Semantic and Channel-Aware Image Transfer in Satellite Networks

Abstract

—Non-Terrestrial Networks (NTNs) are expected to support future large-scale remote sensing, but transmitting high-resolution satellite imagery is challenged by limited feeder links, time-varying channels, and constrained onboard computation. We propose AITACS (Adaptive Image Transmission with Asymmetric Computation over Satellites), an end-to-end framework for satellite-ground image transmission. AITACS combines semantic feature extraction and Deep Joint Source Channel Coding (DJSCC) to transmit task-relevant latent representations instead of pixels. To address satellite-ground computational asymmetry, encoder stages are compressed via structured pruning and quantization, while more expressive decoders are retained at the receiver. A semantic-feedback mechanism adapts representation size and transmission rate to channel conditions. Evaluated on ship detection using a realistic feeder-link model, AITACS improves accuracy by up to 3.7× and reduces transmission rate by over 50% compared to conventional DJSCC. Transmitter complexity and memory usage drop by 74% and 95%, respectively, with minimal performance loss.