Abstract Stellar population measurements in integral field unit surveys are often limited by low signal-to-noise ratios (S/Ns) in low-surface-brightness spaxels. Using controlled synthetic experiments, we investigate whether a deep-learning-based denoising can recover stellar population information from such spectra without requiring spatial binning. We introduce the Enhanced U-Net Transformer (EUT), a one-dimensional convolutional neural network–transformer model trained on 90,000 synthetic spectra constructed from MILES simple stellar population (SSP) models following J. H. Lee et al., with wavelength-dependent noise injected on the fly to emulate SAMI-like data (S/N ≃ 5–20, measured in a 4484.77–4573.12 Å continuum window). Utilizing an independent test set of 10,000 spectra, the EUT reduces the full-spectrum rms residual by ≃96.5% at S/N = 5 (and by ≃94% at S/N = 20), achieving recovery rates of ≥99.8% (the Pearson correlation coefficient between the noise-free and comparison spectra expressed in percent). In fixed windows around Ca ii H, H δ , H β , Fe i 4383, Mg b, and Na D, residuals decrease by ≳88% while preserving line-profile structure. In downstream analysis with p PXF we assess parameter recovery using the Pearson correlation coefficient R p and the rms scatter: the scatter in recovered mass-weighted age decreases from ≃0.41 to ≃0.25 dex at S/N = 5 and from ≃0.32 to ≃0.22 dex at S/N = 10; the corresponding mass-weighted global metallicity, M/H, scatter decreases from ≃0.45 to ≃0.36 dex and from ≃0.32 to ≃0.28 dex. At S/N = 20, denoising yields results consistent with those from the noisy inputs within the synthetic-test uncertainties. These controlled experiments suggest that hybrid CNN–transformer denoisers can enhance the usable low-surface-brightness area for stellar population studies, although further validation with observed spectra will be needed before practical application.
Kim et al. (Mon,) studied this question.