Trip Sampling Doctrine: Fisheries Monitoring Without EM, With Conventional EM, and With AI Analytics
In brief
- Regime 1 (No EM): Severe observational scarcity and observer Hawthorne effects require paired statistical audits and risk-stratified trip sampling.
- Regime 2 (Conventional EM): Massive video review backlogs demand sensor-gated activity triage to decouple 100% camera recording from human review capacity.
- Regime 3 (AI + Analytics): Avoid full AI automation with human in the loop review; edge models structure 100% of footage into events, sampling verification via random controls, uncertainty, and disagreement.
Fisheries monitoring has historically been governed by what is realistically observable, across different fleets and regulatory jurisdictions, monitoring technologies fall into three distinct evolutionary regimes:
- Monitoring without EM devices: Relying entirely on intermittent onboard human observers, captain self-reporting via logbooks, and dockside inspections.
- Monitoring with conventional EM: Operating continuous onboard cameras and vessel activity sensors, but relying on manual post-trip video review.
- Monitoring with EM and AI analytics: Deploying edge compute devices (such as Blurgs Meenu) directly aboard vessels to process continuous video streams into structured, auditable events.
Each regime confronts a distinct operational bottleneck, introduces unique failure modes, and requires a dedicated sampling doctrine. The central policy question is not simply: What percentage of fishing trips should be audited? Rather, a robust policy must define: What evidence should be observed continuously, what should be filtered automatically, and how should scarce human review capacity be sampled across each regime?
This policy perspective breaks down the current operational issues within each regime and articulates three focused doctrines to address them.
1. Regime 1: Trip Sampling Without EM Devices (Observers & eLogbooks)
Where electronic monitoring hardware does not exist, independent observation is extraordinarily scarce. Regulators rely on physical human observers, electronic vessel trip reports (eVTR) or paper logbooks, and opportunistic dockside offload checks.
Current operational issues in Regime 1
- The 95% unobserved void: Physical observer deployments are severely constrained by bunk limitations, observer safety liabilities, and public monitoring budgets. Across international commercial fisheries, most notably industrial pelagic longline fleets governed by Regional Fisheries Management Organizations (RFMOs) mandated observer coverage is legally set at a minimal 5% of hooks or fishing days (Ewell et al., 2020; van Helmond et al., 2020; WCPFC, 2018). In practice, this regulatory baseline leaves 95% of total fishing effort completely unobserved. From a statistical sampling perspective, observing $\le 5\%$ of fishing effort produces extreme estimation variance for non-target bycatch and rare interactions with endangered, threatened, and protected (ETP) species, rendering catch and discard accounting statistically unrepresentative (Gilman et al., 2014).
- The Observer Hawthorne Effect: Vessel behaviour changes fundamentally when an independent observer is standing on deck. Fishing masters navigate more conservatively, avoid boundary lines of Marine Protected Areas (MPAs), handle gear deliberately, and suppress discards. Consequently, catch and discard profiles from observed trips cannot be assumed to represent unobserved trips.
- Within-trip subsampling variance: Observers cannot work 24 hours a day on multi day commercial voyages. To manage fatigue, observers sample alternate hauls (e.g. 30-50% of sets) and draw random sample baskets from the sorting table. An interaction with an endangered sea turtle or seabird that occurs outside the sampled baskets is completely missed, producing false-negative compliance records.
- Retrospective self-reporting bias: Captain logbooks and eVTRs are self declared, followd by dockside inspections can verify what is landed in port, but they cannot independently confirm what was caught and discarded at sea.
Doctrine 1: Trip Selection and Verification Doctrine Without EM
To address these vulnerabilities, fisheries authorities operating without electronic monitoring must abandon naive random trip draws in favour of a risk stratified and discrepancy focused doctrine:
1. Risk stratified fleet allocation
Observer sea days must be allocated across fleet strata defined by gear risk (e.g. bottom trawls vs longlines), target stock vulnerability, and proximity to spatial closures. Fleets with high potential for non-target bycatch must receive prioritized coverage quotas, while low risk sectors receive lower baselines.
2. Irreducible random control baseline
A mandatory random fraction of observer sea days must be reserved for routine, supposedly low risk vessels. Without an untargeted control sample, authorities develop blind spots to emerging non-compliance in unmonitored sectors.
3. Paired statistical auditing (The Divergence Test)
Authorities must treat human observer records as a calibrated baseline rather than a stand-alone dataset. Compliance algorithms must continuously compare the statistical profile of a vessel’s unobserved trips against its observed trips:
\[\Delta_{\text{trip}} = \left| \text{CPUE}_{\text{observed}} - \text{CPUE}_{\text{unobserved}} \right| + \left| \text{DiscardRate}_{\text{observed}} - \text{DiscardRate}_{\text{unobserved}} \right|\]Statistically significant divergence in catch-per-unit-effort (CPUE), species composition, or discard frequency triggers immediate dockside audits and prioritized observer assignment on subsequent voyages.
4. Standardized deck subsampling protocols
Observers must follow strict, randomized basket selection and timed discard-conveyor monitoring protocols rather than opportunistic grab sampling, minimizing human sampling bias during heavy catch processing.
5. Smart eLogbook photo-verification (Blurgs elogbook)
Where permanent overhead electronic monitoring systems are absent; such as on small-scale vessels, artisanal fleets, or vessels awaiting EM retrofits, authorities should mandate Smart eLogbooks that combine operator-assisted photographic capture with automated metadata binding:
- Every captured photograph is automatically timestamped, georeferenced via vessel GNSS, and assigned an immutable cryptographic hash linked to the specific trip and haul number. This prevents retrospective record manipulation or the re-uploading of past catches.
- Integrated mobile vision models assist operators by automatically segmenting individual specimens, identifying species taxanomy, and estimating fork length against calibrated measuring boards or visual tags.
- Because vessel operators cannot photograph every single fish in bulk volume catches, regulations must establish clear sampling triggers:
- Mandatory 100% photo-capture for all quota limited choke species, high value pelagics (e.g. tuna, swordfish), and any interacting Endangered, Threatened, and Protected (ETP) species prior to release or discard;
- Stratified random basket capture (e.g. photographing 2 to 3 randomly selected sorting baskets per haul) for bulk commercial finfish.
- These verified digital records produce an evidentiary trail that compliance officers and dockside inspectors can cross-reference against landed offload tallies and logbook declarations, bridging the observational void without requiring capital-intensive permanent camera installations.
2. Regime 2: Video & Activity Sampling With Conventional EM (Cameras & Sensors, Not an end to end system)
Conventional electronic monitoring resolves the physical observer constraint by mounting tamper evident cameras, GNSS receivers and few other sensors aboard (NOAA Fisheries, 2024). Cameras record 100% of sea days, generating complete visual archives.
Current operational issues in Regime 2
- The Video Deluge Paradox: While recording 100% of footage is technically straightforward, manual 1:1 video review is economically and humanly impossible. Reviewing a 10-day fishing trip with four camera angles in real time would require 960 hours of human labor.
- Arbitrary static review quotas: To manage costs, regulators mandate arbitrary review quotas (e.g. “review 10% of trips” or “review 35% of video”). These static percentages are information-blind: they waste expensive reviewer hours watching routine, highly compliant trips while under-auditing complex, ambiguous, or high-risk hauls.
- Reviewer cognitive fatigue: Scrubbing through hundreds of hours of raw deck video at accelerated speeds (4x to 8x playback) induces severe visual fatigue. Reviewers frequently miss brief, opportunistic discard events, small non-target bycatch, or occluded species at the sorting table.
- Logistical latency and storage expense: Physical hard-drive swapping at ports, secure postal shipping, and manual video queues result in review delays of 60 to 180 days. This lag destroys immediate deterrence. Furthermore, storing and transmitting raw video files constitutes one of the largest recurrent costs of EM programmes (Pierre et al., 2024).
Doctrine 2: Activity Gated Haul Selection Doctrine for Conventional EM
To extract defensible regulatory evidence from conventional EM without incurring unsustainable video review costs, authorities must implement an activity-gated sampling doctrine:
1. Decouple observation coverage from review coverage
Regulatory frameworks must formally separate monitoring coverage from manual review mandates:
\[\text{Observation Coverage} \; (100\%) \neq \text{Review Coverage} \; (20\text{-}50\%)\]Cameras and sensors must operate across 100% of fishing voyages to eliminate the observer Hawthorne effect, but human auditors must inspect only a selectively sampled fraction.
2. Sensor-gated operational demarcation
Ideally, auxiliary sensor telemetry (drum rotation counters, hydraulic pressure thresholds, and GNSS velocity filters) must programmatically segment voyage video into operational states:
- Non-fishing states (Transit, Steaming, Mooring) represent 70% to 80% of total voyage duration, these segments should automatically be bypassed from manual video review.
- Only footage falling within sensor-demarcated fishing windows should be ingested into the review population.
3. Haul level dynamic subsampling
Instead of auditing complete trips (watching all hauls on 20% of trips), authorities should sample at the haul level across the entire fleet (e.g. reviewing 1 in 3 active hauls across all vessels, as evaluated by NOAA Groundfish audit programmes (NOAA Fisheries, 2024)). Haul level sampling distributes audit deterrence evenly across the entire fleet rather than leaving 80% of vessels completely unreviewed.
4. Standardized playback acceleration limits
Review protocols must define maximum permissible playback speeds: no more than 4x playback during active catch sorting, with mandatory single frame stepping whenever fish pass secondary discard chutes.
3. Regime 3: Human Verification Sampling With EM and Edge Analytics (Edge AI / Blurgs EMS)
AI-assisted edge electronic monitoring changes the fundamental nature of the assurance problem. Rather than storing hundreds of gigabytes of unindexed video for retrospective manual screening, onboard edge compute devices (such as Blurgs EMS and Meenu) process video streams in real time as catch is landed on deck.
Deep learning models detect gear deployment, track fishes, segment bounding boxes, and extract structured candidate events:
\[e_i = \bigl\{ t_i,\, x_i,\, v_i,\, h_i,\, q_i,\, \hat{y}_i,\, p_i,\, r_i \bigr\}\]where:
- \(t_i\) is the UTC timestamp;
- \(x_i\) represents tamper-resistant GNSS coordinates;
- \(v_i\) identifies the vessel and \(h_i\) identifies the haul;
- \(q_i\) identifies the physical catch item (digital QR tag or RFID batch tote);
- \(\hat{y}_i\) is the edge model classification (species identification, length estimate, disposition such as retained vs discarded);
- \(p_i \in [0, 1]\) expresses model confidence or predictive uncertainty; and
- \(r_i\) is a cryptographically hashed pointer to the underlying raw video snippet.
Current operational issues in Regime 3
- The Fallacy of Full Automation: Expecting computer vision to act as an autonomous enforcement judge is legally unviable, ethically flawed, and technically vulnerable to lens occlusion, glint, heavy sea spray, and novel biological variations.
- The Fallacy of Full Manual Annotation: Forcing human reviewers to manually re-verify every single detected bounding box defeats the economic purpose of artificial intelligence and re-creates the human review bottleneck.
- Model-Induced Blindness: If auditors review only what an AI flags as anomalous, they examine only the model’s existing preconceptions of irregularity. Novel evasion tactics, undetected species, and subtle system corruptions remain unobserved.
- Conflating model predictions with evidence: Treating automated detector labels as final evidence without preserving data provenance, independent sensor health metrics, or due process for vessel operators.
Doctrine 3: Human Assurance Doctrine With Edge Analytics (Never Full AI, Never Full Manual Review)
An AI-assisted electronic monitoring architecture must maintain strict operational boundaries:
It should bever be fully autonomous AI enforcement: Machine observation must be comprehensive; human verification must be selective, targeted, and legally authoritative.
Under this doctrine, humans never sample video footage to locate fishing activities. Machine models observe and structure 100% of fishing operations continuously. The role of human sampling is shifted entirely from sampling video footage to sampling review decisions, With this in place framework would look something like:
[ 100% Continuous Edge Observation ]
│
[ Edge Event Extraction & Tracking ]
│
[ Automated Multi-Source Cross-Check ]
(eLogbook vs AI vs Catch ID Tags)
│
┌──────────────────────────┼──────────────────────────┐
▼ ▼ ▼
[ Random Control Sample ] [ Uncertainty Review ] [ Disagreement Audit ]
Irreducible 2–5% baseline Softmax entropy high eLogbook vs AI mismatch
│ │ │
└──────────────────────────┼──────────────────────────┘
│
▼
[ Critical Escalations ]
100% Mandatory Review:
ETP species, closures, tampering
1. The random baseline control sample (\(A_{\text{random}}\))
An irreducible, statistically calibrated proportion of routine, high-confidence events (e.g. 2% to 5% of compliant hauls) is randomly routed to human auditors. This random control sample is essential to:
- calculate unbiased empirical false-negative rates (discards or species undetected by the vision models);
- monitor camera angle shifts, lens scratches, or subtle changes in crew handling workflows; and
- establish the ground-truth baseline needed to detect model drift across different seasons and fishing grounds.
2. Uncertainty directed verification (\(A_{\text{uncertainty}}\))
Edge models evaluate their own predictive confidence for every detected specimen. Whenever classification entropy is elevated due to motion blur, partial occlusion, glint, heavy sea spray on the camera housing, or visual ambiguity between sister species (e.g. Yellowfin vs. Bigeye tuna), the event is flagged. Instead of sending hours of video, the system packages a lightweight 5 second video clip and keyframe bounding box directly to an auditor’s review queue.
3. Cross source disagreement auditing (\(A_{\text{disagreement}}\))
The edge platform automatically cross-references machine inferences against independent data feeds in real time:
- the captain’s electronic logbook declarations;
- physical QR tags or RFID batch codes assigned at sorting; and
- vessel kinematic and boundary data from GNSS.
Whenever an objective conflict occurs; such as a divergence between the logbook tally and the optical counter, a species classification dispute, or an unrecorded haul in a restricted area, the system automatically flags the event for audit. Disagreements do not predetermine guilt; they identify where human verification is objectively warranted:
\[\text{Count}_{\text{video}} \neq \text{Count}_{\text{logbook}} \quad\lor\quad \text{Species}_{\text{AI}} \neq \text{Species}_{\text{declared}} \quad\lor\quad \text{Catch}_{\text{recorded}} = 1 \;\land\; \text{Tag}_{\text{created}} = 0\]4. Deterministic critical escalations (\(P(\text{Review} \mid c) \approx 1.0\))
Certain occurrences carry regulatory or conservation consequences that demand deterministic, 100% review rather than probabilistic sampling:
- Interactions with Endangered, Threatened, and Protected (ETP) species;
- Retention of prohibited or quota exhausted species;
- Fishing within closed boundaries or Marine Protected Areas;
- Apparent unrecorded discard events of quota managed catch; and
- Hardware tampering, camera masking, or unannounced power disconnection.
5. Treating equipment failure as an observable compliance event
An electronic monitoring framework must never treat missing sensor data as neutral absence. When an observation is absent (\(\text{Observation}_t = \varnothing\)), the audit pipeline must distinguish between:
\[\text{State A: No fishing event occurred} \quad \not\equiv \quad \text{State B: No reliable observation was available}\]Absence of evidence must not silently become evidence of absence. Equipment health logs, watchdog heartbeat signals, and lens clarity metrics belong inside the regulatory evidence graph, ensuring that machine inference assists assurance rather than substituting for primary evidence.
6. Closed-loop active learning updates
Human audit determinations are not dead-end compliance verdicts. Every reviewed clip, corrected species label, and false-positive resolution is fed back into the central model training repository. These high-value edge cases are incorporated into active learning pipelines to fine-tune edge weights, calibrate Bayesian uncertainty thresholds, and progressively compress human review requirements without degrading evidence quality.
4. Comparison of the Three Monitoring Regimes
The operational shift across monitoring paradigms can be compared across core architectural dimensions:
Primary constraint: Observational scarcity.
Sampling unit: Entire fishing trip or vessel sea day.
Human role: Manual on-deck tallying, dockside inspection, operator smart eLogbook photo-capture.
Analytics role: Post-hoc statistical anomaly checks and spot-check image verification.
Primary risk: Observer Hawthorne effect; 90%+ fleet blind spot.
Proposed doctrine: Paired divergence testing, smart eLogbook verification & risk-stratified trip allocation.
Primary constraint: Human video review capacity.
Sampling unit: Stored video archives & active fishing hauls.
Human role: Scrubbing accelerated footage (4x-8x playback).
Analytics role: None (raw video storage & manual review).
Primary risk: Reviewer fatigue; months of video backlog; high transmission cost.
Proposed doctrine: Sensor-gated activity triage & haul-level subsampling.
Primary constraint: Evidence governance & verification design.
Sampling unit: Discrete, structured catch/handling event.
Human role: Targeted verification of uncertainty, disagreement, and random controls.
Analytics role: Continuous real-time detection, tracking, and cross-reconciliation.
Primary risk: Model-induced blindness.
Proposed doctrine: Never full AI, never full manual; tripartite assurance sampling.
5. Summary Policy Doctrine
An effective fisheries assurance programme must match its sampling doctrine to its technological regime. The objective of technology in fisheries monitoring is not to replace human auditors with algorithms. It is to eliminate observational blindness, dismantle video backlogs, and ensure that human judgment is deployed precisely where it delivers the highest evidentiary integrity —
References and standards
- NOAA Fisheries (2024). Electronic Monitoring in U.S. Fisheries. National Oceanic and Atmospheric Administration.
- NOAA Fisheries (2024). Electronic Monitoring Program: Northeast Multispecies Groundfish Audit Model. Greater Atlantic Regional Fisheries Office.
- Marine Stewardship Council (2024). Evidence Requirements Framework (ERF). MSC Fisheries Standard Toolbox.
- Global Dialogue on Seafood Traceability (2024). GDST Standards and Implementation Materials. Standards for interoperable seafood traceability.
- Ewell, C., Hocevar, J., Mitchell, E., Snowden, S. and Jacquet, J. (2020). An evaluation of Regional Fisheries Management Organization at-sea compliance monitoring and observer programs. Marine Policy, 115, 103842.
- van Helmond, A. T. M., Mortensen, L. O., Plet-Hansen, K. S., Ulrich, C., Needle, C. L., Oesterwind, D., et al. (2020). Electronic monitoring in fisheries: Lessons from global experiences and future opportunities. Fish and Fisheries, 21(5), 990–1007.
- Gilman, E., Passfield, K. and Cate, K. (2014). Performance of regional fisheries management organizations: Governance of bycatch, discard, and post-release mortality. Ocean & Coastal Management, 94, 99–126.
- Babcock, E. A., Pikitch, E. K. and Hudson, C. G. (2003). How much observer coverage is enough to adequately estimate bycatch?. Pew Institute for Ocean Science / Oceana.
- Western and Central Pacific Fisheries Commission (2018). Conservation and Management Measure for the Regional Observer Programme (CMM 2018-05). WCPFC.
- Pierre, J. P., Dunn, A., Snedeker, A., Wealti, M., Cozza, A. and Carovano, K. (2024). Optimising the review of electronic monitoring information for management of commercial fisheries. Reviews in Fish Biology and Fisheries, 34, 1707–1732.
- Kay, M. and Merrifield, M. (2021). Fishnet Open Images Database: A Large-scale Dataset for Fish Detection and Identification. arXiv:2106.09178.
- Yan, B., Ding, P., Zhu, J. and Wu, F. (2026). Research progress on electronic monitoring in tuna longline fisheries. Frontiers in Marine Science, 13, 1830102.