ISSN 2305-7254________________________________________PROCEEDING OF THE 34TH CONFERENCE OF FRUCT ASSOCIATION

Dark Activity Detection in AIS-Based Maritime

Networks
Bekir Nazmi Görkem, Burak Çağlayan, Erkam Karaca, Candar Karabulut and Ömer Korçak
Computer Engineering Department
Marmara University
Istanbul
Email: omer.korcak@marmara.edu.tr

Abstract—The Automated Identification System (AIS) is an to Ron Crean, vice-president for commercial at Windward
indispensable tracking system employed in the maritime in- Maritime Analytics [3]. However, we believe that vessels
dustry for vessel identification, location tracking, and collision working cooperatively can overcome this issue; obviously, not
avoidance. While AIS messages provide essential information for
maritime traffic management, they also present challenges when if the AIS was turned off before sailing in the first place.
vessels aim to conduct operations discreetly or evade observation. Recognizing the significance of AIS signals for maritime
This phenomenon, referred to as ”dark activity”, involves inten- security, we embarked on a mission to develop a rule-based
tional AIS deactivation by vessel operators seeking to conceal decision-making algorithm. This algorithm is designed to help
their actions, often related to illicit or illegal maritime activities vessels meeting specific range criteria determine if a neigh-
such as smuggling, piracy or illegal fishing. The detection and
monitoring of dark activities pose significant challenges for boring vessel has entered a state of dark activity. To realize
law enforcement and security agencies. This paper explores our vision, we crafted a simulation environment that vividly
innovative approaches to address this issue by harnessing AIS portrays vessel movements, including scenarios involving dark
data and incorporating rule-based techniques, as well as machine activity. We harnessed machine learning models to enhance
learning techniques to enhance maritime security efforts. We the accuracy of predictions, utilizing data gathered from our
adopted a local approach where a dark activity of a vessel is
detected by nearby ships depending on the previous signals. We simulation experiments.
implemented a detailed simulation environment based on real and This paper stands as a fundamental tool, contributing to the
realistic data to run the proposed algorithms. Simulation results enhancement of maritime security by combatting smuggling
show that while rule-based approach is successful in detecting and curbing illegal activities on the high seas. Moreover, it
dark activities, it tends to produce false alarms, and ML-based plays a pivotal role in safeguarding marine life and preserving
approach provides better overall accuracy.
the delicate ecological balance by detecting illegal fishing ac-
tivities. Our primary objective revolves around the detection of
I. I NTRODUCTION
illegal vessel activities within specific sea regions. Leveraging
In the vast expanse of open seas, vessels communicate crit- AIS signals, we pursued the following key steps:
ical information to ground stations through satellites or direct • Creating a realistic simulation environment with a user-
means. This communication takes the form of an AIS signal, friendly interface, enabling the visualization of vessel
which contains essential data like the vessel’s current position, movements and the adjustment of parameters related to
speed, course over ground, and heading [1]. These parameters dark activity.
are updated automatically and broadcasted throughout specific • Designing and implementing a rule-based decision-
time intervals depending on vessel’s movement and message making algorithm to identify vessels entering a state of
type. However nothing can stop vessels from turning off dark activity.
their AIS transmitters and going ”dark” intentionally. The • Gathering a dataset from the simulation environment to
authorities cannot tell for sure whether a ship switched off fuel machine learning algorithms.
its AIS to hide its location for some illicit dark activities • Simulating both algorithms within the environment and
(e.g., smuggling, fishing in restricted areas, or unauthorized assessing their accuracy
waste disposal) [2], or its AIS signal cannot be received due Through these endeavors, we aim to strengthen maritime
to natural reasons such as heavy whether condition or signal security, safeguard the environment, and ensure responsible
congestion. Particularly in the open seas, where satellite-based and legal conduct on the high seas.
AIS (S-AIS) is used, the ships can always deny this dark
activity because the signal can be lost the way to the satellite A. Related Work
especially in congested waters. Thus, only a nearby ship can In the literature, there are several studies on dark activity
be aware of the dark activity. The ratio between lost signals detection. Shahir et al. [4] addressed the critical issue of
unintentionally due to conditions beyond AIS and deactivating maritime domain awareness, emphasizing its significance in
the transmitter on purpose is something between 1:10 and preventing smuggling and safeguarding vital sea-based struc-
1:20 depending on ship type and geographical area according tures. Their solution consisted of three phases: i) Engage-

---------------------------------------------------------------------------- 35 ----------------------------------------------------------------------------
ISSN 2305-7254________________________________________PROCEEDING OF THE 34TH CONFERENCE OF FRUCT ASSOCIATION

ment Detection: Vessels in close proximity were clustered time with AIS data from 30 minutes before and after the image
together; ii) Detection of Candidates: Candidates for engage- capture. Utilizing the K-nearest neighbors (KNN) method,
ment were identified based on kinematic features, particularly they matched satellite images with AIS data, identifying ships
slow speeds and converging courses or close proximity; iii) present in images but absent from AIS records as potential
Scenario Detection: Leveraging the results of engagement participants in Dark Activity.
detection, scenarios were represented by left-to-right Hidden This paper introduces a distinct approach from previous
Markov Models and classified using Support Vector Machines. studies. Due to the fact that only nearby ships can notice
Moreover, an additional phase was introduced for anomaly a vessel turning off its AIS transponder, ships working co-
detection, extending the scope of scenario detection to rectify operatively can be crucial in detecting dark activities. Thus,
misclassified scenarios. instead of depending on external sources like satellite imagery,
Mantecon et al. [5] address the challenge of maritime buoys with 360-degree cameras, or X-band radar systems, we
threats and illegal activities, and employ convolutional neural solely rely on local AIS signals. We put forward rule-based and
networks (CNN) to derive navigation patterns based on ship machine-learning-based algorithms to identify dark activities
speed, direction, and maneuverability, using a dataset called of nearby vessels in real-time when their AIS transmitters
DeepMarine, derived from historical AIS data. Then both AIS are deactivated. Our approach offers a cost-effective solution
and Radar trajectories can be used to identify various vessel compared to alternative methods and can be used alongside
behaviors. The proposed method requires both the positional more expensive solutions to enhance the overall efficacy of
data and ship information to detect illegal activities such as illegal activity detection.
fishing in non-allowable areas. The authors in [2] presented
an anomaly detection methodology to discriminate between II. M ETHODOLOGICAL BACKGROUND
AIS messages that are not received by base stations due to We consider a system model where every vessel may verify
communication channel-related effects and those that were not the activities and status of nearby vessels through AIS signals.
broadcasted at all to cover dark activities. The strength of the The horizontal range of vessel-to-vessel AIS signals is 20-30
received signal RSSI is analyzed to detect On/Off switching of nautical miles under most atmospheric conditions [9]. AIS is
the ship’s transponder. A training set of known good AIS data obligatory for ships that meet specific criteria. According to
is used for comparison with received data, an alert is triggered IMO (International Maritime Organization), passenger vessels
when signal dropouts exceed a defined threshold. Eaton et al. irrespective of size, all ships engaged on international voyages
[6] introduced a novel dark activity detection concept called with size of 300 gross tonnage, and cargo ships of 500 gross
Sensors and Platforms for Unmanned Detection of Dark Ships tonnage are required to have AIS transmitter[10]. AIS signals
(SPUDDS), which combines hardware and software compo- should be transmitted at intervals ranging from 2 to 180
nents. SPUDDS involves an autonomous buoy equipped with seconds, with the specific interval determined by the velocity
various sensors and software called CROWSNEST, designed and the change in the course of the vessels [11].
for ship identification and classification. The system accurately When a vessel broadcasts an AIS signal, nearby vessels
categorizes detected ships, including sailboats, merchant ships, equipped with AIS receivers within the coverage area will
and fishing vessels, using a highly precise machine learning receive it. Therefore, if vessel A receives the AIS signal from
algorithm. CROWSNEST relies on a convolutional neural vessel B but then stops receiving it, there could be three
network and data-driven ODF for ship classification. The possible reasons: i) Vessel A has moved out of the coverage
integration of a 360-degree camera on the buoy enhances its area of vessel B due to mobility, ii) Signal collision and
capabilities for safeguarding maritime security. Paolo et al interference have occurred, iii) Vessel B has entered dark
[7] suggests using Synthetic Aperture Radar (SAR) images activity.
and automated machine learning in order to detect illegal In our system model, vessels have the capability to verify
fishing activities. They constructed and released xView3-SAR the activities of nearby vessels and determine if they have
dataset for maritime object detection and characterization, and entered dark activity. These vessels are referred to as detector
combine AIS and human annotations for labeling the data. vessels. To provide a clear explanation, we will concentrate on
Bereta et al. [8] employed satellite imaging techniques to a scenario involving a single detector, represented as vessel
achieve a 95% accuracy rate in detecting Dark Activities. They 5 in Fig.1. As depicted in the figure, each vessel has a
emphasized the limitations of relying solely on AIS signals, predefined broadcasting range for its AIS signal, which varies
leading them to propose a hybrid approach using satellite depending on its type. The green zones indicate areas where
data, specifically Copernicus Sentinel imaging, in conjunction vessels’ signals can be received by the selected detector vessel,
with Marine Traffic AIS data to monitor ship density in areas while the red zones signify that their broadcasting range is
of potential Dark Activity. The project involved acquiring insufficient to transmit their AIS signal to the detector vessel.
data from the Alaska Satellite Facility and Copernicus Data At time t, vessel 5 receives signals from vessels 1 and 4 but
Hub, followed by preprocessing to remove irrelevant image cannot receive signals from vessels 2 and 3. In the subsequent
portions and filtering out cloud-obscured images, reducing time step, vessel 3 comes within range, while vessel 4 goes out
data volume from terabytes to gigabytes. The data fusion step of range. Even though vessel 4’s signal is no longer received,
synthesized satellite and AIS data, aligning the satellite image vessel 5 refrains from making a dark activity decision because

---------------------------------------------------------------------------- 36 ----------------------------------------------------------------------------
ISSN 2305-7254________________________________________PROCEEDING OF THE 34TH CONFERENCE OF FRUCT ASSOCIATION

)LJ  'DUN $FWLYLW\ 6FHQDULR 9HVVHO  HQWHUV GDUN DFWLYLW\

GHWHFWHG E\ YHVVHO 

)LJ  )ORZFKDUW IRU WKH SURFHVV RI UHFRUGLQJ QHDUE\ $,6 VLJQDOV
this situation was expected based on the vessel’s location,
E\ WKH GHWHFWRU YHVVHO
course, and speed in the previous time step. However, at time
t+2, the signal from vessel 1 is unexpectedly not received,
prompting a dark activity decision. Even if vessel 1 reactivates absent in the current temporal segment, a verification process
its AIS signal after a period of dark activity, the decision of is conducted concerning the estimated vessel location. This
”possible dark activity” persists. estimation is derived from previously acquired AIS parameters
and utilizes kinematic principles for computation.
III. DARK ACTIVITY D ETECTION
To accurately detect dark activities, a critical challenge lies
in predicting a ship’s future position based on the AIS data
received at the current moment. A straightforward method
involves utilizing vessel kinematics, considering factors like
location, speed, and course. However, this approach may yield
incorrect results if the vessel alters its course, speed, or other
parameters. To enhance prediction accuracy, a machine learn-
ing approach becomes imperative. The subsequent subsections
will elaborate on the proposed methods.
A. Rule-based Dark Activity Detection (R-DAD) Algorithm
A detector vessel continuously monitors AIS signals from
nearby vessels, recording their transmitted parameters. It cal-
culates the expected positions of these vessels in the next
time step by applying fundamental physics principles to the
received AIS data. These calculated positions, along with the
corresponding AIS parameters, are stored. During each time
step, the detector vessel checks whether the estimated positions
of nearby vessels from the previous time step fall within its
reception range. If a vessel’s estimated position is within the
detector vessel’s reception range but no corresponding AIS
signal is received, it triggers a potential alert for dark activity Fig.
3
Flowchart
of
the
Rule-based
Dark
Activity
Detection
(R-
DAD)
algorithm
detection.
To monitor nearby vessels, the detector vessel maintains a
database of received AIS signals, as depicted in the flowchart In this context, we employ the Haversine formula to
presented in Fig. 2. Subsequently, based on the AIS signal compute both the expected vessel location and the distance
intervals, a designated time period is established to assess between vessels. The Haversine formula serves as a precise
the potential occurrence of dark activities. Fig. 3 displays method for calculating the distance d between two points on
the procedural workflow of the R-DAD algorithm. When the the surface of a sphere, based on their respective latitudes (φ1 ,
AIS signal from a vessel was previously acquired but is φ2 ) and longitudes (λ1 , λ2 ). It is defined as follows:

---------------------------------------------------------------------------- 37 ----------------------------------------------------------------------------
ISSN 2305-7254________________________________________PROCEEDING OF THE 34TH CONFERENCE OF FRUCT ASSOCIATION

TABLE
I
FEATURES
USED
IN
MACHINE
LEARNING


φ2 − φ1 λ2 − λ 1 Features Target
a= sin2 ( ) + cos φ1 · cos φ2 · sin2 ( ) (1) ais range
2 2 distance from nearby vessel
nearby vessel heading
d = 2R · arcsin(a) (2) nearby vessel course
nearby vessel speed
where longitudes and latitudes are defined in radians and R nearby vessel distance to turn point
dark activity
is the radius of the Earth. heading
To find the estimated location of a vessel, we first calculate course
speed
the covered distance with its given speed, assuming that distance to turn point
the speed is constant. Then given the initial latitude (φ1 )
and longitude (λ1 ), initial bearing (θ, clockwise from north),
and the covered distance, the estimated location (φ2 ,λ2 ) is the presence of dark activity. Fig. 4 illustrates the ML-DAD
calculated with the following formula [12]: algorithm. Instead of location estimation, dark activity decision
is done according to the trained model.
φ2 = arcsin(sin φ1 · cos δ + cos φ1 · sin δ · cos θ) (3)
λ2 = λ1 + arctan 2(sin θ · sin δ · cos φ1 , cos δ − sinφ1 · sin φ2 )
(4)
where δ is the angular distance d/R.
If the estimated location is out of range of the detector
vessel, in other words if the distance calculated by (2) is
more than the AIS range, this scenario is categorized as a
typical operational condition. However, when the estimated
location resides within the range, yet no AIS signal is detected,
then this is interpreted as a potential dark activity, and an
alert is triggered. It is noteworthy that one can establish a
probabilistic assessment of the likelihood of dark activity
based on the proximity of the anticipated position to the
boundary of the AIS signal range. This approach necessitates
further investigative exploration and research.
B. Machine Learning Based Dark Activity Detection (ML-
DAD) Algorithm
In order to improve the precision of detecting dark activities,
we employ a machine-learning technique. Initially, we create a
dataset through a realistic simulation detailed in Section IV-A, )LJ  )ORZFKDUW RI WKH 0DFKLQH/HDUQLQJEDVHG 'DUN $FWLYLW\
which encompasses the attributes listed in Table I. Each data 'HWHFWLRQ 0/'$'  DOJRULWKP
instance is also paired with a target value denoting whether it
corresponds to a dark activity or not. The dataset comprises
IV. S IMULATION
information about the detector vessel’s course, heading and
speed, as well as those of the selected nearby vessel. Addi- A. Simulation Setup
tionally, the distance between two vessels and the AIS range To overcome the cost associated with obtaining real-time
of the nearby vessel is also used. Notably, a unique feature S-AIS data, we created a realistic dataset by leveraging actual
included in the dataset is the distance to the turn point, a port locations, established routes, and real-time vessel location
parameter not typically conveyed in current AIS messages. It data from MarineTraffic [10]. This dataset is stored in JSON
is worth mentioning that the AIS protocol accommodates a format and comprises three static data types: ports, routes, and
total of 64 message types, with 27 of them already allocated vessel data.
for specific purposes. The introduction of a novel message type The ports data represents widely used ports and includes
featuring the distance to the turn point holds the potential to three fields: name, latitude (lat), and longitude (long). We
enhance maritime traffic control, and this study presupposes initially compiled this static data by observing real maritime
its utilization for improved accuracy. networks.
Utilizing the dataset we constructed, we train a machine- The routes data, also static, contains five fields: ”from,”
learning model through a supervised learning algorithm. Sub- ”to,” ”density,” ”noise,” and ”coordinates.” ”From” and ”to”
sequently, for each nearby vessel whose AIS signal was denote the starting and final ports, respectively. ”Coordinates”
received in the previous time step but is currently not being is an array containing the latitude and longitude of intermedi-
received, we employ the model to make predictions regarding ate turning points along the route. ”Density” and ”noise” are

---------------------------------------------------------------------------- 38 ----------------------------------------------------------------------------
ISSN 2305-7254________________________________________PROCEEDING OF THE 34TH CONFERENCE OF FRUCT ASSOCIATION

Fig. 5 A snapshot from the simulation map

arrays in which density[i] and noise[i] specify vessel density )LJ  $ VQDSVKRW IURP WKH VLPXODWLRQ VFUHHQ GHSLFWLQJ WKH $,6
and noise levels between coordinates[i] and coordinates[i+1]. VLJQDOV GHWHFWHG E\ D GHWHFWRU YHVVHO
Vessel data is randomly generated along the predefined
routes. The number of vessels between coordinates is de- (with bagging), Support Vector Machines, Logistic Regression,
termined by the corresponding density value from the route and AdaBoost Random Forest. 30% of the data is allocated
data. These vessels are randomly distributed between two for the training set, while the remaining 70% is designated for
coordinates, with their distance from the route randomly the test set.
chosen within the range of -noise to +noise. Vessel data en-
compasses attributes such as Maritime Mobile Service Identity B. Numerical Results
(mmsi), type, AIS range, AIS Interval, course, heading, speed, Following the execution of the simulation involving both R-
longitude, and latitude. The course of the vessel is calculated DAD and ML-DAD algorithms, with the latter employing var-
according to the following equation: ious machine learning models, we assess the outcomes based
sin Δλ · cos φ2 on the accuracy of dark activity detections. We categorize these
θ = arctan( ) (5) assessments into the following cases:
cos φ1 · sin φ2 − sin φ1 · cos φ2 · cos Δλ
• True Positive (TP): A vessel is not involved in dark
where φ1 , λ1 and φ2 , λ2 are latitudes and longitudes (in activity, and the prediction is accurate.
radians) of starting point and ending point respectively, and • False Negative (FN): A vessel is not involved in dark
Δλ is the difference in the longitudes. Additionally, a binary activity, yet the prediction is incorrect.
variable named ”dark activity” is defined to simulate vessel • False Positive (FP): A vessel is involved in dark activity,
dark activity. but the prediction is incorrect.
We focused our study on the Marmara Sea area and sim- • True Negative (TP): A vessel is involved in dark activity,
ulated a scenario involving 87 vessels following predefined and the prediction is correct.
routes. Circular routes were designated for fishing vessels,
Then, the accuracy is calculated according to the following
and Fig. 5 provides a snapshot of a specific time instance
equation:
within this scenario. Our simulation environment allowed for
the selection of any vessel as the detector vessel, while any TP + TN
Accuracy = (6)
vessel other than the detector could be chosen to enter dark TP + FP + TN + FN
activity. Accuracy ratio for R-DAD algorithm is found to be 0.892,
Fig. 6 shows another snapshot from simulation screen, while the accuracy of the ML-DAD algorithm can be seen
where the detector vessel is selected and shown by red color, in Table II. AdaBoost random forest emerges as the best-
and the nearby vessels whose signals are detected are shown performing supervised machine-learning algorithm with an
by yellow color. AIS signals of the vessels that are shown in
dark blue color do not reach to the detector vessel, either due
to long distance, or due to switching of the AIS transmitter
and entering to dark activity.
Fig. 7 illustrates the basic system architecture of the simu-
lation environment.
The simulation was executed through a total of 10 iterations,
each spanning 1000 time steps. Every 10 time steps, a detector
vessel was randomly chosen, and during each time step, a
vessel was selected at random to engage in dark activity. We
conducted experiments employing both R-DAD and ML-DAD
algorithms. Within the ML-DAD algorithm, we utilized a va-
riety of supervised machine learning techniques, including K-
Nearest Neighbors, Decision Trees, Artificial Neural Networks Fig.
7
System
architecture
of
the
simulation
environment

---------------------------------------------------------------------------- 39 ----------------------------------------------------------------------------
ISSN 2305-7254________________________________________PROCEEDING OF THE 34TH CONFERENCE OF FRUCT ASSOCIATION

TABLE
II
ACCURACY
RESULTS
OF
ML-DAD
FOR
predictions in lawful cases, provided an appropriate supervised
VARIOUS
MACHINE
LEARNING
ALGORITHMS
model was selected.
ML Algorithm Accuracy This research contributes valuable insights into the detection
K-Nearest Neighbors 0.942 of dark activities in maritime networks, shedding light on
Decision Tree 0.952
Artificial Neural Networks 0.898
the effectiveness and limitations of rule-based and machine-
Suppor Vector Machines 0.942 learning-based approaches. Further exploration of these meth-
Logistic Regression 0.864 ods and their integration into real-world maritime security
AdaBoost Random Forest 0.961 systems holds promise for enhancing safety and security
at sea. A potential avenue for future research involves the
accuracy ratio of 0.961. Table III displays the confusion development of a collaborative system wherein multiple de-
matrix for the R-DAD algorithm, while Table IV presents tector vessels engage in intercommunication to collectively
the confusion matrix for the ML-DAD algorithm utilizing the identify proximate illicit activities. Another promising area
AdaBoost random forest model. It is worth noting that, in the of investigation entails the utilization of machine learning
machine-learning-based approach, although there is a slight techniques to ascertain regions with elevated susceptibility to
increase in false positive cases, there is a significant decrease covert activities. Subsequently, this spatial information can be
in false negatives, resulting in a substantial improvement in incorporated into dark activity detection algorithms to enhance
the overall accuracy ratio. the overall accuracy of the obtained results.
TABLE
III
CONFUSION
MATRIX
FOR
R- ACKNOWLEDGMENT
DAD
ALGORITHM
This work is supported by Marmara University, Faculty of
Actual Positive 504 74 Engineering.
Condition Negative 5 151
Positive Negative R EFERENCES
Predicted condition
[1] D. Yang, L. Wu, S. Wang, H. Jia, and K. Li, “How big data enriches
maritime research – a critical review of automatic identification system
TABLE
IV
CONFUSION
MATRIX
FOR
ML-DAD
ALGORITHM (ais) data applications,” Transport Reviews, vol. 39, pp. 1–19, 07 2019.
[2] F. Mazzarella, M. Vespe, A. Alessandrini, D. Tarchi, G. Aulicino, and
Actual Positive 574 4 A. Vollero, “A novel anomaly detection approach to identify intentional
Condition Negative 20 136 ais on-off switching,” Expert Systems with Applications, vol. 78, 02
Positive Negative 2017.
Predicted condition [3] J. B. at Global Trade Review. (2020) Dark activity: As sanctions pressure
rises, maritime trade turns to tech. Retrieved 12 September 2023.
[Online]. Available: https://www.gtreview.com/magazine/volume-18-
V. C ONCLUSION issue-3/dark-activity-sanctions-pressure-rises-maritime-trade-turns-tech/
[4] H. Y. Shahir, U. Glasser, A. Y. Shahir, and H. Wehn, “Maritime situ-
This paper has addressed the crucial issue of detecting dark ation analysis framework: Vessel interaction classification and anomaly
activities within AIS-based maritime networks, with a primary detection,” in 2015 IEEE International Conference on Big Data (Big
focus on enhancing security and safety at sea, particularly Data). Ieee, 2015, pp. 1279–1289.
[5] T. Mantecón, D. Casals, J. J. Navarro-Corcuera, C. R. del Blanco, and
in identifying illicit engagements. Our examination primarily F. Jaureguizar, “Deep learning to enhance maritime situation awareness,”
centers on a localized detection scenario, where vessels that in 2019 20th International Radar Symposium (IRS). IEEE, 2019, pp.
deactivate their AIS transmitters are identified by nearby 1–8.
[6] R. S. Eaton, S. German, and A. Balasuriya, “Maritime border security
vessels based on historical signals received from them. using sensors, processing, and platforms to detect dark vessels,” in 2018
We introduced two distinct approaches to address this chal- IEEE International Symposium on Technologies for Homeland Security
lenge. Firstly, we proposed a rule-based method grounded in (HST). IEEE, 2018, pp. 1–5.
[7] F. Paolo, T.-t. T. Lin, R. Gupta, B. Goodman, N. Patel, D. Kuster,
kinematic estimations. Subsequently, we presented a machine D. Kroodsma, and J. Dunnmon, “xview3-sar: Detecting dark fishing
learning approach leveraging a dataset created from a realistic activity using synthetic aperture radar imagery,” Advances in Neural
simulation environment, which was designed with a graphical Information Processing Systems, vol. 35, pp. 37 604–37 616, 2022.
[8] K. Bereta, I. Karantaidis, and D. Zissis, “Vessel traffic density maps
user interface using real data. Both approaches were evaluated based on vessel detection in satellite imagery,” in IGARSS 2022-2022
within a scenario wherein a random subset of vessels engaged IEEE International Geoscience and Remote Sensing Symposium. IEEE,
in dark activity. 2022, pp. 2845–2847.
[9] D. Kocak and P. Browning, “Real-time ais tracking from space expands
The rule-based dark activity detection algorithm demon- opportunities for global ocean observing and maritime domain aware-
strated strong performance in identifying dark activities but ex- ness,” 10 2015, pp. 1–7.
hibited limitations, particularly in instances where AIS signals [10] marinetraffic.com. Marine traffic. Retrieved 15 September 2023.
[Online]. Available: https://marinetraffic.com
were not received due to changes in vessel movement param- [11] M. Redoutey, E. Scotti, C. Jensen, C. Ray, and C. Claramunt, “Efficient
eters, leading to false alarms. On the other hand, the machine- vessel tracking with accuracy guarantees,” 12 2008, pp. 140–151.
learning-based dark activity detection approach yielded im- [12] M. T. Ltd. Calculate distance, bearing and more between
latitude/longitude points. Retrieved 12 September 2023. [Online].
proved prediction accuracy and a reduced occurrence of false Available: https://www.movable-type.co.uk/scripts/latlong.html

---------------------------------------------------------------------------- 40 ----------------------------------------------------------------------------