Deep Learning Based Symbol Detection in NLOS Ultraviolet Communication with Imperfect Channel State Information

Ultraviolet communication (UVC) is a promising alternative to conventional optical wireless communication (OWC) for non-line-of-sight (NLOS) scenarios, offering enhanced security and operation in an unlicensed spectrum. However, atmospheric turbulence and imperfections in channel state information (CSI) estimation introduce channel estimation errors (CEE), degrading system performance. Traditional maximum-likelihood (ML)-based detection methods are highly sensitive to CEE, limiting their reliability. This paper investigates deep learning (DL)-based symbol detection for UVC systems using on-off keying (OOK) modulation under weak turbulence and imperfect CSI conditions. A deep neural network (DNN) is trained to learn signal patterns and mitigate CEE without requiring explicit channel models. Simulation results demonstrate that while MLbased detection performs optimally under perfect CSI, the DLbased approach consistently outperforms it as CEE increases. Notably, for a severe CEE scenario where <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\rho=0.1$</tex>, the DL-based method achieves an SNR gain of approximately 9 dB at a target ASER of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">${10}^{-3}$</tex>.