Deep Alternating Direction Networks for UAV-RIS-assisted Channel Estimation

Reconfigurable intelligent surfaces (RISs) have garnered considerable attention for extending wireless coverage, including non-terrestrial networks. Accurate channel estimation is crucial to fully leverage RISs, while maintaining low complexity and pilot overhead. In this paper, we propose two model-driven deep neural networks for gridless estimation with low pilot overhead. The proposed deep neural network, termed DADU-Net, unfolds the iterations of the alternating direction method of multipliers, incorporating a spectral shift module to approximate optimization constraints. To adaptively manage layers based on convergence, we extend this approach with learnable fixed-point iterations, resulting in the DADF-Net. Simulation results demonstrate the effectiveness of the proposed methods.