Bycatch of protected species in fisheries

Small-scale gillnet fisheries, in particular, pose a high risk, but monitoring these interactions is challenging due to the sporadic nature of bycatch events.

Sweden’s electronic monitoring (EM) programme, led by the Swedish University of Agricultural Sciences (SLU), has equipped over 17 vessels with onboard camera systems to document fishing activity and Pets bycatch. However, manual review of this extensive video data is time-consuming, costly, and raises privacy concerns. This project develops advanced machine learning (ML) methods to automate bycatch detection, improve data quality, and make EM a scalable, efficient, and privacy-safe tool for monitoring fisheries’ impacts on protected, endangered, and threatened species.

This is what we want to achieve

Our overall goal is to enhance EM-based monitoring of protected, endangered, and threatened species bycatch in small-scale fisheries, enabling faster, more accurate, and more secure data analysis. Specifically, we aim to:

• Develop and refine ML models to automatically detect and identify bycatch events involving fish, seabirds, and marine mammals in EM footage.
• Optimise data handling by creating models that identify hauling operations and assess video quality, allowing irrelevant footage to be excluded before analysis.
• Expand and diversify training datasets through collaborations with international partners and by generating new images from both real and artificial bycatch events.
• Implement federated machine learning (FedML) to enable institutions to jointly train models without sharing raw data, protecting privacy while improving model accuracy.
• Integrate ML tools into daily workflows, providing analysts with intuitive interfaces to support efficient bycatch review and validation.

The expected outcome is a significant reduction in the time, cost, and data storage demands associated with EM analysis—ultimately enabling broader monitoring coverage and better-informed management of protected, endangered, and threatened species in Swedish and European waters.

This is what we are doing

Building on SLU’s established EM programme, the project is structured into four interconnected work packages (WPs):

1. WP1 – Data Collection: Expands current bycatch training datasets (~10,000 annotated images) by harmonising data standards between SLU and DTU, ongoing collection of new footage from EM-equipped vessels, and recording artificial bycatch scenarios during marine mammal release events. Additional training data will also be sourced through collaborations and systematic searches for public and private image collections.
   
   
2. WP2 – Model Development: Retrains and improves existing ML models using the expanded dataset to enhance bycatch detection accuracy. It also develops models for identifying hauling operations, combining EM footage with vessel GPS data to reduce unnecessary video storage and processing.
   
   
3. WP3 – Federated Machine Learning (FedML): Establishes a decentralised training network across partner institutions, allowing each to train models locally on their own EM data. Aggregated model updates are then shared to produce a powerful global model—improving accuracy while ensuring data privacy and GDPR compliance.
   
   
4. WP4 – Practical Usage: Integrates ML models into a user-friendly analysis interface, enabling automatic registration of bycatch events. Analysts will verify rather than manually review footage, supported by a detailed methods manual and training video to ensure smooth implementation in future EM workflows.

Together, these efforts will create the technological foundation for large-scale, privacy-aware, and cost-effective bycatch monitoring—supporting sustainable fisheries management and the conservation of protected marine species.