The health of our oceans—and the communities and economies that depend on them—hinges on our ability to sustainably manage industrial fisheries. Yet for decades, fisheries managers have lacked timely and reliable data to keep pace with the complexity of modern fishing operations. Electronic monitoring (EM) systems have significant potential to improve fisheries management and transparency by providing independent verification of catch, but the cost and logistics of data review have hindered global uptake of EM. The Nature Conservancy and our partners are addressing these barriers to scaling EM with research and development focused on embedding computer vision into the EM footage review process for longline tuna vessels; the transparency gap in longline fisheries is particularly large with independent observation rates commonly under 5%. This initiative demonstrates how artificial intelligence (AI) can be harnessed not just to streamline data collection on catch activity, but to fundamentally reshape how we manage industrial fishing.
Our goal is to leverage these technologies to move to a world where verified catch documentation is available to fishery managers and companies before product enters the supply chain. By deploying an AI-powered system capable of analyzing EM footage directly onboard longline vessels in near-real time — also called edge computing, a form of AI that makes processing data faster — this initiative brings visibility to one of the most opaque corners in the seafood supply chain.
We prototyped the AI-powered system in the Eastern Tropical Pacific with a computation approach that takes historically siloed components of fisheries operations – e.g. EM footage, electronic logbooks, GPS data, etc. – and makes them interoperable with edge devices that host computer vision models that detect, track, and classify catch brought on board the vessel, establishing a new system that is greater than the sum of its parts (Figure 1). The AI-powered system consists of a modular pipeline deployed on a NVIDIA Jetson computer that processes live footage using a YOLOv11-medium fish detector, the BoT-SORT tracking algorithm, and a rule-based counter to determine whether fish are retained or discarded.
This configuration enables rapid comparison of the captain’s electronic logbook report to the AI predictions for every longline set to produce a fishing risk profile based on alignment between these data sources, and other variables including GPS location and AI predictions of retained sensitive species. Vessels are equipped with Starlink terminals to facilitate daily transmission of a summary report that contains isolated AI predictions of fishing activity, the associated video, and the associated risk profile for prioritized rolling expert analyst verification, emphasizing the importance of scaling Wi-Fi use at sea (Figure 2). Importantly, the edge solution architecture eliminates the need to ship hard drives to review centers and significantly reduces the volume of raw video sent over satellite links to cloud data centers. By transmitting only the small subset of frames associated with catch events detected on the edge, we reduce transmission costs and improve the carbon efficiency of the solution by lowering energy-intensive video data transfer and reliance on centralized compute infrastructure.
The end-to-end AI-powered system architecture focuses on the most operationally valuable task—accurate count of catch brought onboard—achieving 98% precision, 94% recall, and 6% miss rate relative to the ground truth dataset, which is comprised of footage selected to test model performance under strenuous conditions (Figure 3). A more fine-grained evaluation of model performance of species classification shows strong detection performance for several target species, including 100% recall for Yellowfin tuna and mahi-mahi, while lower precision and recall observed for Indo-Pacific sailfish, Swordfish, Striped marlin, and Blue marlin (Figure 4); billfishes likely reflect inter-species confusion within the billfish group, whose similar visual characteristics made them difficult to distinguish.
The implications for conservation are profound. With species-level catch counts that rival expert human reviewers and risk profiles delivered within hours of fishing activity, managers are no longer forced to rely on delayed, incomplete, or inaccurate self-reported catch data from logbooks. Instead, they can act swiftly to enforce quotas and deter illegal, unreported, and unregulated fishing. Beyond compliance, these rapid insights open the door to more dynamic management strategies—such as adjusting effort limits based on catch trends, identifying high-risk fishing zones for targeted monitoring or spatial planning, and coordinating with market actors to align supply chain decisions with sustainability goals. In short, the system equips decision-makers with the timely, granular intelligence needed to move from data-poor management to proactive stewardship.
The system is designed to work alongside human experts, not replace them. EM reviewers remain in the loop to validate AI predictions, ensuring that final catch assessments are grounded in professional judgment. This design optimizes between machine efficiency and expert oversight to solve for optical challenges encountered by the computer vision model, such as high occlusion rates that impact the model’s ability to consistently track discards and species classification complexities within the billfish group. The daily reports include relevant clips of footage to allow human reviewers to rapidly verify species ID and whether a given catch event is retained or discarded. By handling the initial, time-intensive review of EM footage for catch events, the system allows managers to quickly verify AI outputs and redirect their attention to on-the-water improvements. By combining machine efficiency with expert oversight, the system delivers both scalability and accountability, reinforcing trust in the data that underpin fisheries management.
The Nature Conservancy and our partners demonstrated that actionable data can be generated in minutes and delivered to managers daily, a stark contrast to legacy EM programs that might deliver EM data review months after the completion of a fishing trip. Delivering actionable insights at this speed has implications not only for management effectiveness, but also for the economics of electronic monitoring programs—prompting a closer look at what it could cost to adopt this technology and how those costs may evolve over time.
We partnered with CEA Consulting and Working Ocean Strategies to evaluate the potential cost benefits of integrating edge AI into EM programs for longline fisheries. Their analysis suggests that the greatest direct value of edge AI comes from reducing the cost and effort of video review—one of the most resource‑intensive components of EM. Under reasonable assumptions, deploying edge AI across a 100‑vessel longline fleet could generate meaningful EM video review cost savings compared to traditional EM systems, with potential net benefits on the order of tens of thousands of dollars per vessel and up to USD $2 million across the fleet over a five‑year project lifetime. Because this is an emerging technology, the analysis explored a wide range of cost assumptions—including conservative, upper‑bound scenarios designed to test financial performance under different assumptions. In some scenarios with higher AI development and transfer learning costs (i.e., >USD $20k/vessel), the net benefits could become negative. While these higher-cost scenarios are considered lower‑probability—especially when considering the likelihood that AI development and transfer learning costs will come down over time—they underscore the importance of better understanding the real‑world costs of adapting edge AI to new fisheries and scaling it across fleets. To address these remaining uncertainties, The Nature Conservancy is advancing a next phase of work focused on measuring actual transfer and scaling costs, helping ensure that future EM programs can adopt edge AI with clear expectations and confidence in both financial and operational outcomes.
Taken together, the strong AI performance and early economic signals underscore the importance of making the system publicly available—so that stakeholders can adopt and adapt the technology with confidence. Its modular architecture and reliance on open-source tools enable it to be reconfigured for different target species and monitoring goals. All components, from the models to the reporting logic, were designed with reuse and scalability in mind to lower barriers to adoption and enable broad, equitable access across regions and sectors.
By making the solution freely available, The Nature Conservancy and our partners aim to advance equitable access and empower global longline fisheries operations to harness AI-enabled electronic monitoring for smarter, more sustainable management practices. To this end, we have rigorously documented the development process to enable stakeholders to quickly adopt the solution stack. To get started, visit: nature.org/edgeai
Over the past three years, The Nature Conservancy had the privilege of collaborating with a broad group of partners to develop and test this prototype. In the first phase, technical partners—Thalos, Bureau Veritas, ProductOps, Ai.fish, OnDeck, Integrated Monitoring, InMotion, Wholechain and Deckhand—laid a strong foundation for a highly successful second phase where Tryolabs built the full-stack solution that is now publicly available. Instituto Costarricense de Pesca y Acuicultura (INCOPESCA) and Cámara Nacional de la Industria Palangrera (CNIP) helped shape the project by embracing state‑of‑the‑art technology and supporting project design and engagement, and the Patrick J. McGovern Foundation was an essential collaborator on this endeavor. Importantly, this work would not have been possible without the vessel owners, captains, and crew who carried the effort forward at sea, committing their time, expertise, and trust. Thanks to the dedication of many, this application of modern AI earned recognition from the Bezos Earth Fund, where it was selected as a winner of their AI for Climate and Nature Grand Challenge. This award – alongside continued collaboration with the Patrick J. McGovern Foundation – is supporting The Nature Conservancy’s efforts to scale the AI-powered system to the Western and Central Pacific Ocean, the heart of the planet’s tuna fisheries.
In an era of accelerating environmental change and growing demand for sustainable seafood, this AI-powered system demonstrates meaningful advancement in fisheries monitoring and management. It’s a promising step toward a future where EM programs deliver near real-time insights to markets and fishery managers before products enter the supply chain – creating an environment where verified, on-the-water activities and actionable data are not just an aspiration but a standard we achieve together.
-Vienna Saccomanno, Senior Scientist, Large-Scale Fisheries Program, The Nature Conservancy
-Ben Gilmer, Director, Large-Scale Fisheries Program, The Nature Conservancy
