Star Photometry for DECam Legacy Survey and Sloan Digital Sky Survey Images Based on Convolutional Neural Networks

Abstract

Flux is one of the most fundamental parameters in astrophysics, and aperture photometry and point-spread function (PSF) photometry are commonly used methods to obtain the flux. With the continuous development of astronomical equipment that has generated massive data, researchers have to find more effective methods to obtain stellar fluxes. However, current photometric software such as SExtractor are very sensitive to the configuration parameters and are difficult to configure. We propose a new photometric model based on deep learning called sf-convolutional neural network (CNN) to extract aperture fluxes and PSF fluxes. For the simulated data including 5727 stars, the experimental results show that sf-CNN can predict fluxes better than SExtractor. The mean absolute error (MAE) values of sf-CNN and SExtractor for predicting PSF fluxes are 0.0034 and 0.0134, respectively. On the 6293 mixed stars in DECam Legacy Survey Data Release (DR) 9, the MAE values of the predicted PSF fluxes are 0.0075 and 0.0177, respectively. The PSF accuracy of the sf-CNN model is significantly higher than that of SExtractor. Additionally, the MAE values of the predicted aperture fluxes on 6215 mixed stars and 1341 blends of stars in Sloan Digital Sky Survey DR 12 illustrate that the accuracy of sf-CNN is still the highest. Meanwhile, the results indicate that sf-CNN outperforms VGG16 and ResNet50. Furthermore, sf-CNN is 100–200 times faster than Photutils on RTX 3070 GPU and 20–40 times faster than Photutils on I7 12700 CPU. sf-CNN can calculate fluxes efficiently and accurately only by setting a few parameters and may thus become a fundamental tool for the era of big data in astronomy.

Keywords

• Stellar photometry
• Convolutional neural networks
• Surface photometry
• Fundamental parameters of stars
• CCD photometry