COVID-19 lockdowns reveal the resilience of Adriatic Sea fisheries to forced fishing effort reduction

The COVID-19 pandemic provides a major opportunity to study fishing effort dynamics and to assess the response of the industry to standard and remedial actions. Knowing a fishing fleet’s capacity to compensate for effort reduction (i.e., its resilience) allows differentiating governmental regulations by fleet, i.e., imposing stronger restrictions on the more resilient and weaker restrictions on the less resilient. In the present research, the response of the main fishing fleets of the Adriatic Sea to fishing hour reduction from 2015 to 2020 was measured. Fleet activity per gear type was inferred from monthly Automatic Identification System data. Pattern recognition techniques were applied to study the fishing effort trends and barycentres by gear. The beneficial effects of the lockdowns on Adriatic endangered, threatened and protected (ETP) species were also estimated. Finally, fleet effort series were examined through a stock assessment model to demonstrate that every Adriatic fishing fleet generally behaves like a stock subject to significant stress, which was particularly highlighted by the pandemic. Our findings lend support to the notion that the Adriatic fleets can be compared to predators with medium-high resilience and a generally strong impact on ETP species.

fishing days for purse seiners and pelagic pair trawlers, and (ii) applying temporary spatial closures to protect nursery and spawning areas. Demersal stocks are managed by a separate and more recent multi-annual plan 47 , whose overarching goal is to reach by 2026 the maximum sustainable yield (MSY), i.e., the maximum catch of a fish stock that can be taken indefinitely without depleting the population, for the target species. The plan is mainly aimed at high-effort fleets and involves two phases. The objective of the first phase is to progressively reach a fixed percentage effort reduction for each fleet by 2021 (e.g., -12% for otter trawlers and -16% for beam trawlers), with respect to the average 2015-2018 effort. The second phase (2022-2026) envisages setting yearly effort quotas for each fleet depending on the status of its target stocks. The plan will also regulate fleet size and set up restricted areas to maintain a sustainable biomass. Altogether, the GFCM plans ensure that the fishing effort is commensurate to stock abundance. Their success therefore requires accurate inputs from scientists. The COVID-19 pandemic has involved restrictions for all world fisheries to stem the spread of the infection. In Italy, restrictions were imposed in early March in the North-East, but by March 12 a national lockdown was in place for all commercial activities that involved retailing, catering and personal services 48 . On March 23, all working activities except essential services were suspended until May 18 49 . In the eastern Adriatic, non-essential services and commercial activities were shut down from March 16 in Slovenia 50 and from March 19 in Croatia 51 . As a result, the demand for seafood collapsed, prices plummeted, and fish markets closed, causing reduced fishing activities (see also Table 1 in the main document). Moreover, fishing activities were directly restricted by the practical difficulties of applying social distancing and safety protocols on board. Several operators suspended their activities altogether throughout the lockdown period.

Data sources 2.1 Data for coarse-scale pattern analysis
The Google's Global Fishing Watch 52 (GFW) is a Web portal managed by Google in partnership with Oceana and SkyTruth. The GFW produces a global view of commercial fishing activities by analysing VMS, AIS and Infrared Imaging Radiometer Suite (VIIRS) data. Its global scale data tables report vessel activity information aggregated at 0.01°and 0.1°resolutions that can be accessed for scientific purposes. The published datasets report fishing activity cells classified by machine learning models 22 , which identify fishing activity locations mainly based on speed information. Although the GFW data are well suited to large-scale analyses and averages (e.g., at 0.5°resolution), their large amount and heterogeneity make them less accurate for smaller-scale analyses 53 .
The GFW data of March-April 2020 and 2019 have consistently been used in ecological and economic analyses [54][55][56] . We therefore did not question their quality, also because we employed them for a coarse resolution pattern recognition that reduced data noise and uncertainty biases.

Data for Adriatic-scale pattern analysis
The Astra Paging data collector 57 was our primary source to conduct a detailed analysis of fishing activity in the Adriatic Sea. Astra Paging hosts one of the largest worldwide AIS databases, which goes back to 2009. In particular, we used the complete census of fishing vessel positions, sampled at 5-minute intervals, collected by terrestrial AISs in the Adriatic Sea. The relevant records were those falling within GFCM geographical sub-areas (GSAs) 17 and 18 58 . AIS data between 2015 and 2020 were used to normalise fleet size and reduce the bias due to the massive increase in AIS data recorded after May 2014, when the minimum length of AIS-equipped EU vessels was reduced to 15 m. AIS records report information every few seconds on vessel position, speed over the ground, course and rate of turn. They also include trip-related information (e.g., destination and arrival time) and identifiers such as call sign (IRCS), vessel type, name, Maritime Mobile Service Identity (MMSI) and International Maritime Organisation (IMO) number.

Data for AIS data quality evaluation
We used the data of the EU Fleet Register 59 (EUFR), a public database containing the data of the fishing vessels of all EU member States, to make a quantitative evaluation of AIS data quality. The EUFR contains vessel administrative information (name, registration port), technical specifications (lengths, fishing gear) and relevant events (construction, modifications, decommissioning). These data allowed estimating the size of EU Adriatic fleets over time and to compare it to the one estimated from the AIS dataset. This operation allowed validating AIS data completeness and enabled their reasonable use in our experiment.

AIS data quality control
We employed vessel registration port to estimate the approximate fleet size around a port. Such approximation can be tolerated when validating big data such as the AIS data we used 60 . The EUFR data were filtered by retrieving only vessels of at least 15 m LOA registered in the Italian, Slovenian and Croatian ports of GSAs 17 and 18. In the validation process, we linked AIS data to EUFR events by matching the identifiers stored in both datasets, i.e., IRCS, MMSI, vessel names and IMO number. This heuristic process sought an exact match of one identifier at a time (primary identifier), using another identifier as a secondary control. The match between two records was considered successful if the primary identifiers matched exactly and the minimum edit distance between the secondary identifiers was < 3 (high similarity). Unsuccessful matching for all combinations of primary and secondary identifiers prevented identification (unidentified vessel). This process enriched the AIS data with information on gear licence, flag country, registration site, LOA and relevant events. The final validation of AIS data completeness against EUFR data was obtained by introducing three metrics to estimate the annual overlap between the number of operational EU Adriatic vessels recorded in each dataset. To do this, for each year from 2015 to 2020 we calculated the number of EUFR vessels that were equipped with an AIS transponder (expected vessels) and the number of EU vessels that broadcast AIS data (observed vessels). Finally, we defined the following metrics: Expected coverage year = expected vessels Total EUFR vessels Observed coverage year = observed vessels Total EUFR vessels Representativeness score year = observed vessels expected vessels These scores measure the amount of complementary information found in the EUFR and AIS data, hence the extent to which the AIS data represent all the EU vessels registered in the Adriatic Sea.

Gear classification
The fishing gear types used by the AIS-equipped vessels were automatically inferred using a revised version of the workflow published by Galdelli et al. 61 . The source code is included in the Zenodo repository indicated at the beginning of this document. Our workflow estimates the gear used by a vessel in a fishing trip. It uses speed and position reference clusters, previously learned through cluster analysis from historical annotated data of the Adriatic Sea, for each gear. It identifies the gear that a vessel used in a trip among five types: bottom otter trawl (OTB), pelagic pair trawl (PTM), beam trawl (TBB), purse seine (PS) and other gear (OTHER). As a first step, it estimates which gear types are consistent with each trip based on the overlap between the speed and position distributions and the reference clusters. As a second step, the most probable gear type is assigned to the vessel's trip by a Random Forest classifier 62 based on (i) the frequency of potential gear types the vessel has used the same month and (ii) the number of trips it made that month. The analysis is based on monthly data because fishing strategies change depending on the season. As a final step, the workflow labels all trajectory points in a trip as fishing/non-fishing activity points and calculates the fishing time for each point. To do this, it sets a speed threshold for each trip to estimate when the gear was/was not deployed, i.e., when the vessel was/was not fishing, and the trajectory intervals associated with fishing. The threshold is calculated by detecting the point speeds that are entirely encompassed in the gear's speed cluster. The process also detects non-fishing trips and excludes them from further analysis. Finally, fishing time per point is calculated. For the towed gear types (OTB, PTM, TBB) it is computed as the time difference between two consecutive fishing points; for PS -where gear identification is less accurate due to its complex spatial patterns -it is approximated by dividing the total fishing time by the number of fishing points.
We applied this workflow to the 2015-2020 AIS dataset and assessed its classification performance against EUFR-Adriatic data. Since the classification confidence of our workflow was higher for longer trips with uniform sampling 61 , we processed only trips with at least 12 transmissions and a percentage of data gaps not exceeding 90% of trip duration. The classification performance was estimated according to a confusion matrix that reported the overlap between estimated vessel gear and the corresponding gear reported in the vessel's EUFR licence. Considering the EUFR gear as the ground truth, we calculated the following evaluation parameters: For each gear type g: calculate true positives as the number of AIS-estimated g trips corresponding to EUFR g types; calculate false positives as the number of AIS-estimated g trips not corresponding to EUFR g types;

4/
calculate true negatives as the number of AIS-estimated non-g trips that correspond to EUFR non-g types; calculate false negatives as the number of AIS-estimated non-g trips that correspond to EUFR g types.
Based on these parameters, classification performance was assessed according to the following metrics: Overall agreement with respect to chance was measured using Cohen's Kappa 63 .

Spatio-temporal analysis of fishing effort
We implemented an algorithm to classify the fishing effort cell-by-cell, at 0.5°and 0.1°resolutions. The process is based on the assumption -supported by empirical observation of the data at hand -that the distribution of fishing hours over the cells of a large area is approximately log-normal. Indeed, in this distribution, cells characterised by a higher number of hours tend to be farther from the geometric mean than those with fewer hours. Fishing activity per cell was therefore classified based on the confidence limits assuming a log-normal distribution of fishing hours, through the following rules: For each cell in the area: total fishing hours above the upper confidence limit → high-effort cell total fishing hours between the lower and upper confidence limits → medium-effort cell total fishing hours under the lower confidence limit → low-effort cell This classification allowed studying the spatial distribution of high-effort locations, to discover high-intensity patterns in the Italian seas and to compare these patterns with the trends highlighted in the Adriatic Sea. The analysis was conducted on both GFW (at 0.5°resolution) and AIS data (at 0.1°resolution). In the GFW dataset, it highlighted the difference between the fishing effort identified in the Italian seas in March-April 2019 and in March-April 2020. In the AIS dataset, it yielded patterns of fishing activity change in the Adriatic Sea over the years and months, from 2015 to 2020. In particular, annual aggregations were used to highlight the overall change of fishing activity in 2020, due to the pandemic, compared with the previous years. This change was further explored using monthly aggregations, which were also used to estimate the recovery speed of the different fleets in relation to the stress factors and to understand the response of each fleet to the pandemic.

Barycentre calculation
Barycentre calculation of each gear over time was performed to explore the fishing activity patterns in the Adriatic Sea. The barycentre is the mean location of the observed fishing effort, thus it is not necessarily a high fishing effort location. We weighted each 0.1°cell by the number of fishing hours to calculate the coordinates of the barycentre: f ishing e f f ort i,t where t is the unit of the time aggregation (month of the year) and n is the number of fishing cells at time t.
The sequence of fishing effort barycentres over months of a given fleet is a spatial time series that allows studying fishing vessel movements in the various seasons 6 . In our experiment, this sequence also allowed assessing how the pandemic altered fishing activities in 2020 compared with the previous years.

Potential impact on endangered, threatened and protected species
The presence of ETP species was established by extracting IUCN species observation records from the Ocean Biodiversity Information System 64 (OBIS), which has been demonstrated to be suitable for this type of large-scale biodiversity analyses 65 . The possible observation sampling bias was managed through a statistical analysis of the density of ETP species. In particular, the fishing cells were categorised as low-, medium-or high-impact assuming a log-normal distribution of the number of ETP species per cell, through the following rules: For each cell in the area: number of ETP species in the cell above the upper confidence limit → potential high-impact cell number of ETP species in the cell between the lower and upper confidence limits → potential medium-impact cell number of ETP species in the cell under the lower confidence limit → potential low-impact cell We conducted this analysis in the Italian seas at 0.5°resolution using the GFW dataset, and in the Adriatic Sea at 0.1°r esolution using AIS dataset. In the Adriatic Sea, we performed further spatial aggregation of cells by density-based clustering (with DBScan 66 ). This step generated larger aggregates of impact areas, whose composition over time highlighted more clearly the pandemic-related change in the potential impact of fishing activities.

Fleet assessment model
An approximate relationship 67 between MSY , r, and k is MSY ≈ k · r/4. Thus, knowing r and k allows estimating MSY . The Schaefer model 68 correlates stock biomass dynamics (B t ) with catch (C t ) by the following analytical formula: The idea behind data-limited stock assessment models is that, if the catch is known and r and k can be estimated with acceptable approximation, it is possible to calculate the biomass evolution in time. Catch Maximum Sustainable Yield 67 (CMSY), one of the most widely used data-limited methods, can accurately reproduce the biomass time series based on a Monte Carlo approach, which generates viable r-k pairs from the Schaefer formula. CMSY estimates prior ranges of r-k pairs from prior qualitative information on the general resilience of the species to fishing mortality. An alternative version of CMSY, the Abundance Maximum Sustainable Yield 69 (AMSY), uses a rewrite of the Schaefer formula with an abundance index A t instead of biomass B t : A t can provide a fishery indicator like the catch per unit of effort or another index of species abundance. AMSY reconstructs the catch time series C t multiplied by catchability q, a coefficient that reflects fishery efficiency, i.e., how much biomass becomes catch. AMSY also estimates MSY , r and kq (the carrying capacity multiplied by catchability). If A t is a monthly time series (i.e., t is a monthly index), then r is its monthly growth and kq is the stock abundance in the absence of fishing pressure. The main underlying hypotheses of AMSY are that (i) the abundance index A t , the catch Cq t , r and kq are correlated through Schaefer dynamics and (ii) q, r and kq do not change over time in the stock's area, although they may be different for the same species in different areas.

Representation of AIS data in EUFR data
In the Adriatic Sea EUFR data identified an ∼11% reduction of registered vessels of more than 15 m LOA between 2015 and 2020 ( Table 2 in the main document). The reduction was largest between 2017 and 2018 (from 1018 to 932 vessels) and was mainly due to the EU fleet capacity limitation plan 70 , which set caps on kilowatts and gross tonnage for each EU member State and allowed new vessels to operate in a given area only after their equivalent capacity had been removed (e.g., by decommissioning). The measure exerted its maximum effect in 2018, when 87 vessels were decommissioned and 2 vessels deregistered from the Adriatic ports. Another contributing factor was the limited number of new vessels entering the EU Adriatic fleets (Table 2-New and Entered columns), since 20 new vessels (2 newly constructed and 18 registered to an EU port) were entered in 2017, but only 5 were entered in 2018. Notably, the contribution of new constructions to the Adriatic fleets 6/ was always low, with only 3 vessels built in 6 years. A comparison of AIS and EUFR data disclosed that 1114 out of 1169 vessels (95.3%) included in the AIS dataset belonged to EU Member States and 55 (4.7%) to non-EU States (30 were Albanian). The AIS dataset contained a total number of 10 million vessel trajectory points (of which 98% belonged to European vessels) corresponding to 2 million estimated fishing hours. The change in the expected coverage over the years was low (1%, ranging between 78 and 79%, Table SI-1), indicating that the relative number of EUFR vessels carrying an AIS transponder was almost constant, also during the pandemic. In contrast, the number of vessels that broadcast AIS data was more variable, i.e., the observed coverage showed a greater variability (5%, from 74% in 2016 to 69% in 2020), with a slowly decreasing trend. The phenomenon also affected the annual representativeness score, which fell from 93% in 2016 to 89% in 2020, although there was no reduction in 2020 compared with 2019. Overall, these metrics indicate that the EUFR data are well represented by the AIS data and vice versa, reflecting a reasonably good quality and consistency of the AIS data used in our analyses.

Gear classification algorithm reliability
To summarise our gear classification process, we drew a representation of the total spatial extension and effort of each fleet in the Adriatic Sea. The effort of EU and non-EU vessels was aggregated in 0.1°cells and summed over 2015-2020 ( Figure 2 in the main document). A year by year summary of the AIS vessels recorded on these maps is reported in Table SI-2, which also shows their distribution according to the gear categories. The vessels that were estimated to use at least two gear types in the same year are counted in the mixed category. Vessel and trip numbers decreased over time, especially in 2018 after the decommissioning of EU vessels. The estimated vessel distribution in the gear categories varied little over time. OTB was the richest category with an almost stable average number of vessels (∼389), followed by PS (∼97 average) and TBB (∼45 average); PTM was the category with the smallest number of vessels (∼26 average). Trips with OTHER gear types were the least numerous (∼8 average) and mostly used gillnets and longlines. These gear types showed a steep increase of 20 vessels in 2016 and an average increase of 5 vessels after 2017. To assess the reliability of this classification, we calculated a confusion matrix between AIS-estimated and EUFR-licensed gears (Table SI-3). Since beam trawling is generally practised by vessels with a licence for bottom otter trawling, and the EUFR does not report Rapido trawling licenses, the OTB category incorporated TBB gear. Classification precision ranged from 66% (PTM) to 99% (OTB). The lower precision for PTM was due to its large confusion with OTB, whereas for PS (84% precision) the misclassification was balanced between OTB and OTHER. Our classification algorithm showed the highest precision for OTB, which was also the category with the largest number of registered and estimated gear types. Most of the OTB misclassification fell within OTHER and was due to gillnets and longlines, which share similar speed and spatial patterns. The high sensitivity for PTM, PS and OTB involved a generally small number of false negatives, except for OTHER (53% sensitivity), where the confusion was greater but uniform. The lower performance for OTHER was principally due to its under-representation in the data. Finally, the generally high specificity across categories (96-98%) indicates a very high true negative detection, i.e. agreement on the fact that a trip did not use a given gear. Based on the confusion matrix, Cohen's Kappa was 0.82, indicating excellent agreement 71 between AIS-estimated and EUFR-licensed gear over the five gear categories, hence a reasonable reliability of our classification algorithm.

Coarse-scale pattern analysis on Italian seas
The log-normal analysis of the GFW data of the Italian seas, described in Section 3.3, highlighted March-April effort pattern differences in 2019 and 2020 at a resolution of 0.5°( Figure 1 in the main document). Effort reduction can be generally observed in all Italian seas, but some locations still report a high fishing effort with thousands of hours. A higher resolution analysis of the AIS data for the Adriatic further explored this pattern. First, we visually compared the distributions of summed estimated fishing hours in 0.1°cells during the lockdown and non-lockdown months (Figure 3-a-b in the main document). In particular, we compared the annual aggregated February data (Figure 3-a) to those of March-May (Figure 3-b) in 2015 and 2020, and classified the high-effort locations using the 2015 log-normal ranges to produce comparable distributions over the years. The maps thus obtained show that, up to February 2020, fishing activity was equal to or higher than the previous years with a localised increase in the south-western Adriatic. In contrast, the high-effort cells were much fewer in the 2020 map, reflecting the lockdown effect in the March-May comparisons. The high-effort cells are consistent with those detected by the analysis of GFW data, but the higher resolution of the analysis indicates that these cells are actually more scattered.

Annual and monthly pattern analysis
In a further analysis, we compared the summed distributions of annual and monthly fishing hours in the whole Adriatic and in the high-effort cells (Figure SI-1-a-c). The comparison highlighted an increasingly intense fishing activity up to February 2020 ( Figure SI-1-a-All Fleets), the effort being comparable to the level recorded in 2019 in the high-effort locations ( Figure  SI-1-b-All Fleets). The lockdowns reduced fishing activity proportionally both in the high-effort locations and in the whole Adriatic ( Figure SI-1-a-b). The total number of fishing hours was lower in 2020 than in 2019 (-4.7%, from 985,356 to 939,229 hours, Figure SI-1-c-All Fleets).

7/
The monthly analysis identified a sharp reduction of fishing hours after the start of the March 2020 lockdowns (Figure SI-2-All Fleets) and a general recovery in May 2020, although not to the levels of 2019. The monthly patterns changed over the years due to regulations and decommissioning (especially in 2018, see Section 4.1). The scenario of February 2020 was very similar to the one of February 2019. Clearly, the lockdowns strongly reduced fishing hours from February to March 2020 (-37%, from 87,672 to 54,798 hours, Figure SI-2-All Fleets), but their recovery in May 2020 was fast (-22% compared with February, from 87,672 to 68,246 hours). The patterns in the high-effort locations reflect those of the whole Adriatic, with a stronger reduction between February and March 2020 (-35%, from 34,794 to 22,657 hours) and a smaller difference between February and May 2020 (-7%, from 34,794 to 32,432 hours). Thus, the fisheries recovered faster in the most intensely exploited areas than in the whole Adriatic Sea. In the second half of 2020 (Section S1), the total number of fishing hours returned to the average levels of 2019, indicating that the restrictions affected fishing activities only in the lockdown period.

Annual and monthly pattern analysis per fleet
To establish whether the patterns observed for all the fleets were shared by each fleet, we performed a spatio-temporal analysis for each fishing gear, which demonstrated a variety of patterns. The most important variations per fleet are summarised in Table  SI- The OTB fleet returned to pre-lockdown levels only at the end of 2020 (Section S1) and was therefore strongly impacted by the lockdowns.

Barycentre shifts during the 2020 lockdowns
We calculated the barycentres of monthly fleet activity between 2015 and 2020 ( Figure 4 in the main document and Section S2) based on the AIS dataset. For each fleet, we traced a bounding box around the barycentre to indicate the area encompassing its mean location in the period analysed. The TBB, OTB, and PS barycentres had a focus area of ∼1°latitudinal range, whereas the latitudinal range of the PTM barycentre was ∼2°. The barycentre shifts over time suggested the following considerations: the TBB fleet was active mainly in the north-western Adriatic and close to the Italian coasts (Figure 4-TBB), because this gear is commonly used by Italian fleets in shallow and sandy bottoms to catch common sole and some target shellfish species. In August, during the summer fishing ban, the barycentre moved southward. The 2020 lockdowns altered the TBB barycentre pattern and extended its range compared with the previous years. As regards the PS fleet, since these vessels are more abundant in Croatia, the barycentres are very close to the Croatian coast (Figure 4-PS). The barycentre usually moved southward, away from the coast, from February to May and then returned to the north at the end of the year (Section S2). However, during the 2020 lockdowns it moved northward and returned to the south only at the end of the year. The inversion compared with 2017-2019 was likely due to the pandemic restrictions. The barycentre of the PTM fleet was concentrated in the northern-mid Adriatic and shifted north from February to May (Figure 4-PTM). During the 2020 lockdowns it moved in a more northward direction, with slightly longer monthly shifts compared with the previous years, reflecting a greater geographical range. Moreover, it was more uniformly distributed over the latitudinal range compared with the previous years. Finally, the main location of the OTB barycentre was in the south-middle Adriatic (especially in August), with small shifts (Figure 4-OTB in the main document). During the 2020 lockdowns the pattern changed in March and April due to an increased activity range compared with the previous years and remained in this spatial range for the rest of 2020 (Section S2).

Impact on ETP Species
Several ETP species found in the OBIS database and the IUCN Red List live in the Adriatic Sea. They include the common smooth-hound (Mustelus mustelus), the loggerhead sea turtle (Caretta caretta), the European eel (Anguilla anguilla), and the spurdog (Squalus acanthias). In February and March-May the high-effort fishing areas and those inhabited by ETP species show a strong overlap (Figure 1 . Although after the lockdowns the total fishing effort reverted to 2019 levels (Section 4.3.2), the fishing hours at the locations at higher risk of impact remained below pre-lockdown levels throughout 2020 (Sections S3 and S4), indicating a general attenuation of impact on ETP species in the Adriatic Sea during the pandemic.
However, this was not true of all gear types. As summarised in Table SI

Project Information
The present study is part of the Snapshot-CNR project (of which M. Sprovieri and F. Trincardi are initiators and project leaders), whose aim is to provide a quantitative assessment of the effects of the reduced anthropogenic pressure on marine systems during 9/ the lockdowns that responded to the COVID-19 pandemic 72 . The 2020 restrictions generated unprecedented, and partially unexpected, human and marine ecosystem dynamics at various levels besides those related to fisheries. By analysing these dynamics in the Italian marine ecosystems, specific cause-effect relationships can be identified and extended to other world ecosystems. The aim of the project is to measure these relationships and the multiple factors involved -including pollution, the economy, fisheries and ecosystem services -to design novel strategies for a more sustainable future.
Table SI-1. Summary of the metrics used to assess the representativeness of EU Fleet Register (EUFR) vessels by the AIS dataset in the Adriatic Sea: Total number of vessels registered in the EUFR; Number of EUFR vessels that broadcast at least one AIS message (Expected vessels); Number of EUFR vessels included in the AIS dataset (AIS-observed vessels); Expected vessels divided by the total number of EUFR vessels (Expected coverage); AIS-observed vessels divided by the total number of EUFR vessels (Observed coverage); and AIS-observed vessels divided by the number of Expected vessels (Representativeness score).

Year
Total EUFR vessels Expected vessels