Photometric Redshift Estimation via a Spectra-guided Gaussian Mixture Density Framework

Abstract

Accurate and computationally efficient photometric redshift estimation is essential for cosmological analyses and studies of the large-scale structure of the Universe. However, applying existing template-fitting and machine learning methods to rapidly growing survey data sets can involve trade-offs between predictive performance and computational throughput. Moreover, feature extractors developed for natural images may not be well suited to the sparse, source-centered, and multiband characteristics of galaxy images. We present PRISM (Photometric Redshift Estimation via a Spectra-guided Gaussian Mixture Density Framework), a multimodal framework that jointly exploits multiband image cutouts and catalog-level photometric measurements. PRISM incorporates feature-extraction modules tailored to source-centered astronomical images and a dedicated fusion module that captures complementary information from the imaging and tabular modalities. During training, a frozen spectral encoder provides an auxiliary representation-alignment objective, encouraging selected photometric features to retain information associated with galaxy spectra without requiring spectroscopic data at inference time. A Gaussian-mixture density head models the conditional redshift distribution, enabling PRISM to produce a full predictive probability density function in addition to a point estimate. We evaluated PRISM through five-fold cross-validation on 1.98 million galaxies constructed from the Dark Energy Spectroscopic Instrument Data Release 1 and the DESI Legacy Imaging Surveys. Within this benchmark, PRISM reduces the mean absolute error and catastrophic-outlier fraction by 17.1% and 20.0%, respectively, relative to the best baseline value for each metric. Our PRISM ablations showed that combining multiband images with catalog flux measurements improved predictive performance relative to either input alone. Spectra-guided supervision provided a smaller additional gain in the full multimodal configuration. The code and data catalog are publicly available at https://github.com/qintianjian-lab/PRISM.