Image-catalog complementarity in quasar-galaxy ambiguous regimes of photometric source classification
- 1School of Modern Logistics, Zhejiang Technical Institute of Economics , Xuezheng Street, Hangzhou ,
- 2Research Center for Computational Earth and Space Sciences , Zhejiang Lab, No. 2880 Wenyi West Road, Hangzhou ,
Abstract
The photometric separation of stars, galaxies, and quasars is important for large imaging surveys, but quasar–galaxy ambiguity remains challenging when colors and optical appearances overlap. We present a rule-guided image–catalog fusion analysis to identify regimes in which optical image cutouts provide complementary information to catalog features. The study uses a balanced medium-hard benchmark of 12 000 spectroscopically labeled sources combining SDSS labels, Legacy Survey measurements, WISE variability, Gaia features, and optical cutouts. We compare catalog ensemble models, a ConvNeXt image classifier, and probability-level image–catalog fusion in quasar–galaxy ambiguous regimes. These regimes are selected using local color-space overlap from training-set nearest-neighbor class probabilities and a label-free catalog-uncertainty score from a reference catalog classifier. In the rule-selected color-overlap and high-uncertainty regimes, fusion reaches macro-F1 scores of 0.643–0.696, improving catalog-only baselines by 2.6–3.1 percentage points. The quasar F1 score increases in all three fixed ambiguous subsets, and fusion corrects more catalog errors than it introduces without increasing quasar–galaxy confusion. With validation and test sets fixed, increasing the image-training set from 8 400 to 84 000 sources raises image-only macro-F1 from 0.7486 to 0.7856. Ablation tests, nearest-neighbor-scale tests, threshold sensitivity analyses, and bootstrap resampling associate these gains with the joint selection of local color mixing and catalog uncertainty. The results provide evidence for a modest positive trend, suggesting that optical cutouts may be particularly useful for sources that are both locally mixed in color space and uncertain under catalog-only prediction.
Related
- The Chandra-Gaia Catalog of Counterparts: Resolving ambiguous Gaia matches to X-ray sources in the Chandra Source Catalog using Machine Learning
- photo-3x2-pt: Cosmology from cosmic shear and galaxy clustering with a single photometric galaxy catalog
- A Machine Learning Quasar Candidate Catalog with Photometric Redshifts from KiDS DR5 and AllWISE
- A Homogeneous Catalog of Rossiter-McLaughlin Systems: Distinct e-λ Trends in Three Gas-Giant Mass Regimes
- FOMO-Z. Optimal galaxy classification, photometric redshifts, and reconstruction of population distributions
BibTeX
@article{an2026imagecatalog,
title = {{Image-catalog complementarity in quasar-galaxy ambiguous regimes of photometric source classification}},
author = {Ruqiao An and Chuan He and Yijun Pan and Hao Chen},
journal = {Monthly Notices of the Royal Astronomical Society},
year = {2026},
month = nov,
doi = {10.1093/mnras/stag2091},
url = {https://academic.oup.com/mnras/advance-article/doi/10.1093/mnras/stag2091/9012718},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Image-catalog complementarity in quasar-galaxy ambiguous regimes of photometric source classification
AU - Ruqiao An
AU - Chuan He
AU - Yijun Pan
AU - Hao Chen
PY - 2026
DA - 2026/11/04
JO - Monthly Notices of the Royal Astronomical Society
DO - 10.1093/mnras/stag2091
UR - https://academic.oup.com/mnras/advance-article/doi/10.1093/mnras/stag2091/9012718
AB - The photometric separation of stars, galaxies, and quasars is important for large imaging surveys, but quasar–galaxy ambiguity remains challenging when colors and optical appearances overlap. We present a rule-guided image–catalog fusion analysis to identify regimes in which optical image cutouts provide complementary information to catalog features. The study uses a balanced medium-hard benchmark of 12 000 spectroscopically labeled sources combining SDSS labels, Legacy Survey measurements, WISE variability, Gaia features, and optical cutouts. We compare catalog ensemble models, a ConvNeXt image classifier, and probability-level image–catalog fusion in quasar–galaxy ambiguous regimes. These regimes are selected using local color-space overlap from training-set nearest-neighbor class probabilities and a label-free catalog-uncertainty score from a reference catalog classifier. In the rule-selected color-overlap and high-uncertainty regimes, fusion reaches macro-F1 scores of 0.643–0.696, improving catalog-only baselines by 2.6–3.1 percentage points. The quasar F1 score increases in all three fixed ambiguous subsets, and fusion corrects more catalog errors than it introduces without increasing quasar–galaxy confusion. With validation and test sets fixed, increasing the image-training set from 8 400 to 84 000 sources raises image-only macro-F1 from 0.7486 to 0.7856. Ablation tests, nearest-neighbor-scale tests, threshold sensitivity analyses, and bootstrap resampling associate these gains with the joint selection of local color mixing and catalog uncertainty. The results provide evidence for a modest positive trend, suggesting that optical cutouts may be particularly useful for sources that are both locally mixed in color space and uncertain under catalog-only prediction.
SN - 0035-8711
SN - 1365-2966
ER -