DeepDISC-photoz: Deep Learning-Based Photometric Redshift Estimation for Rubin LSST

TL;DR

DeepDISC-photoz is a deep learning model designed for accurate photometric redshift estimation in LSST, outperforming traditional methods by leveraging object detection, segmentation, and probabilistic redshift predictions, with robustness to systematics.

Contribution

It introduces a novel deep learning framework that combines object detection and redshift estimation, improving accuracy and efficiency over traditional catalog-based photo-z methods.

Findings

DeepDISC photo-z outperforms traditional estimators on simulated LSST data.

Photo-z accuracy is primarily affected by image signal-to-noise ratio.

The method is robust against systematics like galactic extinction and blending.

Abstract

Photometric redshifts will be a key data product for the Rubin Observatory Legacy Survey of Space and Time (LSST) as well as for future ground and space-based surveys. The need for photometric redshifts, or photo-zs, arises from sparse spectroscopic coverage of observed galaxies. LSST is expected to observe billions of objects, making it crucial to have a photo-z estimator that is accurate and efficient. To that end, we present DeepDISC photo-z, a photo-z estimator that is an extension of the DeepDISC framework. The base DeepDISC network simultaneously detects, segments, and classifies objects in multi-band coadded images. We introduce photo-z capabilities to DeepDISC by adding a redshift estimation Region of Interest head, which produces a photo-z probability distribution function for each detected object. On simulated LSST images, DeepDISC photo-z outperforms traditional catalog-based estimators, in both point estimate and probabilistic metrics. We validate DeepDISC by examining dependencies on systematics including galactic extinction, blending and PSF effects. We also examine the impact of the data quality and the size of the training set and model. We find that the biggest factor in DeepDISC photo-z quality is the signal-to-noise of the imaging data, and see a reduction in photo-z scatter approximately proportional to the image data signal-to-noise. Our code is fully public and integrated in the RAIL photo-z package for ease of use and comparison to other codes at https://github.com/LSSTDESC/rail_deepdisc