Objective Underwater wireless optical communication ( UWOC) has a longer transmission distance and a higher data rate compared with underwater radio frequency communication and underwater acoustic communication. However, the absorption, scattering, and turbulence effects in the marine environment seriously affect the transmission quality of the optical signals, resulting in a limited transmission rate and an increased bit error rate (BER) of the UWOC system. Autoencoders can achieve end- to-end UWOC performance by using deep neural networks to jointly optimize the transmitter and receiver. However, as one of the most important data representation methods in autoencoders, the one-hot vector has a low data transmission rate. In order to solve these issues, in this paper, we propose an adaptive transmission scheme for underwater autoencoders based on deep neural networks on a joint channel that considers Gamma- Gamma turbulence and transmission path loss. This scheme can effectively suppress the impacts of underwater turbulence, absorption, and scattering on the performance of UWOC systems, improve the data rate of underwater autoencoders, and reduce the BER of the system. Methods In this paper, an adaptive transmission scheme for underwater autoencoders with mean square error (MSE) performance constraints was proposed by using the deep neural network. The UWOC channel model was established by using the path loss of the Beer- Lambert law and the probability density function of the Gamma-Gamma underwater turbulence distribution. By simulating the performance of the autoencoder's non-adaptive one- hot vector and comparing it with that of the adaptive transmission scheme under different UWOC channel conditions, the effects of different turbulence intensities, received signal-to-noise ratios (SNRs), and training parameter ensembles on the non- adaptive and adaptive transmission performance of the underwater autoencoder were discussed, respectively. Results and Discussions In this paper, an adaptive transmission scheme for underwater autoencoders is proposed to solve the problem of limited data rate caused by the one-hot vector of underwater autoencoders. The autoencoder is trained and tested under different ocean channels, as well as under different network training conditions, and the optimal transmission vectors are adaptively selected according to the set MSE performance constraints. Compared with nonadaptive transmission, the adaptive transmission scheme of the underwater autoencoder maximizes data transmission rate, reduces the BER, and improves communication performance (Fig. 5 and Fig. 7). At the same time, for different types of water bodies, instead of using a single training condition parameter, using a training parameter set for underwater autoencoders can obtain a more robust neural network model, making the autoencoder have a certain degree of generalization ability ( Fig. 8). Conclusions The adaptive transmission scheme for underwater deep autoencoders proposed in this paper can adaptively select the optimal vector for transmission according to the MSE constraints under different UWOC channel conditions, so as to maximize the data transmission rate. Under the joint influence of Gamma- Gamma turbulence and transmission path loss, the BER and data rate of the autoencoder using non-adaptive one-hot vector and adaptive transmission schemes are simulated and analyzed, respectively. The results show that the underwater autoencoder not only simplifies the system model but also has better BER performance compared with conventional communication systems. The autoencoder has different network loss performances under different training conditions, and the autoencoder trained by utilizing training parameter sets can obtain a more robust performance than that trained by utilizing a single training parameter. In addition, under the same training conditions, the BER and data rate of the adaptive transmission scheme adopted by autoencoders are better than those of the non-adaptive scheme. The proposed adaptive transmission scheme for underwater autoencoders provides a new approach to improving the performance of the UWOC system, and its feasibility has been verified through simulation.
