Deep Learning Techniques for Underwater Fish Classification

Deep learning algorithms have transformed the field of underwater monitoring by automating the detection and classification of fish species in complex aquatic environments. Convolutional neural networks (CNNs) are the most widely adopted architecture, owing to their ability to learn hierarchical feature representations from raw pixel data. Advanced models incorporate attention mechanisms, residual and squeeze‐and‐excitation modules to enhance discrimination in low‐contrast, turbid waters. Object detection frameworks such as You Only Look Once (YOLO) and region‐based CNNs generate species‐specific bounding boxes, while semantic and instance segmentation techniques provide pixel‐level delineation of fish contours. Recent trends combine motion cues extracted through optical flow with spatial features to address challenges posed by occlusion and dynamic backgrounds. Transfer learning from large‐scale terrestrial datasets has accelerated training on limited underwater imagery, and data‐augmentation strategies simulate varied lighting, turbidity and viewpoint changes to improve robustness. Emerging approaches integrate sonar and optical data streams to exploit complementary modalities in fish passage monitoring. Collectively, these deep‐learning techniques enable reliable species identification, abundance estimation and ecological inference, supporting sustainable fisheries management, biodiversity conservation and habitat assessment on a global scale.

Research from Nature Portfolio

In a pivotal study, automated processing of long‐term cabled video‐observatory data was achieved by evolving genetic‐programming classifiers capable of tracking fish abundance under variable light and turbidity conditions. The resulting system demonstrated strong agreement with manual counts across hourly, daily and monthly scales, offering a template for continuous ecosystem surveillance. A more recent contribution introduced a large‐scale, multi‐habitat benchmark comprising nearly forty thousand underwater images with classification, point‐level and segmentation labels. Evaluations with leading CNN architectures exposed persistent generalisation gaps when models were transferred between distinct habitat types, emphasising the need for adaptive network designs and comprehensive annotation schemes to improve classification accuracy in diverse marine settings.

Research from all publishers

One study presented a two‐stage approach combining YOLO for initial fish detection with a squeeze‐and‐excitation‐enhanced CNN for species classification. By leveraging ImageNet pre‐training and fine‐tuning on local datasets, the methodology delivered high accuracy in temperate marine environments despite limited annotations. Another investigation fused high‐resolution imaging sonar and optical video in a unified deep‐learning framework, integrating an adapted object detector with Kalman‐filter tracking to identify and count multiple species in real time. This multisensor system maintained robust classification under low visibility and high turbidity, illustrating the potential of combined modalities for reliable monitoring in complex fish‐passage infrastructures.

Deep Learning Techniques for Underwater Fish Classification publication trend

The graph below shows the total number of articles in deep learning techniques for underwater fish classification across all publications each year (not limited to Nature Index journals).

Technical terms

Convolutional neural network (CNN): A class of deep learning models that apply convolutional filters to extract spatial hierarchies of features from images.

Squeeze‐and‐excitation (SE) module: An architectural unit that recalibrates channel‐wise feature responses by explicitly modelling interdependencies.

You Only Look Once (YOLO): A real‐time object detection framework that predicts bounding boxes and class probabilities from full images in a single evaluation.

Transfer learning: A technique where a model pre‐trained on one task is adapted to a related task to improve performance and reduce training data requirements.

Semantic segmentation: The process of assigning a class label to every pixel in an image, allowing precise delineation of object shapes.

References

1. Automatic fish detection in underwater videos by a deep neural network-based hybrid motion learning system. ICES Journal of Marine Science (2019).
2. Temperate fish detection and classification: a deep learning based approach. Applied Intelligence (2021).
3. A realistic fish-habitat dataset to evaluate algorithms for underwater visual analysis. Scientific Reports (2020).