Leveraging Deep-Learning for Adaptive Coding and Modulation for LEO Satellite-Terrestrial Networks

Low-earth-orbit (LEO) satellite-terrestrial network (LSTN) is expected to play a significant role in the next-generation wireless networks ($6 \mathbf{G}$), providing global coverage and facilitating various 6G applications. However, the high dynamics of channel state information (CSI) in LSTN have posed challenges to modulation and coding schemes (MCS), which adversely affects the correctness of data transmission and the quality of service (QoS) expected by users. To enhance the communication reliability of LSTN, in this paper, we investigate the adaptive coding and modulation (ACM) method under the digital video broadcasting (DVB) standards. We first establish a channel model and then introduce an ACM method involving channel state estimation, channel state prediction, and MCS switching. Subsequently, the least squares (LS) algorithm is leveraged for channel estimation and we propose an ARIMA-Bidirectional Long Short-Term Memory (AB-LSTM) algorithm specifically for forecasting time-series signal-to-noise ratio (SNR) values for precise channel prediction. Then MCS switching is conducted based on the predicted SNR and switching threshold based on DVB standards. We conduct extensive simulations to show that our proposed algorithms can achieve higher estimation and prediction accuracy, and the introduced ACM method can effectively optimize the usage of channel capacity.