COBAC: An Adaptive Transhipment Station Localization for Reducing IUU Fishing Practices

The global consumption of �sh products is increasing on the scale of millions of tonnes every year. This makes the aquaculture industry as one of the leading sectors to provide food, employment, and ensuring a sustainable livelihood. The implication of rapid growth in global �sh production and massive consumption is causing productivity burden on the �sheries management to meet the market demands. This eventually leads to aggravated competition within the �shing networked community. For surviving the competition and increased pressure, few people from �shing community often indulge in various kinds of illegal �shing activities. Illegal, Unreported and Unregulated (IUU) �shing happens to be a major problem plaguing the �sh production. Our research proposes a solution based on o�cial transhipment station that solves the problems of illegal transhipment activities, thereby allowing transhipment to continue in a legal and safe manner. We have proposed Cost Optimisation Based Adaptive clustering (COBAC) algorithm that takes into consideration various operational cost and provides the location of establishment of the wirelessly operating transhipment stations in the ocean. The performance of the proposed transhipment was compared with random, greedy and heuristic approaches. Also, the experimentation results show that our proposed COBAC algorithm consumes one-tenth execution time as compared to Brute force clustering and produced result with 0.1% relative error.


Introduction
The consumption of shes has increased substantially all around the globe during the past few decades.The annual population growth rate of the world stood at 1.6 per cent while the annual sh food consumption growth rate is recorded as 3.1 per cent for the period 1961-2017(FAO, 2020)).Per capita sh consumption has increased from 9.0 Kg in 1961 to a record breaking of 20.5 kg in 2018 (FAO, 2020).
The increased demand has put enormous pressure on the shing community.As a consequence of which cases of involvement of shing boats operating outside the permitted region and resorting to tactics to avoid reporting sh catch has become rampant these days.In the light of increased demand on sh production, Illegal, Unreported and Unregulated (IUU) shing happens to be one of the major stumbling blocks in the process (Berveridge et al, 2013).

IUU Fishing & Its Adverse Implications
Illegal(I) shing refers to the practice of catching shes in disregard to shery regulation and violating norms of Exclusive Economic Zones (EEZ) (Camilleri).Unreported(U) shing is the practice of hiding or misreporting actual amount of sh landing by a shing vessel.Unregulated (U) shing refers to the act of operating in the region of a Regional Fisheries Management Organisation (RFMO), to which the country of the concerned shing vessel does not belong, in non-conformance with the management and conservation norms of RFMO.IUU shing accounts for 26 million tonnes of shes being caught, and consequently is responsible for a global loss of 10 billion to 23 billion USD per year (Agnew et al., 2009).
There are several other terrible consequences of IUU shing such as over-exploitation of marine resources and heightened threat to endangered species.
Transhipment refers to the process of transferring the shes from wirelessly operating shing vessels to reefer cargo ships.During this phenomena, the shing vessels often requires refuelling tanks and storing the catches.This allows the shing vessels to stay longer in the ocean without frequently returning to the home port.Though, the transhipment of shing vessels seems to be a great strategy to allow economical operation, however, it often gets indulged in hiding and smuggling illegal catches, thus leading to IUU shing in seas and ocean (Ewell et al., 2017).Sometimes, shing vessel also trap crew members forcefully on board, thereby violating labour and human rights.The situation is particularly worse in the countries where law enforcement is weak.For instance, incompetent monitoring of waters by countries of West Africa makes it a breeding ground for drugs and weapons smuggling (Bondaroff et al., 2015).The oceans areas susceptible to illegal transhipment activities include Indian Ocean, waters around Southeast Asia (Yea, 2016), Atlantic off West Africa (EJF, 2010) and Western Paci c. Fish species like tuna, Russian pollock, salmon along with crab and wild shrimp are specially threatened by unregulated transhipment (Ewell et al., 2017).Further, not only the monetary aspect of the problem is colossal, but also the disincentive to the small shermen is huge as the amount of legal catch is cut short by the illegal entrants.If such issues are left untouched or improperly handled they could lead to provoking those who follow rules aptly to indulge into the same.Therefore counteractive measures against IUU shing is the need of the hour (Azzi et al, 2019, Sunny et al., 2020).

Related work
A lot of work has been done to tackle IUU shing.Initially most of the work was in form of policies and laws stipulated by national, international sheries management bodies.One of the most widely used methods is to decrease the incentive of IUU shing using trade and market-related policies (Latun et al, 2013;Hosch, 2016).Higher prices are paid for shes that have proper documentation according to Catch Documentation Scheme of shery management body.Secondly, shes caught in the country supporting IUU shing, are imposed with high tariffs so as to make them unattractive to customers (Le Gallic et al., 2006).Adoption of measure on regional as well as national level along with strict compliance with international regulatory framework could also go long way in reducing the incidents of IUU shing (Vince et al., 2007;Johns, 2013).In 2009 the United Nations Food and Agriculture Organization (FAO) released Port State Measure Agreement (PSMA), under which the port states have to deny services to shing vessels involved in IUU shing (Flothmann et al., 2010).Port measures were effective but suffered from lack of universal implementation.Another way to force the shing vessels to not involve in IUU shing is to make amendments to insurance policy.Fishing vessels with history of illegal shing should be denied liability insurance (Soyer et al., 2018).The problem of IUU shing could also be seen from the lens of criminology.Rational choice theory and Situation crime prevention theory allow reducing the case of IUU shing in a systematic and scienti c manner.These theories try to nd a pattern in the crime and investigate the environmental facts that promoted the crime.Finding patterns in the crime allows coming up with prevention measures.A few broad ways of prevention could be to reduce the access to resource, to increase the risk of being caught (Petrossian, 2015).On the similar lines crime script analysis could be used for response generation (Petrossian et al., 2018).Using crime script analysis, IUU shing is treated as a crime and it is divided into various stages like Preparation, Entry, Target selection, Doing, Exit.Subsequently various responses in the form of policies are devised corresponding to each stage of the crime.
Game theory also nds its application in reducing IUU shing.The problem is modelled using Green Security Games (GSGs) approach which provides defender strategy to provide response to attacker (Fang, 2015).Algorithms could be developed for resource allocation and scheduling like patrol boat scheduling.Reinforcement learning is also used to provide solution (Akinbulire, 2017).It is used to model the problem as pursuer-evader game.Using various episodes of training, autonomous agent is able to learn strategies to chase the absconding shing vessel.Reinforcement learning could be used to provide autonomous patrol boats which could be used in response to an event of illegal shing.Data is considered the new gold, therefore, information sharing using tools like common database could help the countries cooperate.Regional Fishing Vessel Record database was created so that Association of Southeast Asian Nations (ASEAN) member states could access and provide information regarding shing vessels that are operating in the EEZ of another country or engaging in IUU shing (Matsumoto et al., 2012;Saraphaivanich et al, 2016).Since, it is not feasible to comprehensively span across the ocean with patrolling boats for tracking the illegal practices, hence, computational techniques are employed to classify and track the shing activities through wireless communication.
Classi cation algorithms are used to track and categorize a particular shing vessels activity as illegal or legal based on certain oceanographic parameters.These oceanographic variables include sea surface temperature, chlorophyll, and seascape (Woodill et al., 2020).Detection of shing gear of the vessel could help in deciding whether a shing vessel is exhibiting unexpected behaviour at sea.Detection of shing gear also helps in verifying whether a shing vessel is operating by registering wrong credentials.Such suspicious vessels are likely to be involved in IUU shing.Vessel Monitoring System (VMS) trajectories are used to predict the shing gear with the help of classi ers like Support Vector Machine (SVM) and Random Forrest (RF) (Marzuki et al, 2015(Marzuki et al, , 2017)).A wireless distributed Automatic Identi cation System (AIS) is used by vessels for the purpose of navigating and avoiding collision in sea.It can be used to provide vital information about the vessel.Global Fishing Watch used AIS messages to develop a classi cation model that predict when and where a shing vessel is engaged in shing and for how long.This allowed the regulatory agencies to know when a shing vessel is operating in a prohibited area or if it is involved in over shing (Merten et al, 2016).Most robust detection of illegal shing could be accomplished by combing AIS, VMS and Synthetic Aperture Radar (SAR) imagery.Vessels without VMS or with turned off AIS transceiver could be detected using SAR imagery (Longépé et al., 2018).The detection of illegal activity using SAR and Multi Spectral Imager (MSI) data from satellite in conjunction with AIS and VMS has provided a complete solution for surveillance (Kurekin et al., 2019).AIS messages could also be used to detect potential transhipment events at sea (Chuaysi et al., 2020).AIS messages allow us to measure speed of the vessels as well as duration for which the vessel is travelling at that speed.Using certain threshold for speed of the vessel and duration of the event, transhipment events could be identi ed (Miller et al, 2018).

Proposed Solution Approach
According to the literature survey, most of the methods are employed to tackle IUU shing for strengthening the wireless surveillance.A lot of work has been done using AIS and VMS data to detect illegal activity.Transhipment events can be detected using machine learning algorithms on AIS messages.Vessels identi ed as part of transhipment event could be marked as being involved in illegal activities and face the necessary action from the regulatory authority.But there is a caveat to the above approach.Transhipment is a necessary evil.Transhipment is useful as it allows shing vessels to make their operation economically e cient and at the same time save time.Therefore, a safe way out of this situation is to establish Transhipment station in high seas and ocean.Transhipment stations will work as the authorized station only where transfer of shes from one vessel to another vessel is allowed.If the transhipment happens at any other location by the very de nition of this approach, such transhipment is illegal.We propose Cost Optimisation-Based Adaptive Clustering (COBAC) technique to nd the optimal locations for setting up the wireless transhipment stations by clustering the shing events.

Research Highlights
The global consumption of sh products is increasing on the scale of millions of tonnes annually.
For surviving aggravated competition, shing community people often indulge in IUU shing.
We have proposed Cost Optimisation-Based Adaptive Clustering (COBAC) to address the issue.
COBAC considers operational cost to compute the optimal location of wireless transhipment stations in the ocean.
Our algorithm addresses IUU shing problems for sustainable sheries management.

Methods & Models
Two obvious questions arise because of this approach: 1)Where will the Transhipment stations be established?2) Is the operation cost of Transhipment stations feasible?The answer to the above question can be found using a machine learning optimization algorithm that will provide as output the number and location of Transhipment stations.The algorithm will take as input various cost that will be incurred due to this solution.First is the economic cost of setting up stations, and then comes the annual cost of operation, followed by the cost of extra fuel required by shing vessel to travel to the Transhipment stations.The algorithm provides a complete solution.

Solution Formulation
The solution involves using a clustering algorithm.The algorithm needs data regarding the shing activity in seas and ocean.It is provided by Global Fishing Watch's shing effort data (Kroodsma et al, 2018).The dataset contains information about 2.83 million shing events in a single year.In order to nd location for Transhipment stations, there is a requirement of organising the shing events into grids based upon the latitude and longitude of shing event.The coordinates of a grid's centre will be the location of Transhipment station, if that grid is selected for setting up Transhipment station.Further there is a requirement of minimizing cost function F n st along with clustering (Eq.1).
minF n st = C ves × D total + n st × C st s. tn min ≤ n st ≤ n max (1) C ves is the cost of covering unit distance in ocean by a shing vessel.D total is the sum of distance travelled by all shing vessels to reach to transhipment station.C st is the annual cost of operating a transhipment stations.n st is the number of transhipment stations.n min and n max is the minimum and maximum number of transhipment stations allowed respectively.If n st is increased then annual cost of operating transhipment stations ( n st × C st ) is increased and if n st is decreased, cost of fuel ( C ves × D total ) increases.Therefore the trade-off between both the costs is non-trivial.

Proposed COBAC & Other Algorithms
In order to reduce the time complexity of the clustering algorithm the world map is divided into 2 ∘ × 2 ∘ grids.Consequently 16200 (360/2 × 180/2) grids were generated using Cluster grid generation algorithm (Algorithm 1).Since approximately 70% of earth surface is covered by ocean and not all places on ocean witness shing activity, we nally get n max = 2360 grids for clustering.The rfmo() function takes a grid as input and returns the RFMO that should be responsible for the grid based upon country ag of ships in that grid.The location() function provides the coordinates of the centre of the grid.The ship_count() function provides the count of shing ships active in the grid in Algorithm 1.After creation of grids, information regarding the grid centre coordinate is available.Using grid centre location, clusters of grids could be formed.Now the important task is to determine the number of clusters.There are several ways to determine the number of clusters.One way is to try every number of clusters from 1 to n max , which, however, is quite time consuming.Therefore, proposed Cost Optimization-Based Adaptive clustering (COBAC) algorithm (Algorithm 5) was proposed to determine the optimal number of clusters in less number of iterations (Eq.2-8).For each cluster of grids formed by the algorithm, there is a transhipment station.Here, n st (t) refers to the number of transhipment stations att th iteration of proposed COBAC algorithm.
F n st (t − 1) (3) n st (t) = min max transform(π(t)) + n st (t − 1), n min , n max ( 8) floor(x) = ⌊x⌋(10) The COBAC algorithm after the initial iteration will provide a guess for the number of transhipment stations.The only thing preventing us from using K-means is that we did not know the number of clusters.Now that number of clusters is known, we should be able to use K-means.However, usually the clustering algorithms use commutative distance measure (Eq.11), but because of grids the distance measure is no longer commutative (Eq.12).Therefore, we need a clustering algorithm that works on a distance measure that follows Eq. 12, to minimize the cumulative cost of clustering, D total (Eq.13).
Here, n grids is the total number of grids with shing activity.G i (Fig. 1 blue blocks) refers to the i th grid and C i (Fig. 1, yellow blocks) refers to the grid that is the centre of the cluster of grids.
Further, centre(x) returns the latitude and longitude of a centre of the gridx.DistEarth(u, v) refers to distance in kilometres between two locations, u and v according to the Geographic Coordinate System (GCS) (Eqs.14-17).u lat and u lon represents the latitude and longitude of the location u.Δ lat is the difference between the latitude and Δ lon is the difference between longitude of the two locations, u and v.
The proposed COBAC algorithm is given as Algorithm 5.It takes as input various costs associated with clustering (C ves , C st ), data regarding grids, initial estimate of number of transhipment stations and number of iterations to search for optimal number of transhipment stations.The Algorithm 5 outputs optimal number of transhipment station, complete information (in charge RFMO, location) regarding transhipment station and total optimal cost of operation (Eq.1).The algorithm also returns the transhipment station responsible for a particular grid.The algorithm uses station_generation() method given as Algorithm 2 to nd the best transhipment stations corresponding to the number of transhipment stations allowed.Algorithm 3 also nds transhipment station by selecting rst the grids with most shing vessel as transhipment station in a greedy fashion.Algorithm 4 nds transhipment station by random selection of grids as transhipment station.However, the best performance is achieved with Algorithm 2.
( ) ( ) ( ) ( ) ( ) fori ∈ 0,1⋯n grids − 1 : forj ∈ 0,1⋯n grids − 1 : end for fori ∈ 0,1⋯n grids − 1 : forj ∈ 1⋯n grids − 1 : fori ∈ 0,1⋯n grids − 1 : positive or negative for a number of consecutive iteration, ρ(t) is used to have compounded effect on the change in the number of stations,π(t).In short, momentum factor ρ(t) is used to accelerate increase or decrease in the number of stations.The nal update takes place using Eq. 8. transform() is used to ensure effective change in the number of station is always a non-zero integer.In the expression for ρ(t) in Eq. 6, an arbitrary value of 0.5 is used to prevent ρ(t) from becoming zero and stalling the algorithm.

( ) ( )
One of the most important datasets used in the study is "Daily Fishing Effort at 10th Degree Resolution by MMSI, version 1.0 (2012-2016)" (Dataset A) (Kroodsma et al, 2018) 27 .This dataset provides information regarding daily shing activity all around the global.It provides the number of shing hours spent by a shing vessel identi ed with Maritime Mobile Service Identity (MMSI) at a particular latitude and longitude.For the purpose of developing clustering algorithm, only data for the year 2012 was taken.
"Fishing vessels, version 1.0 (2012-2016)" (Dataset B) (Kroodsma et al, 2018) 27 dataset provided information pertaining to MMSI of shing vessel and the ISO 3166-1 alpha-3 code of country to which the vessel is registered.In order to get comprehensive data regarding MMSI and the country is associated with, several other datasets were "Identifying Global Patterns of Transhipment Behavior" (Dataset C) (Miller et al., 2018).The Global View of Transhipment: Revised Preliminary Findings (Dataset D) (Kroodsma et al., 2017) also provided MMSI and country ag information.In few of the datasets complete name of the country in place of ISO code of the country corresponding to an MMSI is present.Therefore a dataset from Github repository is downloaded to get the mapping from complete country name to ISO code of the country (Dataset E) (https://github.com/lukes/ISO-3166-Countries-with-Regional-Codes/tree/master/slim-3).The nal dataset prepared is called FishTank (Fig. 2). [IMAGE-C:\Workspace\ACDC\ImageHandler\d3 Initially, dataset of shing activity with all required attributes is generated by adding standard dataset.The nal dataset, has the following attributes: latitude, longitude, shing hours, MMSI, ISO code.ISO code of the country with which MMSI is associated with, will help in deciding which RFMO is responsible for a particular transhipment station.The size of dataset required to get the location of transhipment station such that Eq. ( 1) is minimized, is huge.Therefore, there was a requirement to divide the entire world map into grids of dimension2 ∘ × 2 ∘ .For this purpose, Cluster grid generation algorithm (Algorithm 1) was used that the dataset FishTank as input with dimension, dim = 2 ( g. 3).The cluster grid generation algorithm returned 2 ∘ × 2 ∘ grids and information regarding the grids.Information regarding the grids includes the RFMO responsible for the grid, total number of unique MMSI ( shing vessels) present in the grid and coordinates of the centre of the grid.Fig. 5 shows the area of jurisdiction for a limited set of RFMOs for the sake of clarity.Fig. 4 plots the 2 ∘ × 2 ∘ grids and intensity of colour denotes the number of shing vessels present.Further

Experimental Results
The proposed COBAC algorithm performs well in terms of time complexity when compared with other clustering algorithm (section 3.1).Its performance is also compared with the Brute force clustering algorithm.In order to arrive at the best version of proposed COBAC algorithm, several version of wireless transhipment station generation algorithms are also explored.

Time Complexity Analysis
There are various clustering algorithms available which can be used to cluster the shing events.Disadvantage with a lot of clustering algorithms is their large time complexity (Table 1) (Xu et al., 2015) 28 .Algorithms with large time complexity take a lot of time to generate clusters.et al., 2015) 28 .However, there is a signi cant shortcoming even with algorithms with low time complexities.These algorithms either take as input the number of cluster or nd the number of clusters without minimizing objective function (Eq.1).    4. Algorithm 2 outperforms all the other algorithms (Algorithm 3, 4) used to get the best transhipment stations corresponding to a particular value of n st (t) (Table 4).Fig. 9 shows the nal distribution of transhipment station along with the 2 ∘ × 2 ∘ grids on map of world, where blue dots represent transhipment stations and feeble reds dots represents grids.

Discussions
Our research attempts to address the issue of IUU shing by focusing on the problem of illicit usage of transhipment in oceans.Though, transhipment solves several problems encountered by shing vessels functioning in the oceans and seas, however falls under the grasp of illegal activities.Our study offers a solution by proposing to organize the transhipment activity being carried out in oceans.O cial wireless transhipment stations could be set up at appropriate locations in oceans and will be responsible to arrange interaction between cargo vessels and shing vessels.The interaction will be monitored by appropriate regulatory authority and consequently this will put a check on all the illegal activities being carried out in lieu of necessary activities.This approach will also make it easier to detect illegal transhipment in oceans.Several methods have already been developed to detect transhipment activity in ocean.Since, the transhipments will only be allowed on o cial stations, all the rest of the transhipment activity apart from that on o cial station, could easily be termed as illegal.Automatically, there is a requirement of a system that could provide the location for establishing o cial transhipment station, so that their operation is economically e cient.

Conclusion
Our proposed COBAC algorithm provided by this study delivers the locations of the wireless transhipment stations keeping in consideration the cost of operation of stations as well as the extra cost incurred by shing vessels to reach the stations.The stations could be managed by appropriate RFMO.Therefore, the algorithm also assigns a RFMO to a transhipment station depending upon the location of transhipment station and country ag of active shing vessels.COBAC algorithm takes one-tenth execution time as compared to Brute force clustering algorithm and produces result with 0.1% relative error.Our COBAC algorithm was capable to locate optimal number of stations in one-eighteenth number of iterations as compared to Brute force clustering.The algorithm also enjoys a time complexity of O(1) because of using grid structure.For our experimentation, shing activity of only year 2012 was considered.The algorithm could be made more e cient by using shing activity data for more than one year.Further, information regarding the routes followed by cargo vessels could also be incorporated to make the whole arrangement more economic for cargo vessels as well.Moreover, the data regarding anchorage points could also be used for deciding whether station should be set up in ocean-bed or onland.

Declarations
, dim) sort data in ascending order according to latitude divide data into binsof dim degree according to latitude for each b ∈ bins: sort b in ascending order according to longitude divide b into gridsof dim degree according to longitude for each g ∈ grids : grids[g][ ' rfmo ' ] = rmfo(g) grids[g][ ' location'] = location(g) grids[g][ ' ship_count] = ship_count(g)

Fig 3 :
Fig 3: Flow diagram of proposed COBAC algorithm proposed COBAC algorithm (Algorithm 5) is invoked.The input parameters include, Distance cost (C ves ), Annual operation cost (C st ), Information regarding grids, initial change in number of transhipment station (π(0)), initial number of transhipment station(n st (0))and number of iterations (n iter ) to execute our proposed COBAC algorithm.The proposed COBAC algorithm (Algorithm 5) returns as output the minimized cost of implementing well-localized transhipment station strategy.It also returns the optimal number of transhipment Fig 4: Distribution of 2 ∘ × 2 ∘ grids & density of shing vesselsstations and assignment of all the 2 ∘ × 2 ∘ grids to a transhipment station.PICESICESIATTCICCATCCAMLRCACFishIOTCFig 5: Distribution of RFMOs on world map

Fig 6 :
Fig 6: Optimal total cost of proposed COBAC

Table 1
Symbols used in Proposed COBAC algorithm π(t)Change in number of transhipment station at t th iteration ρ(t) Momentum factor for change in number of transhipment stations at t th iteration ( )

Table 1
Clustering algorithms and their time complexity

Table 3
Performance of Proposed COBAC & Brute Force Search Clustering Algorithm

Table 4
Performance of different station generation algorithms