1 INTRODUCTION

Global seaborne trade is doing well, but is facing a

number of political (inward-looking policies and

protectionism), economical (merger and alliances of

large shipping lines), and environmental challenges

(carbon footprint and sulphur cap) making its long-

range outlook quite uncertain

. This calls the maritime

administrations and industries, among others, for

maximizing efficiency of the shipping routes.

Shipping routes are specific tracks that vessels follow

when traveling between ports and form a global

maritime exchange network. Their spatial

characteristics are not static though, as over time

https://unctad.org/en/PublicationsLibrary/rmt2018_

en.pdf

shipping responds to a number of external

parameters. The actual tracks may for instance result

from either security reasons (e.g. avoiding risk of

exposure to piracy), or economic reasons leading to

an optimization process for avoiding rough marine

weather (Arguedas et al., 2018). To what extent tracks

are already optimized in actual business is not

straightforward to assess, but at the same time it

would represent an extremely valuable piece of

information.

We contribute to this topic via an inter-comparison

exercise between reported vessel tracks as collected

through the Automatic Identification System (AIS)

and tracks planned through the ship route

optimization model VISIR. This paper is part of a

broader strategy for a comprehensive verification and

evaluation of the VISIR ship routing model. The

verification (i.e., checking if its equations are solved

Preliminary Inter-comparison of AIS Data and Optimal

Ship Tracks

G. Mannarini & L. Carelli

CMCC, Lecce, Italy

D. Zissis, G. Spiliopoulos & K. Chatzikokolakis

arine Traffic, London, United Kingdom

ABSTRACT: Optimal ship tracks computed via the VISIR model are compared to tracks recorded by the

Automatic Identification System (AIS). The evaluation regards 43 tracks in the Southern Atlantic Ocean, sailed

during 2016-2017 by different bulk carriers. In this exercise, VISIR is fed by wave analysis fields from the

Copernicus Marine Environment Monitoring Service (CMEMS). In order to reproduce vessel speed loss in

waves, a new methodology is developed, where kinematic information from AIS is fusioned with wave

information from CMEMS. Resulting VISIR tracks are analyzed along with AIS tracks in terms of their

topological features and duration. The tracks exhibit quite diverse topological shapes, including orthodromic,

loxodromic, and other paths with complex and dynamic diversions. The distribution of AIS to VISIR track

durations is analyzed in terms of several parameters, such as the AIS to VISIR track length and their Fréchet

distance. Model features of VISIR affecting the results are discussed and future developments suggested by the

results are outlined.

http://www.transnav.eu

the International Journal

on Marine Navigation

and Safety of Sea Transportation

Volume 13

Number 1

March 2019

DOI: 10.12716/1001.13.01.04

correctly) included comparison to both static

(Mannarini et al., 2016) and time-dependent

(Mannarini & Carelli, 2019) analytical solutions as

well as comparison to the outcomes of another path

planning model (Mannarini et al., 2018). The first step

in evaluation (i.e., assessing whether the model fairly

represents reality) is presented in this work.

The core of the paper is organized into a

description of the evaluation methodology (Section 2)

and a presentation of the results (Section 3), which

precede conclusive remarks and outline (Section 4).

2 METHODOLOGY

In this Section we describe the methodology

employed for this evaluation experiment. The

description comprises the extraction of AIS tracks

from reported raw data (Section 2.1) and the

computation of optimal tracks through VISIR (Section

2.2). The key development enabling the comparison of

AIS and VISIR tracks is the computation of the vessel

speed loss in waves out of AIS kinematic information,

VISIR vessel performance model, and sea state

analysis fields (Section 2.3).

Abbreviations and symbols employed throughout

this text are defined in Table 1.

Table 1. Some abbreviations and symbols employed in this

manuscript.

_______________________________________________

Symbol Meaning Units

_______________________________________________

SOG Speed Over Ground kts

STW Speed Through Water kts

COG Course Over Ground deg

HDG Heading deg

EOT Engine Order Telegraph %

UKC Under Keel Clearance m

HWHM Half Width Half Maximum [-]

wl Length at the waterline m

wl Beam at the waterline m

avg Average draught m

max (fitted) maximum engine rating kW

max (fitted) maximum speed in calm water kts

s Significant wave height m

α Wave direction relative to vessel heading deg

_______________________________________________

2.1 AIS tracks

Nowadays, a multitude of tracking devices and

systems produce massive amounts of maritime data

on a daily basis. The most commonly used of such

tracking systems is the Automatic Identification

System (AIS), a collaborative, self-reporting system

that allows vessels to broadcast their identification

information, characteristics and destination, along

with other information originating from on-board

devices and sensors, such as location, speed and

heading (M.1371, 2014). AIS messages are broadcast

periodically and they are received by other vessels

equipped with AIS transceivers, as well as by on-

ground stations and satellites.

A growing body of literature describes methods of

exploiting AIS data for safety and optimisation of

seafaring, traffic analysis, anomaly detection, route

extraction and prediction, collision detection, path

planning, weather routing and many more (Tu et al.,

2018). As the amount of available AIS data grows to

massive scales though, researchers are realising that

computational techniques must contend with

difficulties faced when acquiring, storing, and

processing the data. Applying traditional techniques

to data processing can lead to processing times of

several days, if applied to global data sets of

considerable size. Additionally, algorithms are

challenged by difficulties related to the datasets

themselves; including highly skewed, not uniform,

and uncertain data. For example, the update interval

for AIS is not constant, but dependent on a ship’s

behaviour; as such it is common for a vessel to

broadcast its data every three minutes if moving no

faster than 3 knots, while every two seconds if

travelling above 14 knots and changing course. In

such occasions, the collected positions are spatially

and temporally closer. In addition to this, depending

on AIS coastal coverage, there are geographical areas

where huge amounts of data are collected (e.g. busy

ports) while others with much less data. Furthermore,

human errors during data entry generate

discrepancies for example in naming of ports and

areas, generating further uncertainty and ambiguity.

In our previous work (Spiliopoulos et al., 2017), we

proposed an efficient big data approach for building a

global network of sea routes from AIS data. As a first

step of this process we reassign departure and arrival

information for each vessel trajectory as although this

data is existent in AIS it is error prone and often an

issue of confusion. In the present work we consider

vessel trajectories travelling between the ports of

Buenos Aires (Argentina) and Port Elizabeth (South

Africa) from July 2016 until the end of December

2017. These trajectories amounted to more than

160,000 AIS messages in total. Each of these messages

consists of the coordinates of the vessel transmitting

the message, the corresponding kinematic

characteristics of the vessel (such as SOG, COG, and

HDG), and a timestamp. For identification purposes

an anonymised vessel’s identifier and voyage data are

bound to each message.

Within this dataset, we identified 61 different

trajectories in total, each of them being performed by

a different vessel. Moreover 51 of these trajectories

correspond to dry bulk carriers and the remaining 10

trajectories to wet bulk carriers. For each voyage, the

vessel’s dimensions, engine power and maximum

historical speed information were included in our

analysis. The summarised characteristics of the

anonymised vessels in this dataset are shown in Table

Table 2. Vessel characteristics.  and  stand for average

and standard deviation of the vessel sample.

_______________________________________________

Dry bulk Wet bulk

_______________________________________________

Number of vessels 51 10

Length [m]  201 182

 19 2

Width [m]  32 31

 2 2

Engine Power [kW]  8600 9000

 1500 2200

Max Speed [knots]  12 12

 2 3

_______________________________________________

2.2 VISIR tracks

VISIR (an acronym for "discoVerIng Safe and effIcient

Routes") is the ship routing model resulting from the

prototype first published in Mannarini et al. (2013).

The model eventually evolved to compute least-time

tracks in presence of time-dependent fields from

wave models (Mannarini et al., 2016), and has been

recently extended to deal also with ocean currents in

Mannarini et al. (2019).

VISIR is based on a graph-search algorithm, with

graph edges accounting for vessel COG, and edge

weights depending on the sailing time between graph

nodes. Graph edges crossing the landmass are

pruned, enabling computation of tracks even in

coastal waters or in vicinity of islands. Furthermore,

vessel intact stability can be accounted for through

checks on parametric roll, pure loss of stability, or

surfriding/broaching-to (IMO, 2007). Either

intentional speed reduction (EOT<1) or course change

can be exploited by VISIR for fulfilling the stability

constraints.

2.2.1 Path planner setup

For this actual evaluation experiment, the VISIR

model configuration is described through the

parameters provided in Table 3. The graph resolution

parameters are chosen to compromise between spatial

and angular accuracy on the one hand and

computational effort on the other one. UKC is not

checked in this exercise, as the employed GEBCO

bathymetry

would not allow, for the actual vessel

draught, a UKC>0 at the Western end of the AIS

tracks (located at the estuary of Rio de la Plata, in

South America).

Safety constraints and intentional speed reduction

are also disabled as, according to Mannarini et al.

(2019), for this actual route, they do not to

significantly impact the results. Conversely, disabling

them allows reducing the computer RAM allocation

of the computations and, thus, increasing the

maximum number of time steps considered for VISIR

paths. In fact, the vessel type considered in this

experiment is (both dry and wet) bulk carrier, which

top speeds (Table 2) are generally lower than

container ships considered instead in Mannarini et al.

(2019). A lower speed implies a longer sailing time,

which is represented in VISIR through a larger

number of time steps and, thus, requires higher RAM

allocation.

Table 3. VISIR configuration for the computations of this

experiment.

_______________________________________________

Feature Value

_______________________________________________

Grid resolution 1/7 deg ( = 8.6 nmi in meridional

direction)

Angular resolution 8.1 deg

Ocean currents neglected

Safety constraints shoreline only, UKC>0 and intact

stability checks disabled

Intentional speed disabled

reduction

_______________________________________________

https://www.gebco.net/data_and_products/gridded_b

athymetry_data/)

For the same reason, ocean currents are neglected

in this first version of the evaluation experiment. The

role of this approximation with respect to the self-

consistency of the whole methodology is discussed in

Section 2.3.1.

Finally, CMEMS

three-hourly wave fields are

averaged into daily fields before being employed by

VISIR. Again, as discussed in (Mannarini et al., 2019),

the reason for this approximation is the reduction of

RAM allocation.

2.3 Vessel response function

As mentioned above, the critical modeling piece for

the evaluation of VISIR vs. AIS tracks is the vessel

response function. It defines the involuntary speed

loss in waves due to the added resistance R

(Bertram and Couser, 2014). Because of speed loss,

path diversions may allow sailing at an higher speed

than along the least-distance track. Thus, by taking a

diversion, destination may be reached earlier. Since

the objective of the track optimization is to minimize

such a sailing time, diversions that optimally

compromise between reduced speed loss and

increased track length are chosen by the algorithm.

The vessel response function is defined within

VISIR as the STW sustained at specific values of

significant wave height H

s (Mannarini et al., 2016). Its

functional form is obtained from a power balance at

the ship propeller and makes use of a parametrization

of R

aw based on a statistical reanalysis of numerical

simulations via the Gerritsma and Beukelman's

method (Alexandersson, 2009). The parametrization

depends on just three main geometrical parameters:

vessel length L

wl, beam Bwl, and draught Tavg. When

applied to the VISIR power balance, it results into a

Gaussian-shaped vessel response function, with a

peak value given by vessel top speed in calm water

max and a HWHM proportional to maximum engine

break power P

max.

The dependence on wave direction (relative to

vessel heading) is presently neglected within the

VISIR vessel model. The impact of this approximation

is discussed in the subsection below.

2.3.1 Developments for the current experiment

For the current evaluation experiment, the five

parameters (L

wl, Bwl, Tavg, Pmax, Vmax) needed for

computing the vessel response function are identified

in the following way:

1 Hull geometry parameters (L

wl, Bwl, Tavg)

correspond to the vessel details identified through

the IMO-number provided along with the AIS

record;

2 Propulsion and performance parameters (P

max,

max) are fitted to the data of speed loss in waves,

generated as in the following.

A speed loss diagram (scatter plots of SOG vs.

corresponding H

s experienced by the vessel) is first

produced.

http://marine.copernicus.eu/

s is not part of the AIS record. However, it can be

estimated from sea state model analysis. Analyses are

the best available reconstructions of an environmental

state, making use of both observations and

geophysical model outputs. We employ CMEMS

analysis fields which assimilate observations of

significant wave height from Jason 2 & 3, Saral and

Cryosat-2 altimeters into the MFWAM sea state

model

. For each AIS track leg, we extract the spatially

and temporally nearest CMEMS gridpoint value. This

value represents our best estimation of H

encountered by the vessel at that specific AIS spatial

location and time.

Resulting data of speed loss in waves are

displayed in Figure 1, where they are plotted vs.

either H

s or relative wave direction α. Despite the fact

that, even for a specific vessel, data are strongly

scattered, a general trend for speed loss with

increasing H

s is recognized. Also, maximum loss is

achieved for head waves. This is consistent with

calculations and towing tank data of wave added

resistance provided by Tsujimoto et al. (2013).

SOG

Figure 1. Speed loss in waves out of AIS kinematic data and

CMEMS sea state fields. Panel a) displays the dependence

on significant wave height, while b) the dependence on

wave-vessel relative direction (α = 0 deg means head seas, α

> 0 refer to waves from the starboard). Marker grey tones

refer to individual voyages/vessels.

Datapoint scattering in Figure 1 can be attributed

to:

http://cmems-

resources.cls.fr/documents/PUM/CMEMS-GLO-PUM-001-

027.pdf

1 neglect of ocean currents, which combines with

STW for producing SOG;

2 limited skill of the CMEMS analysis fields in the

reconstructing the sea state.

The difference between SOG and STW, being

related to ocean current magnitude and direction

(Mannarini & Carelli, 2019), can be as high as several

knots. However, identifying STW with SOG is

consistent with the fact that, for this experiment, we

are neglecting ocean currents also for the computation

of VISIR optimal tracks. A more accurate treatment is

planned for future developments of this evaluation

methodology.

009

002

051

016

SOG

Figure 2. SOG dependence on significant wave height. AIS

data are represented as gray dots, while the VISIR response

function for fitted P

max and Vmax parameters is shown as a

dashed line. Each panel refers to a different voyage/vessel,

identified by the 3-digit code in the top-right.

Both the response function resulting from AIS

kinematical data and CMEMS wave fields and the

fitted VISIR speed loss curves are displayed in Figure

2 for selected vessels. The data have been previously

filtered by pruning data in the vicinity of the track

endpoints (harbours). There, due to vessel

acceleration, SOG can strongly vary even in

correspondence of a constant H

s value.

VISIR speed loss curve generally fits well to the

data, but in the case when they include more than a

branch at larger H

s. If this is the case, the VISIR curve

is fitted in between the branches (Figure 2b).

The presence of multiple speed loss branches may

be due to vessel performance in head waves, which

are not accounted by the present ship model. In some

cases, even a speed increases with H

s is observed

(Figure 2d), which may be due either to EOT changes

or to ocean currents.

In Fujii et al. (2017) too, a speed loss curve out of

AIS and model wave data is displayed. The authors

considered tracks of container ships and pure car

carriers in the North Pacific. However, their data do

not support a clear relation between SOG and H

Finally, we note that the information in step b) of

the present procedure just represents the P

max and

max fit parameters and does not necessarily agree

with the actual vessel parameters for maximum

engine brake power and speed.

3 RESULTS

In this Section, results relative to systematic

application of the methodology of Section 2 to

transatlantic passages between Buenos Aires

(Argentina) and Port Elizabeth (South Africa) and

vice-versa during the years 2016 and 2017 are

presented and discussed. The results for a few

individual tracks (Section 3.1) precede the analysis of

the ensemble of the tracks from the topological

viewpoint (Section 3.2) and by investigation of the

dependence of track duration on several variables

(Section 3.3).

3.1 Individual tracks

A one-to-one comparison of VISIR to AIS track

topology is displayed for a selection of all the voyages

in Figure 3.

The tracks are portrayed on top of the CMEMS

wave fields. The latter are generally taken at the

timestamps of the AIS waypoints and displayed as

adjacent vertical stripes. VISIR instead employs daily

averages of the CMEMS waves. These are shown in

Figure 3.b-c in the portion of the map containing the

VISIR trajectory.

Finally, each panel displays also the geodetic track

computed by VISIR. This track is, in the open ocean,

an arc of great circle joining the track endpoints. In

that case, it is identical to the orthodromic path. The

main features of the various panels of Figure 3 are

described in the following:

1 For this ship voyage, both VISIR and AIS tracks

take an initial Northbound diversion and then sail

to destination along a trajectory close to the

loxodromic (i.e., constant bearing) path. The

diversion is slightly anticipated by VISIR, which in

the second part of the voyage computes a

diversion approaching the geodetic track.

2 A significant track disagreement is noticed, with

AIS close to orthodromic navigation and VISIR

taking a wide Northbound diversion. VISIR

diversion is instrumental in avoiding the rough

seas between 25-10

W and 5-15

E. This

corresponds to the second part of the voyage, at 6

and 12 days since departure respectively. AIS track

may result from the fact that the rough sea in the

second part of the voyage could not be exactly

forecast at the time of departure.

3 AIS track diverts North while VISIR track diverts

even South of the geodetic. The different wave

field stripes in the upper and lower part of the

map make clear that the diversion computed by

VISIR is instrumental in avoiding rough seas at the

latitude of the geodetic and above.

4 Both AIS and VISIR tracks sail a path between the

rhumb-line and the least-distance track. VISIR

track is closer to the latter in vicinity of both

departure and arrival locations, resulting in a

shorter track length than AIS data.

In both Figure 3.b-c an appreciable difference

between three-hourly and daily-averaged wave fields

can be noticed. The observed divergent AIS and VISIR

trajectories might be ascribed to this fact.

3.2 Tracks ensemble topology

The original 61 AIS tracks were reduced to 43 due to

the fact that some tracks presented an anomalous

length or that the response function, upon spatial

pruning, contained too few datapoints. Each of the 43

voyages in the AIS record is sailed by a different

vessel, for which a speed loss curve was computed as

described in Section 2.3.1.

It is employed for computing an optimal track via

VISIR between the actual AIS endpoints, for their

specific departure date and time.

In Figure 4 all Eastbound tracks from both AIS and

VISIR between July 31 2016 and December 13, 2017

are displayed. The tracks form a bundle with an

appreciable meridional extent. In the middle of the

passage, the bundle extent is about 12

for AIS and

about 20

for VISIR, which bundle extends even South

of the geodetic track. For both AIS and VISIR, there is

a general trend to larger diversions for tracks sailed

during (Southern hemisphere) winter months. A

significant meridional dispersion of the AIS tracks is

also noticed in the data published by Fujii et al.

(2017).

[m]

051

VISIR

AIS

002

VISIR

AIS

VISIR

AIS

009

VISIR

AIS

016

Figure 3. Tracks of voyages referenced also in Figure 2 displayed on top of CMEMS three-hourly wave fields. The white

arrows denote wave direction. In each panel, AIS, VISIR, and geodetic tracks are displayed as a green solid line, a red

dashed line, and a blue dotted line respectively. Panels b) and c) are split vertically, with the daily-averaged wave fields

displayed in correspondence of the VISIR tracks and the three-hourly fields in correspondence of the AIS tracks.

3.3 Tracks ensemble – key metrics

The track ensemble is also analyzed in terms of track

duration (sailing time), which is the optimization

objective for VISIR. Track duration could be the

guiding principle for shipmaster decisions as well,

since costs relative to bunker and onboard personnel

are proportional to duration.

First of all, durations of VISIR optimal tracks (T

are compared to durations on the geodetic track (T

G).

They are computed by VISIR accounting for speed

loss in waves. Figure 5a confirms that VISIR’s

optimization works as expected, saving up to about

two days with respect to orthodromic navigation.

V are also compared to AIS durations (TA) in

Figure 5b, finding T

V < TA always. Again, VISIR

savings exceed two days in some cases.

In order to get a deeper insight on the reasons for

the better performance of VISIR, the relative duration

saving -

va of VISIR to AIS trajectory (va = TV/TA -1)

is displayed in Figure 6. in dependence of four

different variables:

VISIR

AIS

009

016

051

002

009

016

051

002

Figure 4. Route departing from Buenos Aires (Argentina) arriving in Port Elizabeth (South Africa). a) displays AIS tracks,

while b) shows VISIR simulations departing on the same dates. Dashed bold tracks with numeric labels refer to the selected

voyages referenced also in Figure 2.

Figure 5. Analysis of VISIR optimal track durations TV. Panels a) and b) compare to the duration of VISIR geodetic TG and

AIS tracks T

A respectively. In every panel, the codes of the voyages are put in evidence, which speed losses in waves are

displayed in Figure 2.

Vessel maximum speed (fit parameter Vmax). There

possibly is some correlation between 

va and Vmax.

This is one of the fit parameters in Figure 2 and is

related but not identical to the actual vessel top

speed. This trend should be confirmed by a larger

track statistics;

The length ratio of AIS to VISIR tracks (L

A/LV). It

is close to unity with a high precision (within 5%) in

most cases. When departing from unity, a trend of -



va increasing with LA/LV seems to be supported by

the data;

1 The month of departure of the track. No clear

seasonal trend is recognized, but a larger -

variability in (Southern Hemisphere) summer

months. Before making hypothesis on its origin,

the statistics should be enlarged;

Figure 6. Analysis of relative VISIR to AIS track duration savings -va =1-TV/TA with respect to several parameters. Panel a)

displays the dependence on the V

max fit parameter; b) dependence on the AIS to VISIR track length ratio LA/LV; c)

dependence on month of departure; d) dependence on Fréchet distance F

d between VISIR and AIS tracks. In every panel, the

codes of the voyages are put in evidence, which speed losses in waves are displayed Figure 2.

2 The Fréchet distance

Fd, representing VISIR and

AIS track similarity. It seems that the maximum -



va increases with Fd, but the data are quite

scattered.

4 CONCLUSIONS

We have completed a preliminary evaluation of

VISIR ship routing model with respect to vessel

tracks reported through AIS data. The evaluation has

included the development of a new methodology for

determining speed loss in waves from fusion of

empirical data (AIS) and meteo-oceanographic model

output (wave analysis fields). A case study of 43

voyages in the Southern Atlantic Ocean has been

considered. The results indicate that:

1 VISIR track topology generally agree well with

AIS one, though in some cases completely

different tracks, with a large Fréchet distance from

AIS, are computed;

2 There is a significant annual variability of the

spatial structure of the tracks, both in the AIS

records and in the VISIR simulations. A number of

http://www.kr.tuwien.ac.at/staff/eiter/et-

archive/cdtr9464.pdf

AIS tracks are close to the loxodromic track and

the VISIR tracks at times extend even South of the

orthodromic track;

3 VISIR sailing times are always shorter (4-14%)

than AIS ones, while their total length remains

quite close to the AIS one in most cases;

4 VISIR time savings are larger for AIS tracks longer

than VISIR ones. Furthermore, the duration

savings could also be related to top vessel speed

and seems to vary most during (Southern

Hemisphere) summer months.

These conclusions are supportive of the fact that

some kind of optimization (either automated or

manual) took place in most of the actually sailed

tracks.

Discrepancies between VISIR and AIS both in

terms of duration and spatial structure can be

ascribed to at least two factors:

1 VISIR model features and approximations. For

this first exercise:

 three-hourly wave fields have been averaged

into daily fields, for the sake of reducing the

computer RAM allocation, cf. Mannarini &

Carelli (2019). As noted in Section 3.1, this may

be one of the reasons for the larger

discrepancies between AIS and VISIR tracks;

 ocean currents have not been considered,

neither for the optimal track computation, nor

for reconstructing STW out of SOG;

 a directional wave response is not yet modeled.

Its impact depends on how the a-directional

VISIR response function was fitted to the AIS-

CMEMS data in Figure 2. Whenever VISIR

curve is adjusted to the steepest branch of the

data, its speed loss is overestimated to always

be the one relative to head waves. On the other

hand, adjusting to the most constant branch

would underestimate the speed loss in head

wave. Overestimation of speed loss would

induce an overestimation of the spatial

diversion along the optimal tracks;

2 Unavailability, for the ship officer in charge of

track planning, of long enough wave forecasts.

Since forecasts are usually limited to a few days

lead time (e.g. 5 days for CMEMS

, 7.5 days for

NCEP

), a sub-optimal track may result from this

fundamental knowledge gap at the time of vessel

departure. This is especially relevant when rough

seas are encountered several days after departure

(cf Figure 3.b,c). This may induce a conservative

approach by the shipmaster, for instance avoiding

diversions towards circumpolar latitudes (cf

Figure 4). VISIR tracks instead, employing

reconstructions of the sea-state (analysis fields),

are not affected by this fundamental limitation.

We consider these preliminary results quite

encouraging. A larger track statistics and

consideration of routes in other parts of the global

ocean should enable an even deeper insight into

optimization choices by actual vessels. Also, this

approach provides further indication of what VISIR

model features need to be developed more urgently.

ACKNOWLEDGEMENTS

This work has received funding from the European Union

Horizon 2020 research and innovation programme under

grant agreements No. 633211 (AtlantOS) and No. 732310

(BigDataOcean) and the Italy-Croatia Interreg V-A

programme under project ID 10043587 (GUTTA).

REFERENCES

Alexandersson, M. A study of methods to predict added

resistance in waves. Master’s thesis, KTH Centre for

Naval Architecture, 2009.

Arguedas et al. 2018. Maritime Traffic Networks: From

historical positioning data to unsupervised maritime

traffic monitoring. IEEE-ITS, 19(3): pp.722732.

Bertram V. and Couser. P. 2017. Computational methods

for seakeeping and added resistance in waves. In B.

Volker (ed.), 13th International Conference on Computer

and IT Applications in the Maritime Industries, Redworth,

12-14 May 2014, pages 8–16. Harburg: Technische

Universität Hamburg.

GLOBAL_ANALYSIS_FORECAST_WAV_001_027

https://polar.ncep.noaa.gov/waves/product_table.s

html?-hs-multi_1-atlantic-latest-

Fujii M., Hashimoto H., Taniguchi Y. 2017: Analysis of

Satellite AIS Data to Derive Weather Judging Criteria

for Voyage Route Selection. TransNav, the International

Journal on Marine Navigation and Safety of Sea

Transportation, Vol. 11, No. 2, doi:10.12716/1001.11.02.09,

pp. 271-277

IMO. MSC.1/Circ.1228 Revised guidance to the Master for

avoiding dangerous situations in adverse weather and

sea conditions. Technical report, International Maritime

Organization, London, UK, 2007.

M.1371: Technical characteristics for an automatic

identification system using time-division multiple

access in the VHF maritime mobile band.” [Online].

Available: https://www.itu.int/rec/R-REC- M.1371/en

Mannarini G., Coppini G., Oddo P. and Pinardi N. 2013. A

prototype of ship routing decision support system for

an operational oceanographic service. TransNav, the

International Journal on Marine Navigation and Safety of

Sea Transportation, 7(1): 53–59.

Mannarini G., Pinardi N., Coppini G., Oddo P. and Iafrati

A. 2016. VISIR-I: small vessels – least-time nautical

routes using wave forecasts. Geoscientific Model

Development, 9(4): 1597–1625. https://www.geosci-model

dev.net/9/1597/2016/gmd-9-1597-2016.pdf

Mannarini G. & Carelli L. 2019. VISIR-I.b: waves and ocean

currents for energy efficient navigation. Geoscientific

Model Development Discussions, (under review)

Spiliopoulos, G., Chatzikokolakis, K., Zissis, D., Biliri, E.,

Papaspyros, D., Tsapelas, G. and Mouzakitis, S. 2017.

Knowledge extraction from maritime spatiotemporal

data: An evaluation of clustering algorithms on Big

Data. In Big Data (Big Data), 2017 IEEE International

Conference on (pp. 1682-1687). IEEE.

Tu E., Zhang G., Rachmawati L., Rajabally E., and Huang

G. B. 2018, Exploiting AIS Data for Intelligent Maritime

Navigation: A Comprehensive Survey From Data to

Methodology. IEEE Trans. Intell. Transp. Syst., vol. 19,

no. 5, pp. 1559–1582.

Tsujimoto, M., Kuroda, M. and Sogihara, N. 2013.

Development of a calculation method for fuel

consumption of ships in actual seas with performance

evaluation. In ASME 2013 32nd International

Conference on Ocean, Offshore and Arctic Engineering

(pp. V009T12A047-V009T12A047). American Society of

Mechanical Engineers