Highlights What are the main findings? A Heat Wave-Optimized Physics-Informed Neural Network (HW-OPINN) is proposed that integrates ocean mixed-layer heat budget dynamics with adaptive deep learning, achieving a test MSE of 0.009138 and RMSE of 0.0956 degrees C, representing improvements of 43.9% and 25.1% over the ConvLSTM baseline and 44.8% and 25.7% over standard PINN, respectively. The framework successfully reproduces the spatial heterogeneity of marine heatwave characteristics in the Mediterranean Sea, achieving the lowest MAE for event frequency (0.822 events/year) and mean duration (3.999 days) among all compared methods. What are the implications of the main findings? The dual-adaptive optimization strategy-combining Boltzmann-based collocation point sampling, region-specific adaptive weight optimization, and Bayesian hyperparameter tuning-provides a systematic methodology for balancing physical constraints against data fitting in Physics-Informed neural networks. This work demonstrates the potential of integrating physical knowledge with data-driven methods for reliable extreme ocean event prediction, supporting climate adaptation and marine resource management.Highlights What are the main findings? A Heat Wave-Optimized Physics-Informed Neural Network (HW-OPINN) is proposed that integrates ocean mixed-layer heat budget dynamics with adaptive deep learning, achieving a test MSE of 0.009138 and RMSE of 0.0956 degrees C, representing improvements of 43.9% and 25.1% over the ConvLSTM baseline and 44.8% and 25.7% over standard PINN, respectively. The framework successfully reproduces the spatial heterogeneity of marine heatwave characteristics in the Mediterranean Sea, achieving the lowest MAE for event frequency (0.822 events/year) and mean duration (3.999 days) among all compared methods. What are the implications of the main findings? The dual-adaptive optimization strategy-combining Boltzmann-based collocation point sampling, region-specific adaptive weight optimization, and Bayesian hyperparameter tuning-provides a systematic methodology for balancing physical constraints against data fitting in Physics-Informed neural networks. This work demonstrates the potential of integrating physical knowledge with data-driven methods for reliable extreme ocean event prediction, supporting climate adaptation and marine resource management.Abstract Marine heatwaves (MHWs) are prolonged extreme warming events that pose severe threats to marine ecosystems and coastal communities, necessitating reliable prediction capabilities for climate adaptation and marine resource management. Traditional numerical models, while physically grounded, are constrained by computational costs and error accumulation, whereas purely data-driven approaches often lack physical consistency and generalize poorly to extreme events. To address these challenges, this study proposes a Heat Wave-Optimized Physics-Informed Neural Network (HW-OPINN) that synergistically integrates ocean mixed-layer heat budget dynamics with adaptive deep learning techniques. The proposed framework introduces three methodological innovations. First, an adaptive sampling strategy grounded in Boltzmann distribution theory dynamically reallocates physical collocation points toward high-gradient regions based on historical loss patterns. Second, a residual-based adaptive weight update mechanism automatically modulates physical constraint contributions across spatially heterogeneous regions during training. Third, a Bayesian optimization framework employing Gaussian process surrogates systematically balances physical constraints against data fitting objectives. The framework is validated through comprehensive experiments in the Mediterranean Sea using multi-source reanalysis data spanning over two decades. Results demonstrate that HW-OPINN achieves superior performance in sea surface temperature (SST) prediction, with a test MSE of 0.009138 and RMSE of 0.095595, representing improvements of 43.9% and 25.1%, respectively, compared to the ConvLSTM baseline (MSE: 0.016275, RMSE: 0.127575), and 44.8% and 25.7% improvements over standard PINN (MSE: 0.016550, RMSE: 0.128661). Based on the predicted SST fields, the model successfully reproduces the spatial heterogeneity of key MHW characteristics, including event frequency, duration, and intensity distributions, demonstrating its effectiveness for downstream MHW detection and analysis.

Sea surface temperature (SST) plays a central role in regulating ocean and atmosphere interactions and influencing extreme climate events such as marine heatwaves. However, the inherent complexity and nonlinearity of SST dynamics present major challenges for achieving accurate and interpretable forecasting. To address this problem, we propose a novel interpretable framework named Multitemporal Scale Fusion Transformers (MTSFT), which provides a solution for prediction accuracy and explanatory power. MTSFT incorporates Enhanced Multitemporal Scale Periodic Features to decouple overlapping temporal patterns at daily, seasonal, and interannual scales, improving the model's ability to capture key temporal structures. Based on an improved Temporal Fusion Transformers, the framework integrates static covariates, historical environmental inputs, and known future indicators into a unified architecture. In addition, MTSFT supports multilevel interpretability by identifying dominant drivers, detecting SST anomalies, and characterizing periodic patterns across various time scales. Experimental results across typical coastal regions of China show that MTSFT consistently achieves reliable prediction performance and offers meaningful scientific insights to support marine risk assessment and climate-informed decision-making.

Graph neural networks (GNNs) have been widely applied in recommender systems due to their strong relational modeling capability. However, existing methods remain limited in heterogeneous multi-source information uti lization and higher-order collaborative signal capture. To address these limitations, we propose HR-HGCL for recommendation. HR-HGCL explicitly models the user-item interaction view, the user-user social view, and the item-item association view, and further constructs a user-item-user-item higher-order path view to enrich collaborative signals. Additionally, we employ a dual-neighborhood-enhanced personalized meta network that integrates information from both direct and higher-order neighbors to achieve cross-view semantic alignment. To improve training effectiveness, we design an efficient hard-negative sampling strategy that provides strong dis criminative supervision. Experiments on four public datasets demonstrate that HR-HGCL consistently outperforms representative baselines across multiple evaluation metrics. The results further suggest that jointly modeling het erogeneous multi-source relations and higher-order collaborative signals can benefit recommendations in sparse interaction scenarios.

[Objective]Detecting dense and small aquaculture net cages in complex backgrounds is difficult, the purpose of this study is to build a specialized dataset and design a targeted detection model that enhances recognition accuracy and robustness for practical aquaculture management.[Methods]A dataset of aquaculture net cages was constructed using high-resolution remote sensing imagery collected from seven representative farming regions (Australia, Canada, Chile, Croatia, Greece, China, and the Faroe Islands), and Cage-YOLO, a deep learning model based on YOLOv5, was proposed for detecting dense and small aquaculture net cages. First, an adaptive dense perception algorithm was introduced, which automatically selects and generates feature maps that reflect the high-density distribution of small aquaculture net cages. Second, an enhanced module based on spatial pyramid pooling fast was integrated to effectively reduce background noise interference and improve global feature extraction capabilities. Finally, a mixed attention block was incorporated to further enhance the model's perception of dense and small objects.[Results and Discussions]Experimental results showed that the proposed Cage-YOLO achieved improvements over the original YOLOv5 in terms of precision, recall, and mean average precision by 5.6, 21.8, and 17.4 percentage points, respectively. The model size was maintained at 16.9 MB, demonstrating both strong performance and deployment advantages.[Conclusions]This study provides a new approach for dense and small object detection and offers technical support for the intelligent management of marine cage aquaculture.

Pre-trained Language Models (PLMs) empower Knowledge Graph Completion (KGC) but suffer from a compounding failure chain rooted in topological heterogeneity: hub nodes introduce noise while long-tail nodes starve for context, coarse-grained aggregation dilutes the resulting embeddings, and static negative sampling cannot probe the fine-grained ambiguities that remain. We propose HACKGC, an end-to-end framework that breaks this chain through a tightly coupled sample → encode → challenge pipeline. First, Topology-Aware Curriculum Sampling dynamically adjusts random walk horizons by node degree to construct high-SNR subgraphs. Second, a Hierarchical Attention Encoder refines contexts across token, cross-view, and batch granularities, with a training-only regularizer that transfers cross-encoder interaction patterns into a bi-encoder via gradient backpropagation. Third, an Adversarial-Contrastive module — operating on the structure-enhanced embeddings rather than static geometric representations—synthesizes topology-conditioned hard negatives targeting decision-boundary ambiguities. This closed-loop design ensures each stage’s output directly fuels the next. Experiments on WN18RR, FB15k-237, and Wikidata5M confirm state-of-the-art performance (e.g., +4.0

Context Fine-grained classification of cephalopod beaks faces significant challenges because of small inter-class variations and large intra-class variations, limiting the accuracy of existing methods for morphologically similar species.Aims To develop a fine-grained classification method that can automatically identify discriminative anatomical regions and improve species identification accuracy.Methods We constructed a dataset of 14,000 images from 700 beak pairs of seven cephalopod species. Building on ResNet101, we combined region grouping with beta prior constraints to achieve adaptive segmentation of key anatomical regions and attention-weighted feature learning.Key results Our method achieved classification accuracies of 97.00 and 97.36% for upper and lower beaks respectively, improving on the best existing model by 2.79-3.79%. The model automatically focused on discriminative features, including lateral wall posterior margin curvature, hood development patterns, and wing structures.Conclusions The proposed weakly supervised regional segmentation approach significantly improves both classification performance and interpretability, with results being highly consistent with biological knowledge.Implications This method provides an efficient solution for cephalopod species identification, with broad application prospects in ecological research and fisheries management.

Underwater object detection is of great value in marine resource development, but the traditional fully supervised methods are limited due to the scarcity of labeled data. Currently, semi-supervised learning is an efficient approach to addressing this issue. However, the complex underwater environment may exacerbate issues such as pseudo-label noise, bias, and insufficient feature extraction in traditional semi-supervised methods. To address these issues, we propose an innovative framework for underwater semi-supervised object detection. Firstly, we design a CARAFE-enhanced PAFPN (CE-PAFPN) model that can solve the problem of high-level features lacking low-level details in FPN while fully integrating contextual information and improving the model’s ability to learn target features. Secondly, we propose an adaptive threshold adjustment mechanism to mitigate the pseudo-label noise, which dynamically adjusts the threshold to filter high-quality pseudo-labels according to the sample difficulty. Finally, by incorporating a class-balance loss to downweight overly confident pseudo-labels, thereby mitigating the pseudo-label bias caused by class imbalance. The proposed method ultimately achieves 56.8 mAP_50:95 on the DUO dataset with only 10

The scheduling of Earth observation satellites presents a formidable multi-objective optimization challenge, characterized by inherent trade-offs among task completion rate, execution timeliness, and the temporal uniformity of revisits. To address this, we introduce the Multi-Satellite Observation Task Scheduling (MSOTS) framework, a novel end-to-end approach based on Multi-Agent Reinforcement Learning (MARL). This framework formulates the scheduling process as a Markov game, employing the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm within a Centralized Training, Decentralized Execution (CTDE) paradigm to effectively navigate these competing objectives. Furthermore, to ensure a balanced evaluation, we propose a Composite Multi-Objective Performance Score grounded in a weighted harmonic mean. Comprehensive empirical evaluations conducted on large-scale, simulated orbital scenarios demonstrate that MSOTS significantly outperforms both traditional heuristics and existing deep reinforcement learning methods in comprehensive performance and robust efficiency. This research provides a highly effective and intelligent approach to modern satellite task scheduling.

Unsupervised anomaly detection on multivariate time series remains challenging due to the concealed nature of anomalies and the high dimensionality of the data. Although reconstruction-based methods have recently demonstrated impressive progress, they often suffer from data contamination and insufficient learning of diverse normal patterns. To address these issues, we propose a novel multivariate time series anomaly detection model that integrates a stochastic attention mechanism with a learnable memory module. Specifically, we introduce stochastic pooling and a rank-adjustment strategy into self-attention to mitigate overfitting. Moreover, the learnable memory module is designed to capture and record diverse normal patterns, thereby enhancing the model’s ability to discriminate anomalies. Extensive experiments on four public benchmark datasets demonstrate that our approach achieves state-of-the-art performance.

Remote sensing image super-resolution (RSISR) aims to enhance low-resolution remote sensing images to restore high-quality details. Currently, methods based on Swin Transformer have achieved remarkable success due to their excellent contextual attention capabilities and moving window mechanism. However, the limitations of this window mechanism make it difficult for the model to achieve cross-window information interaction, resulting in the loss of important complementary information. To address this issue, inspired by the adaptive token dictionary (ATD), we propose a new RSISR method, HAMD-RSISR, and for the first time, introduce the use of a "multidictionary mechanism" to break through the limitations of traditional window mechanisms. Unlike the single-dictionary ATD, our multidictionary mechanism enhances the network's global context modeling capabilities at both coarse and fine granularities. Specifically, one dictionary is used for preliminary redivision of feature maps, while the other dictionary performs further refinement on the previous dictionary to process feature maps at deeper scales. In addition, we propose a hybrid attention mechanism to further effectively utilize channel, spatial, and frequency information in remote sensing images. Finally, the hybrid attention and multidictionary mechanisms process the input feature maps in parallel and perform feature fusion. Experimental results demonstrate the excellent performance of our HAMD in RSISR. On the CLRS x 2 dataset, it achieves a peak signal-to-noise ratio (PSNR) improvement of approximately 0.44 dB compared to the state-of-the-art RGT in recent years, and on the UC Merced x 4 dataset, it achieves a PSNR improvement of approximately 0.4 dB compared to HAT.

Accurate coastal zone change detection is crucial for coastal urban planning and marine resource development. To address the specificity of coastal zone change detection and the category imbalance issue in the model, we propose a multi-scale coastal zone change detection method (AMMNet) based on the attention mechanism. The method leverages multi-scale features extracted by the ResNet backbone, which are then optimized and integrated through high-frequency attention and spatio-temporal difference modules. These modules allow the model to focus on both global and local changes, enhancing its ability to detect variations in coastal zones. Additionally, the foreground attention module refines the model’s attention on relevant regions, ensuring improved performance. The experimental results show that our method achieves the highest scores in several evaluation metrics, demonstrating significant advantages in accuracy and generalization and effectively addressing the category imbalance problem. It provides a robust solution for coastal zone change detection.

Ocean data quality control (QC) is important for ocean scientific research and resource management, which aims to identify data errors and ensure the reliability of ocean data. Modern ocean observation data are characterized by large volumes and complex association patterns. This results in low efficiency of manual QC and low accuracy of automatic QC. Therefore, the use of unsupervised deep learning models for ocean data QC has become a major trend in recent years. However, it faces some challenges: the powerful learning ability of deep learning models could learn too well from anomalous data; the determination of anomaly threshold has an important effect on the results of QC. To address these issues, we propose a QC method for ocean data based multi-time scale downsampling and dynamic threshold strategy (MTSDTS-QC). A Multi-time Scale Downsampling module is developed to extract the distribution of normal data at both large and small time scales. This process can reduce the interference of anomalous data in the subsequent process. Then, TimesBlocks are used to reconstruct the ocean data considering their multi-periodic characteristics. The difference between the original and the reconstructed data represents the degree of abnormality of the data. Furthermore, we propose a dynamic threshold strategy that can determine the optimal thresholds for different marine data based on the intrinsic characteristics of anomaly scores. Experimental results on marine meteorological data show that the MTSDTS-QC method outperforms 14 existing baseline methods. The generalizability of our model is also validated on a public industrial control dataset WADI.

The unrestricted exploitation of marine resources has triggered severe ecological crises. As the primary carriers of the fishing industry, fishing vessels play a crucial role in marine fishery resource management and ecological conservation. Identifying their operation types is essential for maintaining marine ecological balance, combating illegal, unreported, and unregulated (IUU) fishing, and promoting sustainable fisheries development. However, the accuracy of existing AIS-based identification methods is limited by the inherent contradiction: majority-class non-fishing data exhibit high kinematic homogeneity across different vessels, whereas minority-class fishing data, which serve as the primary basis for discrimination, display significant heterogeneity due to variations in fishing gear. Therefore, this paper proposes MM-FishingNet, a multi-modal adaptive identification model that integrates time-frequency features with trajectory images. First, a time-frequency dual-domain feature extraction module is designed to enhance the sensitivity to trajectory discrepancies. Second, to compensate for the insufficient discriminative capacity of AIS data, trajectory images are introduced as a supplement, and a knowledge-guided fusion mechanism is constructed to adaptively adjust multi-modal fusion. Experimental results show that on three fishing vessel datasets from the East China Sea with real annotations, MM-FishingNet significantly outperforms the single-modal baselines and the model using simple feature fusion methods, achieving an average accuracy of 92.23% and an F1-score of 92.18%. This provides a new technical paradigm for the intelligent monitoring of fishing vessel operation type.

Ocean fronts are central to upper-ocean dynamics and ecosystem processes, yet recognizing their evolutionary trends from satellite data remains challenging. We present a 3D U-Net-based multi-task framework that jointly performs ocean-front detection (OFD) and ocean-front evolutionary trend recognition (OFETR) from sea surface temperature gradient heatmaps. Instead of cascading OFD and OFETR in separate stages that pass OFD outputs downstream and can amplify upstream errors, the proposed model shares 3D spatiotemporal features and is trained end-to-end. We construct the Zhejiang–Fujian Coastal Front Mask (ZFCFM) and Evolutionary Trend (ZFCFET) datasets from ESA SST CCI L4 products for 2002–2021 and use them to evaluate the framework against 2D CNN baselines and traditional methods. Multi-task learning improves OFETR compared with single-task training while keeping OFD performance comparable, and the unified design reduces parameter count and daily computational cost. The model outputs daily point-level trend labels aligned with the dataset’s temporal resolution, indicating that end-to-end multi-task learning can mitigate error propagation and provide temporally resolved estimates.

Accurately subseasonal daily sea surface temperature prediction (SSTP) is significant for forecasting and mitigating extreme climate events related to sea surface temperature (SST). However, this scale's forecasting lies at the transitional zone between short-term forecasting and long-term climate prediction, requiring simultaneous consideration of small-scale variations crucial for the former and large-scale variations fundamental to the latter. Thus, achieving precise subseasonal daily SSTPs is challenging. In this study, we introduce a novel Multi-Scale Periodic Transformer (MSPT) to predict subseasonal daily SST, which can account for temporal variations at various scales. Initially, MSPT integrates Fast Fourier Transform and multi-layer perceptron to extract all potential periodic scales and adaptively identify critical ones. Each periodic scale features an independent branch composed of patch embedding and Transformer encoder, dedicated to specifically learning temporal variations at that scale. Only the outputs of critical branches are weighted and aggregated to obtain effective multi-periodic scale characteristics. This approach effectively decouples complex temporal patterns, enabling the model to capture reliable dependencies that are beneficial for improving subseasonal forecasting. Furthermore, by introducing additional multivariate attention, our improved Transformer Encoder can capture the inherent multivariate correlations of SST dynamics, perfecting the representation of temporal variations at specific periodic scales. Extensive subseasonal forecasting experiments conducted at four locations in the South China Sea demonstrate that MSPT achieves state-of-theart performance in 10-30d subseasonal daily SSTPs, validating the effectiveness of multi-scale periodic information in improving subseasonal forecasting.

Multimodal amed entity recognition (MNER) is a natural language-processing technique that integrates text and visual modalities to detect and segment entity boundaries and their types from unstructured multimodal data. Although existing methods alleviate semantic deficiencies by optimizing image and text feature extraction and fusion, a fundamental challenge remains due to the lack of fine-grained alignment caused by cross-modal semantic deviations and image noise interference. To address these issues, this paper proposes a dual-enhanced hierarchical alignment (DEHA) framework that achieves dual semantic and spatial enhancement via global–local cooperative alignment optimization. The proposed framework incorporates a dual enhancement strategy comprising Semantic-Augmented Global Contrast (SAGC) and Multi-scale Spatial Local Contrast (MS-SLC), which reinforce the alignment of image and text modalities at the global sample level and local feature level, respectively, thereby reducing image noise. Additionally, a cross-modal feature fusion and vision-constrained CRF prediction layer is designed to achieve adaptive aggregation of global and local features. Experimental results on the Twitter-2015 and Twitter-2017 datasets yield F1 scores of 77.42% and 88.79%, outperforming baseline models. These results demonstrate that the global–local complementary mechanism effectively balances alignment precision and noise robustness, thereby enhancing entity recognition accuracy in social media and advancing multimodal semantic understanding.

With advances in graph computing, an increasing amount of real-world data is now stored in computers as graphs. However, the time and resource costs associated with directly processing large graphs are escalating. As a result, graph summarization techniques, which provide a concise representation of graphs, have been extensively studied. While current graph summarization methods are effective in reducing processing time and minimizing storage overhead, we observe that the hierarchical and overlapping structures prevalent in graphs are not adequately explored, and the resulting summaries lack clear meaning. In this study, we present a new graph summarization model that uses a linear iterative hierarchical clustering algorithm combined with a lexicon-based structure extraction algorithm to generate relational patterns with clear hierarchical and overlapping meanings. In addition, we use the Minimum Description Length (MDL) principle for summarization encoding to save storage space. Experiments show that our approach achieves superior compression ratios on most datasets compared to existing methods and is able to uncover a larger number of relational patterns.

Heterogeneous graph neural networks (HGNNs) often face challenges in efficiently integrating information from multiple views, which hinders their ability to fully leverage complex data structures. To overcome this problem, we present an improved graph-level cross-attention mechanism specifically designed to enhance multi-view integration and improve the model's expressiveness in heterogeneous networks. By incorporating random walks, the Katz index, and Transformers, the model captures higher-order semantic relationships between nodes within the meta-path view. Node context information is extracted by decomposing the network and applying the attention mechanism within the network schema view. The improved graph-level cross-attention in the cross-view context adaptively fuses features from both views. Furthermore, a contrastive loss function is employed to select positive samples based on the local connection strength and global centrality of nodes, enhancing the model's robustness. The suggested self-supervised model performs exceptionally well in node classification and clustering tasks, according to experimental data, demonstrating the effectiveness of our method.

Accurately subseasonal daily sea surface temperature prediction (SSTP) is significant for forecasting and mitigating extreme climate events related to sea surface temperature (SST). However, this scale's forecasting lies at the transitional zone between short-term forecasting and long-term climate prediction, requiring simultaneous consideration of small-scale variations crucial for the former and large-scale variations fundamental to the latter. Thus, achieving precise subseasonal daily SSTPs is challenging. In this study, we introduce a novel multiscale periodic transformer (MSPT) to predict subseasonal daily SST, which can account for temporal variations at various scales. Initially, MSPT integrates fast Fourier transform and multilayer perceptron to extract all potential periodic scales and adaptively identify critical ones. Each periodic scale features an independent branch composed of patch embedding and Transformer encoder, dedicated to specifically learning temporal variations at that scale. Only the outputs of critical branches are weighted and aggregated to obtain effective multiperiodic scale characteristics. This approach effectively decouples complex temporal patterns, enabling the model to capture reliable dependencies that are beneficial for improving subseasonal forecasting. Furthermore, by introducing additional multivariate attention, our improved Transformer encoder can capture the inherent multivariate correlations of SST dynamics, perfecting the representation of temporal variations at specific periodic scales. Extensive subseasonal forecasting experiments conducted at four locations in the South China Sea demonstrate that MSPT achieves state-of-the-art performance in 10–30 d subseasonal daily SSTPs, validating the effectiveness of multiscale periodic information in improving subseasonal forecasting.

Marine heatwaves (MHWs) refer to a phenomenon where the sea surface temperature is significantly higher than the historical average for that region over a period, which is typically a result of the combined effects of climate change and local meteorological conditions, thereby potentially leading to alterations in marine ecosystems and an increased incidence of extreme weather events. MHWs have significant impacts on the marine environment, ecosystems, and economic livelihoods. In recent years, global warming has intensified MHWs, and research on MHWs has rapidly developed into an important research frontier. With the development of deep learning models, they have demonstrated remarkable performance in predicting sea surface temperature, which is instrumental in identifying and anticipating marine heatwaves (MHWs). However, the complexity of deep learning models makes it difficult for users to understand how the models make predictions, posing a challenge for scientists and decision-makers who rely on interpretable results to manage the risks associated with MHWs. In this study, we propose an interpretable model for discovering MHWs. We first input variables that are relevant to the occurrence of MHWs into an LSTM model and use a posteriori explanation method called Expected Gradients to represent the degree to which different variables affect the prediction results. Additionally, we decompose the LSTM model to examine the information flow within the model. Our method can be used to understand which features the deep learning model focuses on and how these features affect the model’s predictions. From the experimental results, this study provides a new perspective for understanding the causes of MHWs and demonstrates the prospect of future artificial intelligence-assisted scientific discovery.