Coastal fish populations are increasingly monitored using underwater cameras, which provide rich, non-invasive data on marine life. However, analysing these videos has traditionally required human experts to manually identify, count, and sometimes size fish, frame by frame. This is time-consuming and costly, and it limits how widely camera-based monitoring can be deployed. The CoastVision project set out to address this bottleneck by developing computer vision methods that automate the analysis of underwater images and video using modern artificial intelligence.
Over the course of the project, CoastVision has demonstrated how deep learning can be used to extract detailed biological information from underwater video at scales not previously possible. The project developed methods to automatically detect and classify multiple species of coastal fish with expert-level accuracy, based on large, carefully curated image and video datasets collected along the Scandinavian coastline. These datasets are openly available and form an important resource for future research and innovation.
A central focus of CoastVision has been the development of deep learning methods for (re-) identifying individual fish based on their natural visual “fingerprints”, such as body shape and colour patterns. This approach makes it possible to follow the same individuals over time without physical tagging, opening new opportunities to study growth, survival, behaviour, and population dynamics in the wild. CoastVision has produced unique long-term image datasets of Atlantic cod, corkwing wrasse, and ballan wrasse, and has developed re-IDentification methods that account for natural changes in appearance over time.
As the project progressed, rapid advances in artificial intelligence created new opportunities that were incorporated into the work. CoastVision therefore expanded its scope to include automated classification of nesting behaviour in corkwing wrasse, and to develop models that can distinguish sexes even when this is not possible for human observers. This is because some males develop as female mimics, which effectively deceive and confuse the scientists, whereas the computer vision achieves accuracy above 95%. Accurate sexing is crucial both in stock assessment and studies of mating behaviour.
Beyond individual methods, CoastVision has laid the groundwork for fully automated pipelines that can detect, track, and identify individual fish in long-term video monitoring systems. These tools are now being integrated into large-scale coastal monitoring programmes in Norway, supporting a transition from labour-intensive manual video review to scalable, standardised analysis. This will improve our ability to monitor biodiversity, detect changes in fish communities, and respond to environmental pressures such as climate change.
By combining marine ecology with state-of-the-art artificial intelligence, CoastVision has contributed new tools, datasets, and expertise that significantly advance camera-based monitoring of coastal ecosystems. The project has strengthened interdisciplinary collaboration and laid the foundation for increased application of computer vision in marine research.
The CoastVision project has delivered several of the key outcomes anticipated in the original grant application and has also generated additional results with broader and longer-term impact.
A major outcome of the project is the development of large, high-quality, labelled image datasets of shallow-water coastal fish species collected from underwater camera surveys. These openly available datasets provide a critical foundation for further development and application of computer vision methods in marine monitoring. The datasets have enabled CoastVision to developed computer vision methods for automated detection, tracking, and re-identification of fish in underwater imagery. In particular, the project has made considerable progress in demonstrating how deep learning, combined with biological information, can be used to re-identify individual fish among several hundred candidates based solely on natural visual characteristics.
Although the current pipeline for fish detection and quantification still requires refinement before full operational deployment, it represents a key step towards automated analysis pipelines that will substantially reduce the manual effort required to process underwater video data. Continued development of the pipeline is supported by the Institute of Marine Research and is expected to result in fully automated analysis by 2027. This will support a fundamental shift from labour-intensive expert review towards scalable, standardised analysis, allowing monitoring programmes to reallocate effort from basic data processing to interpretation and decision support, and to deliver results more rapidly to management and policy contexts.
While fully automated sizing on stereo imagery was deprioritised during the project, the methods developed for detection, tracking, and individual identification provide a strong foundation for future integration of size estimation and other phenotypic traits. In addition, the project expanded beyond its original scope to include automated classification of behaviour and sex, enabling new ecological analyses such as sex-specific behaviour, reproductive strategies, and fishing mortality. Several of the methods developed in CoastVision are now being integrated into large-scale coastal monitoring programmes in Norway. In the longer term, widespread adoption of camera-based monitoring combined with computer vision has the potential to transform marine biodiversity monitoring.
Beyond specific tools and datasets, CoastVision has strengthened interdisciplinary collaboration, contributed to the development of multiple follow-up research proposals to national and international funding bodies, and built lasting competence at the interface between marine ecology and artificial intelligence, supporting sustained scientific and societal impact beyond the project period.
Effective monitoring and management of coastal ecosystems is limited by observation methods. Underwater cameras are increasingly being used to monitor and study coastal fish communities; a major bottleneck for upscaling their use is dependence on human experts for image and video analysis. CoastVision will use the power of deep learning to refine and extend a computer vision pipeline for detecting, classifying and sizing the key fish species in shallow water coastal ecosystems, facilitating a transition to fully automated video analysis. Our models will be trained on data sets from several different surveys, ensuring cost-efficient development of routines that will be widely applicable. Computer vision for re-identifying (re-ID) individuals solely based on their unique visible features will also be developed. This novel aspect of CoastVision could ultimately provide new opportunities to obtain detailed knowledge about behaviour and population dynamics in wild fish populations, with minimal negative impact on animals and habitats and at a low cost. Our focal species for re-ID are Atlantic cod, ballan wrasse and corkwing wrasse, commercially important species with complex, high-contrast skin patterns. To generate the necessary training data for re-ID we will use synchronized radio frequency identification and camera systems. CoastVision’s automated video analysis pipeline will be integrated into ongoing ecosystem surveys and case studies whose main objective is to better understand the factors that affects the reproduction, recruitment and survival of commercially important coastal species. As such, CoastVision will contribute to independent, but complementary, research objectives. The project will advance the international research front for applied machine learning in marine ecology, which ultimately can revolutionize our ability to observe, understand and respond to ecological change at scales far more refined than is currently possible.
