Multiscale context-aided neural detection for underwater optical NOMA under turbulence-induced SIC mismatch

Power-domain nonorthogonal multiple access (NOMA) is attractive for underwater optical wireless communication, but the near-user detector becomes sensitive to residual interference when the gain used for successive interference cancellation (SIC) does not track turbulence. This paper develops a multiscale context-aided neural detector (MCNet) for a two-user IM/DD NOMA link with Beer–Lambert attenuation, temporally correlated lognormal turbulence, and additive electrical Gaussian noise. MCNet uses the center received sample together with amplitude statistics computed over several nested windows, and it is not supplied with instantaneous channel gain, scintillation index, distance, water type, or temporal correlation. A context-matched block-energy SIC receiver is introduced as a fair conventional benchmark, alongside nominal-CSI SIC, perfect-CSI SIC/ML, a scalar MLP, and a plain 1-D CNN. At 8 m in clear-ocean water, 28 dB reference SNR, SI = 0.08, and log-amplitude correlation ρ = 0.95, three independently trained MCNet models give an average BER of (9.19 ± 0.19) × 10 −3 , compared with 3.45 × 10 −2 for nominal SIC, 1.12 × 10 −2 for block-energy SIC, and 4.31 × 10 −3 for perfect-CSI SIC. The neural advantage over block-energy SIC increases as the fading varies faster within the observation window, while conventional processing remains preferable in the nearly static or very weak-mismatch regimes.