Deep Learning-Based Channel Estimation for Satellite OTFS Systems in High-Mobility Scenarios

Abstract

This paper addresses the challenge of channel estimation for orthogonal time-frequency space (OTFS) modulation in low-Earth-orbit (LEO) satellite systems experiencing extreme Doppler spreads (>1 kHz). We propose a deep learning framework integrating a hybrid convolutional-recurrent network with physics-informed meta-learning, trained on multi-fidelity datasets combining ray-traced simulations and Iridium-NEXT channel measurements. The architecture employs 3D convolutions for delay-Doppler feature extraction and gated recurrent units for temporal coherence, while domain knowledge (e.g., cross-symbol interference patterns) is embedded via regularization constraints. Evaluations demonstrate 41.8% lower normalized mean squared error than compressed sensing baselines at 18 dB SNR, with 28% fewer FLOPs than transformer-based models. The solution achieves 92.4% pilot overhead reduction for 64-QAM transmissions, yielding 3.1×10⁻³ bit error rate at 23 ms latency under ITU-R M.2460-0 mobility conditions. Cross-validation across 10-30 dB SNR and 500-1200 km orbits shows <9% performance variance in Ricean channels, confirming robustness. Computational complexity analysis reveals a 2.7× improvement in FLOPs/accuracy trade-offs versus least-squares estimators. These advancements establish the framework as a scalable solution for 6G non-terrestrial networks, with immediate applications in LEO mega-constellations requiring high spectral efficiency. The work bridges OTFS theory and practical deployment through machine learning, providing measurable performance gains in dynamic satellite environments while maintaining backward compatibility with existing beamforming protocols.