Feature Configuration

3.1. Observational Features

To characterize the observing conditions and data quality, we included four observational features:

• Point Spread Function (PSF) of the tile,
• Galactic extinction $E(B-V)$,
• $E(B-V)$ error,
• Object flags indicating observational issues.

The PSF is characterized as the mean full width at half maximum (FWHM) of point-like sources in each tile, estimated with PSFEx [18]. The color excess $E(B-V)$ is the integrated reddening due to Milky Way dust along the line of sight, provided at the position of each source by the Bayestar17 three-dimensional dust map [19,20].

3.2. Photometric Features

We used a total of 120 photometric features derived from the 60 J-PAS bands. Specifically, we considered MAG_AUTO magnitudes and their associated uncertainties in each of the 60 bands from MagABDualObj. Bands without detection were assigned a magnitude of 99, while the magnitude errors encapsulate valuable information about the strength and reliability of each detection.

Although MagABDualPointSources provides PSF-corrected photometry (MAG_APER_COR_3_0) optimized for point-like objects–shown to outperform MAG_AUTO for spectral energy distribution (SED) fitting and photometric redshift estimation–our classification tests indicate a different trend. For the photometry-only model, MAG_AUTO yields better results, with ROC AUC = 0.992 and star purity-completeness AUC = 0.984, compared to 0.984 and 0.965 for MAG_APER_COR_3_0 (ROC stands for receiver operating characteristic, and AUC denotes the area under the curve). For galaxies, the improvement is smaller but still in favor of MAG_AUTO (AUC = 0.997 vs. 0.994). When morphology is included, the two definitions perform nearly identically–0.1% better performance for MAG_APER_COR_3_0. Given the superior performance of MAG_AUTO in the photometry-only case and its comparable performance when morphology is included, we adopt MAG_AUTO as our default choice in this work.

3.3. Morphological Features

We adopted four morphological parameters derived from columns relative to the detection i band, available in the MagABDualObj catalog:

• Ellipticity A/B = A_WORLD/B_WORLD, where A_WORLD and B_WORLD are the root mean square values of the light distribution along the major and minor axes of the source, respectively.
• Concentration c_r = MAG_APER_1_5 − MAG_APER_3_0 (in the r-band), calculated from magnitudes measured within circular apertures of 1.5 and 3.0 arcsec.
• Full Width at Half Maximum FWHM_WORLD, defined assuming a Gaussian intensity distribution.
• Normalized peak surface brightness μ = MU_MAX/MAG_APER_3_0 (in the r-band), where MU_MAX represents the peak surface brightness above the local background.

The spatial variation of the PSF across the J-PAS tiles introduces multimodality into the distributions of concentration, FWHM, and peak surface brightness. To correct for this effect and achieve uniform morphological features, we normalize each parameter by subtracting the corresponding median stellar value computed within the same tile:

$$\begin{array}{c}\hfill X_i\to X_i-{\langle X\rangle }_{\mathrm{stars},\mathrm{tile}},\end{array}$$

(1)

where X denotes concentration, FWHM, or peak surface brightness. This correction effectively removes the multimodal behavior, as illustrated in Figure 7, which compares the distributions before and after correction. In addition, the separation between stellar and galactic populations becomes more pronounced following this normalization.

We do not apply this correction to the ellipticity, as it does not display multimodal features and can take spurious values for stars. Due to the finite pixel scale, a narrow PSF may lead SExtractor to assign $A/B\ne 1$ to unresolved sources, introducing artificial shape distortions caused by limitations in the PSF fitting procedure. See (Bonoli et al. [1], Appendix C) for the SExtractor configuration adopted.

We use the normalized morphological parameters for the representativeness analysis presented in the next section. However, for the classification task, we retain the uncorrected catalog values. Our tests show that including the PSF as an additional feature is sufficient to account for its spatial variations and prevent confusion, while the normalization procedure introduces additional complexity to the pipeline without yielding a significant improvement in classification performance.

3.4. Feature Sets

We perform three separate classifications. The first uses the full set of 128 features, which includes 60 magnitudes with their associated errors, four morphological parameters, the PSF information, Galactic extinction with its error, and detection flags. The second is restricted to 123 features, consisting of the 60 magnitudes with their errors, Galactic extinction with its error, and detection flags. We refer to these two configurations as ‘morphology + photometry’ and ‘photometry only’, respectively. The rationale for testing the latter is that morphological parameters may be biased in tiles produced by combining exposures taken under different observing conditions; in such cases, the tile-averaged PSF may not adequately represent individual sources, making a photometry-only classification potentially more robust.

In addition, we evaluate a third model based solely on morphological information, comprising the four morphological parameters, PSF, and detection flags. This setup serves as a benchmark to compare with the other models and, in particular, with the SGLC classifier, thereby quantifying the performance gain provided by machine learning.

4. Supervised Classification

4.1. Representativeness

The miniJPAS and J-NEP surveys were conducted using the same instrument and filter system, though under slightly different observing conditions. To evaluate the reliability and generalization capacity of the labeled datasets assembled in Section 2, we assess their representativeness along two complementary directions:

• the internal consistency between the labeled miniJPAS and labeled J-NEP subsets;
• the external coverage of the full miniJPAS and J-NEP datasets by the combined labeled set.

For this purpose, we adopt the Uniform Manifold Approximation and Projection (UMAP) algorithm [21], a non-linear dimensionality reduction technique designed to preserve both local and global data structure in low-dimensional projections. Importantly, our use of UMAP is diagnostic: we do not attempt classification in the embedded space, but rather evaluate whether the labeled samples span the same feature manifold as the full catalogs.

We apply UMAP to the 124 photometric and morphological features described in Section 3, excluding the observational features listed in Section 3.1, as these are non–object-specific or discrete and could introduce spurious clustering. Figure 8 shows the two-dimensional UMAP projection of the labeled miniJPAS and J-NEP sources. The J-NEP subset is largely embedded within the distribution of miniJPAS, indicating that the two labeled samples are consistent and can be reliably combined for model training. Figure 9 compares the combined labeled dataset against the full miniJPAS and J-NEP populations. Note that the UMAP projection differs from that in Figure 8, so the two plots are not directly comparable in terms of their coordinate positions. The labeled set broadly spans the full feature space, confirming that it is sufficiently representative to serve as a training base for supervised classification of the entire catalog. It is important to note, however, that while the labeled set spans the overall feature space, this does not guarantee that labeled stars (or galaxies) fully cover the star (or galaxy) subspaces of the full catalogs.

4.2. TPOT Pipeline for XGBoost Optimization

Based on the results of von Marttens et al. [13], who compared several machine learning algorithms, we adopt the eXtreme Gradient Boosting algorithm (XGBoost v3.0.0 [22]) for classification. Given the relatively modest size of our labeled dataset (fewer than 15,000 objects), XGBoost is a suitable choice. While neural networks may outperform tree-based models for substantially larger datasets, they are unlikely to offer significant advantages in our case. XGBoost is an ensemble learning method based on decision trees and gradient boosting. It iteratively adds new trees to correct the errors of the existing ensemble while minimizing a regularized objective function, which balances prediction accuracy and model complexity. This process results in a robust and efficient classifier that is well-suited for structured tabular data. For technical details, see Chen and Guestrin [22].

The final hyperparameter configuration was selected using the Tree-based Pipeline Optimization Tool (TPOT v0.12.2 [23], github.com/EpistasisLab/tpot (accessed on 23 November 2025)) an automated machine learning framework that optimizes machine learning pipelines through genetic programming. The number of pipelines evaluated by TPOT depends on three key hyperparameters: generations, population size, and offspring size. These respectively control the number of optimization iterations, the number of top-performing pipelines retained in each generation, and the number of new pipelines generated per generation.

The total number of pipelines analyzed is given by Population Size + Generations × Offspring Size. In this work, we set all three hyperparameters to 200, resulting in the evaluation of 40,200 distinct pipelines. The final hyperparameter configuration is listed in Appendix B. Model selection was performed exclusively within the training set, comprising 80% of the combined miniJPAS and J-NEP labeled samples. TPOT used 5-fold cross-validation within this training set to optimize pipeline architectures and hyperparameters. After optimization, the best-performing pipeline was retrained on the same 80% split, while the remaining 20% was held out as an independent test set. This test set was used only once for final evaluation, ensuring that the reported performance provides an unbiased estimate of the model’s generalization ability.

4.3. Performance Metrics

The XGBoost classifier outputs a probabilistic score between 0 (star) and 1 (galaxy), requiring the choice of a classification threshold $p_{\mathrm{cut}}$ to separate the two classes. Once $p_{\mathrm{cut}}$ is specified, the model’s performance can be evaluated using a $2\times 2$ confusion matrix. From its entries, we derive standard classification metrics such as the receiver operating characteristic (ROC) curve and the purity–completeness curve, which provide graphical assessments of classifier performance.

The ROC curve is a parametric plot of the true positive rate (TPR) versus the false positive rate (FPR) as a function of $p_{\mathrm{cut}}$:

$$\begin{array}{c}\hfill \mathrm{TPR}\left(p_{\mathrm{cut}}\right)={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}},\mathrm{FPR}\left(p_{\mathrm{cut}}\right)={\displaystyle \frac{\mathrm{FP}}{\mathrm{FP}+\mathrm{TN}}},\end{array}$$

(2)

where TP, FN, FP, and TN are the entries of the confusion matrix. TPR corresponds to recall, commonly referred to as completeness in astronomy. A commonly used scalar summary of the ROC curve is the area under the curve (AUC), which ranges from 0.5 (random classifier) to 1 (perfect classifier).

To account for class imbalance in the training set, we also consider the purity-completeness curve. It is a parametric plot of the purity (or precision) versus the completeness (or recall) as a function of $p_{\mathrm{cut}}$:

$$\begin{array}{c}\hfill \mathrm{Purity}\left(p_{\mathrm{cut}}\right)={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}},\mathrm{Completeness}\left(p_{\mathrm{cut}}\right)={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}}.\end{array}$$

(3)

A compact summary of this curve is given by the average precision (AP), defined as the area under the purity–completeness curve. AP values range from 0 to 1, with higher values indicating better classification performance.

5. Results

5.1. Metrics

Figure 10 shows the ROC curves for the three XGBoost models: photometry only (123 features, top), morphology only (6 features, middle), and morphology + photometry (128 features, bottom). For comparison, we also include the performance of SGLC and CLASS_STAR, the neural-network classifier implemented in SExtractor [24]. The photometry-only model outperforms CLASS_STAR and achieves performance comparable to, and slightly better than, SGLC. It also surpasses the morphology-only model, which performs as SGLC. The morphology + photometry model is the best overall, yielding the highest ROC AUC and thus the most accurate classifications. The same figure also shows the purity and completeness curves for galaxies (second column) and stars (third column). Despite the class imbalance in the training set, stars are well classified, with only a modest reduction in average precision (area under the purity-completeness curve).

Figure 11 shows the confusion matrices for the models based on photometry only (top row) and on photometry combined with morphology (bottom row), evaluated in three r-band magnitude bins. A classification threshold of $p_{\mathrm{cut}}=0.5$ is used throughout. As expected, classification performance deteriorates at fainter magnitudes, primarily due to stars being increasingly misclassified as galaxies. The inclusion of morphological features significantly mitigates this effect, confining most misclassifications to the faintest bin ($r>22.5$). In contrast, the classification of galaxies remains consistently accurate across all bins.

Figure 12 shows the purity and completeness for galaxies and stars as functions of the r-band magnitude, evaluated at a fixed target value of 95%. In each magnitude bin, this is achieved by selecting the probability threshold $p_{\mathrm{cut}}$ that yields the desired classification level. Results are presented in discrete magnitude bins (top panels) and cumulatively (bottom panels), for XGBoost models trained with photometry only and with photometry plus morphology, compared against SGLC. Galaxy classifications remain highly accurate across all magnitudes. For stars, performance degrades at faint magnitudes, but improves with the inclusion of morphology, surpassing SGLC. The apparent drop at $r>22.5$ in the binned results is explained by the very small number of faint stars in the training set (see Figure 5) and does not significantly affect the cumulative statistics.

5.2. Misclassifications

Misclassifications are dominated by faint stars ($r>21$) erroneously labeled as galaxies. Two effects drive this: (i) a brightness imbalance in the training set (relatively brighter stars and fainter galaxies) and (ii) color/SED degeneracy in photometry-only models at low SNR. Adding morphology largely mitigates the problem—star→galaxy errors drop sharply and galaxy→star errors are essentially removed. At the faint end, however, stellar and galactic morphologies converge, increasing ambiguity. This is evident in Figure 13, where the normalized peak surface brightness $\mu$ drifts toward the stellar reference; the same trend appears in the photometry-only model. Thus, morphology reduces the rate of mistakes but does not fully break the underlying degeneracies, indicating that part of the confusion arises from photometric/SED similarity in faint sources. These results align with von Marttens et al. [13] and suggest that more sophisticated treatment of photometry (e.g., improved uncertainty handling) could yield further gains.

To investigate whether additional factors contribute to the misclassifications, Figure 14 shows the fraction of stars classified as galaxies by survey of origin within the test set. No clear trends are observed, excluding the possibility that stellar misclassifications are mainly due to incorrect truth labels. We also examined misclassifications as a function of flags, but the numbers were negligible. Finally, Appendix D presents visual inspections of representative misclassifications across both models and magnitude ranges.

5.3. Stellar Locus

Figure 15 shows the stellar locus. The dashed line is a fourth-degree polynomial fit to all labeled stars, overplotted on the 2D histogram to guide the reader’s eye. Red markers show the XGBoost predictions from the full value-added catalog (VAC) using the model with photometry (left) and photometry and morphology (right): each point corresponds to the median value in a bin, with error bars indicating the 16th and 84th percentiles. The predicted locus closely follows the reference across the magnitude range, consistent with the high stellar purity reported in Figure 11.

5.4. Feature Importance

To evaluate the relative contribution of each feature group, we computed permutation importance on the test set. This metric measures the change in ROC AUC when the values of a given feature group are randomly permuted, thereby breaking its relationship with the target. A larger drop corresponds to higher importance. Small negative values (i.e., a slight increase in AUC after permutation) do not imply that a feature ’confuses’ the classifier; they typically arise from sampling noise on a finite test set and/or correlations with other features. We therefore interpret such negative importances as statistically consistent with zero (non-informative) importance.

Figure 16 shows the permutation importance results for the XGBoost model trained with photometry only. The most informative features are the three broad bands i, g, and r, together with three narrow bands: J0460 (4603 Å), J0680 (6812 Å), and J0390 (3904 Å). These results are consistent with those of Baqui et al. [5] (Figure 13), who found peaks in feature importance around 4000 and 6900 Å. This suggests that the model exploits these wavelengths to estimate the slope of the spectral energy distribution.

Figure 17 shows the results for the XGBoost model trained with photometry and morphology. The most informative features are the concentration index $c_r$ and the normalized peak surface brightness $\mu$, followed by the PSF. FWHM and photometry also contribute, while ellipticity $(A/B)$, extinction $E(B-V)$ with its error, and detection flags have negligible impact. This reflects the fact that miniJPAS and J-NEP each cover only about $1{deg}^2$, where Milky Way extinction varies very little. The results therefore confirm that morphological features carry the dominant information for effective star–galaxy separation in these fields.

6. Conclusions

The machine learning-based classification of J-NEP and miniJPAS sources presented in this paper improves star-galaxy separation compared to previous methods. Using an XGBoost classifier with hyperparameters tuned through automated optimization, trained on a robust labeled dataset derived from crossmatching with multiple spectroscopic and photometric catalogs, we provide reliable classifications both from purely photometric data and from combined photometric and morphological features.

Our key findings can be summarized as follows:

• The XGBoost model combining photometric and morphological features achieves a higher area under the ROC curve and outperforms both the SGLC and CLASS_STAR classifiers across all magnitude ranges.
• Using photometry alone yields competitive performance, slightly surpassing SGLC, but leads to misclassifications of faint stars as galaxies at $r>22$. This ambiguity is reduced when morphological features are included.
• The permutation importance analysis identifies the concentration, normalized peak surface brightness, and PSF as critical morphological parameters for classification. Photometric bands at approximately 3900, 4600, and 6800 Å are consistently among the most informative, highlighting their relevance for characterizing stellar and galactic spectral slopes.
• The UMAP representativeness analysis demonstrates the robustness and consistency of the labeled training set, validating its applicability for reliable classification across the full miniJPAS and J-NEP datasets.
• Our ML pipeline, publicly released as a value-added catalog (VAC), offers reliable star-galaxy classifications that can directly support downstream scientific analyses, including studies of galaxy evolution and cosmological investigations reliant on precise object classifications.