Abstract
Underwater wireless optical communication (UWOC) has been widely studied as a key technology for high-speed underwater data transmission in ocean exploration and exploitation. As the demand for image and video transmission grows in such scenarios, the limited bandwidth of UWOC becomes a critical bottleneck, under which traditional separate source-channel coding suffers substantial efficiency loss when handling high-dimensional image sources. Underwater wireless optical semantic communication (UWOSC) has accordingly been explored, which employs joint source-channel coding (JSCC) to directly transmit source semantics and improve both transmission efficiency and robustness. However, the performance of existing UWOSC systems degrades significantly under oceanic turbulence, whose non-stationary and intensity-varying characteristics pose a major challenge to robust JSCC design. In this paper, we propose gated-JSCC, an enhanced vision-transformer-based JSCC (ViT-JSCC) architecture that suppresses turbulence-induced unreliable attention propagation through dual-end attention-output gating and a depthwise-convolution-enhanced feed-forward network (DWConv-FFN). The gating mechanism confines the influence of deeply faded tokens within the attention output, while the DWConv-FFN reinforces local spatial coherence under spatially correlated fading. Simulation results across a wide range of SNR, turbulence strength, and propagation-distance settings show that gated-JSCC achieves consistent PSNR gains of approximately 1.1–1.7 dB over the ViT-JSCC baseline. Experiments under the same symbol budget further show that the proposed scheme outperforms conventional separate source-channel coding by 6.1–14.4 dB under turbulence-induced bubble and temperature fading. These results demonstrate the effectiveness of dual-end attention-output gating together with DWConv-FFN for robust semantic transmission over turbulent underwater optical channels.