Hybrid STA With FNN and CNN Models for Robust Channel Estimation in Turbulent FSO and Doppler‐Affected OFDM transmission for LEO Satellite Communication

This paper introduces a hybrid framework for deep learning—based channel estimation tailored to free‐space optical (FSO) and FSO‐based low Earth orbit (LEO) OFDM systems. Unlike prior STA–DL approaches, the proposed method explicitly addresses two key challenges: atmospheric turbulence in terrestrial FSO links and Doppler‐induced carrier‐frequency offset (CFO) in high‐mobility LEO channels. The framework integrates spectral temporal averaging (STA) with three deep learning architectures—a feed‐forward neural network (STA‐FNN), a super‐resolution convolutional neural network (STA‐SRCNN), and a denoising convolutional neural network (STA‐DNCNN)—in a two‐stage process where STA supplies a coarse, noise‐resilient estimate that is refined by the deep network to capture nonlinear distortions and rapid channel variations. A pilot tone–based CFO synchronization module, incorporated ahead of the DL stage, provides explicit Doppler mitigation, a critical feature rarely combined with learning‐based estimators. Extensive simulations demonstrate that STA‐SRCNN and STA‐DNCNN achieve state‐of‐the‐art performance, delivering NMSE of , BER of , and EVM of 8% for 64‐QAM at 11 dB SNR in terrestrial FSO channels, while maintaining similar robust performance in LEO scenarios, as compared to conventional LS and STA. These results confirm that the proposed hybrid STA‐DL framework offers a computationally efficient and practically deployable solution that outperforms classical least‐squares and MMSE estimators for next‐generation high‐throughput optical satellite communications.