Photometric redshift estimation of galaxies with Convolutional Neural Network

Abstract

The abundant photometric data collected from multiple large-scale sky surveys give important opportunities for photometric redshift estimation. However, low accuracy is still a serious issue in the current photometric redshift estimation methods. In this paper, we propose a novel two-stage approach by integration of Self Organizing Map (SOM) and Convolutional Neural Network (CNN) methods together. The SOM-CNN method is tested on the dataset of 150 000 galaxies from Sloan Digital Sky Survey Data Release 13 (SDSS-DR13). In the first stage, we apply the SOM algorithm to photometric data clustering and divide the samples into early-type and late-type. In the second stage, the SOM-CNN model is established to estimate the photometric redshifts of galaxies. Next, the precision rate and recall rate curves (PRC) are given to evaluate the models of SOM-CNN and Back Propagation (BP). It can been seen from the PRC that the SOM-CNN model is better than BP, and the area of SOM-CNN is 0.94, while the BP is 0.91. Finally, we provide two key error indicators: mean square error (MSE) and Outliers. Our results show that the MSE of early-type is 0.0014 while late-type is 0.0019, which are better than the BP algorithm 22.2% and 26%, respectively. When compared with Outliers, our result is optimally 1.32%, while the K-nearest neighbor (KNN) algorithm has 3.93%. In addition, we also provide the error visualization figures about ΔZ and δ. According to the statistical calculations, the early-type with an error of less than 0.1 accounts for 98.86%, while the late-type is 99.03%. This result is better than those reported in the literature.