Deep Learning Approach to Photometric Redshift Estimation

Abstract

Data-driven approaches play a crucial role in space computing, and our paper focuses on analyzing data to learn more about celestial objects. Photometric redshift, a measure of the shift of light towards the red part of the spectrum, helps determine the distance of celestial objects. This study used a dataset from the Sloan Digital Sky Survey (SDSS) with five magnitudes alongside their corresponding redshift labels. Traditionally, redshift prediction relied on spectral distribution templates (SEDs), which, though effective, are costly and limited, especially for large datasets. This paper explores data-driven methodologies instead of SEDs. By employing a decision tree regressor and a fully connected neural network (FCN), we found that the FCN outperforms the decision tree regressor in RMS. The results show that data-driven estimation is a valuable tool for astronomical surveys. With the adaptability to complement previous methods, FCNs will reshape the field of redshift estimation.