J-PLUS: Spectral classification and photometric redshifts for 79 million sources in the fourth data release

Abstract

We present spectral classifications and photometric redshifts for 79.2 million sources up to an r-band magnitude of 22 in Data Release 4 of the Javalambre Photometric Local Universe Survey (J-PLUS). Leveraging the 12-band J-PLUS filter system, we compare a template-fitting approach (LePhare) against LeMoNNADE, a morphology-blind machine learning pipeline that uses spectral mixing augmentation to overcome training set limitations. LeMoNNADE consistently outperforms template fitting in precision, robust scatter, and outlier rates. Including WISE infrared photometry breaks optical degeneracies between stars and quasars, reducing the catastrophic outlier rate for quasars from ~40% to ~23% and constraining systemic redshift bias to <1% up to z = 4. We find LeMoNNADE is also less susceptible to redshift aliasing, particularly when adopting the probability density function median. Because the spectroscopic training samples severely under-represent stars, we apply an Expectation-Maximization Bayesian calibration to recover unbiased class probabilities for the magnitude-limited sample. This reveals that extragalactic counts agree with the literature down to the r ~ 20.5 completeness limit. The inferred redshift distribution for r < 21 extragalactic sources peaks at z ~ 0.3, showing broad agreement with existing literature up to z ~ 0.6. The resulting catalogues represent a significant milestone for local Universe science, offering probabilistically calibrated classifications and distances while explicitly characterising faint-end limits and contamination.