With the dramatic rise in high-quality galaxy data expected from Euclid and Vera C. Rubin Observatory, there will be increasing demand for fast high-precision methods for measuring galaxy fluxes. These will be essential for inferring the redshifts of the galaxies. In this paper, we introduce Lumos, a deep learning method to measure photometry from galaxy images. Lumos builds on BKGnet, an algorithm to predict the background and its associated error, and predicts the background-subtracted flux probability density function. We have developed Lumos for data from the Physics of the Accelerating Universe Survey (PAUS), an imaging survey using a 40 narrow-band filter camera (PAUCam). PAUCam images are affected by scattered light, displaying a background noise pattern that can be predicted and corrected for. On average, Lumos increases the SNR of the observations by a factor of 2 compared to an aperture photometry algorithm. It also incorporates other advantages like robustness towards distorting artefacts, e.g. cosmic rays or scattered light, the ability of deblending and less sensitivity to uncertainties in the galaxy profile parameters used to infer the photometry. Indeed, the number of flagged photometry outlier observations is reduced from 10 to 2 per cent, comparing to aperture photometry. Furthermore, with Lumos photometry, the photo-z scatter is reduced by ≈10 per cent with the Deepz machine-learning photo-z code and the photo-z outlier rate by 20 per cent. The photo-z improvement is lower than expected from the SNR increment, however, currently the photometric calibration and outliers in the photometry seem to be its limiting factor.

中文翻译：

---

PAU 调查：用深度学习估计星系光度

随着 Euclid 和 Vera C. Rubin 天文台预期的高质量星系数据的急剧增加，对测量星系通量的快速高精度方法的需求将不断增加。这些对于推断星系的红移至关重要。在本文中，我们介绍了 Lumos，这是一种从星系图像中测量光度的深度学习方法。Lumos 建立在 BKGnet 的基础上，这是一种预测背景及其相关误差的算法，并预测减去背景的通量概率密度函数。我们为来自加速宇宙调查 (PAUS) 的数据开发了 Lumos，这是一项使用 40 个窄带滤光片相机 (PAUCam) 的成像调查。PAUCam 图像受到散射光的影响，显示出可以预测和校正的背景噪声模式。一般，与孔径测光算法相比，Lumos 将观测的 SNR 提高了 2 倍。它还结合了其他优点，例如对扭曲伪像（例如宇宙射线或散射光）的鲁棒性、去混合能力以及对用于推断光度测量的星系轮廓参数的不确定性的敏感性较低。实际上，与孔径测光法相比，标记的测光异常观测值的数量从 10% 减少到 2%。此外，使用 Lumos 光度计，使用 Deepz 机器学习 photo-z 代码，photo-z 散射减少了约 10%，photo-z 异常值率减少了 20%。从 SNR 增量来看，photo-z 改进低于预期，但是，目前光度校准和光度测量中的异常值似乎是其限制因素。