http://arxiv.org/abs/2301.00313
We exploit the statistical independence of stellar features and atmospheric adversarial effects in stellar spectra, to remove the latter from observed signals using a fully unsupervised data-driven approach. Concretely, we first increase the inter-observation entropy of telluric absorption lines by imposing a random, virtual radial velocity to the observed spectrum. This novel “trick” results in a non-standard form of “whitening” in the atmospheric components of the spectrum, decorelating them across multiple observations. Then we use deep convolutional auto-encoders, to learn a feature-space in which the two “sources” of information, stellar and atmospheric, are easily separable, leading to removal of the latter. We apply the process on spectra from two different data collections: ~250,000 HARPS spectra and ~660,000 from SDSS. We compare and analyze the results across datasets, as well as with existing tools, and discuss directions for utilizing the introduced method as a fast and more reliable tool in the future.
N. Sedaghat, J. Kalmbach, B. Smart, et. al.
Tue, 3 Jan 23
44/49
Comments: N/A