Spectroscopic and Photometric Redshift Estimation by Neural Networks For the China Space Station Optical Survey (CSS-OS)

TL;DR

This paper demonstrates that neural networks, specifically 1-d CNN and MLP, can accurately estimate spectroscopic and photometric redshifts for the CSS-OS survey, outperforming traditional methods in efficiency and outlier suppression.

Contribution

It introduces neural network models for redshift estimation in the CSS-OS survey, achieving high accuracy and reducing outliers compared to existing methods.

Findings

Redshift accuracy ~0.001 for spec-z and ~0.01 for photo-z.

Neural networks outperform traditional template-fitting methods.

Efficient training and suppressed outliers demonstrate feasibility for future surveys.

Abstract

The estimation of spectroscopic and photometric redshifts (spec-z and photo-z) is crucial for future cosmological surveys. It can directly affect several powerful measurements of the Universe, e.g. weak lensing and galaxy clustering. In this work, we explore the accuracies of spec-z and photo-z that can be obtained in the China Space Station Optical Surveys (CSS-OS), which is a next-generation space survey, using neural networks. The 1-dimensional Convolutional Neural Networks (1-d CNN) and Multi-Layer Perceptron (MLP, one of the simplest forms of Artificial Neural Network) are employed to derive the spec-z and photo-z, respectively. The mock spectral and photometric data used for training and testing the networks are generated based on the COSMOS catalog. The networks have been trained with noisy data by creating Gaussian random realizations to reduce the statistical effects, resulting in similar redshift accuracy for both high-SNR (signal to noise ratio) and low-SNR data. The probability distribution functions (PDFs) of the predicted redshifts are also derived via Gaussian random realizations of the testing data, and then the best-fit redshifts and 1-sigma errors also can be obtained. We find that our networks can provide excellent redshift estimates with accuracies ~0.001 and 0.01 on spec-z and photo-z, respectively. Compared to existing photo-z codes, our MLP has similar accuracy but is more efficient in the training process. The fractions of catastrophic redshifts or outliers can be dramatically suppressed comparing to the ordinary template-fitting method. This indicates that the neural network method is feasible and powerful for spec-z and photo-z estimations in future cosmological surveys.