Cross-Telescope Generalization of Deep Learning RFI Segmentation

Radio frequency interference (RFI) is a critical challenge for radio astronomy, and deep learning methods have shown excellent in-domain segmentation performance on individual telescopes. However, whether these models generalise across telescopes remains largely untested. We present a controlled-architecture multi-domain benchmark for deep learning RFI segmentation that is, to our knowledge, among the first designed specifically to quantify cross-domain generalization under a unified evaluation matrix, evaluating a fixed U-Net baseline (approximately 1.93 million parameters) across 12 training configurations and 4 evaluation domains spanning three radio telescopes on three continents: HERA (South Africa, 100--200 MHz), LOFAR (Netherlands, 115--189 MHz), and NSRT (China, 1--1.5 GHz). For the newly introduced NSRT domain, we develop an automated preprocessing pipeline incorporating median-subtraction baseline removal and frequency-only flagging strategies to protect pulsar signals. Our 12 by 4 evaluation matrix reveals three principal findings: transfer between the evaluated domains is severely asymmetric, with a LOFAR-trained model achieving a Dice coefficient of 0.544 on HERA while the reverse transfer yields only 0.242; however, this ordering should be read as an empirical pattern under the present data, label, split, and training-recipe conditions rather than as a single-factor causal ordering. Sequential fine-tuning to a new domain consistently triggers catastrophic forgetting of the source domain, with performance drops of up to 93%. Target-calibrated median absolute deviation (MAD) baselines provide useful offline diagnostic references for telescopes lacking labelled training data, but the MAD and deep-learning operating points use different calibration regimes and therefore are not a fully matched blind zero-shot contest. These results demonstrate that single-telescope evaluations substantially overestimate the practical utility of deep learning RFI methods, and we recommend cross-domain evaluation as standard practice.