Deep Learning Improves Photometric Redshifts in All Regions of Color Space

Image-based deep learning methods outperform photometry-based methods for photometric redshift (photo-z) estimation for SDSS galaxies, but it is not understood why. We investigate this mystery by partitioning galaxy color space with a self-organizing map (SOM). Deep learning methods outperform classical machine learning methods in all regions of color space. Much of this difference is due a color-dependent attenuation bias for photometry-based methods that is subtle when considering the entire dataset but clear for neighborhoods of color space. Our findings suggest that photometry is less optimally combining the redshift information from individual pixels than deep learning models.