Robust and efficient image transmission in power-constrained underwater visible light communication systems using neural architecture search Authors: Qiujun Jin, Bohua Deng, H Y Fu, Faisal Nadeem Khan Published: 2025 ABSTRACT AbstractUnderwater visible light communication (UVLC) is a method that uses visible light as a carrier to transmit information through an underwater channel and offers the key advantages of fast speed, low cost and large capacity. The main challenges faced by UVLC systems include the power constraints of underwater terminals and signal distortions due to light attenuation in complex and changing underwater environments. In order to improve the efficiency and robustness of image transmission process in UVLC systems, we propose a novel image compression and reconstruction solution that jointly optimizes coding complexity and image reconstruction quality using neural architecture search (NAS) mechanism. Experimental validations on an UVLC system demonstrate that the proposed scheme achieves increased transmission efficiency with reduced algorithmic complexity and offers robustness to errors caused by the underwater channel distortions. For bit error rate (BER) values as high as 1.58 × 10− 2, the peak signal-to-noise ratio (PSNR) remains above 20 dB while structure similarity index measure (SSIM) value is above 0.67, for a compression ratio (CR) of 0.33.Keywords: Underwater communication, Visible light communication, Internet of underwater things (IoUT), Autonomous underwater vehicles (AUVs), Machine learningSubject terms: Engineering, Mathematics and computing, Optics and photonics, Physics Introduction In the past decade or so, underwater wireless communication networks have undergone rapid development. Underwater communication methods usually include acoustic and radio frequency (RF) communication. However, traditional acoustic communication has limitations in underwater scenarios due to its low data transmission rate, slow propagation of acoustic waves underwater, and the typically large size, high cost, and high energy consumption of acoustic transceivers. On the other hand, traditional RF communication has a limited transmission distance of only a few meters at ultra-low frequencies due to the conductive nature of seawater. Additionally, it requires expensive transmitter antennas and transceivers1–3. Underwater visible light communication (UVLC) is a communication technology that uses visible light for underwater data transmission. In this scenario, water, the medium through which the light waves travel, absorbs the light and causes it to attenuate. In addition, the suspended and dissolved substances in water cause light scattering, making light diffusely reflective in water, thus limiting the transmission distance and data rate. Especially, in complex underwater environments, such as the ocean, the water quality and topography bring great challenges to UVLC. Moreover, underwater turbulences further complicate the propagation of light, leading to additional signal degradation and variability in the transmission path. As shown in Fig. 1, UVLC can be used to build a long-range, high-speed underwater data link between platforms such as autonomous underwater vehicles (AUVs), underwater wireless sensors, and surface vessels and buoys. In such scenarios, both transmitter and receiver are located in an underwater environment. At the transmitter, data encoding is first required to convert the data to be transmitted into a digital signal and then modulate it over an optical carrier of visible light using certain modulation techniques. A suitable light source, such as a laser diode (LD), is often used to generate a visible light carrier signal. Typically, red light, which has a longer wavelength and thus lower scattering and attenuation in water, is chosen for enhanced performance over longer distances4. The modulated optical signal can be generated by changing the intensity, frequency or phase of the visible light carrier. Optics such as beam shapers or optical lenses can be used to focus the resulting light signal into a beam to increase the directionality and energy focus of the transmitted signal. At the receiver, a photodetector (PD) first converts the visible light signals, transmitted through the water, into corresponding electrical signals. Subsequently, the signals undergo processing operations such as amplification, filtering, and demodulation to extract the original transmitted data. An example UVLC scenario in an ocean. Compared to traditional acoustic and RF communications, UVLC has gradually received a great deal of attention since it can achieve higher transmission rates and longer transmission distances while also providing illumination. Therefore, it is expected to become a key technology for real-time, high-speed and wireless transmission of images and videos in underwater communication scenarios. However, in an UVLC system, the propagation of light beams underwater is susceptible to water molecules’ absorption and scattering, which in turn affect the quality of communication. Therefore, an UVLC system faces the following major challenges in image transmission scenarios. (i) The attenuation of light in water is high and in addition, the optical link is vulnerable to water waves, turbulences, suspended solids in water, air bubbles and other effects5. As a result, the signal distortion becomes a major problem in image transmission process. (ii) The UVLC systems have power constraints as their end devices are deployed underwater and because of that, it is not easy to recharge the power batteries too frequently, which imposes a limit on power consumption. Therefore, reducing the actual amount of data to transmit the images is important for transmission efficiency and thus power saving purpose. In recent years, the research on internet of underwater things (IoUT) and underwater wireless optical communication (UWOC) has gradually evolved from single-technology optimizations to more systematic designs. For instance, Zayed et al.6 proposed a lightweight and efficient convolutional neural network (CNN) model for modulation scheme recognition in UWOC. Similarly, the authors in7 proposed an underwater optical multiple-input multiple-output (MIMO) method, leveraging spatial multiplexing and parallel transmission to significantly enhance link capacity, thus enabling high-speed data transmission in IoUT. On the other hand, the authors in8 explored multihop UWOC links with various routing protocols to overcome severe attenuation and limited transmission distance of optical signals in water. Similarly, Ali et al.9 proposed a green Li-Fi communication framework for remotely operated vehicles (ROVs), in which energy optimization and renewable energy integration are used for enabling sustainable long-term operations underwater. Furthermore, in order to reduce the effect of signal distortions or to increase the transmission efficiency in UVLC systems, various researches have been conducted. For example, Zhao et al.10 proposed a novel channel emulator based on a two-tributaries heterogeneous neural network (TTHnet) which can mitigate the effects of attenuation and is suitable for estimating both single-carrier and multi-carrier modulated channels in UVLC systems. Cai et al.11 proposed a post-equalization method based on bi-directional gated recurrent units (BGRUs) that uses structural amplitude phase-shift keying to mitigate the distortion problem of UVLC systems, however, this method consumes significant power. Shen et al.12 proposed a high-speed and robust UVLC technique that utilizes an adaptive diversity combining technique with single input and multiple outputs to improve the transmission efficiency. To combat underwater interference, maximum ratio combining method is used at the receiver end while a multilayer perceptron (MLP) neural network is utilized to mitigate the nonlinear effects. However, despite the high complexity of the neural network, the distortions are only partially mitigated. Jiang et al.13 investigated an asymmetric clipped optical orthogonal frequency-division multiplexing and non-orthogonal multiple access (ACO-OFDM-NOMA) scheme for uplink UVLC systems to achieve energy saving and support massive device connectivity. The existing works usually attempt to increase the transmission efficiency by employing advanced modulation/demodulation techniques. However, such an approach would significantly increase the implementation complexity of the end devices. On the other hand, to reduce the channel distortion effects, the existing works typically use methods such as channel estimation (e.g., by sending extra pilot symbols) and/or conventional channel coding schemes (requiring additional coding overhead). However, both these approaches would result in increasing the power consumption of the end devices as well as reducing the transmission efficiency. For image transmission, a classical way to increase the transmission efficiency is to reduce the amount of data by image encoding or compression. Various approaches, such as joint photographic experts group (JPEG) standard14, JPEG2000, and other transforms-based compression techniques have been widely employed for efficient storage and transmission of digital images. These methods typically involve a combination of transform coding, quantization, and entropy coding to achieve high compression ratios (CRs) while preserving visual quality. Moreover, image compression and denoising techniques based on wavelet transform (WT) have been proposed in15,16. Besides, various image compression solutions based on machine learning (ML) have been investigated. Hinton et al.17 first showed image dimensionality reduction using autoencoders (AEs). In this context, variational AE (VAE)18, a variant of AE, further achieved compression of hyperspectral images by introducing hidden variables in AE and generating different images by controlling the values of those hidden variables. A more advanced deep learning (DL) model called CNN can extract spatial and spectral information from a hyperspectral image through convolutional and pooling layers. The extracted features are then mapped into a low-dimensional coding space through a fully-connected layer. Finally, the coding can be remapped back to the original hyperspectral image using an inverse convolutional layer19. Theis et al.20 described the use of spatial long short-term memory (LSTM) model for image compression. Cheng et al.21 used convolutional autoencoder (CAE) for lossy image compression. While some of the above-mentioned DL algorithms have demonstrated remarkable performance for image compression, their direct application in UVLC systems faces significant challenges, primarily due to their prohibitive algorithmic complexity. This is because these DL models typically constitute numerous layers with millions of parameters, requiring intensive computational resources for the encoding process, which will significantly increase the power consumption of the transmitters. This inherent complexity barrier makes such DL models impractical for use in UVLC systems, necessitating the development of lightweight, highly efficient neural network architectures specifically designed to operate within the stringent power and computational resources constraints of underwater environments. Another crucial aspect to highlight is that all the previous works optimize CR and reconstruction quality while not considering the effects of communication channel errors on the quality of reconstructed images. Because of this shortcoming, these methods are not quite suitable for use in practical UVLC systems that have certain bit-error rates (BERs). To address the challenges of power constraints and unreliable underwater channel while transmitting images in UVLC systems, in this work, we propose a novel image compression and reconstruction solution based on neural architecture search (NAS)22. The proposed method achieves comprehensive performance in terms of coding complexity and images reconstruction quality despite bit errors caused by underwater channel. Overall, our solution offers the following key advantages: (i)Higher data transmission efficiency and lower power consumption: Our solution uses a lightweight image encoder to compress images, which can efficiently reduce the amount of transmitted data while requiring lower algorithmic complexity than other conventional compression techniques. As a result, the data transmission efficiency increases and the total power consumption by the end devices decreases. Higher data transmission efficiency and lower power consumption: Our solution uses a lightweight image encoder to compress images, which can efficiently reduce the amount of transmitted data while requiring lower algorithmic complexity than other conventional compression techniques. As a result, the data transmission efficiency increases and the total power consumption by the end devices decreases. (ii)Errors robustness: For the bit errors introduced during data transmission in complex underwater environment, our approach has a reasonable degree of robustness, which means the effects of channel distortion on reconstructed image quality are principally reduced. Errors robustness: For the bit errors introduced during data transmission in complex underwater environment, our approach has a reasonable degree of robustness, which means the effects of channel distortion on reconstructed image quality are principally reduced. (iii)Optimal performance through joint optimization of image compression and reconstruction processes: The conventional approach of separately designing and optimizing individual UVLC system blocks in a modular fashion may lead to suboptimal system performance. In contrast, the proposed solution optimizes image compression and reconstruction processes jointly such that the coding complexity and the difference between original images and the reconstructed images are jointly minimized, thus offering clear performance advantage. Optimal performance through joint optimization of image compression and reconstruction processes: The conventional approach of separately designing and optimizing individual UVLC system blocks in a modular fashion may lead to suboptimal system performance. In contrast, the proposed solution optimizes image compression and reconstruction processes jointly such that the coding complexity and the difference between original images and the reconstructed images are jointly minimized, thus offering clear performance advantage. Operating principle System model Our proposed image transmission system model is shown in Fig. 2, comprising a transmitter, a visible light transmission channel, and a receiver in an underwater environment. In this UVLC system model, we employ a codec for image compression and reconstruction tasks, where the image encoder is deployed in the transmitter while the image decoder can be deployed either in the receiving end device in UVLC system or in the cloud. It is worth mentioning here that deploying the decoder in the cloud could be a preferable choice since the analysis of received images is often performed in the cloud itself. In addition, there will be no strict restrictions on the complexity of the decoder in this case, thereby yielding better decoding performance. Inside the transmitter, image encoder, quantizer, binary encoder and modulator blocks are used to compress the input image and to generate a suitable light signal for image data transmission over underwater channel. The corresponding symmetric demodulator and binary decoder blocks are employed inside the receiver, followed by the image decoder. As Fig. 2 shows, the captured input images are downscaled into feature maps of smaller size by the image encoder at the transmitter. Then, the feature maps are quantized and encoded into bits by quantizer and binary encoder, respectively. After that, bits are converted into light signals by the modulator and are finally transmitted. During the transmission process in UVLC systems, signals will be affected by complex underwater environment, resulting in many erroneous bits at the receiver after demodulation. Afterwards, a binary decoder converts the bitstream back into feature maps and our proposed image decoder then reconstructs images from those feature maps. Schematic diagram of system model with codec blocks shown in green color. Design of codec model architecture through NAS In order to design a comprehensive image compression and reconstruction codec model for multi-level CRs, we use NAS approach for searching an optimal architecture in terms of encoding complexity and images reconstruction quality. The NAS process in our work involves several intricate steps to systematically explore and identify an optimal neural network architecture as shown in Fig. 3. Initially, a comprehensive search space is defined, encompassing numerous possible configurations of network architectures. Once the search space is established, a search strategy is selected to navigate through this vast landscape. In our work, an evolutionary algorithm is chosen to simulate natural selection by iteratively selecting and recombining promising architectures to produce new generations. Following the generation of candidate architectures, each one must be rigorously evaluated for its performance. This process typically involves training an architecture using fewer epochs to get a proxy model and measuring model performance by fitness score. Once the performance evaluations are complete, a few most promising architectures are identified based on their performance metrics, which are then further trained and rigorously tested to confirm their robustness and generalizability. The NAS process can be mathematically modeled as an optimization problem23, as shown by Eq. (1) Flowchart of the complete NAS process employed in our method. where denotes the search space of the neural architectures, denotes the measurement fitness score,means the training dataset and means the evaluation dataset. A candidate architecture A is trained on a training dataset and evaluated on an evaluation dataset to get the fitness score that is being minimized during the NAS process. Search space design In order to improve the transmission efficiency and achieve robust transmission of images in UVLC systems, the images encoder needs to have the capabilities of features extraction and dimensionality reduction while the decoder needs to reconstruct images robustly despite the bit errors caused by the UVLC channel. Considering that, we modularize the search space of NAS with three functional blocks as shown in Fig. 4. For the encoder, we use a basic module derived from a CNN model, which contains a convolutional layer and a max-pooling layer connected by an activation function. However, in order to reduce the computational complexity of the basic module, we replace the classical multi-channel convolution with depthwise separable convolution (DSC). Such a module, called DSC block, has excellent performance in image features extraction and dimensionality reduction while significantly decreasing the computational complexity24. In the DSC block, as shown in Fig. 4 (a), the number of output channels of the pointwise convolutional layer is an important parameter that determines the ability of image features extraction. The stacking of multiple DSC blocks can likewise achieve the purpose of features extraction and data dimensionality reduction with a controllable CR. Therefore, the number of DSC blocks stacked in the encoder and the selection of the parameters of the pointwise convolutional layer in each DSC block are factors that need to be determined by NAS. In detail, DSC is an efficient convolution operation that decomposes the standard convolution operation into two simpler steps, depthwise convolution and pointwise convolution, as shown in Fig. 5, thereby reducing the computation complexity and the number of parameters while maintaining the expressiveness of the model as much as possible. This method has been widely applied in architectures such as MobileNet to accommodate real-time processing requirements on mobile and embedded devices25. In depthwise convolution, each input channel undergoes convolution operation independently with its own kernel to extract spatial features. On the other hand, in pointwise convolution, a 1 × 1 convolution kernel is used to combine these spatial features and adjust the number of output channels, thus realizing cross-channel information interaction and dimensional transformation. Instead of standard convolution’s multiply accumulate (MAC) operations of , the total MAC operations of DSC are reduced to 24, where and mean the number of input channels and output channels, respectively, H and W denote the height and width of input image, respectively, and K denotes the convolutional kernel size. This method significantly decreases the number of model parameters and computations, making it more suitable for deployment in devices with limited computational resources such as those used in practical UVLC systems. In addition to convolution, a rectified linear unit (ReLU) is used in a DSC block as well as a max-pooling layer to reduce the spatial size of the feature map while retaining important features. Overall, the above-mentioned operations in DSC block constitute the important process of image features extraction and the dimensionality reduction with a limited computational complexity, making the DSC block a suitable candidate module for a lightweight image encoder. Architectures of the basic modules included in the defined search space. (a) DSC block for encoder, (b) residual block for decoder , and (c) deconvolutional block for decoder. Operating principle of DSC. For the decoder, in order to achieve image reconstruction, a deconvolution step symmetrical to the encoder is required. Hence, a deconvolutional block, comprising a transpose convolutional layer and an activation function, is the basic module for images reconstruction, and the way the deconvolutional blocks are stacked is then determined by the searched encoder architecture. In addition, to enhance the robustness of the decoder to the bit errors introduced by the UVLC channel, we use classical residual blocks that are stacked at the beginning of the decoder to learn the distribution of the distortions introduced into the compressed representations, where the optimal number of stacked residual blocks needs to be determined by NAS. Therefore, the basic modules in the search space for decoder include residual and deconvolutional blocks, as illustrated in Fig. 4 (b) and (c), respectively. These components serve distinct purposes, i.e., the residual block enhances error robustness while the deconvolutional block facilitates reconstruction of images to their original size. Residual blocks are particularly effective for robustness to bit errors because they leverage skip connections to fit a residual mapping, which is the core concept of residual learning26. For image reconstruction, the deconvolutional block acts as the inverse operation of the DSC block, enabling the up-sampling of feature maps and the images reconstruction from low-dimensional representations. After the designing of basic modules in the search space, we define the initial population of codec architecture configurations in the search space as shown in Table 1. For each CR, we define three lists (shown in second column of Table 1) to represent different configurations of DSC blocks forming the encoder component of initial population of candidate architectures. Here, the size of the list indicates the number of DSC blocks stacked in the encoder, and each number in the list indicates the number of output channels of the pointwise convolutional layer in the corresponding DSC block. For example, the list [8,16] means that two DSC blocks are stacked in the encoder, where the first DSC block has 8 output channels and the second DSC block has 16. The number of input channels in the second DSC block are the same as the number of output channels in the first DSC block. During the process of defining the initial search space, the number of DSC blocks and the number of output channels in each DSC block are initialized empirically. For each CR, the number of DSC blocks is considered from 1 to 3, which gives an initial restriction on the encoding complexity. The number of output channels of the last DSC block is defined according to the corresponding CR and for the other DSC blocks, it follows the empirical design patterns for convolutional layers, gradually increasing with network depth. The mapping relation between CR and encoder architecture is given as Table 1Initial population of codec architecture configurations in the search space for different CRs considered.CRConsidered DSC blocks configurations for the candidate architecturesConsidered number of residual blocks for the candidate architectures0.83[10], [20, 40], [40, 80, 160]1, 2, 3, 4, 50.66[8], [16, 32], [32, 64, 128]1, 2, 3, 4, 50.50[6], [12, 24], [24, 48, 96]1, 2, 3, 4, 50.33[4], [8, 16], [16, 32, 64]1, 2, 3, 4, 5 Initial population of codec architecture configurations in the search space for different CRs considered. where, denotes the number of output channels of the last DSC block and n denotes the number of DSC blocks. Apart from the fixed number of output channels of the last DSC block, the intermediate DSC block’s output channel numbers are selected from integer multiples of 4 up to 128 (i.e., 4, 8, 12, 16, …, 128) and all DSC layers use a fixed stride of 1, kernel size of 3 3, and a ReLU activation function. In addition, for each CR, we consider five different number of residual blocks for the decoder component of initial population of candidate architectures. Adopted evolutionary NAS strategy After the establishment of above-mentioned search space, we adopt an evolutionary NAS strategy to navigate through it. During this process, random genetic cross-overs, with a cross-over probability of 0.8, generate new lists of channels, representing new architectures of the encoder. As shown in Fig. 6 (a), the evolutionary NAS process starts with a provided initial population, taking the two architectures with the highest fitness scores from K candidates (where K is the tournament size) as parents and combining parts of the structures of the two parent architectures to generate new child architectures that are then added into the new population and evaluated again23. The traditional evolutionary NAS also includes a random mutation process for the parent architectures. However, in our image compression and reconstruction task, in order to keep the encoder CR controllable, we added an architecture examination procedure after the random mutation and genetic cross-over process, during the evolution of the population, so as to avoid the generation of architectures that do not meet the CR requirements. An example evolution process is shown in Fig. 6 (b), where the two selected parent architectures with different configurations of DSC blocks perform genetic cross-over and generate a child architecture with a new configuration of DSC blocks. More specifically, the genetic cross-over happens in the first crossover point of the configurations of parents. Parent A with configuration of [8,16] contributes its second number of output channels and parent B with configuration of [4] contributes its first number of channels to the child architecture, which gets a configuration of [4,16] as a result. In our evolutionary NAS process, the maximum population size is set to 30, the maximum number of generations is 20, and the mutation rate is 0.2. Parent architectures are selected through tournament selection, where the tournament size K = 3 specifies the number of candidate architectures randomly sampled from the population to compete in each selection round. The two best-performing architectures among these K candidates are then chosen as parents. A larger tournament size increases the selection pressure on lower-performing architectures by giving higher-performing architectures a greater advantage in being selected. In addition, if the best validation performance does not improve for 5 consecutive generations,the search process terminates early. Such an evolutionary process allows to select top-performing architectures from the initial population and generate new possibilities, helping to automatically find an optimal architecture for each CR. (a) The flowchart and (b) an example evolution operation of adopted evolutionary NAS algorithm. To evaluate the comprehensive performance of the architectures in terms of coding complexity and images reconstruction quality in UVLC systems, we devise a fitness score, given by Eq. (3) where, the MSE denotes the mean squared error between the original and reconstructed images I and K, respectively. MSE is calculated as the average of the squared differences between corresponding pixels, as shown by Eq. (4), where H and W denote the height and width of images, respectively. The second term in Eq. (3) represents the complexity penalty, where NPE denotes the total number of parameters in the candidate encoder architecture, which can be calculated from the proxy model after partial training process. NPE is a classic metric to evaluate the computational complexity of ML models. Furthermore, denotes the number of parameters of the most complex encoder architecture in the initial search space and is used as a reference. Similarly, denotes the number of parameters of the encoder being evaluated while is a scaling factor defined to scale the complexity penalty term to an order of magnitude similar to MSE. After multiple rounds of evolution and evaluation processes, we found the optimal codec architectures for each of the four considered CR levels of 0.83, 0.66, 0.5, and 0.33, as shown in Fig. 7. It can be seen from the figure that the more complex encoder architectures present in the initial search space, consisting of three stacked DSC blocks, are not selected due to the constraint on encoder complexity imposed by the search process. Instead, most of the optimal architectures originate from the single DSC block architecture in the initial population, and only for a CR as low as 0.33, where a single DSC block is not sufficient to learn the deep features of the images, the number of DSC blocks in the optimal architecture increases to two. Furthermore, the optimal architecture in this case is a child architecture generated in the evolutionary process. Hence, the optimal architectures searched in this way reduce the complexity of the encoder as much as possible while retaining the required quality of the reconstructed image. The selected best architectures for four considered CR levels, i.e., (a) CR = 0.83, (b) CR = 0.66, (c) CR = 0.50, and (d) CR = 0.33. Note that the latent representation generated by the encoder will encounter distortions during transmission in a UVLC system. Therefore, in order to improve the robustness of the decoder to these distortions, we include a bit errors introduction process during both partial and complete training processes of our codec models to emulate the UVLC channel conditions, which will be discussed in detail in next section. In short, our evolutionary NAS process can be summarized as follows: an initial search space is first defined (based on domain knowledge and UVLC system constraints) that includes a number of possible candidate architectures. During the search process, the algorithm evaluates the candidate architectures in terms of both coding complexity and image reconstruction performance based on the designed fitness score. Next, using the genetic inheritance theory, it selects two best-performing architectures to combine and generate a new child architecture, completing the population updating process once. After several iterations, a number of newly-created architectures are evaluated and the one with the highest fitness score is selected as the optimal architecture. This process results in greatly expanding the initial search space and eventually determining the best possible architecture in a finite search time span. We would like to emphasize here that the architectures searched using our NAS mechanism comprehensively address the above-mentioned three challenges faced by the real-world UVLC systems. Firstly, the constraint on the complexity of the encoder during the search process enables a lightweight image encoder, improving the transmission efficiency, reducing the energy consumption, and facilitating easy deployment in underwater transmitter devices. Secondly, the bit errors introduced during the training process ensure that the chosen architecture is robust to errors in UVLC channels and has the capability to reconstruct images with good quality in unreliable UVLC systems. Thirdly, in contrast with conventional modular design approach, our search process jointly optimizes coding complexity and channel errors robustness, resulting in a searched architecture that offers comprehensive optimal performance in UVLC systems. It is also worth mentioning that in our evolutionary NAS process, each model requires an approximately 54 min of training and evaluation, including the overhead from emulation of bit errors caused by UVLC channel. Leveraging two NVIDIA RTX A6000 GPUs, each generation takes about 7 h, leading to an estimated total runtime of 6 days. The entire NAS pipeline is computationally feasible within one-week timeframe on moderate hardware. It shall be noted that the above-mentioned NAS process must be performed offline and prior to codec deployment in an actual UVLC system, therefore, the computation time/complexity of the NAS process does not matter much in this case. Hence, our proposed approach is practically viable for real-world deployment and research prototyping. Codec model training and testing After optimal architecture selection for each CR, a complete training process is needed for the codec models. The images dataset employed in our study is a subset of Enhancing Underwater Visual Perception (EUVP) dataset27, comprising 5500 underwater RGB images, each with a dimension of 256 × 256 pixels, out of which 5000 images were used for training and validation processes while the remaining 500 images were employed for testing purpose. The reason we chose this dataset to verify the codec models’ effectiveness in reconstructing images is that it comprehensively covers various underwater ecological environments and creatures. As shown in Fig. 8, to emulate the channel errors in UVLC systems, we included an error introduction process in both partial training process during NAS and the complete training process of the optimal architectures. In detail, a latent representation generated by the encoder is first denormalized, which maps the range of values to 0−255 and then 8-bit quantization is performed and the corresponding bitstream is generated. Next, bit errors corresponding to a given signal-to-noise ratio (SNR) are introduced into the bitstream and the SNR-to-BER mapping is given by Eq. (5)28. Illustration of the codec model learning process. Channel modelling for realistic UVLC systems is a complex task. Recent works on UVLC channel measurements have investigated turbulence-induced scintillation, colored noise, and channel memory effects29,30. In29, the authors performed controlled-tank experiments with temperature/salinity gradients and air bubbles, and they ascertained slow fading nature of UVLC turbulent channels. Similarly, in30, the authors studied various UVLC channel characteristics including scattering, absorption, turbulence-induced fading, channel’s spatial and temporal spread, etc., and derived closed-form expressions for system BER based on additive white Gaussian noise (AWGN) model, assuming the incoming optical signal to be a deterministic quantity while various noise components as an aggregated additive Gaussian noise. In existing UVLC-related research works31,32, the BER is typically considered in the range of to . The above-mentioned works provided references for selecting channel model and considered BER range in our study. Hence, we used an AWGN channel model to simulate the UVLC channel for the sake of simplicity since accurate channel modelling is difficult to achieve in practical UVLC systems. Moreover, using more complex channel models will greatly slow down the codec model training process. Therefore, we employed an AWGN channel model in this case and relied on the robustness and generalization ability of our ML model to mitigate the effects of other types of channel degradations as well. The results for an experimental UVLC system (discussed in next section) will further validate such an ability of our NAS-based codec models. After detection, the received bitstream with errors is converted back to feature map in the form of a tensor. Then, the decoder is used to reconstruct the image from the feature map. After reconstruction, the loss function between the reconstructed image and the original image is calculated and the weights of the codec model are updated according to the gradient of loss. The adopted learning process is described in Algorithm 1, where θ denotes the parameters of the codec model, denotes the 8-bit quantization, DQ means the binary decoding process, N means normalization and DN means denormalization. During the model training process, the batch size is set to 32 and the optimizer is Adam optimizer with a learning rate of 0.001. The training and validation loss curves are shown in Fig. 9. It can be seen from the figure that the loss function curves gradually converge, and the validation loss is not significantly higher than the training loss, indicating that there is no overfitting problem. Pseudo-code for the training process. Training and validation loss vs. number of epochs during the codec model training process. In order to demonstrate the feasibility of our codec models in UVLC systems, we evaluated the models from the perspectives of coding complexity and images reconstruction performance for different BERs and compared them with those of commonly-used JPEG standard, which is widely deployed in the edge devices of wireless communication systems offering excellent compression and reconstruction performance with much lower coding complexity than most existing ML-based image compression algorithms. Firstly, to compare the encoding complexity of the developed neural network architectures and the JPEG algorithm, we used number of MAC operations and encoding time on the same computing platform as two evaluation metrics. Secondly, to evaluate the image reconstruction quality, we employed two classical metrics, i.e., peak signal-to-noise ratio (PSNR) and structure similarity index measure (SSIM). Figure 10 shows the encoding time distribution of our codec models for different CRs. Note that due to complexity of modern computer systems and their resource management mechanisms, the inference times may fluctuate a bit even for the identical algorithms, data, and computing platforms. However, the algorithm complexity remains the primary factor influencing the inference time. In our case, we tested the inference time for each model architecture five times and the results are presented in Fig. 10. The sample points shown in the lower right corner are the cases where there is only one DSC block used in the encoder while the sample points in the upper left corner are the cases where the encoder contains two DSC blocks. In the case of one DSC block, the number of output channels of the pointwise convolutional layer gradually increases as the CR increases, leading to a rise in number of MAC operations along with an increase in encoding time. On the other hand, when the number of DSC blocks increases to 2 for a CR of 0.33, the number of MAC operations of the encoder increases significantly, resulting in longer encoding time. Encoding complexity (measured in terms of inference time) distribution of the proposed codec models for different CRs. The results of encoding complexity comparison with several other image compression algorithms, in a scenario where there are no channel errors introduced, are shown in Table 2. It is worth mentioning here that entropy coding is an important part of JPEG and JPEG2000 compression algorithms, whose computation complexity mainly comes from bit manipulation and logical judgment33, which cannot be measured in terms of multiplication and addition operations. Therefore, the actual coding complexities of JPEG and JPEG2000 algorithms are much higher than the ones defined in terms of number of MAC operations. However, as can be seen from Table 2, despite disregarding the computation complexity of the entropy coding part, our encoder model requires fewer number of MAC operations than JPEG and other image compression algorithms while the inference time on the same CPU is much lower than them for a similar CR level, which clearly demonstrates that our image encoders are fully deployable in underwater transmitter devices with power constraints. Due to excessively-high computation complexity of AE and CNN, they remain impractical for power-constrained UVLC systems despite their superior image reconstruction capabilities. Therefore, we selected JPEG and JPEG2000 as the primary baseline algorithms for further comparative analysis at different BERs. It can also be seen from the last column of Table 2 that the images reconstruction performance, measured in terms of PSNR, of the proposed codec model is comparable to that of JPEG for a similar CR. Table 2Performance comparison of proposed and other approaches without channel bit errors.EncoderCRNumber of MAC operationsEncoding time (s)PSNR (dB)10.83 27.5320.66 26.8830.50 25.2240.33 23.58JPEG0.55 1.7226.07JPEG20000.51 2.4328.92AE with convolutional layers0.50 29.03CNN0.50 29.87 Performance comparison of proposed and other approaches without channel bit errors. Next, we test and compare the images reconstruction performance of our codec models with JPEG and JPEG 2000 algorithms, when there are bit errors introduced, and the evaluation results in terms of PSNR and SSIM metrics are shown in Figs. 11 and 12, respectively. The BER is set to 5 different values in a wide range, i.e., , , , , and . For better visualization, BER is expressed in logarithmic form of , which is abbreviated later in the text as simply . It can be seen from Figs. 11 and 12 that as the BER increases, the JPEG algorithm is vulnerable to bit errors and the quality of the reconstructed image decreases dramatically, whereas our models show good robustness to bit errors for all CRs considered. Even for the extreme case of log(BER) = −2, the PSNR can still be maintained between 21 and 25 dB for our codec models. In contrast, a high BER has a devastating effect on the JPEG algorithm, leading to its complete inability to decode images. It is clear from Fig. 11 that the reconstructed image quality, in terms of PSNR, using JPEG algorithm decreases dramatically as the BER increases and when the log(BER) reaches − 3.3, the image reconstruction quality of the JPEG with a CR of 0.55 is lower than that of our model with a comparable CR of 0.50. The SSIM curve for JPEG algorithm in Fig. 12 shows a similar trend with BER and the image reconstruction quality of the JPEG is lower than that of our codec model with a comparable CR of 0.50 when the log(BER) reaches − 3. It can also be observed from Fig. 11 that JPEG2000 offers significantly improved robustness to bit errors compared to the conventional JPEG algorithm and performs well under relatively low BERs. However, when the log(BER) reaches − 2, the image reconstruction quality, in terms of PSNR, of JPEG2000 with a CR of 0.51 is lower than that of our model with a comparable CR of 0.50. Similarly, it can be seen from Fig. 12 that the image reconstruction quality, in terms of SSIM, of the JPEG2000 is lower than that of our codec model when the log(BER) reaches − 2.3, for a comparable CR. More importantly, as clear from Table 2, the algorithmic complexity of JPEG2000 (despite neglecting the complexity of entropy coding part) is approximately twice that of our proposed method for a comparable CR. Thus, it can be concluded from Figs. 11 and 12, and Table 2 that in the absence of any bit errors as well as for relatively low BERs, the JPEG and JPEG2000 have some advantage in image reconstruction quality but their vulnerability under high BERs and their significantly higher computational complexity compared to our codec models limit their effectiveness for deployment in practical UVLC systems. PSNR as a function of log(BER) for different CRs using the proposed codec models, JPEG, and JPEG2000 algorithms. SSIM as a function of log(BER) for different CRs using the proposed codec models, JPEG, and JPEG2000 algorithms. Hence, our image compression and transmission solution demonstrates the ability to increase the transmission efficiency and reduce the power consumption of the end devices by compressing images with low computational complexity, and it solves the unreliable image transmission problem encountered in UVLC systems by reconstructing images robustly even when the BERs are extremely high. Experimental analysis and results To validate the feasibility of the proposed codec models and evaluate their performance in an actual UVLC scenario, experiments were conducted. The corresponding experimental setup and the offline digital signal processing (DSP) chain are shown in Fig. 13. In the experiments, the encoder compressed 20 sample images and transformed them into corresponding bits sequence. Standard orthogonal frequency-division multiplexing (OFDM) technique was employed, with the sequence of bits first mapped to quadrature amplitude modulation (QAM) symbols. These symbols in series were transformed into parallel streams via serial-to-parallel (S/P) conversion and organized with Hermitian symmetry to facilitate the inverse fast Fourier transform (IFFT). Following the IFFT operation, a cyclic prefix (CP) was added to each frame to reduce inter-symbol interference (ISI). This was succeeded by converting the signals from parallel to serial (P/S) to synthesize the OFDM signal. The OFDM signal was configured with a total of 512 subcarriers, of which 120 data subcarriers were actively modulated using 16-QAM. The selection of 16-QAM modulation format in our UVLC system is a strategic compromise between achievable data rate and robustness to channel conditions. The baud rate used in our experiments was 3.75 GBaud. Subsequently, an arbitrary waveform generator (AWG, Tektronix AWG7101) with a sampling rate of 4.0 GSamples/s was utilized in our system. It translated this signal into an analog waveform, which modulated a red laser light source (THORLABS, HL6544FM). The emitted light beams were then collimated using an aspherical lens and transmitted through a 1 m long water tank as shown in Fig. 13. At the receiving end, optical signals were focused on a specially-designed PD module (APD 210, Menlo System) using an aspherical lens. The detected electrical signal was then recorded by a digital storage oscilloscope (DSO, Tektronix MSO73304DX) and subjected to standard DSP techniques, such as synchronization, removal of the CP, applying the fast Fourier transform (FFT), channel equalization, etc. Subsequently, the detected QAM symbols were de-mapped into a bitstream for evaluating the BER performance. Finally, the bits sequence was transformed into the required input format of the decoder for the purpose of images reconstruction. Experimental setup of UVLC system. By adjusting the LD-APD alignment, bits sequence data with four different BER values was collected. Subsequently, the image reconstruction quality was measured using PSNR and SSIM metrics. The results shown in Fig. 14 and Fig. 15 illustrate the PSNR and SSIM values, respectively, obtained for each BER value for four different CRs. The experimental results curves demonstrate similar trend with the simulation results, i.e., the image reconstruction quality shows reasonable robustness to different BERs. Even when the log(BER) reaches a very high value of −1.8, the PSNR remains above 20 dB for a CR of 0.33. PSNR as a function of log(BER) for different CRs using the proposed codec models in an experimental UVLC system. SSIM as a function of log(BER) for different CRs using the proposed codec models in an experimental UVLC system. Finally, Fig. 16 shows some sample reconstructed images by the proposed codec models in our experimental UVLC system and it is evident from the figure that the obtained images have a fair quality from a visual perspective. Hence, our experimental results for an actual UVLC system clearly show that the proposed NAS-based codec models offer significant advantages in terms of algorithmic simplicity and transmission errors robustness, enabling good PSNR values even at very high BERs. These findings suggest that the proposed codec models can be a valuable tool for enhancing the efficiency and reliability of image transmission in challenging underwater environments, thus paving the way for more robust and energy-efficient future UVLC systems. Examples of the reconstructed images using the proposed codec models for different log(BER) values and CRs in an experimental UVLC system. Conclusion In this paper, we proposed and experimentally demonstrated a novel NAS-based image transmission method in UVLC systems that not only increases the transmission efficiency by compressing the images but also shows robustness to channel errors in complex underwater communication scenarios. Our lightweight image encoder can be employed inside the transmitters of UVLC systems, which typically have limited power and computational resources. The proposed solution offers great potential for energy-efficient and reliable image transmission in future UVLC systems. In the future, we will explore extending our work to enhance video compression for underwater transmission environments. Moreover, we will consider embedded systems implementation of our NAS-based codec models so as to evaluate their performance in real-time communication scenarios as well as to demonstrate compatibility with existing communication infrastructure. Author contributions Study design: F.N.K., and Q.J.; Experiments: Q.J., and B.D.; Data analysis and algorithms development: Q.J.; Writing of the manuscript: Q.J., F.N.K., and H.Y.F.; Approval: All authors. Funding This work was supported by Guangdong Innovation and Entrepreneurship Team Fund under Project No. 2023ZT10X137; Scientific Research Startup Fund of Tsinghua Shenzhen International Graduate School under Project No. QD2022004C; National Natural Science Foundation of China under Project No. W2432041; and Tsinghua Shenzhen International Graduate School-Shenzhen Pengrui Young Faculty Program of Shenzhen Pengrui Foundation under Project No. SZPR2023008. Data availability The dataset used in this research is available in Ref27. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The dataset used in this research is available in Ref27.