Suitability of deep JSCC for rapidly varying satellite communication channels

Determine whether deep-learning-based joint source-and-channel coding architectures that integrate source compression, channel coding, and modulation into a single neural encoder–decoder are suitable for rapidly varying satellite-to-ground RF communication channels characterized by multipath propagation and shadowing, rather than only for standard terrestrial channel models such as additive white Gaussian noise (AWGN).

Background

Recent work has proposed deep-learning-based joint source-and-channel coding (JSCC) that combines source compression, channel coding, and modulation within a single neural encoder–decoder. These methods have shown promising results when trained and evaluated on standard terrestrial channel models, typically AWGN.

Small satellite links in low Earth orbit (LEO) present more complex and rapidly varying propagation conditions, including shadowing and multipath. The paper highlights that whether deep JSCC methods generalize from simple terrestrial models to these more challenging satellite channels remains explicitly open, motivating the authors to study a realistic satellite channel model and attention-based adaptability within a single network.