Electronic monitoring data requirements: what to collect, why it matters, and what AI changes

In brief

• EM requirements should start from management questions, not camera specifications.
• Video, sensor and review metadata must be synchronised into event-level evidence with provenance.
• Computer vision can reduce review cost, but sampling design is what determines whether the resulting evidence is trustworthy.

NOAA defines electronic monitoring (EM) broadly: video cameras, GPS, transmitters, gear sensors and automated image processing can be combined to capture fishing location, activity, catch, discards and protected-species interactions. That definition is useful because it prevents EM from being reduced to “put a camera on a boat.”

The data requirement is determined by the decision requirement.

Start with the question

Different objectives need different evidence:

Catch accountingspecies/taxon, count or weight, retained/discarded state, event time, review provenance.

Bycatch / ETPinteraction event, condition, handling, disposition, coverage and missed-event audit.

Compliancegear deployment, spatial/temporal restrictions, retention/discard behaviour, chain of evidence.

Operational analyticsreview workload, event density, equipment health, data latency and failure rates.

A camera layout, frame rate or retention policy should be justified against those objectives.

Minimum synchronised data

For each sensor stream I propose to require a common clock and immutable identifiers:

• vessel and trip ID;
• device/camera/sensor ID and firmware/configuration;
• UTC timestamps with drift/health checks;
• video/image segment URI plus integrity hash;
• GPS/VMS/AIS position where legally/operationally appropriate;
• gear/activity sensor events;
• event start/end and event type;
• machine output: model/version, label, confidence/uncertainty;
• review output: reviewer/version, correction, disposition;
• retention/access classification.

The objective is not to maximise schema size, but to ensure that every reported claim can be traced to its source evidence, processing version and review history.

Review design is part of the data design

The largest recurring costs in EM programmes include manual video review, transmission and storage. NOAA explicitly identifies computer vision and machine learning as potential ways to reduce those costs while extracting species, weight/length, gear counts and activity.

But a model that filters 95% of video also changes the statistical sample seen by humans. That means the review policy must contain a random audit stream in addition to AI-prioritised review. Otherwise the programme may become very efficient at reviewing what the model already knows how to find.

The fisheries EM literature makes the same point from a programme-design perspective: coverage allocation across vessels, space and time is fundamental to information quality, and optimisation of review must preserve the monitoring objective.

What deep learning changes

Deep learning shifts EM from “record then manually watch” toward detect → aggregate → prioritise → verify.

The model stack can include:

1. object detection / segmentation for catch and gear;
2. tracking to connect repeated views of the same object;
3. temporal models to turn detections into interaction events;
4. representation learning for species similarity and domain shift;
5. calibrated uncertainty to decide what requires human review.

The hard problem is generalisation. Cameras, vessels, decks, weather, catch handling and species mix all change. A deployment claim should therefore include vessel-held-out and temporal evaluation, not only random frame splits.

Governance and retention are requirements, not afterthoughts

EM produces sensitive operational imagery. NOAA has published specific policy on EM record retention and information law; federal EM imagery has a five-year records schedule when it becomes a federal record, and confidentiality/FOIA handling is explicitly addressed. Other jurisdictions will differ, but the system should treat retention class, access control and deletion date as data fields.

For standards-driven fisheries, MSC’s Evidence Requirements Framework focuses on information quality and how evidence is collected, reported, handled and analysed. In October 2026 the framework is under active revision, including proposed changes to independent-observation expectations, so product requirements should be versioned against the applicable standard rather than hard-coded around a single coverage percentage.

Impact metrics

I propose to report impact on two axes:

Monitoring quality

• event recall and false-alert rate;
• unflagged-event audit recall;
• coverage by vessel/trip/season;
• reviewer agreement and calibration;
• evidence completeness.

Operational efficiency

• review minutes per hour of footage;
• time from landing to validated record;
• storage/transmission per trip;
• reviewer corrections per event;
• system/device downtime.

The claim that matters is not “AI processed X hours of video.” It is: for a defined monitoring objective, we produced an equally or more reliable evidence set with lower cost or latency, and we can quantify what was still missed.

References and white papers

• NOAA Fisheries — Electronic Monitoring: programmes, policy, retention and ML direction.
• NOAA Fisheries — Electronic Technologies.
• Optimising the review of electronic monitoring information for management of commercial fisheries (2024).
• Research progress on electronic monitoring in tuna longline fisheries, Frontiers in Marine Science (2026).
• MSC — Evidence Requirements Framework revisions (2026).