Abstract:

Driven by bio-inspired propulsion mechanisms， bionic robotic fish exhibit high propulsion efficiency， superior maneuverability， and low navigation noise during underwater locomotion， enabling stable operation in complex aquatic environments. This provides new technical pathways for marine environmental monitoring， fishery resource surveys， and underwater engineering operations. This paper systematically reviews the research progress of AI-enabled soft bionic robotic fish from 2005 to 2024， summarizing key developments across core dimensions including biological prototype mechanisms， bionic design theories， structural and actuation approaches， intelligent control algorithms， and power supply systems. Comparative analyses are conducted based on propulsion modes， actuation types， and typical application scenarios.The results show that the BCF （Body and/or caudal fin） propulsion mode delivers superior efficiency for long-distance， high-speed cruising， while the MPF （Median and/or paired fin） propulsion mode offers enhanced stability for low-speed maneuvering and near-bottom precision operations. Significant breakthroughs have been achieved in artificial muscles and flexible actuation for low-noise， highly biomimetic locomotion， yet bottlenecks remain in cycle life， output consistency， and engineering costs. Phased progress has been made in multimodal perception， autonomous decision-making， and distributed cooperative control， yet cross-platform adaptability， performance in complex sea conditions， and reproducibility in real-sea trials remain insufficient.Future research should prioritize three core breakthrough directions： deep integration of artificial intelligence and intelligent control， development of novel smart materials， and expansion of cross-medium operational capabilities. Key priorities include establishing unified testing benchmarks， long-term sea trial verification systems， and standardized engineering design specifications to comprehensively enhance the engineering application level of soft bionic robotic fish in marine exploration， fishery resource surveys， and underwater operation scenarios.

Abstract:

As global climate change intensifies and uncertainty in the spatiotemporal distribution of fishery increases， artificial intelligence （AI） is becoming an important force inadvancing marine fisheries forecasting technologies， owing to its robust capability in representing high-dimensional spatiotemporal data. This paper systematically reviews the application and research progress of AI in marine fisheries forecasting. First， a standardized AI-based forecasting framework is outlined， encompassing multi-source data integration and preprocessing， model development and training， and model performance evaluation. Second， representative studies of AI in forecasting fisheries for major oceanic commercial species， particularly squid and tuna， are reviewed， highlighting the advantages of ensemble learning and deep learning in modeling complex nonlinear responses to environmental variability. Finally， three key scientific issues are discussed in depth， including spatiotemporal scale optimization， optimal model selection， and response mechanisms to climate change. With the continuous advancement of AI， future fisheries forecasting is expected to evolve toward multi-task joint forecasting and probabilistic forecasting， thereby providing theoretical support for the development of intelligent fisheries forecasting systems with high accuracy， strong robustness， and enhanced ecological interpretability， as well as autonomous decision-making agents.

Abstract:

Against the backdrop of the vigorous development of artificial intelligence and big data technologies， knowledge graphs， which integrate multidisciplinary knowledge， have achieved remarkable results in fisheries and many other professional fields. Relevant research is conducive to addressing issues such as the fragmentation of professional knowledge and insufficient visualization. Therefore， this paper systematically sorts out the basic theories of knowledge graphs， including key technologies such as construction methods and processes， ontology construction， and knowledge extraction， as well as the application progress and challenges in the fishery field. Meanwhile， this paper explores the research hotspots， specific application situations and existing deficiencies in this field based on bibliometric methods. The research indicates that while the application of fisheries knowledge graphs demonstrates a growing adoption and has achieved progress in core domains such as bibliometric analysis， information retrieval， intelligent Q&A， and recommendation systems， it still faces widespread limitations. These include a narrow research scope， outdated retrieval algorithms， inadequate complex reasoning capabilities， and ineffective handling of long-tail content. Furthermore， although knowledge graph construction is transitioning from manual to automated processes， with technologies for ontology development， knowledge extraction， and storage maturing， significant shortcomings persist in knowledge fusion， reasoning， and visualization， posing challenges such as data heterogeneity and domain complexity. The research suggests that the future application of knowledge graphs in fisheries should focus on areas such as data integration and standardization， construction of multimodal fishery knowledge graphs， and graph neural network prediction based on fishery knowledge graphs， so as to promote the intelligent development and sustainable development of fisheries.

Abstract:

To address the challenges of restricted operating space， variable hull curvature， and increasingly stringent environmental regulations in the underwater maintenance of small special-purpose vessels， this paper systematically reviews and evaluates key technologies applicable to underwater maintenance robots for such vessels. It analyzes the current technical status of underwater maintenance robots both domestically and internationally， with special emphasis on comparing the stability of adhesion mechanisms， namely magnetic adhesion， negative pressure adhesion， and mechanical gripping， on surfaces with variable curvature and non-magnetic substrates. The motion performance of various drive systems under complex operating conditions is assessed， and mainstream cleaning techniques such as brushing and water jetting are summarized. The study reveals the evolutionary trends of existing technologies in adaptive adhesion， high-precision actuation， and environmentally friendly operations. It further explores detection and intervention approaches tailored to non-magnetic composite-material vessels and examines the performance requirements specified in current technical standards for biofouling prevention and pollutant capture. The findings indicate that future underwater maintenance robots for small special-purpose vessels should advance toward AI-based path planning， multi-modal SLAM navigation， and modular multi-functional payloads to achieve efficient， precise， and environmentally sustainable maintenance operations. This work provides a reference for the development of intelligent underwater maintenance equipment for small special-purpose vessels and holds significant importance for enhancing maintenance efficiency of small vessels in China while meeting aquatic ecological protection requirements.

Abstract:

To address the trajectory deviation of remotely operated vehicles （ROVs） during operations in complex environments-caused by the nonlinear towing of a negative-buoyancy umbilical cable and the transient impact of ocean waves-the dynamic response of the umbilical cable and an anti-disturbance trajectory tracking control method are investigated. First， a dynamic model of the umbilical cable is established using the lumped-mass method and the Morison equation， and the time-varying tension and towing-disturbance characteristics under typical operating conditions are analyzed with the OrcaFlex software. Second， an improved line-of-sight （ILOS） guidance law is designed to achieve dimensionality reduction， decoupling， and guidance of the three-dimensional trajectory； a linear extended state observer （LESO） is introduced to perform real-time estimation and feedforward compensation of the unknown composite disturbances arising from cable towing forces and the wave-current field， and a power-rate sliding mode control （PRSMC） is combined to smooth the control output and suppress high-frequency chattering. Finally， a co-simulation platform based on MATLAB and OrcaFlex is built for closed-loop verification. Simulation results show that， in tracking a complex polyline trajectory， the proposed control method reduces the mean absolute errors （MAEs） of the three degrees of freedom by 28.40%， 27.95%， and 49.78%， respectively， compared with conventional sliding mode control （SMC）.The results indicate that， under complex trajectories and multi-source composite disturbances， the proposed control algorithm can accurately observe and compensate in real time for the unknown current-cable coupled disturbances， effectively reducing steady-state error and system chattering， and achieving good tracking accuracy and strong robustness. This study can provide a reference for the design and application of anti-disturbance control systems for ROVs in strongly disturbed environments.

Abstract:

In order to deeply explore the complex hydrodynamic characteristics of bigeye tuna （Thunnus obesus ） and provide a theoretical basis for the bionic design of low-drag fishing gear and the optimization of longline fishing operation parameters， this study takes bigeye tuna as the research subject. Based on its body shape characteristics， a three-dimensional numerical model was constructed using Solidworks software， and a silicone physical model of the fish was produced. By combining numerical simulation with flume model experiments， the study systematically analyzed the hydrodynamic properties， flow velocity distribution， and surface pressure distribution of bigeye tuna under different flow velocities and attack angles. The results indicate： （1） at a 0° angle of attack， the velocity field is symmetrically distributed， characterized by high velocities in the mid-body region and lower velocities at the head and tail. As the angle of attack increases， the high-velocity region shifts towards the leeward side of the fish body， while the high-velocity region on the windward side shifts rearward. When the angle of attack exceeds 25°， the high-velocity region on the leeward side expands， and the low-velocity region on the windward side encompasses the windward surface. Simultaneously， a distinct high-velocity region emerges between the dorsal fin and the caudal fin； （2） as the angle of attack increases， the region of positive pressure on the windward side of the bigeye tuna expands， and the region of negative pressure on the leeward side also increases. When the angle of attack exceeds 20°， positive pressure predominates over nearly the entire fish body； however， negative pressure intensifies in the dorsal and abdominal regions and at the trailing edge of the first dorsal fin. On the leeward side of the fish body， the head region exhibits positive pressure at angles of attack below 15°. However， as the angle of attack increases （from 0° to 45°）， the positive pressure area coverage decreases from 14.64% to 5.61%， while the negative pressure region expands. In contrast， the caudal peduncle consistently maintains positive pressure. This study indicates that the flow velocity and pressure distributions of the bigeye tuna are closely related to its body morphology and swimming posture， providing a significant theoretical basis for further exploration of efficient swimming mechanisms in fish. Moreover， this study conducts an in-depth analysis of the hydrodynamic role of fins， offering a reference for the optimization of bionic robot fish design and the improvement of fishing gear and methods.

Abstract:

The jumbo flying squid （Dosidicus gigas）， widely distributed in the southeastern Pacific Ocean， serves as an important target species for China's pelagic fisheries. To understand its selectivity toward different jig colors and improve fishing efficiency， this study analyzed fishing video data collected by the pelagic squid jigging vessel Hongrun 107 during its operations in the southeastern Pacific from January to February 2025. Using an improved YOLOv8n model integrated with the ByteTrack multi-object tracking algorithm， we achieved automatic identification of jig colors and squid hooking behavior， thereby determining the color selectivity and hooking rates of Dosidicus gigas. The results show that by optimizing input resolution， adjusting the loss function， and introducing data augmentation strategies， the improved YOLOv8n model achieves faster convergence in box loss， with the mAP50-95 increased by about 15 percentage points compared to the original model. It can accurately identify squid jigs of different colors and squid targets under complex backgrounds. When combined with a multi-object tracking algorithm， the system can automatically determine which jigs have caught squid in video footage.Statistical results showed that Dosidicus gigas exhibited an obvious selectivity toward jigs of different colors. The jig colors with relatively higher selectivity were yellow， orange， blue， red， pink， dark， and light yellow， in descending order， but the differences were not significant （P>0.05）. In contrast， cyan and white jigs showed the lowest selectivity （P<0.05）.The study revealed that Dosidicus gigas exhibited a certain preference in its selection of jig colors. Warm-colored jigs （such as yellow， orange， and red） are more effective in eliciting its predatory response， whereas cool-colored jigs （such as cyan and white） show weaker attraction effects.This study demonstrates that AI-based image recognition technology enables systematic quantification and fully automated monitoring of jig colors and catch status in pelagic squid fisheries， providing technical support for optimizing jigging operations and improving fishing efficiency. It also supplies valuable data for building a digital twin system for squid jigging， contributing significantly to the intelligent and digital development of China's pelagic squid fishing industry.

Abstract:

To address the problems of insufficient automatic species detection accuracy and difficulty in edge deployment caused by complex deck backgrounds， illumination variations， fish occlusion， and fine-grained differences among similar species in shipborne camera images of tuna targets in distant-water fisheries， a lightweight tuna species detection algorithm based on YOLO11n， named ACLB-YOLO， is proposed. YOLO11n is used as the baseline model. In the backbone network， the multi-cognitive visual adapter Mona is integrated with C2PSA to enhance robust feature representation under strong illumination， local overexposure， and complex backgrounds. In the neck network， a BiFPN-GLSA module is constructed to strengthen feature association for densely occluded targets through a global-local spatial attention mechanism. In the detection head， the Detect_AFPN_P345 progressive feature pyramid is introduced to alleviate semantic conflicts in multi-scale feature fusion. Furthermore， layer-adaptive magnitude-based pruning （LAMP） is adopted for lightweight model compression， and embedded deployment verification is conducted on the NVIDIA Jetson Orin NX platform. The results show that， on the self-built tuna dataset， the pruned ACLB-YOLO achieves an mAP@0.5 of 86.15%， which is 9.03 percentage points higher than that of the YOLO11n baseline model， while the model size is reduced from 5.1 MB to 2.4 MB. On the embedded platform， the pruned model achieves an average inference speed of 46.5 FPS and a total latency of 21.51 ms per frame. ACLB-YOLO can effectively improve tuna species detection performance in complex shipborne image scenarios， and the pruned model significantly reduces model complexity while maintaining good detection accuracy. This study can provide technical support for automatic tuna classification， catch statistics， and real-time deployment on shipborne edge devices in electronic monitoring systems for distant-water fisheries.

Abstract:

To efficiently extract fishing status from automatic identification system （AIS） data and provide critical technical support for combating illegal， unreported， and unregulated （IUU） fishing， this study proposes a hybrid deep learning model that integrates a multi-scale convolutional neural network （CNN） and a long short-term memory （LSTM） network， enhanced with a multi-head self-attention （MHSA） mechanism. The model takes AIS trajectory segments from four typical vessel types （longliners， purse seiners， trawlers， and squid jiggers） in the Northwest Pacific Ocean as input， comprehensively extracting features such as vessel speed， heading， spatiotemporal location， and their dynamic changes. To address the issue of class imbalance， a class weighting mechanism was incorporated during model training. The experimental results， based on 2020 AIS data， show that the model achieves recognition accuracies of 85.4% for non-fishing and 84.8% for fishing states， demonstrating good stability and discrimination ability. The model exhibits good recognition performance for trawlers， purse seiners， and squid jiggers， with accuracies exceeding 88%. However， the recall rate for longliners is relatively low （73.12%）， indicating that their complex operational behavior poses a challenge for the model. The analysis also reveals that different vessel types exhibit significant variations in fishing speed distribution， diurnal patterns， and seasonal spatial distribution. The developed model effectively identifies the fishing status of multiple vessel types， compensating for the limitations of traditional rule-based or clustering methods in complex behavior recognition. This research can provide technical support for improving dynamic monitoring and regulatory efficiency， enabling IUU fishing behavior detection， and supporting fishery resource assessment.

Abstract:

To overcome the visual degradation of optical images caused by the complexity of underwater environments and address the challenge of deploying existing high-precision models on resource-constrained devices， this study aims to enhance the optical perception and autonomous operational capabilities of intelligent fishery equipment such as autonomous underwater vehicles. This paper proposes a lightweight backbone-based underwater dual-task network， UED-Transformer （Underwater Efficient Dual-task Transformer）. The network integrates monocular depth estimation and semantic segmentation tasks， utilizes two single-task teacher models to generate pseudo-labels， and trains the UED-Transformer using the generated underwater image dataset. Experimental results on the public underwater multi-label dataset MIMIR-UW show that the model achieves a mean Intersection over Union （mIoU） of 79.3% and a pixel accuracy （PA） of 93.7% in the semantic segmentation task， while the root mean square error （RMSE） for the monocular depth estimation task is reduced to 4.975. The research indicates that this lightweight multi-task network exhibits robust multi-task performance while maintaining low computational costs， which will effectively improve the intelligent scene perception and understanding capabilities of fishery equipment in complex underwater environments. This study provides an efficient and lightweight vision algorithm system for intelligent fishery devices， reducing the dependence on high-computational hardware and promoting the practical application of artificial intelligence in the intelligentization of fishery equipment.

Abstract:

To address the problems of fragmentation， misclassification， and omission that often occur in high-resolution UAV remote sensing-based island waterline segmentation， this study proposes MF-MaxViT（Moga-Semantic FPN-MaxViT）， a dedicated model for instantaneous island waterline segmentation from UAV remote sensing imagery. The model employs MaxViT as the backbone and integrates both the Moga module and Semantic FPN， with key improvements as follows： Multi-axis attention mechanism： incorporating block attention and grid attention within MaxViT to enable the joint extraction of fine-grained local details and global semantic features； Moga（Multi-order gated aggregation） module integration： embedding the Moga module within MaxViT blocks to decompose features， extract multi-order contextual information， and perform feature aggregation， thereby enhancing the representation of multi-scale context； Semantic feature pyramid： employing a Semantic FPN（Semantic feature pyramid network） structure for multi-scale feature fusion， which strengthens the collaborative representation of deep semantic and shallow boundary information.Experimental results demonstrate that MF-MaxViT achieves an mIoU（Mean intersection over union，mIoU） of 94.40% on a self-constructed UAV remote sensing dataset， representing an improvement of 0.95% over the original MaxViT. The model also attains an mPA（Mean pixel accuracy，mPA） of 97.15% and an OA（Overall accuracy，OA） of 97.14%， outperforming mainstream models including DeepLab v3+， CrossFormer， U-Net， and TransUNet. In addition， MF-MaxViT also demonstrates strong cross-dataset generalization capability on two public satellite remote sensing datasets， the China Coastal Region dataset and the YTU-WaterNet dataset.In summary， MF-MaxViT provides a high-precision， robust approach for UAV-based instantaneous island waterline segmentation， offering reliable technical support for dynamic waterline monitoring and ecological risk assessment applications.

Abstract:

To address the safety risks posed by the blockage of water intakes by cold-source organisms in coastal nuclear power plants and to tackle the challenges of insufficient spatial and temporal coverage and short warning times in existing monitoring methods， this study explored the feasibility of utilizing commercial Echosounders mounted on fishing boats to construct a mobile monitoring network. A simulation study was conducted to assess the resource quantity of shrimp in the waters outside the water intake of Yangjiang Nuclear Power Plant. Based on the shrimp resource quantity data intercepted by the water intake net and Gaussian random field theory， a simulation was performed to construct a shrimp biomass distribution field with real spatial heterogeneity. The resource monitoring efficiency of traditional acoustic survey assessment methods for fishery resources and random acoustic sampling by fishing boats with different numbers of routes was compared. The results showed that monitoring frequency is a key factor determining estimation accuracy. Under high-frequency monitoring conditions， the absolute error of shrimp biomass estimation by the fishing boat network can approach the level of low-frequency traditional systematic surveys （ε=4）. The increase in the number of fishing boat routes is limited by the spatial overlap between trajectories， and the improvement in accuracy follows a law of diminishing marginal returns. The optimal number of fishing boat routes during the shrimp outbreak period （January） is eight， beyond which the improvement in survey accuracy slows down. This study provides a scientific basis for the construction of a monitoring system that combines "precise monitoring at fixed stations along the embankment of nuclear power plant water intakes" with "wide-area early warning by mobile fishing boat networks in external waters"， thereby enhancing the early warning capability for cold-source organisms in nuclear power plants.

Abstract:

To explore whether optical fiber sensing technology can be applied to monitoring vibration signals generated by small fish targets with body lengths of 10-30 cm in aquatic environments， and to compensate for the limitations of traditional monitoring methods in meeting the demands of modern fishery management， a dual M-Z （Mach-Zehnder） fiber-optic sensing system was established based on dual M-Z fiber interferometric technology. A biomimetic fish with a length of 20 cm and a weight of 90 g was used as the vibration source to experimentally verify the system ability to detect and locate vibration signals generated by the biomimetic fish. Four filtering algorithms were employed to comparatively analyze their noise reduction performance on vibration signals. The filtered signals were then processed using a windowed cross-correlation algorithm to calculate the propagation time delay， thereby achieving vibration source localization. Repeated experiments were conducted at three vibration source positions （10 km， 11 km， and 12 km）， with 20 trials performed at each position， to evaluate the localization capability and accuracy of the system.The results showed that the Savitzky-Golay filtering algorithm improved the signal-to-noise ratio by 4.4363 dB compared with the original noisy signal and achieved a correlation coefficient of 0.7830. For the vibration source located at 11 km along the sensing fiber， the windowed cross-correlation algorithm yielded an offset value of m=6 474， a time delay of τ=51.79 μs， a localization position of 10 821 m， and an error of approximately 179 m. Multiple experiments conducted at the three vibration positions resulted in mean absolute localization errors of 302.04 m， 282.44 m， and 256.63 m， respectively， with localization accuracies of 97.48%， 97.65%， and 97.86%.The results indicate that the dual M-Z fiber-optic sensing system can accurately locate vibration signals generated by biomimetic fish and can provide a novel monitoring approach for modern fishery monitoring and management.

Abstract:

To support the low-carbon and digital transformation of fisheries and to address the strong intercorrelation of vertical dissolved oxygen （DO） profiles in fishing ground prediction，this study focuses on the Dosidicus gigas fishery off Peru in the Southeastern Pacific Ocean and investigates the applicability of objectively identifying key DO layers for intelligent fishing ground prediction. Based on fishing operation data from 2014 to 2021 and DO profile data from 0 to 200 m depth， a self-organizing map （SOM） combined with hierarchical cluster analysis was employed to perform nonlinear dimensionality reduction on 31 vertical DO layers， extract the nonlinear features of the vertical DO structure， and identify representative key layers. On this basis， climate factors， including the Pacific Decadal Oscillation （PDO） and Ni?o 1+2 index， were further integrated to construct Random Forest （RF）， Gradient Boosting Regression Tree （GBRT）， and TabPFN models， and to compare the predictive performance and spatial distribution characteristics of CPUE under full-profile and reduced-feature DO representation schemes. The results demonstrate that： （1） two key DO layers at depths of 23 m and 147 m can be stably identified through vertical DO structure analysis； （2） the reduced-feature scheme enables effective simplification of DO information without a significant loss of prediction accuracy； and （3） the spatial distribution of DO in the key layers is consistent with the high-value areas of jumbo flying squid CPUE， and CPUE is positively correlated with DO in the key layers. These findings provide a reference for intelligent fishing ground prediction and fishing decision-making based on ocean environmental monitoring data.

Abstract:

To identify suitable fishing areas for Argentine shortfin squid （Illex argentinus） in the Southwest Atlantic and to compare the applicability and differences of different types of fishing-vessel activity data in habitat suitability index （HSI）-based fishing-ground prediction， this study selected the Southwest Atlantic as the study area. Automatic identification system （AIS） data and visible infrared imaging radiometer suite （VIIRS） boat detection （VBD） data from January to May during 2017–2023 were used as experimental datasets to represent squid-jigging fishing effort. These data were combined with multi-source and multi-depth marine environmental variables. A boosted regression tree （BRT） model was used to screen monthly key environmental variables and determine their weights. Based on the weighting results， a weighted arithmetic mean model （WAMM） and a weighted geometric mean model （WGMM） were constructed to predict suitable fishing areas from January to May 2024. The prediction results were validated using AIS data， VBD data， and fishery production statistics in 2024. The results showed that the BRT model was able to identify the key environmental factors affecting the distribution of fishing activity. The cross-validated area under the curve （AUC） of the AIS-based BRT model ranged from 0.91 to 0.96， while that of the VBD-based BRT model ranged from 0.81 to 0.83. The variable-selection results from the BRT models constructed using the two types of data indicated that seawater temperature， seawater salinity， and sea surface height were the key environmental factors influencing the fishing activity distribution of I. argentinus. HSI model validation showed that the average proportions of suitable habitats under the WAMM based on AIS and VBD data were 81.01% and 71.15%， respectively， both higher than those under the WGMM. From January to May 2024， highly suitable habitats were mainly distributed in a belt-like pattern along the outer Patagonian Shelf and the shelf-slope zone， with localized high-value areas occurring to the north and northeast of the Falkland Islands. The results suggest that the BRT-weighted WAMM is more suitable for identifying suitable fishing areas for I. argentinus. AIS data are advantageous for characterizing the spatial continuity and overall distribution pattern of fishing activity at the monthly scale， whereas VBD data show higher coverage of high-catch and high-catch per unit effort （CPUE） areas during the middle and late fishing season. This study provides a methodological reference for fishing-ground identification， dynamic monitoring， and the joint application of multi-source fishing-vessel activity data for I. argentinus in the Southwest Atlantic.

Abstract:

Drifting fish aggregating devices （DFAD） have been widely used in tuna purse seine fisheries， and their large-scale deployment may alter fish aggregation behavior and influence the distribution of unassociated tuna schools. However， systematic analyses of the impacts of DFAD on the distribution of unassociated tuna schools remain lacking， and related predictive models are still limited. Based on purse seine fishery data from the Western and Central Pacific Ocean， this study assumes relatively stable fishing efficiency and adopts catch per unit effort （CPUE） as an indicator of the catch level of unassociated schools， proposing a “background field + residual” decomposition framework. First， a machine learning model integrating environmental variables and temporal features was developed to construct the environmental background field of unassociated school CPUE， and the residual was defined as the difference between observed CPUE and the background field. DFAD features with different temporal lags and spatial neighborhood scales were then constructed， and a random forest model was used to identify spatiotemporal feature combinations with greater predictive contribution for residual prediction. Based on the selected features， a spatio-temporal multi-task model was further developed to predict CPUE residuals， and ablation experiments were conducted to evaluate the contribution of DFAD features. The results showed that DFAD features exhibited varying predictive performance at different temporal lags， with features with features at lags of 3， 4 and 5 months showing higher importance than those from the 2nd previous month. The proposed model achieved an accuracy of 71.54% in the residual binary classification task， outperforming the comparison model without DFAD related features. These findings indicate that the proposed decomposition framework can integrate environmental and DFAD features， effectively improving the prediction of anomalous variations in unassociated school CPUE.