Machine learning in marine science is helping researchers process growing volumes of imagery and acoustic data from surveys, underwater cameras and fishing-vessel monitoring systems. These methods can automate tasks such as detecting objects, classifying species and measuring fish. Their value, however, depends on more than algorithm choice: data quality, model validation and integration into the working process all shape whether automated analysis can support reliable scientific decisions.
Why marine science is turning to machine learning
Advances in sensors, imaging systems and autonomous platforms have expanded the amount of data collected at sea. Manual review remains important, but it can become a bottleneck when datasets are too large to analyze promptly using conventional methods. Machine learning (ML) offers a way to automate parts of that analysis and help turn collected data into usable information.
In the workshop report, ML describes computational methods that learn patterns from data rather than relying on a predetermined set of instructions. In supervised learning, examples are labeled with the desired output, such as an image marked with the species present. Unsupervised learning looks for patterns in data without those output labels, for example by grouping similar observations. Deep learning, including convolutional neural networks (CNNs), is a subset of ML used in many image-analysis applications.
These approaches are not interchangeable, and more data does not automatically produce a better model. Training data must be relevant to the task, and performance needs to be tested independently. That is particularly important when model outputs may contribute to abundance estimates, stock assessments or management decisions.
Machine learning in marine science: practical applications
Automating image and video review
Computer vision systems can help locate and classify organisms in still images and video. The open-source Video and Image Analytics for Marine Environments (VIAME) toolkit was developed to support marine and fisheries image analysis. Its pipeline can bring together tasks such as object detection, tracking and classification, along with tools for visualizing and evaluating results.
Applications described in the report include processing imagery from scallop and reef-fish surveys, identifying seals in aerial imagery and analyzing fish captured by underwater cameras. In these workflows, automated annotations can reduce the amount of material requiring manual review. They can also provide a basis for human checking and subsequent improvement of the training data and models.
Analyzing acoustic survey data
Acoustic surveys create another significant analysis challenge. Echosounder data are commonly reviewed to identify fish schools and allocate acoustic signals to species or species groups. Researchers have explored deep-learning segmentation models that classify image pixels as fish or background. Species-level classification remains more difficult: differences in instrumentation settings or inconsistent annotations can make datasets less uniform, affecting model development.
This illustrates why a model’s output must be considered in the context of the data collection process. A classifier trained on one set of surveys may not perform as expected on data collected in a different area, at another depth or during another season. Testing how a model generalizes beyond its training data is therefore a practical requirement, not just a technical detail.
Supporting electronic monitoring
Electronic monitoring systems use cameras and sensors to record fishing activity and catch. ML is being investigated to automate video review and tasks such as species identification, activity detection and fish measurement. The aim is to make review more efficient and extract information in a timely way. Reliable performance depends on training examples that represent the images and conditions encountered in the fishery, as well as standardized data collection and metadata.
Build the workflow around data and validation
Successful ML implementation involves a chain of connected activities, from collection through interpretation. Teams need to consider how data will be stored and accessed, how observations will be labeled, where processing will take place, and how results will enter existing analysis tools. A technically capable model may still be difficult to use if data formats, software components and operational procedures do not work together.
Several priorities follow from the workshop’s findings:
- Improve data accessibility. Consistent metadata and accessible datasets make it easier to develop and assess models. Sharing non-confidential data can also support broader collaboration.
- Develop representative training data. Expert-reviewed labels help teach models what to detect or classify. Synthetic data and methods that use unlabeled data may help address shortages, but their suitability needs to be assessed for the specific task.
- Validate independently. Test data should be used to evaluate performance apart from the data used to train the model. The acceptable level of accuracy and precision depends on how outputs will be used.
- Plan for workflow integration. Teams must connect data handling, model execution, review and reporting. Modular tools can support different tasks, but the overall process still needs to be coordinated.
- Account for transparency. Some deep-learning models can be difficult to interpret. Where users need to understand why a model produced a result, model explainability and human review should be considered.
Tools and partnerships help build capability
Open-source software and shared libraries can make ML methods more accessible and support collaboration. The report describes hands-on training in GitHub and VIAME, as well as the use of programming languages and ML frameworks. Cloud resources may also support processing, although organizations need to consider data access, infrastructure and workflow requirements when choosing where computation should take place.
Partnerships among scientists, software developers, government agencies, academic institutions and private-sector organizations can bring together domain knowledge and technical expertise. Such collaboration is especially useful when developing labeled datasets, adapting tools to a particular application and establishing a process for model updates. Workforce training and scientific exchange also help teams assess evolving methods and maintain confidence in the resulting scientific products.
Conclusion: treat ML as part of the measurement system
Machine learning can help marine science teams address the growing volume of imagery and acoustic observations, from survey analysis to electronic monitoring. The practical lesson is to treat ML as one part of a broader measurement and analysis system. Data quality, representative labels, independent testing, clear workflows and appropriate human oversight all contribute to whether automated outputs are useful. With those elements in place, ML can support more timely analysis while preserving the reliability needed for science-based decisions.
Access Official Technical PDF