Abstract:ABSTRACT Deep learning (DL) has recently been proposed as a novel approach for 21cm foreground removal. Before applying DL to real observations, it is essential to assess its consistency with established methods, its performance across various simulation models, and its robustness against instrumental systematics. This study develops a commonly used U-Net and evaluates its performance for post-reionization foreground removal across three distinct sky simulation models based on pure Gaussian realizations, the Lagrangian perturbation theory, and the Planck sky model. Consistent outcomes across the models are achieved provided that training and testing data align with the same model. On average, the residual foreground in the U-Net reconstructed data is $\sim 10~{{\ \rm per\ cent}}$ of the signal across angular scales at the considered redshift range. Comparable results are found with traditional approaches. However, blindly using a network trained on one model for data from another model yields inaccurate reconstructions, emphasizing the need for consistent training data. The study then introduces frequency-dependent Gaussian beams and bandpass fluctuations to the test data. The network struggles to denoise data affected by ‘unexpected’ systematics without prior information. However, after re-training consistently with systematics-contaminated data, the network effectively restores its reconstruction accuracy. Our results highlight the importance of incorporating prior knowledge during network training compared with established blind methods. Our work provides critical guidelines for using DL for 21cm foreground removal, tailored to specific data attributes. Notably, it is the first time that DL has been applied to the Planck sky model being most realistic foregrounds at present.

Star-Image Centering with Deep Learning: HST/WFPC2 Images

Star Image Centering with Deep Learning. II. HST/WFPC2 Full Field of View

Star-Image Centering with Deep Learning II: HST/WFPC2 Full Field of View

Sub-pixel Location Algorithm for Planetary Center Measurement

Accuracy Analysis of Centroid Detection Method Used to Star Sensor

Star Cluster Classification using Deep Transfer Learning with PHANGS-HST

Image deconvolution and PSF reconstruction with STARRED: a wavelet-based two-channel method optimized for light-curve extraction

CSST Dense Star Field Preparation: A Framework for Astrometry and Photometry for Dense Star Field Images Obtained by the China Space Station Telescope (CSST)

Combining statistical learning with deep learning for improved exoplanet detection and characterization

Direct Exoplanet Detection Using Deep Convolutional Image Reconstruction (ConStruct): A New Algorithm for Post-Processing High-Contrast Images

A New Cosmic-Ray Rejection Routine for HST WFC3/UVIS via Label-free Training of deepCR

The stability of deep learning for 21cm foreground removal across various sky models and frequency-dependent systematics

Real-Time Convolutional Neural Network-Based Star Detection and Centroiding Method for CubeSat Star Tracker

Star Centroiding Based on Fast Gaussian Fitting for Star Sensors

Direct Exoplanet Detection using Convolutional Image Reconstruction (ConStruct): A New Algorithm for Post-processing High-contrast Images

High-precision Algorithm for Centroid Location of Star Pattern Recognition

Comparison of Digital Centering Algorithms Based on CCD Images

AutoSourceID-Classifier. Star-Galaxy Classification using a Convolutional Neural Network with Spatial Information

A HIGH-PRECISION ALGORITHM FOR CENTROID LOCATION OF STAR PATTERN RECOGNITION WITH APPLICATION TO OPTICAL IMAGING DISTORTION CORRECTION

Near Infrared Star Centroid Detection by Area Analysis of Multi-Scale Super Pixel Saliency Fusion Map

Scientific Preparation for CSST: Classification of Galaxy and Nebula/Star Cluster Based on Deep Learning