Model-Driven Deep Learning-Based Estimation for Underwater Acoustic Channels With Uncertain Sparsity

Underwater acoustic (UWA) channels present diverse sparsity, which is challenging for UWA communications and underwater networked applications. Conventional methods fail to adapt to the uncertain sparsity due to inflexible adjustment rules or improper assumptions. This paper investigates the estimation of channels with uncertain sparsity and model-driven deep learning (DL) based estimation networks are proposed. We construct the channel estimation as sparse signal recovery model and exploit the channel sparsity based on the approximate message passing (AMP) method. Then the learned AMP network with soft-thresholding shrinkage function (ST-LAMP) is designed as a generalized estimator and suitable to random sparse UWA channels. Moreover, to achieve the Bayesian optimal solution, Gaussian mixture (GM) prior is imposed for UWA channels and the GM-LAMP network is formulated with the derived shrinkage function based on the minimum mean squared error criterion. The proposed networks incorporate the learning ability of DL methods by exploiting marine data and the domain knowledge from the classical model to realize parameter configuration and adaptive optimization. Experimental results verify that the proposed networks outperform benchmarks over UWA channels with diverse sparsity in terms of adaptability and effectiveness. Especially, real tests present the feasibility of proposed networks for on- line deployments.