A vessel track is evidence of positions, not proof of fishing
Automatic Identification System, or AIS, messages help ships share identity, position, speed and course for maritime safety. Vessel Monitoring Systems, or VMS, are reporting arrangements used by fisheries authorities for specified regulated fleets and management purposes. Both can supply location histories, but neither directly records the moment a hook enters the water or a net catches a fish. A study that labels positions as fishing normally uses algorithms or supporting declarations to infer what the vessel might be doing. We need to distinguish a received position, a reconstructed track, apparent fishing activity and independently verified fishing. These are different evidence levels, and moving between them without explanation changes the claim.
An AIS message may contain a vessel identifier, transmission time, position and navigational information. The identifier can be entered incorrectly, shared unexpectedly or changed, and messages can be missing because of receiver limitations and congestion. VMS reporting can be subject to rules for specified permit holders, and its data are often confidential. Neither source automatically covers every fishing vessel, especially smaller vessels or those in regions without suitable coverage requirements. A map with no plotted vessel is therefore not necessarily an empty ocean, and a continuous line through a protected area is not automatically proof that gear was deployed. The spatial and temporal record must be interpreted in its surveillance context.
NOAA Fisheries describes VMS as a satellite-based monitoring tool that supports compliance, management and related research. Global Fishing Watch describes AIS and some partner-shared VMS information as inputs to an algorithm that estimates apparent fishing activity. The word apparent matters. Its model classifies movement patterns and associated features; it is not a physical gear sensor. Public maps are valuable for understanding broad activity patterns, but visibility varies by fleet, vessel characteristics and reception conditions. Accurate translation should say an algorithm estimated apparently fishing positions, not an enforcement agency observed illegal catches. That distinction protects both scientific validity and fair treatment of vessel operators.
A useful reader begins by asking which collection system produced the record, which vessels were eligible to be observed, what date range was covered and what the data processor inferred. A VMS requirement in one national fishery may differ from another jurisdiction’s rules, so legal conclusions require the actual governing conditions. An AIS dataset may contain positions from vessels that voluntarily or legally broadcast for safety, but absent broadcasts do not prove no vessel was present. The strongest summary combines observed track facts, limitations and validation sources. It reserves accusations of wrongdoing for evidence tested against applicable rules, permits, declarations and corroborating records.
From messages to reconstructed tracks
Raw vessel positions arrive at irregular time intervals. To reconstruct a track, analysts generally sort observations by vessel identity and timestamp, check for duplicates and improbable jumps, and examine whether positions are reliable. A line drawn between two widely spaced pings is an interpolation, not a record of the exact route travelled. If a vessel transmits at 08:00 and next at 12:00, the straight connection across the map does not prove that it crossed every point on that segment. Complex coastlines, manoeuvres and alternative routes make such inferences particularly uncertain. Maps should identify gaps rather than silently fill them as though continuous observation were available.
A speed field may report speed over ground, while course represents movement direction. Fishing practices can produce distinctive speed and turning patterns, but these overlap with searching, weather avoidance, port approach, drifting, towing or equipment maintenance. Some gear types involve slow movement during deployment, while others remain stationary or follow different patterns. Therefore an algorithm that infers fishing from speed alone is vulnerable to false positives and negatives. Better systems use context and expert-labelled vessel tracks to learn patterns, yet classification errors remain. An ordinary reader should ask what behaviour the model was trained to recognise and for which gears and locations its accuracy was checked.
Identity resolution is a separate problem from movement classification. Maritime Mobile Service Identity, or MMSI, identifies a radio station for AIS purposes, but a mistakenly configured or reused identifier can produce impossible jumps that look like one vessel crossed enormous distances. Registry, call sign, vessel length and repeated track histories can help researchers identify inconsistencies. But matching records from different sources must be handled carefully to avoid merging two vessels or splitting one vessel into multiple identities. A summary that reports unique devices may not equal the number of unique hulls. The denominator used to estimate fleet activity should say whether it counts devices, vessels, trips or independently verified fishing units.
Receiver coverage varies across coastlines and open ocean. Satellite AIS reception may be affected by transmission timing, overlapping messages and receiver availability, while coastal stations have finite range. A quiet period in the dataset could reflect equipment, geography or reception rather than an operator switching equipment off. VMS operates through a different reporting arrangement and may have known expected reporting intervals, but failures and exemptions can still require context. Any gap metric must include its definition, minimum length and coverage quality criteria. Without those, the phrase dark vessel can exaggerate uncertainty into an allegation and a map can be mistaken for a direct surveillance recording.
Apparent fishing hours and what they count
Apparent fishing activity is a model’s classification of positions or intervals as likely associated with fishing behaviour. It may include searching or gear preparation, depending on the dataset, and is not necessarily a log of gear setting and hauling. Global Fishing Watch explicitly warns that modelled apparent activity should not be described as confirmed fishing. If a dataset reports one hundred apparent fishing hours, read the documentation: how was time assigned to positions, how were gaps treated and which gear classifications were used? Do not assume one hundred hours of gear actually catching fish. Apparent activity hours, vessel-days, effort in hooks or trawl-hours, and tonnes landed are different units.
Suppose a track segment contains five time intervals, each thirty minutes long, of which an algorithm classifies three as apparent fishing. Under a simple interval allocation this represents ninety minutes, or 1.5 apparent fishing hours. It does not reveal kilograms caught, gear deployed or actual fishing time without independent validation. If two adjacent segments overlap in allocated time because of a data transformation, naive summation may double count; the dataset’s time-accounting rules must govern aggregation. A careful article states whether hours reflect intervals, classified positions or grouped events. A count of events cannot be substituted for a duration, because events can differ enormously in length.
Comparing fishing-hour totals across years is hazardous if the underlying algorithm or vessel coverage changes. Global Fishing Watch’s 2026 version 5 pipeline release describes changes that remove false-positive transits and alter the apparent fishing activity estimates from previous versions. A researcher should therefore identify dataset version, classification model and update date before interpreting an apparent trend as behavioural change. A fall in modelled hours can arise partly from revised classification, even when vessel tracks did not change in the same way. The distinction is essential to reproducibility: repeat analyses with consistently versioned data and examine any documented methodological break before making a trend claim.
Apparent fishing effort is also not a universal unit of environmental impact. One hour for a small handline vessel is not necessarily equivalent to one hour of industrial trawling; catch rates, swept area, species interactions and habitat effects depend on gear and context. Aggregating time may be useful for spatial patterns but insufficient for comparing exploitation intensity across fleets. To estimate biological pressure, analysts may combine track information with validated gear attributes, effort metrics and catch or observer data. A non-specialist summary should not convert visible track density straight into fish mortality. It can say where modelled vessel activity was concentrated and why additional evidence would be needed for ecological conclusions.
Position gaps, uncertainty and compliance
A transmission gap is an interval between received messages, not proof of intentional disabling. Technical malfunction, receiver limitations, coverage quality, power arrangements and administrative factors may explain a missing sequence. Intentional interference can also occur, but determining intent requires corroboration. Global Fishing Watch’s data documentation applies specific criteria when it flags AIS gaps and notes strong limits near shore and under poor reception. Analysts should report the gap rule and surrounding conditions rather than suggest every unexplained interval is an illegal event. A responsible inference distinguishes a data-quality observation, a reason for further investigation and a legally established compliance finding.
Suppose one vessel has a three-hour interval with no received AIS message while a second has an eighteen-hour interval. A headline that automatically labels the second a deliberate concealment event goes beyond timing information. One must examine typical reception at the location, device characteristics, vessel activity and the dataset’s documented gap-detection thresholds. Even an algorithmically flagged gap is a screening indicator, not a definitive determination of illegal fishing. A vessel may travel through an area without fishing, or fish legally under a permit, depending on jurisdiction and gear. A map reader needs to know what additional administrative and physical evidence can support or contradict the suspicion.
Spatial closure analysis raises another trap: a reported position near a boundary is not the same as prohibited fishing within it. Boundary coordinates, legal definitions, GPS accuracy, vessel dimension, reporting interval and permitted transit provisions may all matter. A single position inside a closed polygon can have a benign explanation if the vessel was lawfully transiting without gear deployed. Conversely, absence of a tracked position inside the polygon does not prove the vessel never entered between messages. Enforcement decisions require appropriate legal and evidentiary processes, not a speculative reading of a public heat map. Translate location and activity claims separately before drawing conclusions.
The meaning of compliance differs across settings. A tracking system can support monitoring of vessel location, reporting obligations and apparent patterns. Actual fishing regulation may also concern gear type, authorisation, catch declarations, protected species and seasonal restrictions. A track alone does not disclose every requirement or whether it was met. NOAA Fisheries’ VMS information explains its role in supporting law enforcement rather than replacing the full evidentiary system. A technical educational article should therefore identify which condition a dataset could test and which would need logs, inspections, observer records or declarations. This limits false accusations while preserving the legitimate value of surveillance for conservation.
Coverage is part of every activity statistic
An activity map can be visually detailed yet systematically incomplete. Coverage depends on vessels carrying and operating the relevant devices, reception and transmission conditions, record access and data processing. AIS coverage of industrial vessels may differ considerably from coverage of small coastal boats. VMS data often involve fleets required by regulation to report positions, and public availability may be restricted by confidentiality. A gridded apparent fishing-hour dataset is therefore a map of activity inferred among visible vessels under a particular model, not a census of every fishing activity in the sea. The population represented by the dataset must be specified before comparing areas or national fleets.
A coverage fraction should have an explicit denominator. Imagine a reconstructed voyage spans ninety-six one-hour periods, and at least one accepted AIS message is available during seventy-two of them. Under that illustrative definition, observed hourly coverage is 72/96 = 0.75, or 75 percent. The remaining 24 hours are periods without messages under this measure, not proven hours of intentional disabling or unobserved fishing. The reported coverage might be calculated differently by a particular provider, so the documentation must govern interpretation. Percent coverage should not be turned into percent of legal activity unless the study independently verifies the connection.
Comparing two fleets also requires thinking about observation mechanisms. If Fleet A has high AIS adoption and strong reception while Fleet B relies on smaller vessels with limited broadcasting, a map may show far more apparent activity for A even when the true effort difference is unknown. Dividing by the number of visible vessels does not automatically repair unobserved fleet composition. Analysts may need registry data, targeted monitoring or other independent samples to estimate what is missing. The key question is not whether the map is useful; it is what it represents and what it cannot support without correction. A responsible caption names the observed fleet and data limitations.
Coverage is not just geographical. It changes over time as devices, regulations, data partnerships and satellite reception evolve. If a new processing system detects activity that older data missed, an apparent increase may reflect new visibility rather than more fishing. Similarly, changes in vessel identity matching can alter the number of unique vessels in annual summaries. Historical comparisons should use a consistent pipeline version where possible and document breaks. If a reader cannot determine whether the earlier year was measured with the same tools as the later year, the claimed trend needs a stronger caveat than one footnote saying results may vary.
Fishing labels come from a classification model
Apparent fishing labels are assigned from observed or derived track features. Candidate inputs may include speed, turning, vessel identity, gear type, location, bathymetry or distance from shore, depending on the methodology. A supervised model can learn movement patterns from expert-labelled tracks, but classification quality depends on how representative those training examples are. Transit, searching, gear preparation, weather avoidance and actual fishing sometimes overlap. The resulting label is a best-effort estimate of a behaviour class. It is not equivalent to an observer witnessing a net being deployed. The source’s word apparent is an essential technical qualifier, not editorial hesitation.
False positives occur when the system calls a segment fishing even though it was not fishing according to the ground-truth definition. False negatives occur when genuine fishing is missed. Both matter, and their costs differ by application. False positives can make a heat map look busier or trigger unjustified suspicion; false negatives can hide activity from aggregate analysis. A model may improve one error type while worsening the other. A quality report should indicate the validation data, classes considered, gear-specific performance and regions where accuracy is weaker. A global average accuracy does not guarantee the same performance for each fishing gear or coastal environment.
Consider a deliberately invented validation table with one hundred independent track segments: forty truly fishing and sixty truly not fishing. The classifier correctly flags thirty-two of the forty fishing segments, missing eight, and incorrectly flags six of the sixty non-fishing segments. Sensitivity is 32/40 = 80 percent. Specificity is 54/60 = 90 percent. Positive predictive value is 32/(32+6) ≈ 84.2 percent in this sample. These quantities answer different questions. The final value also depends on how prevalent true fishing was in the validation sample, so it should not be transported unchanged to a location where most vessel movements are transit.
Model validation should avoid data leakage. If many segments from the same vessel and trip appear in both training and test sets, the model may appear more accurate than when applied to genuinely new vessels or environments. Stronger evaluations may hold out entire vessels, regions, time periods or gear categories depending on the intended use. The data’s true deployment setting matters. A model trained predominantly on large trawlers may struggle with small passive-gear vessels. Readers should ask whether performance was measured on representative unseen behaviour and whether errors are reported separately for difficult classes. Calling a model artificial intelligence does not answer any of these questions.
Version changes can create false trends
An algorithmic fishing-hour series depends on a versioned processing pipeline. Changes in vessel classification, identity matching, reception filtering, time allocation and fishing detection can change historical totals without any new activity in the historical ocean. Global Fishing Watch’s October 2026 version 5 release explicitly describes revisions to its general fishing detection model, including reductions in false-positive transit classification. Such a release is evidence that activity maps are scientific products undergoing correction. It is not evidence that an older analyst was dishonest or that fishing fleets necessarily reduced effort by the amount the revised dataset changed. Methodological and behavioural changes must be separated.
Suppose an older model estimates 1,200 apparent fishing hours for a region during one fixed historical year, while a revised model estimates 900 hours from the same underlying period. The difference is 300 hours, or 25 percent lower relative to 1,200. It would be invalid to report that vessels fished 25 percent less that year because both results describe the same year. The difference arises from the changed processing method or data. A study comparing genuine years should either reprocess both with one consistent method or quantify what the version break contributes to the trend. This arithmetic example is invented and does not describe the actual version 5 release.
Version metadata should accompany any chart: dataset release, extraction date, spatial resolution, projection if relevant, gear filters, vessel-identification rules, minimum event duration and time-allocation convention. These details let future readers reproduce a result and understand why a recalculated number differs. Without them, a copied heat-map image becomes difficult to audit. The phrase downloaded in 2026 is less informative than a precise dataset identifier. A clear summary can spare the general reader many technical details but should preserve a link to the methods and release notes for those who need to inspect assumptions.
Another source of apparent change is the definition of an event. Some systems group adjacent classified positions into fishing events using time and distance thresholds. An analyst who changes the maximum separation allowed between points can split one long event into several or combine multiple short intervals. Total apparent hours might stay similar while event counts change dramatically. It is therefore wrong to compare a count of events from one methodology with hours from another or treat all events as equal lengths. Ask what constitutes an event, which distances and gaps merge segments, and how overlapping time is handled before constructing a denominator.
Geography, effort and impacts are not synonyms
A coastline heat map may show concentrated apparent fishing activity around a habitat, but it does not establish the biological effect of that activity. Different gears interact with fish stocks and seabed habitats differently. The map may also represent more visible vessels in one area rather than genuinely higher total effort. Ecological conclusions require gear-specific effort measurements, stock distribution, catch or observer data and a suitable impact model. A crowded gridded cell is a spatial summary of classified activity, not a direct measurement of fish mortality. The right next question is which species and habitats were exposed and by what fishing mechanism.
A spatial aggregation must preserve area and time denominators. Suppose Cell A covers 100 square kilometres and records 200 apparent fishing hours over a month, while Cell B covers 400 square kilometres and records 300. The raw total is larger in B, but the simple density is 2 hours per square kilometre in A versus 0.75 in B. Whether density is the appropriate statistic depends on the scientific purpose and the comparability of sampling coverage. Neither density is total biological impact. Gridded data may also include spatial assignment errors or uncertain gear class, so a careful map caption should state cell size, interval and the meaning of the plotted quantity.
Spatial boundaries introduce classification sensitivity. A vessel position can be near the edge of a marine protected area or across a boundary drawn at a particular resolution. Even if the geographic location is accurate, an algorithmically classified fishing position near the boundary is not automatically a violation of law. Rules may allow passage without fishing, permit specific gears, apply seasonally or include exemptions. Vessel dimensions, location uncertainty and gaps between pings can matter. A legitimate compliance assessment uses the governing regulation and appropriate corroborating evidence. Map position and activity inference should be presented separately before any legal conclusion.
To interpret economic consequences, tracked movement also needs to be linked to vessel operations and costs. A closure may displace fishing effort to alternative grounds, causing more transit fuel use, altered catch opportunities or different exposure to weather. Yet a map cannot show vessel profit by itself. Landings, prices, trip costs and market conditions require separate data. An increased density of apparent fishing outside a closure could reflect displacement, stock availability or observation improvements. The related eduKateSG article on spatial closures, BACI evidence and spillover examines causal comparisons; vessel tracks alone do not supply a controlled counterfactual.
Read an AIS gap as uncertainty, not a verdict
The word gap can mean several different things. A raw message gap is a time interval between recorded transmissions. A qualified AIS-off event is an algorithm-defined screening output requiring a particular minimum duration and reception context. An incident finding is the result of a separate evidentiary process. Collapsing these into one category is misleading. Global Fishing Watch documents filters for AIS-gap events and cautions that nearshore reception is especially complex. These details matter because a short or poorly covered interval may be ordinary data loss. The strongest public statement describes the observed gap and the need for verification without declaring intent.
Consider a vessel observed at 06:00 and again at 20:00, giving a fourteen-hour gap in received messages. No recorded location is available for the interval in this illustrative dataset. A straight line between the endpoints cannot reconstruct the vessel’s actual route. Even if the line crosses a closed region, that alone does not establish entry, fishing or illegality. Appropriate follow-up might examine authorised administrative records, other tracking sources and reception quality, subject to law and privacy rules. The safe scientific statement says the available AIS record cannot resolve the vessel’s intervening route. It does not fill the gap with invented activity.
An unexplained absence of AIS data may also coincide with a VMS record available to an authorised regulator but not to the public. The absence of a public map track is not proof that no monitoring record exists. Conversely, some smaller boats may not be subject to a specific VMS requirement. A report should state which fleet and data feeds were included, not imply that all vessels worldwide face uniform reporting obligations. Privacy and confidentiality restrictions on VMS data mean that broad public comparisons may be incomplete. The difference between unavailable information and confirmed absence of information is a fundamental concept across all evidence-based writing.
Readers should avoid turning screening indicators into accusations against identifiable operators. High-risk monitoring can legitimately prioritize further review, but an algorithmic flag has known error rates and may be affected by context the public data cannot show. For a compliance claim, disclose the specific rule, required conduct, time and jurisdiction, then rely on appropriate authenticated records and decisions. A probability of apparent fishing is not equivalent to guilt. The distinction keeps the article technically honest and prevents readers from mistaking a conservation dataset for a universal judicial record.
Checked practice: calculate and interpret the evidence
Exercise one: a voyage spans eighty one-hour blocks. AIS messages appear in sixty blocks under an hourly coverage rule. Calculate coverage and interpret the remainder. Answer: 60/80 = 0.75, so 75 percent of the voyage’s hour blocks contain at least one observed AIS message. Twenty blocks lack a message in the dataset. They are not automatically twenty hours of unreported fishing or intentional deactivation. This fraction is meaningful only because the denominator and hourly rule are specified. If the provider uses a different definition of voyage or acceptable message coverage, its published metric may differ.
Exercise two: a model labels eight half-hour intervals as apparent fishing. Assume each interval is counted once without overlap. Duration equals 8 × 0.5 = 4 apparent fishing hours. A second dataset counts the same sequence as one apparent fishing event. Which is larger? The question is ill-posed because hours and events are different units. Four hours describe duration; one event describes a grouped episode. Neither establishes a catch weight or biological impact. The correct translation must preserve the unit and the modelled status. Apparent does not mean an inspector confirmed gear use.
Exercise three: a study reports a model-based fishing-hour total changing from 2,000 to 1,500 for exactly the same historical period after a software release. Compute the relative change. Answer: (1,500−2,000)/2,000 = −0.25, a 25 percent difference in the reported model output. It cannot be called a 25 percent reduction in actual fishing over time, because the observation period did not change. To evaluate a genuine trend, use a consistent processing method for both years or account for the version break. This is an invented calculation illustrating why data provenance belongs in every chart caption.
Exercise four uses the classification table from the earlier example: forty true fishing segments, thirty-two true positives, eight misses; sixty non-fishing segments, six false positives and fifty-four true negatives. Sensitivity = 32/40 = 0.8, specificity = 54/60 = 0.9, and precision among flagged segments = 32/38 ≈ 0.842. Accuracy = (32+54)/100 = 86 percent. Each answers a different question. A claim that the model is 86 percent accurate does not disclose the eight missed fishing segments or six wrongly flagged non-fishing segments. The class balance in future use can alter the proportion of flagged positions that are genuinely fishing.
Exercise five considers an apparent fishing position just within a protected polygon. Is the vessel proven to have violated the regulation? No. One must establish the operative rule, date, permit status, transit exceptions, position reliability and whether the inferred behaviour truly constituted prohibited fishing. The track may justify careful investigation but not an automatic finding. A stronger analysis could combine data from appropriate authorities and independently recorded activities. The distinction is not leniency toward wrongdoing; it is evidentiary precision. Environmental monitoring remains more credible when readers can tell a modelled suspicion from a verified legal finding.
A reproducible reading protocol before any claim
A transparent analyst should record the data provider, access date, spatial filter, vessel-selection rule, activity model and product version in the same notebook as the final number. Suppose one report uses all vessels classified as fishing vessels, while another uses only vessels with confirmed gear categories. The resulting difference is a changed population, not necessarily a change in fleet behaviour. If a heat map has been aggregated by day, the colour may encode modelled hours summed across vessels; it does not mean a single vessel fished continuously for that many hours. When converting the number into prose, retain the definition of the counted unit and the population represented. A reproducible figure caption should also state whether incomplete tracks were retained or excluded.
Time conversion deserves more care than it receives. An AIS timestamp stored in Coordinated Universal Time should not be relabelled as local port time merely because a map centres on that port. A vessel can cross the International Date Line while fishing, and a report grouped by local dates may allocate effort to different calendar days from a UTC-grouped export. To test the magnitude, take two segments, each one hour, at 23:30 UTC and 00:30 UTC. A UTC aggregation gives one segment to each of two dates. A three-hour advance in local time places both on the following local date. The total duration remains two hours; the daily pattern changes solely because the grouping calendar changes. Report the time basis explicitly.
An additional problem is that the sampled fleet may change while the real fleet does not. Imagine one hundred eligible vessels in a hypothetical region. During year A only forty have usable AIS observations; during year B seventy have them. The measured apparent activity rises from eight hundred to one thousand hours. Without a comparability adjustment, the difference cannot be interpreted as a twenty-five percent behavioural increase: the observation coverage broadened as well. A defensible comparison could restrict both years to a documented common panel, disclose the discarded vessels and explore sensitivity to the restriction. Even the common panel would not represent unobserved small-scale fleets automatically. The numerator must always be interpreted in light of its observable population.
A line across a marine reserve also needs spatial sensitivity analysis. Positions have uncertainty and polygon boundaries can move after legal designation updates. Consider a reported point situated fifty metres inside a boundary, while a conservative horizontal position uncertainty is one hundred metres. It would be premature to describe the precise location as certain. A good analysis would retain the observed coordinate, datum and uncertainty model, evaluate whether the boundary layer matches the operative rule at that date, and check neighbouring pings. A buffered sensitivity assessment might label the point boundary-indeterminate rather than automatically inside or outside. That procedure does not decide legality; it prevents the map from making a stronger spatial claim than the input supports.
A useful chain-of-evidence table distinguishes raw position evidence, derived trajectories, activity classification, rule interpretation and corroborated enforcement findings. Raw evidence may record a time and coordinate. A reconstructed path adds assumptions between times. Activity classification adds a model and its validation limitations. Rule interpretation adds legal geography, gear restrictions, licence scope and effective dates. A finding of non-compliance requires the relevant authority’s evidentiary standard and supporting records. This chain is particularly important when a chart is reused in a news article, translated briefing or educational handout. A short phrase such as ‘suspected fishing activity’ should not be strengthened to ‘illegal fishing confirmed’ without new evidence.
More worked interpretation, tests and error repair
Example six examines spatial aggregation. Two vessels each have three model-labelled half-hour fishing intervals in the same one-degree cell. If every interval contributes its full half-hour and no temporal weighting or boundary splitting is applied, the cell accumulates 2 × 3 × 0.5 = 3 vessel-hours of apparent fishing. The value is not three elapsed clock hours: both vessels could have operated simultaneously. It is also not three vessel-days, tonnes landed or three illegal events. When an export reports vessel-hours, a translator must not shorten the term to hours if the audience might mistake a fleet sum for elapsed time. Other data products can allocate time differently; always reproduce the actual provider’s aggregation rule.
Example seven tests receiver coverage. A study samples one hundred equal one-hour windows. Seventy contain at least one message, but only sixty have enough consecutive observations to assess turning or speed reliably. Under the specified definitions, message presence is 70/100 = 70 percent, while classification-eligible coverage is 60/100 = 60 percent. Neither number says the remaining thirty or forty windows are fishing or non-fishing. If thirty classifier-eligible windows are labelled apparent fishing, the modelled share of eligible windows is 30/60 = 50 percent, not 30/100 = 30 percent without qualification. The first is conditional on sufficient observation; the second uses all scheduled windows. Both may be displayed, but their denominators answer different questions.
Example eight considers precision when prevalence changes. In a validation set containing one hundred actual fishing intervals and nine hundred non-fishing intervals, assume a model detects eighty true fishing intervals but falsely flags ninety non-fishing intervals. Sensitivity equals 80/100 = 80 percent. The false-positive rate equals 90/900 = 10 percent. Yet only 80/(80+90), approximately 47.1 percent, of its flagged intervals are actually fishing in this example. Reporting ‘ninety percent specificity’ alone might leave a reader with unjustified confidence in an individual flag. Any publication of these illustrative numbers must call them hypothetical; real model performance must be obtained from appropriate validation data for the fleets and gears under study.
Example nine repairs a false conclusion from a map gap. Suppose a trawler last transmits at 11:00 UTC and reappears at 20:00 UTC. The nine-hour interval is an observed absence in the received dataset; it is not a traced nine-hour fishing operation. An investigator should first compare nearby receivers and other vessels, test clock and identity consistency, inspect whether the vessel was outside reception coverage and seek independent lawful evidence where available. The appropriate public description is that no AIS messages from the identified vessel were present in this dataset during the interval. Attributing motive requires more than an absence of data. Different datasets may reconstruct the gap differently, so the source must be named.
Example ten asks a model auditor to compare the same ten-day period under two processing releases. Release A reports 4,200 apparent fishing vessel-hours; release B reports 4,830. The version difference is 630 hours, or 630/4,200 = 15 percent relative to A. There has been no real time change: both outputs describe the same ten days. A method comparison should therefore report the discrepancy as a sensitivity of inferred effort to processing choices. If comparing different years as well, use a matched version or quantify the interaction between year and version. This is why archived data provenance, code version and caveats are practical requirements, not merely academic footnotes.
An independent transfer exercise brings the pieces together. A report counts 120 vessels with adequate track coverage, and forty have at least one apparent fishing event. Of the forty, twelve enter a polygon labelled protected in a contemporary map. What can be concluded? Answer: one-third of adequately observed vessels have at least one inferred event under this model, and twelve such vessels have mapped tracks within that polygon. The information does not by itself establish gear deployment, a catch, violation of a rule, the proportion of all real vessels operating in the area or which boundary was in force when each observation occurred. A responsible next step is to resolve vessel eligibility, dates, polygon authority, track uncertainty and any legally available corroboration.
Fisheries tracking is often presented as a single effortless colour layer, but meaningful interpretation is a sequence of constrained translations: messages into positions, positions into tracks, tracks into activity estimates, estimates into spatial and temporal summaries, and summaries into claims suitable for publication. At every transition, record what was added by observation, inference or outside rules. The reader who retains the units, the missed population, the processing version and the evidence grade can draw useful conclusions without either dismissing remote sensing or overstating what it proves. The best result is a statement whose uncertainty travels with it and whose numbers can be reconstructed independently.
Primary methodological sources and related learning
Global Fishing Watch explains the distinction between apparent fishing activity and fishing events and publishes data caveats, including AIS gaps. NOAA Fisheries documents the Vessel Monitoring System in its enforcement work. These sources describe real systems; the numerical exercises above are invented teaching examples, not observed fleet results.
For adjacent methods read the verified eduKateSG guides on fisheries electronic monitoring and logbook discrepancies, fishery-independent survey coverage, and fish eDNA detection limits. Each uses a different evidence source and should not be substituted for vessel-track observation.