Abstract: Spatial channel recognition is a crucial technology in Non-Terrestrial Networks (NTN) communication, providing prior information to meet communication needs in different environments. Due to the complex and rapidly changing propagation environment of long-distance satellite communication, existing channel recognition technologies that only use amplitude features struggle to accurately identify NTN channel types, resulting in low recognition accuracy.To address this issue, a Cascade Merge Convolutional Network (CMC-Net) using the channel impulse response complex signal is proposed. This approach captures independent and joint spatiotemporal features of IQ components without increasing model parameters, which improves NTN channel recognition performance and information utilization. Initially, a one-dimensional convolution extracts local spatial features from each IQ component, which are then combined via a concatenation layer followed by further feature extraction and compression via two cascaded convolution modules. Finally, a temporal classification module is employed to capture global temporal features, leading to the recognition result. Simulation results show that, compared to channel recognition methods using amplitude features, the use of channel IQ data improves the average recognition accuracy by 0.99%, 2.99%, and 8.25% for CNN, CNN-LSTM, and MLP classifiers, respectively. The proposed CMC-Net achieves an average recognition accuracy of 98.15% in the 0～20dB SNR range. In the -20～20dB SNR range, the CMC-Net method improves recognition accuracy by 32.77%, 4.07%, 3.48%, and 1.26% compared to LSTM, CNN, MLP, and CNN-LSTM classifiers, respectively.The results demonstrate that using channel impulse response complex signals effectively improves NTN channel recognition accuracy. The proposed CMC-Net method achieves higher accuracy for NTN channel recognition with fewer parameters, which expands the application of channel recognition technology in the satellite domain.

Key words: spatial channel recognition, NTN channels, deep learning, space communication, convolutional neural network