Deep learning denoising by dimension reduction: Application to the ORION-B line cubes

Abstract

Context. The availability of large bandwidth receivers for millimeter radio telescopes allows for the acquisition of position-position-frequency data cubes over a wide field of view and a broad frequency coverage. These cubes contain a lot of information on the physical, chemical, and kinematical properties of the emitting gas. However, their large size coupled with an inhomogenous signal-to-noise ratio (S/N) are major challenges for consistent analysis and interpretation. Aims. We searched for a denoising method of the low S/N regions of the studied data cubes that would allow the low S/N emission to be recovered without distorting the signals with a high S/N. Methods. We performed an in-depth data analysis of the 13CO and C17O (1–0) data cubes obtained as part of the ORION-B large program performed at the IRAM 30 m telescope. We analyzed the statistical properties of the noise and the evolution of the correlation of the signal in a given frequency channel with that of the adjacent channels. This has allowed us to propose significant improvements of typical autoassociative neural networks, often used to denoise hyperspectral Earth remote sensing data. Applying this method to the 13CO (1–0) cube, we were able to compare the denoised data with those derived with the multiple Gaussian fitting algorithm ROHSA, considered as the state-of-the-art procedure for data line cubes. Results. The nature of astronomical spectral data cubes is distinct from that of the hyperspectral data usually studied in the Earth remote sensing literature because the observed intensities become statistically independent beyond a short channel separation. This lack of redundancy in data has led us to adapt the method, notably by taking into account the sparsity of the signal along the spectral axis. The application of the proposed algorithm leads to an increase in the S/N in voxels with a weak signal, while preserving the spectral shape of the data in high S/N voxels. Conclusions. The proposed algorithm that combines a detailed analysis of the noise statistics with an innovative autoencoder architecture is a promising path to denoise radio-astronomy line data cubes. In the future, exploring whether a better use of the spatial correlations of the noise may further improve the denoising performances seems to be a promising avenue. In addition, dealing with the multiplicative noise associated with the calibration uncertainty at high S/N would also be beneficial for such large data cubes.