Ocean exploration is crucial for utilizing its extensive resources. Images captured by underwater robots suffer from issues such as color distortion and reduced contrast. To address the issue, an innovative enhancement algorithm is proposed, which integrates Transformer and Convolutional Neural Network (CNN) in a parallel fusion manner. Firstly, a novel transformer model is introduced to capture local features, employing peak-signal-to-noise ratio (PSNR) attention and linear operations. Subsequently, to extract global features, both temporal and frequency domain features are incorporated to construct the convolutional neural network. Finally, the image’s high and low frequency information are utilized to fuse different features. To demonstrate the algorithm’s effectiveness, underwater images with various levels of color distortion are selected for both qualitative and quantitative analyses. The experimental results demonstrate that our approach outperforms other mainstream methods, achieving superior PSNR and structural similarity index measure (SSIM) metrics and yielding a detection performance improvement of over ten percent.

Ocean exploration is crucial for utilizing its extensive resources. Images captured by underwater robots suffer from issues such as color distortion and reduced contrast. To address the issue, we propose an innovative enhancement algorithm that integrates Transformer and Convolutional Neural Network (CNN) in a parallel fusion manner. Firstly, a novel transformer model is intro-duced to capture local features, employing peak-signal-to-noise ratio (PSNR) attention and linear operations. Subsequently, to extract global features, both temporal and frequency domain features are incorporated to construct convolutional neural network. Finally, the Fourier’s high and low-frequency information of the original image are utilized to fuse different features. To demon-strate the algorithm's effectiveness, underwater images with various levels of color distortion are selected for both qualitative and quantitative analyses. The experimental results demonstrate that our approach surpasses other mainstream methods, achieving superior PSNR and structural sim-ilarity index measure (SSIM) metrics and leading to a detection performance improvement of over ten percent.

In China’s coastal regions, severe seawater eutrophication has led to frequent occurrences of red tides, causing significant damage to marine fisheries and aquatic resources. Therefore, red tide detection and prediction are of great research importance. Although current deep learning-based red tide detection methods perform well in detecting single-day red tides, they struggle with continuous multi-day detection due to insufficient mining of temporal features and difficulties in accurately capturing dynamic variations, limiting further improvements in detection accuracy. To address these issues, this study proposes a time-series fusion network model (CSF-RTDNet) for red tide detection using time-continuous GOCI data from the East China Sea. By integrating multi-temporal GOCI data, the model comprehensively captures spatiotemporal characteristics of red tides, enhancing dynamic process modeling. The CSF-RTDNet method improves feature discrimination by introducing NDVI to enhance red tide characteristics and increase separability between red tides and seawater. Additionally, an ECA channel attention mechanism is employed to fully exploit spectral features across different bands for deeper feature extraction. A novel feature extraction module, ASPC-DSC, combines atrous spatial pyramid convolution with depthwise separable convolution to effectively fuse multi-scale contextual features while improving computational efficiency. Furthermore, ConvLSTM is introduced to integrate temporal and spatial features, effectively addressing the insufficient mining of sequential characteristics in multi-day red tide detection. Experimental results demonstrate that CSF-RTDNet achieves robust detection of red tides with complex boundaries and continuous temporal patterns, attaining an accuracy of 95.89%, precision of 93.03%, recall of 96.34%, and a Kappa coefficient of 0.95. This method significantly enhances red tide detection accuracy and provides valuable technical support for marine environmental monitoring.

Fish condition detection, including the identification of feeding desire, biological attachments, fence breaches, and dead fishes, has become an important research frontier in fishery aquaculture. However, perception in underwater conditions is less satisfactory and remains a tricky problem. Firstly, we have developed a multimodal dataset based on Neuromorphic vision (NeuroVI) and RGB images, encompassing challenging fishery aquaculture scenarios. Within the fishery aquaculture dataset, a spike neural network (SNN) method is designed to filter NeuroVI images, and the sift feature points are leveraged to select the optimal image. Next, we propose a dual-image cross-attention learning network that achieves scene segmentation in a fishery aquaculture cage. This network comprises double-channels feature extraction and guided attention learning modules. In detail, the feature matrix of NeuroVI images serves as the query matrix for RGB images, generating attention for calculating key and value matrices. Then, to alleviate the computational burden of the dual-channel network, we replace dot-product multiplication with element-wise multiplication, thereby reducing the computational load among different matrices. Finally, our experimental results from the fishery cage demonstrate that the proposed method achieves the state-of-the-art segmentation performance in the management process of fishery aquaculture.

In confined waters such as cylindrical aquaculture cages and ponds, underwater robots for cleaning, harvesting and other tasks that use single-beam mechanical scanning sonar for positioning and navigation are susceptible to multipath interference in complex water environments. Meanwhile, under extreme sea conditions, the severe attitude swaying of the robot and the slow-scanning characteristic of the sonar superimpose on each other, causing range stretching and helical deformation of the acoustic point cloud. To address these problems, this paper analyzes the deformation mechanism of single-beam sonar and proposes a spatiotemporal joint deformation compensation and localization-mapping method. First, an attitude-derived probabilistic confidence model is introduced as a lightweight robustness safeguard to characterize the geometric reliability of sonar echoes and reduce the contribution of low-confidence measurements during subsequent registration. Second, a beam-level spatiotemporal joint de-deformation algorithm is designed: the slant range in polar coordinates is flattened to eliminate nonlinear swaying deformation, and a beam-level displacement back-estimation based on the beam time offset and feedback velocity is employed to remove helical misalignment, thereby enhancing the underlying correction capability for dynamic deformation processes. Finally, a lightweight SLAM architecture that integrates keyframe-based dynamic sub-maps is constructed, where a confidence-weighted ICP is used to estimate the planar position with the heading provided by the compass and provide velocity-based closed-loop feedback, effectively mitigating the problem of global matching divergence caused by underlying dynamic deformations. Real-data-driven semi-physical disturbance tests based on measured pool data show that, under the injected ±45° roll disturbance and translational drift, the proposed method reduces the maximum point-to-reference error MaxAE from 1.059 m to 0.098 m. The reported mean internal registration residual decreases from 0.275 m for the traditional navigation odometry SLAM to 0.158 m for the proposed method, corresponding to a numerical reduction of approximately 42.5%. Under the evaluated conditions, the proposed method effectively mitigates point-cloud deformation and registration instability caused by robot swaying and slow-scanning sonar, while confidence weighting is retained as an auxiliary robustness mechanism for handling low-confidence correspondences.

To address the challenges of severe optical attenuation and dynamic feature extraction for moving target depth retrieval in complex underwater remote sensing environments, this paper proposes a dynamic target depth estimation method based on multi-source data fusion. Taking optical RGB imagery and neuromorphic vision (NeuroIV) data as joint inputs, the proposed method constructs a three-channel feature extraction and fusion network. By leveraging a hypergraph structure, it establishes association weights between dynamic (temporal) and static (spatial) nodes to capture spatiotemporal correlations. To efficiently process the high-dimensional multi-modal data, the traditional dot-product attention is replaced with element-wise multiplication, significantly reducing computational complexity. Furthermore, a lightweight deformable attention pyramid (DAP) and diffusion model is introduced to refine depth image edges, effectively suppressing discontinuities and abruptness in the estimation results. Compared to single-modality optical imagery, the fused multi-modal data yields a superior signal-to-noise ratio and foreground contrast, achieving an improvement of over 20% in the MAE index. These results validate the effectiveness and superiority of the proposed multi-modal fusion strategy for dynamic target observation and depth retrieval in aquatic environments.