With the year 2030 looming on the horizon, efforts to achieve the United Nations Sustainable Development Goal 14.4 are a focus for fishing nations around the world. Global concerns that 1 in 5 wild-caught fish originate from illegal, unreported, and unregulated (IUU) fishing remain. Rapidly advancing technology presents a significant opportunity for cost-effective fisheries management with electronic monitoring (EM) supported by artificial intelligence (AI) being a strong candidate for widespread implementation. However, a large gap exists in knowledge to implement these technologies successfully with a view to future resilience within the program design.
EM systems consist of one or more cameras rated for operation in marine environments and the harsh conditions sometimes experienced by fishing vessels at sea, connected to a data recording device that captures and stores video footage of critical operational areas where fish and gear are handled. Video is recorded either continuously or based on activity triggers, while vessels are at sea. Outside of regulatory limitations and questions, there are a number of cost challenges that contribute to relatively low rates of global adoption of EM systems including: the sheer volume of video data that is generated and requires review; the cost of review services; and the cost of the systems themselves. Additionally, challenges can arise due to the typically long feedback loop from the technology; as video review often requires offboarding hard drives for human review upon a vessel’s return to port, there can be delays to analysing the data and assessing relevant information contained within the footage.
AI offers a promising improvement to the time and cost of video review as well as the potential for near real-time data through onboard low-powered computing technology integrated within the EM system. This opportunity is expanding its reach rapidly with technology advances occurring worldwide, but significant gaps remain for such technology to be successfully commercialized. Further, many fisheries examining the opportunity for technology integration within their fishery management approach lack technical knowledge to consider and evaluate AI opportunities, let alone select vendors and implement such programming.
Following a number of collaborative projects examining diverse opportunities for real impact in the fishing industry through AI-assisted electronic monitoring, Teem Fish Monitoring and Ai.Fish will present best practice guidance to support fisheries considering the adoption of this technology with planning and successful implementation in the hopes that this closes a key gap blocking wider implementation of the technology for global benefit.
An Overview of Project Experiences
Founded in 2019, Ai.Fish is a Hawaii-based startup with a focus on the application of computer vision in marine conservation and commercial fisheries. Computer vision is a discipline within AI focused on the understanding and interpretation of still images and video through machine learning approaches. Specializing in customized AI pipelines, software applications, and cloud-based or edge-computing data analysis, we are interested in technology integration in the fishing industry that maintains or enhances fisher livelihoods while ensuring robust data and surveillance to support responsible and sustainable fishing for generations to come.
We have conducted projects examining AI implementation opportunities in fisheries in Hawaii, Alaska, Oregon, and New England as well as international projects in Canada and Costa Rica. We have examined methods to support data pre-processing to reduce storage demands and costs; AI pipelines supporting electronic monitoring review in longline and trawl fishing; and onboard real-time analysis in artisanal and commercial fisheries. We have overcome one of the key barriers to specialized AI development for fisheries application through the development of a world-leading real-world data collection comprising nearly 3 million annotations of fish, humans, and gear. This vast collection is now also supported by a robust annotation pipeline that supports rapid dataset generation as new fisheries opportunities emerge. Recent research and development has included examining opportunities for AI to support monitoring of discarded catch and interactions with endangered, threatened, and protected (ETP) species.
Teem Fish Monitoring Inc (Teem Fish) is a fisheries monitoring service and technology provider based in Canada that delivers advanced electronic monitoring (EM) solutions for capturing and reporting reliable, accurate and verifiable fisheries data. Unlike other EM solution providers, Teem Fish grew from a social enterprise that leverages advanced technology and local partnerships to help fish harvesters with the safety, security and compliance requirements of their jobs creating an affordable way to scale the use of electronic data capture to ensure ongoing viability of global fish stocks. Using compact, robust, in-house built-for-purpose world class cameras and custom edge-AI enabled data capture systems, Teem Fish collects on-vessel footage and sensor data in a very unobtrusive way, only looking at the required fishing areas necessary for the specific data collection goals. In addition to this, Teem Fish works with world leading AI providers to provide a variety of machine learning and data automation including, but not limited to, fishing event detection, fish size, species identification and by-catch monitoring.
Together, Ai.Fish and Teem Fish have collaborated on 5 project conceptualizations and/or deliveries including novel AI conceptualization and development as well as initial EM implementation.
Lessons Informing Best Practice Guidance
Throughout our project implementations some criteria have emerged that strongly influence the success of EM program implementation. We believe every fishery examining technology integration opportunities should consider these factors as they evaluate the opportunity.
1. Implementation will be most successful when the program is co-designed with industry
While many regulatory entities may view this as challenging, engagement with industry ensures program design considers fishing practices, operational limitations, and mutually beneficial goals to support cost-shared implementations. It is important to remember that the operations and handling on deck will directly impact the quality and effectiveness of the data to be analysed from the EM system collection. Since EM systems are typically used in replacement of scientists on board, it is imperative to understand the realities and limitations of the on-deck operations on the intended data collection, and therefore essential to include the on-vessel personnel in the design process to ensure on-deck realities align with the data priorities.
Co-design with industry has served as the basis of most every EM/AI project to date. In New England, industry collaboration was instrumental in designing sub-sampling fish handling protocols that balance the need for time-consuming, high-touch handling of regulated species for scientific data with on-deck realities. Through collaborative protocol development, we were able to reduce the undue burden of excessive fish handling for every individual discard, while maintaining a method of high quality data collection that we are further accelerating with AI.
2. Involve vendors early in the program lifecycle
Vendor engagement during program design can help to optimize the program’s implementations towards achievable objectives and minimize the program learning curve. The EM vendor can provide important guidance for equipment selection, installation, and configuration to ensure that footage captures everything to fulfill the program objectives from Day 1. The AI vendor can provide insights into incremental steps the program can undertake to acquire data needed to train models and can help to elicit and develop the most impactful use cases for technology integration across stakeholder needs.
Allowing this cross-vendor collaboration, similar to the above note on cross-industry collaboration, ensures that collected data is usable and optimized for all parties from the start. This was noted as a particularly valuable learning for work on-going in the Eastern Canadian fisheries, where collaboration between AI, EM, and review staff vendors is allowing for camera adjustments early and throughout projects to ensure that camera position and angles are adequate for both human and AI review and avoid lost time due to un-usable camera angles.
3. Examine program alignment with fleet patterns; pilot vessels are only helpful if they are representative of typical fishing operations
Particularly for the incorporation of AI, variation in fishing practices, handling of fish onboard vessels, vessel configurations, and vessel sizes/layouts should all be considered at the time of AI design. It is ideal if the fishery can minimize these variations, for example, having all vessels implement handling protocols, but if variations will persist then they need to be represented in the footage supporting AI development as well as testing parameters. The AI development process can be better designed to accommodate these challenges if the AI vendor is involved early in the project and educated on the realities of the fleet in order to assist with understanding what parameters are most variable or consistent across the fleet in order to consider the most representative vessels for participation.
Being prescriptive in vessel participation is not always an option, of course, which is why understanding the specific variables is key. If exploring EM for a multi-gear fishery, it may be best to be selective or utilize a phased approach by gear types, such as we are currently experimenting with in the New England groundfish fishery. For this fishery, there is the potential for up to 5 different gear type choices for any given vessel. To minimize complexity and get an understanding of opportunity for AI, we began our AI assessment focusing on trawl gear first, with consistent fish handling on-vessel to account for variations in vessel layout and therefore camera placement. Alternatively, for opportunity assessments in the longline fishery where vessels are similarly laid-out, focus was instead put on getting similar camera angles and view points across vessels so the AI could understand the process flow a bit easier and there was less need for standardized fish handling on-vessel. This can certainly vary, but the focus here is on understanding patterns.
4. Align program objectives and EM program metrics
Every EM Program will require a video reviewer protocol that ensures standardization in video review and the production of high quality fishery metrics. Depending on the use case, AI may or may not introduce a higher error rate than human review. Consider whether the thresholds you set for errors are feasible. Remember that highly accurate results are possible, but are often achieved with trade offs in processing cost, processing time, dataset development (larger datasets may be required). If you have a 6 month timeline for your initial project report, don’t set an accuracy target of being within 2% as it will be difficult to deliver. If your program has the objective to count discards, don’t also expect that reviewers quantify target catch as this could dramatically increase the review effort. If you want species detail in the AI results, don’t try to implement a 1-camera system on a 65 foot vessel as close-up images to see species differentiators will be required.
Regardless of the agreed acceptable thresholds, the priority in this case should be around focusing on the intended outcome, and co-designing the protocols to allow for flexibility while keeping that goal in mind throughout the data collection process. It should be understood that adjustments will be necessary, but by keeping AI in mind at the start of the project, you can better balance those compromises with intent. This is currently being highlighted through our collaborative work on discard detection in New England, where historical data that was collected prior to AI being considered for the project is proving less valuable for training than more recent data because of complexity in differing camera set ups between vessels and camera angles at installation only considering human observation rather than AI support and therefore the pool of available historical data that is valuable is quite variable.
5. Define data retention and data management expectations up front
At the beginning of the EM program implementation it can be tempting to think that data retention and management are discussions that can be postponed. But these expectations can affect both system design and system implementation. For example, how much data needs to be saved and stored on a per trip basis might dictate hard drive size or video compression settings. If you want to optimize data storage, you might choose to pre-process video with an activity detector to remove video segments of little to no interest (e.g. segments without fishing activity), but choosing to delete these is a permanent data retention choice that needs to be informed by stakeholder expectations. If hard drives will fill up due to trip duration, protocols need to include the exchange of hard drives during vessel turnaround which may require careful planning, or it might drive system design with redundant or backup drives.
For example the data capacity required for a Hawaii longline vessel is quite different from that of a New England groundfish vessel. This is largely because 24/7 video recording for 21 days requires a lot more memory than for 24 – 36 hours.
Best Practice Pathway to Developing Fisheries of the Future
If your fishery is considering an EM implementation with AI optimization, based on our experience here’s the checklist/procedure we recommend you follow:
✓ Identify an industry collaboration for fishery sustainability improvement
✓ Select or tender for an EM vendor and AI partner, this could include market engagement activities prior to tendering- if AI is to be incorporated, it should be considered from the initial design phase
✓ Form a co-design committee with representatives from the EM vendor, AI partner, fishery/industry representatives, and regulatory representatives
✓ Define fishery monitoring objectives and desired metrics to be monitored
✓ Examine fleet characteristics such as vessel sizes, crew sizes, catch handling, vessel configuration, vessel age, and vessel connectivity.
✓ Draft an EM program protocol, with flexibility for adjustment throughout the learning phases
✓ Define AI-assist opportunities and priorities within the EM program protocol
✓ Conduct a vessel survey of one or more participating vessels to evaluate technical implementation, keeping in mind a goal of representative diversification
✓ Select technical configuration and secure system components required to suit vessel characteristics, program objectives, EM and AI protocols
✓ Install 1 – 2 initial vessels and run a pilot trip to test that all data collection and handling is as expected.
✓ Retrieve pilot trip data and run a sample evaluation following the EM program protocol
✓ Assess results against program objectives and metrics. Adjust as may be required and run additional pilot trip(s).
✓ Following additional successful pilot trip(s), roll-out installation to additional vessels and scale the program.
✓ Collect required footage on an ongoing basis to develop the AI dataset
✓ Train or fine-tune candidate algorithms based on the dataset development and objective refinement
✓ Implement AI analysis and assess result outputs against program objectives and metrics. Adjustments and iteration may be required to achieve program objectives or metrics.
✓ Provide quarterly reporting to the co-design committee on an ongoing basis to maintain the project and stakeholder alignment.
In summary, attention to the above guidance and recommendations can significantly reduce time, costs, and confusion in implementing AI-assisted electronic monitoring (EM) in fisheries. By emphasizing early and collaborative program design—including fishers, AI vendors, and EM providers—projects are more likely to avoid rework, system misalignments, and unusable data, thereby saving time and money. Additionally, tailoring EM system installations and AI analysis to match actual fishing practices streamlines deployment and reduces training and adjustment costs. Standardizing protocols and clearly defining objectives early on minimizes confusion and rework during review or AI model training, which can be both resource-intensive and time-consuming. Through clear planning and collaboration, EM programs can benefit from scalable, effective AI integration—ultimately enabling a more efficient and sustainable path to fisheries monitoring.
Jillian Di Maio is the Fisheries Director for Teem Fish Monitoring, Inc. Jimmy Freese is co-founder and CEO, and Megan Johnston is Director of Operations at Ai.Fish.
Industry Focus is an ongoing series of informational articles profiling companies involved in electronic monitoring and other fishtech pursuits. To include your company, please drop us a line at info@em4.fish.
