KF-LSTM-Enhanced Time-Frequency Channel Estimation for OTFS Low-Earth-Orbit Satellite Communication Systems

In low earth orbit (LEO) satellite-ground communication scenarios, orthogonal time-frequency space (OTFS) modulation has been recognized as a highly promising waveform. Many existing OTFS channel estimation methods commonly rely on inserting high-power pilot symbols in the DD domain, which may lead to a high peak-to-average power ratio (PAPR). To mitigate this issue, several studies have proposed placing pilot symbols in the time-frequency (TF) domain and combining them with deep-learning-based techniques for channel estimation. Nevertheless, the offline training and online inference of such deep learning-based schemes typically involve large-scale matrix operations, thereby imposing considerable challenges on satellite-ground terminals with limited computational resources. To overcome these limitations, this paper proposes a TF domain OTFS channel estimation method based on Kalman filter (KF) and long short-term memory (LSTM) networks. The proposed KF-LSTM model first employs a KF to denoise and smooth the initial channel estimates, followed by an LSTM network to capture the nonlinear and temporal correlations of the channel states, thereby achieving a high-precision interpolation across the entire TF grid. Simulation results demonstrate that the proposed method outperforms several typical neural network models in terms of normalized mean square error (NMSE) and bit error rate (BER). Moreover, the KF-LSTM approach achieved faster convergence, lower computational overhead, and stronger generalization capability under varying Doppler and deployment conditions, making it well suited for resource-constrained LEO satellite-ground communication systems.