Hossein Ahmadi, Ali Kuhestani, Mohammad Reza Keshavarzi, Luca Mesin, Semantic-Aware Doppler Shift Estimation for Low Earth Orbit Satellite Communications, Submitted to IEEE Trans. Wireless Commun., 2025

Abstract

Accurate Doppler shift estimation is essential for coherent communication in low Earth orbit (LEO) satellite communication systems, particularly under the high-mobility and low-signal-to-noise ratio (SNR) conditions typical of nextgeneration 6G networks. Traditional methods often fail to generalize across varying channel dynamics and overlook the semantic structure inherent in motion-induced frequency shifts. This paper proposes a novel multi-task, semantic-aware Doppler estimation framework that fuses convolutional and Transformer-based representations, guided by attention-driven semantic masking and supervised contrastive learning. Our architecture jointly performs Doppler regression, Doppler bin classification, and semantic motion classification, enabling both precise estimation and interpretable reasoning. Extensive simulations show that our ensemble model outperforms convolutional neural networks (CNN) and Transformer baselines, reducing normalized mean squared error (NMSE) by up to 35%, and achieving over 85% accuracy in semantic classification across all SNR levels and temporal resolutions. Additionally, the model demonstrates calibrated uncertainty (±2σ coverage of 74.5%) and attentionbased interpretability through semantic heatmaps and confusion matrices. This work bridges physical-layer signal processing with semantic-aware learning, offering a robust foundation for intelligent Doppler tracking in future LEO satellite networks.