Hybrid Deep Learning for Channel Estimation and Tracking in RIS-Assisted UAV Wireless Communications

Abstract

Channel estimation in reconfigurable intelligent surface (RIS)-aided unmanned aerial vehicle (UAV) systems is severely hindered by high-dimensional cascaded channels, UAV-induced fast time variation, and the passive nature of RIS elements that precludes conventional pilot-based acquisition. This paper proposes a hybrid deep learning framework synergistically combining convolutional neural networks (CNN) for spatial feature extraction with bidirectional long short-term memory (BiLSTM) networks for temporal sequence modeling.
The architecture hierarchically decomposes estimation into: CNNs extracting multipath spatial patterns from canonical K-path representations, then BiLSTMs modeling temporal evolution across sequential snapshots, effectively capturing spatial-temporal coupling in RIS-UAV propagation.
We develop comprehensive methodology with DeepMIMO ray-tracing generation, K=10 path selection achieving >95% channel power capture, and systematic preprocessing. Extensive evaluation across SNR -10 to 30 dB demonstrates hybrid CNN-BiLSTM achieves NRMSE 0.018 at 30 dB (21.7% improvement over CNN, 33.3% over BiLSTM, 50-60% over LS/LMMSE/CS-OMP), correlation 0.989, SSIM 0.985, with 3.5M FLOPs and 2.0 ms inference on NVIDIA Tesla V100 enabling real-time operation within 5-10 ms UAV channel coherence time. This validates the hybrid approach as an enabling technology for next-generation 6G aerial communications requiring ultra-reliable, low-latency channel acquisition in highly dynamic three-dimensional environments.