JESSNet: Joint dEconvolution and Sparse Separation Network for HI 21 cm foreground cleaning

Abstract

i 21 cm intensity mapping, here considered over the redshift range z 0.58, particularly in the presence of a chromatic instrumental beam that couples angular foreground structure into the frequency direction. This problem is further complicated by incomplete sky coverage, as expected for realistic Galactic masks and survey footprints. -- i signal in single-dish intensity mapping observations affected by frequency-dependent beam distortions and masked sky regions. JESSNet advances the joint deconvolution and blind source-separation framework of DecGMCA and SDecGMCA through three methodological extensions. First, while the original formulation uses a single effective mixing model over the angular range being cleaned, JESSNet operates in angular windows, with an effective mixing matrix and a channel selection adapted to each scale. Second, JESSNet replaces the fixed starlet thresholding step with a learned spherical-wavelet ``learnlet'' sparse regularization operator. Third, whereas the original spherical implementation assumes full-sky coverage, JESSNet constrains the source estimates to the observed sky region, making the method applicable to Galactic masks and survey footprints. We tested the method on simulated SKA-MID-like observations covering 900 -- ̊m MHz , including H, i emission, Galactic and extragalactic foregrounds, thermal noise, and an oscillating in frequency MeerKAT-inspired beam model. i signal. Compared to applying JESSNet over a single angular range, the multiscale formulation significantly improves the reconstruction at low and intermediate angular scales and reduces residual beam-induced spectral features. The method also remains effective for a synthetic analysis footprint, preserving the H, i power spectra under incomplete sky coverage. i intensity mapping. The code used to produce the analysis and results has been publicly released, allowing the community to reproduce the pipeline and adapt it to future simulations and real observations.