235

1 INTRODUCTION

Marine surveillance and situation awareness is

essential for monitoring and controlling piracy,

smuggling, fishing, irregular migration, trespassing,

spying, traffic safety, icebergs, sea ice, shipwrecks, the

environment (oil spill or pollution), etc. Dark ships are

non-cooperative ships with non-functioning

transponder systems such as the automatic

identification system (AIS). Their transmission may be

jammed, spoofed, sometimes experience erroneous

returns, or simply turned off deliberately or by

accident. Furthermore, AIS satellite coverage at high

latitudes is sparse, which means that other non-

cooperative surveillance systems, including satellite

or airborne systems, are required.

The Sentinel satellites under the Copernicus

program [1] provide excellent and freely available

imagery with pixel resolutions down to 10 m in

multispectral and Synthetic Aperture Radar bands.

The orbital periods are 6 days between the Sentinel-1

(S1) satellites A + B, and 5 days between the Sentinel-2

(S2) satellites A + B. Furthermore, the swaths from

different satellite orbits overlap at higher latitudes,

and the resulting frequent transits over the polar

regions make these satellites particularly useful for

Artic surveillance and for monitoring sea-ice

coverage, icebergs and ships in SAR [2–11] and

multispectral images (see [12-13] and refs. therein),

etc.

Ship and iceberg detection in SAR imagery has

recently been studied in detail for earlier satellites and

TerraSAR-X [2-8] and Sentinel-1 [9-10]. Their different

modes and resolutions lead to interesting differences

for the ship detection lengths [5,6,10,11], classification

in comparison to AIS and false alarm rates [2-11], etc.

A comparison to these results will be made in the

conclusion of this work.

In the following an analysis of the S1 SAR data

and the search for objects in a sea background with

masking of land and sea ice is presented. In addition,

an algorithm for suppressing sidelobes from strongly

Ship-Iceberg Detection & Classification in Sentinel-1

SAR Images

H. Heiselberg

Technical University of Denmark, Kongens Lyngby, Denmark

ABSTRACT: The European Space Agency Sentinel-1 satellites provide good resolution all weather SAR images.

We describe algorithms for detection and classification of ships, icebergs and other objects at sea. Sidelobes

from strongly reflecting objects as large ships are suppressed for better determination of ship parameters. The

resulting improved ship lengths and breadths are larger than the ground truth values known from Automatic

Identification System (AIS) data due to the limited resolution in the processing of the SAR images as compared

to previous analyses of Sentinel-2 optical images. The limited resolution in SAR imagery degrades spatial

classification algorithms but it is found that the backscatter horizontal and vertical polarizations can be

exploited to distinguish icebergs in the Arctic from large ships but not small boats or wakes.

http://www.transnav.eu

the

International Journal

on Marine Navigation

and Safety of Sea Transportation

Volume 14

Number 1

March 2020

DOI:

10.12716/1001.14.01.30

236

reflecting objects as large ships is described, and a

change detection algorithm which identifies

stationary objects such as sea turbines, islands, piers,

etc. Subsequently, the segment classification

algorithm for ships and icebergs is described based on

backscatter polarizations. Results for ships and

icebergs in Denmark and Greenland are shown.

Finally, ship lengths and breadths in S1 data are

analysed and compared to AIS ground truth numbers.

These results are also compared to earlier TerraSAR-X

[4,5], S1 [9-10], Envisat [11], and S2 multispectral

results [12-13].

2 SATELLITE IMAGES AND METHOD OF

ANALYSIS

The S1 SAR images are analyzed using dedicated

software developed specifically to optimize the

classification of smaller ships and icebergs in large

images.

2.1 Sentinel-1 SAR Images

S1 carries the C-band Synthetic Aperture Radar all-

weather day-and-night imager [1]. As we are

interested in small object classification and

discrimination, we will focus on analyzing the

processed level-1 high resolution ground range

detected (GRDH) interferometric wide (IW) swath

images with 16 bit grey levels. The titles above Figs. 1-

5 are the filenames [1] that describe the data set (S1A-

IW-GRD-polarization-date and time).

We analyze 30 S1 images covering several parts of

Denmark and Greenland. These images are

convenient for classification because objects are

abundant and relatively easy to identify at sea. In

Denmark the objects are ships, wind turbines, islands,

and wakes whereas in Greenland there can be

abundant icebergs and floes, some islands but few

ships all depending on region, weather, time of year

and day. The S1 images analyzed here are recorded in

2018 from north of Denmark (Skagerak) down to the

southeast and the Baltic Sea, and along the west coast

of Greenland from Nuuk and up to the Disko Bay.

The S1 images contain polarimetric SAR

backscatter for horizontal (H) and vertical (V)

polarizations transmitted and reflected. The dual

direct polarized backscatter images are HH and HV in

Arctic regions with abundant sea ice, but VV and VH

in sub-Arctic regions because sea and wave reflect

stronger in VV than in HH. The cross polarizations

VH and HV are very similar.

Figure 1. The Disko Bay, Greenland (July 30th, 2018, 10:00

a.m. UTC). The title is the filename (see text). The box is an

iceberg ROI around the iceflow from the Ilulissat Icefjord,

one of the worlds fastest flowing glaciers.

Figure 2. Copernicus Sentinel-1A image [August 21

, 2018 at

05:32 a.m. UTC] covering Skagen, the northern tip of

Denmark. The added numbers refer to the list of objects

found by the classification algorithm, where white numbers

are ships and red numbers are mostly wakes, sidelobes, and

harbor piers.

237

2.2 Sentinel-1 SAR Images

S1 carries the C-band Synthetic Aperture Radar all-

weather day-and-night imager [1]. As we are

interested in small object classification and

discrimination, we will focus on analyzing the

processed level-1 high resolution ground range

detected (GRDH) interferometric wide (IW) swath

images with 16 bit grey levels. The titles above Figs. 1-

5 are the filenames [1] that describe the data set (S1A-

IW-GRD-polarization-date and time).

We analyze 30 S1 images covering several parts of

Denmark and Greenland. These images are

convenient for classification because objects are

abundant and relatively easy to identify at sea. In

Denmark the objects are ships, wind turbines, islands,

and wakes whereas in Greenland there can be

abundant icebergs and floes, some islands but few

ships all depending on region, weather, time of year

and day. The S1 images analyzed here are recorded in

2018 from north of Denmark (Skagerak) down to the

southeast and the Baltic Sea, and along the west coast

of Greenland from Nuuk and up to the Disko Bay.

The S1 images contain polarimetric SAR

backscatter for horizontal (H) and vertical (V)

polarizations transmitted and reflected. The dual

direct polarized backscatter images are HH and HV in

Arctic regions with abundant sea ice, but VV and VH

in sub-Arctic regions because sea and wave reflect

stronger in VV than in HH. The cross polarizations

VH and HV are very similar.

The spatial coordinates (x,y) are the pixel

coordinates (i,j) multiplied by the pixel resolution l =

10m for the S1 high resolution Ground Range

Detected (GRDH) images. The total vertical

backscatter is

(1)

and analogously for the horizontal backscatter

H=HH+HV. Examples are shown in Figs. 1-3.

2.3 Object detection from Background

To detect an object, its backscatter must deviate from

the sea background which varies with satellite

viewing (incidence) angle, wind and waves. Ship, ice,

sea turbines, oil rigs, islands, and other objects

generally reflect more. The next step is to select a

region-of-interest (ROI) and mask large areas of land

and sea ice by a simple algorithm which detects and

connects segments [13] above a given area. Choosing

the ROI such that the sea covers more than half of the

image after land removal, the median backscatter

value for each ROI image provides an accurate and

robust value for the background. The detection

threshold T

B is determined from the cumulative

distribution for the ROI image such that the constant

false alarm rate is 10

-4

. If the ROI has megapixel size,

this can lead to a large number of false alarms –

mostly single pixels. Therefore, we filter objects larger

than three pixels only, which effectively removes

most false alarms from noise. The false alarm rate is at

the same time sufficiently large that most if not all

large and medium size ships are detected.

The total backscatter image V(i,j) or H(i,j) has the

highest resolution and is therefore optimal for object

search and detection. Treating these as matrices (see

[13]), we construct a connectivity matrix in which

pixels with total backscatter above and below the

threshold T

B are assigned 1 and 0 respectively. In this

connectivity matrix, all neighboring entries with value

1 are then connected as a segment (s), and listed s =

1,..., N

s, where Ns is the total number of separate

segments found in the image. Each segment has an

observed area corresponding to the sum over the

pixels in the segment

(

)

As 1

ij s

∈

∑

, (2)

in units of the pixel area (l

). The cumulated

backscatter of that segment is found by summing over

its pixels

( ) ( )

Xs X ,

ij s

∈

∑

(3)

for the co-polarized X=VV or HH, cross polarization

X=VH or HV, or total backscatter X=V or H. We define

the average backscatters

( ) (

) ( )

s /V V s As=

( ) ( ) ( )

s /H Hs As=

, (4)

and the cross polarization ratios

( ) ( ) ( )

C s VH /s Vs=

( ) (

) ( )

C s HV /s Hs=

, (5)

which are very useful classifiers as will be shown

below.

Figure 3. Top: S1 SAR image of the container ship Ivar

Reefer in Fig.2. Ship dimensions are 164x26m but show

strong sidelobes. Bottom: sidelobes have been removed. Red

bars show the resulting and improved ship length, breadth

and direction.

238

2.4 Sidelobe removal

Sidelobes are often encountered in radar backscatter

images. In particular ships and oil rigs, with large

metal areas or corner reflectors can be very bright and

produce strong sidelobes both along (i-direction) and

transverse (j-direction) to the satellite swath direction.

In Fig. 3 we show a typical SAR image of a container

ship (Ivar Reefer) with strong sidelobes. These make

classification difficult and corrupt the automatic

determination of ship length, breadth and orientation.

Therefore we apply a phenomenological correction

algorithm which we have adjusted such that it

effectively removes the sidelobes. Whenever a pixel

(i,j) value is so bright that it exceeds a sidelobe

threshold T

S, the algorithm subtracts a value given by

a simple sidelobe correction function

(

) ( )

', ' V ,

1 | | / 1 | | /

S i j S ij

ii j j

λλ

′



= +



+− + −



, (6)

for pixels in the x-direction and y-direction separately

relative to the bright pixel (i,j). The magnitude

parameter S

0=0.1 is low such that the nearby pixels are

not strongly suppressed such that the ship pixels

remain, otherwise ships may be broken into separate

segments. The sidelobe range parameters

and

are chosen such that the long range

sidelobes are sufficiently suppressed. More precise

modeling of the radiation pattern is difficult due to

the finite pixel resolution, which is comparable to the

oscillation length and implicitly the range parameters

λ. They differ because the sidelobes in S1 images are

stronger in the y-direction along swath. The

Kronecker delta’s

and

'jj

insure suppression

of one row and one column respectively in the image

matrix which cross at pixel (i,j). When several pixels

(i,j) exceed the threshold, stripes are suppressed in the

image as seen in Fig. 3.

The oscillating nature of sidelobes often leads to

separate segments nearby a strongly reflecting object.

Such “collaterals” are automatically picked up in the

search and segment detection algorithm and can be

classified as sidelobe remnant segments.

3 CLASSIFICATION

For all segments their position, length, width, area,

total backscatter and cross polarization are calculated

and listed. The classification scheme will as explained

below identify the segment as an object such as a ship,

iceberg or ice floe, wake, sidelobe, or a stationary

object as an island, wind turbine, or oil rig. Each

segment is assigned a number referring to a list with

details on the calculated spatial and backscattering

parameters. The numbers are plotted at the segment

coordinates as shown in Fig. 2 with a color

classification code.

3.1 Icebergs

Fig. 1 shows a S1 image from the Disko Bay in

Greenland with thousands of icebergs and ice floes

from several glaciers. As icebergs can come in many

sizes and shapes, spatial parameters as area and

length are not good classifiers. Instead the

and

H classification parameters are useful for ice floe and

iceberg classification as can be seen in Fig. 5. Most

icebergs have low

and CH, and are situated below

the dashed line which therefore can be used for

separating into classes by the k-nearest neighbor

method.

3.2 Ships

Ships are identified from AIS ship coordinates and

correlated with their positions in the satellite images.

Fig. 2 shows a S1 image where a number of ships are

anchored in the tranquil sea east of Skagen, the

northern tip of Denmark. Our detection and

classification algorithm finds almost all ships

recorded by AIS.

Figure 4. Scatter plots of all segments (mainly icebergs)

found in the Disko Bay ROI of Fig.1. Backscatter

distribution C

H vs.

, where size is proportional to area A

and shading to B/L. This gives a visual view of the four

classification parameters. Vertical line is the threshold T

Almost one thousand icebergs and ice floes are correctly

classified below the dashed line. According to AIS there are

two ships present of which the large cruise ship Ocean

Diamond is clearly separated on the right whereas a smaller

trawler is not identified.

As described in [12-13], one can for each object

calculate the center of mass coordinates, length (L),

width or breadth (B), orientation angle as well as a

number of other spatial parameters for the segment.

In the multispectral S2 images, these parameters were

exploited for spatial classification of the segments as

objects as ships are elongated and generally have

small breadth to length ratio, B/L. As will be

discussed in Sec. IV, the S1 IW SAR images have

poorer resolution 20x22m. Also ship wakes are

fragmented and not nearly as visible as for the S2

multispectral ship wakes. We therefore find that B/L

is not as useful for a ship classification parameter for

S1 images.

239

Figure 5. As Fig. 4 but for all segments (mainly ships,

sidelobes and wakes) found in Fig. 2 of Skagen. Dashed line

is the k-nearest-neighbor (kNN) classification decision

boundary.

The remaining classification possibilities are the

two backscatter polarizations. We find that the

average object backscatter

and the cross

polarization C

V are good classifiers as shown in Fig. 5.

Ships reflect much stronger than wakes and ice floes,

i.e. their radar cross sections are much larger due to

metal and flat surfaces and possible corner reflectors,

in particular for VH. Therefore large ships show up

on the right in Fig. 5, whereas wakes, sidelobe

segments and unfortunately also some small ships

show up on the left.

In Fig. 2 ships are denoted white numbers,

whereas objects classified as wakes, sidelobes, harbor

quay and a few small ships are denoted other colors.

This allows for a quick identification of objects in the

images with reference to the identification list with

position, size, size after sidelobe correction, length,

breadth, orientation, backscatter, etc.

3.3 Islands, sea turbines, oil rigs

Stationary objects such as islands and wind turbines

are separated by change detection. By comparing to

earlier S1 image(s) of the same region, and checking

whether an object was present at the same place

within a few pixels, we can remove most stationary

objects as islands, sea turbines, oil rigs, harbor quays,

etc. We choose a 5 pixel radius corresponding to 50m,

which is large enough to be robust towards noise and

some change in backscatter with time, but small

enough that accidental position overlap between two

moving objects is improbable.

3.4 Classification Accuracy and Comparison to AIS

The above object classifications indicate that the

backscatter parameters

and C are much

more useful than spatial classification parameters A

and B/L when it comes to discriminating ships from

icebergs in S1 IW images. As shown in Fig. 4+5 large

ships and icebergs can to a large degree be separated

in both

vs CH and

vs. CV plots by the dashed

line.

Smaller ships and boats as well as sidelobes and

wakes, however, tend to be widely spread into the

iceberg classification region - leading to mis-

identification. These results for S1 SAR ship and

iceberg classification are compatible with earlier

analyses based on dual cross polarisations from

RadarSAT, TerraSAR and other satellite data [2].

Note that almost all the Arctic S1 IW images are

HH whereas non-Arctic are VV, which complicates

the classification because there are few ships in HH

and few icebergs in VV. We therefore use the same

dashed line for classification in Figs. 4+5 for lack of

data. A differentiated classification should exploit that

the sea background has more than double reflection

in VV than HH. This will, however, require more

ships with AIS records in the Arctic and records of

icebergs drifting south.

Santamaria et al. [9] have analyzed more than two

thousand S1 extra wide swath (EW) images with

50x50m resolution in the Arctic in which they have

detected 13,312 objects all believed to be ships. On

average 84% of these were correlated to AIS ships.

The detection probability was 80-100% for ship

lengths above 150m but dropped to 60-70% for ship

lengths around the S1 EW resolution of 90m, and

below 20% for ships shorter than 20m. This gave an

average detection of 52% of the AIS ships in the S1

EW images.

We have performed a similar analysis for about 30

S1 IW images which contain about 200 ships in total.

Although our statistics are limited, the about 5 times

better resolution in the IW than EW images clearly

improves the detection probability of large ships. We

find almost 100% correlation between S1 IW and AIS

for ships longer than 100m. For smaller ships,

however, the detection probability drops rapidly to

zero as for S1 EW.

For comparison, the high resolution multispectral

S2 images had a detection probability of almost 100%

even for small ships of length 20m [12-13]. Also, a

greater number of small ships were detected which

did not transmit AIS, probably because AIS reporting

is only required by law for ships above 25m.

4 SHIP LENGTHS AND WIDTHS

For each connected segment in the images we can

calculate its position, heading angle, length and

breadth as described in detail in the ship model

algorithm of [12]. The image processing techniques

are general and apply to both S1 and S2 images.

However, the different backgrounds, noise, speckle

and resolution affect the results. The resulting ship

lengths and breadths are plotted in Figs. 6+7 vs. their

ground truth values as given by AIS. We find that it is

important to correct for sidelobes as they corrupt the

ship images considerably and result in erroneous ship

orientations, exaggerated lengths and breadths. The

sidelobe corrected ship lengths and breadths from S1

data are closer to the ground truth numbers from AIS.

The S1 ship lengths and breadths do, however,

overestimate the ground truth dimensions by a

constant offset of 36m and 26m respectively as shown

in Fig. 6+7. The discrepancy is due to the limited S1

GRDH resolution of 20x22m and less importantly the

pixel resolution of 10m. Correcting for this off-set, the

240

standard deviation σ defined as the root mean square

average of the difference between the S1 satellite data

and the AIS ground truth (L

S1 +B/2− LGT) is σ = 24m for

the ship lengths and σ(B

S1 − BGT) = 9m for the ship

breadths. Note that one can observe the effect of the

ship sterns which extends the ship length by of order

half the ship width as explained in [12].

In Fig.6+7 the S2 optical data ship lengths and

breadths from [12] are also shown for comparison.

These agree much better with ground truth AIS

values because the resolution is only the pixel length

l=10m and there are four high resolution bands. Also

the background noise is less and no sidelobes are

present. The resulting standard deviations were

σ(L

S2+B/2 − LGT) = 10 m and σ(BS2 − BGT) = 4 m and a

very small offset of order 2m between S2 and AIS ship

lengths and breadths.

As described above the ship images are corrected

for sidelobes, and the resulting ship lengths and

breadths calculated depend on the parameters

(

, , )

λλ

. The stronger the correction, the smaller

the length and width, but at the same time the off-set

decreases. The resulting offset corrected ship lengths

and breadths are found to be robust, i.e. almost

independent of the sidelobe parameters within a

reasonable range around the parameters chosen.

Figure 6. Ship lengths. Squares and circles are from S1 and

S2 data respectively vs ground truth lengths and breadths

from AIS ship records. Blue lines indicate best off-set fit to

S1 data and dashed blue line to S2 data, see text.

Figure 7. As Fig. 6 but for ship breadths.

5 COMPARISON TO OTHER WORK

Brusch et al. [5] have performed a similar analysis of

ship lengths for TerraSAR satellite images, which

have resolution down to 3-4m. They find a good

linear correlation for ship lengths between 50-300m

with a standard deviation of 22m and an offset

(negative bias) of 7m. The offset is smaller due to the

better resolution as could be expected, and is

comparable to the resolution as is also the case for our

Sentinel-1 data. Surprisingly, the standard deviation

in TerraSAR ship lengths is much larger than the

resolution and comparable to that found in the S1

analysis above. Part of the reason could be the

sidelobe reduction performed on the S1 data. We also

suspect that the higher detail in TerraSAR resolves

high and low reflective parts of the ships, which

complicates the spatial analysis and dimension

determination. Therefore, the S1 images are useful for

ship detection and they also cover larger areas and

thus allow for faster wide area search.

Bentes et al. [4] have analyzed a wide range of

TerraSAR-X modes and found relations between

resolution and minimum ship length detection. At

wide incidence angle and low wind speed they are

similar, but at low incidence angle and/or higher

wind speed the large backscatter from the sea makes

detection more difficult and the minimum detection

length can increase by an order of magnitude

compared to the resolution. Extrapolating to the

resolution and inclination angles in our S2 data, we

find that our ship length off-set and standard

deviation are compatible with their ship size

detectability lengths.

Stasolla and Greidanus [10] have analysed 127

ships in S1 GRDH images from the Panama Canal.

Their method exploits the topology information of the

ship backscatter geometry to place a rectangle for ship

length and width estimates. Their methodology

avoids an off-set. Comparing to AIS data they find an

absolute error of 30m and 11m respectively. Both are

compatible with the standard deviations found in this

analysis.

Hajduch et al. [11] analyse a large number of ships

in Envisat ASAR/WSM/VV images. Besides

geometrical information they also exploit the ship

normalized radar cross section to improve their ship

length and width estimates. Their methodology seems

to remove off-set lengths. A comparison to AIS data

shown in a plot seems to have a variation as in Fig. 6

but no standard deviation is given.

6 SUMMARY

Sentinel-1 SAR data was analyzed in detail including

search and detection of objects above sea background

with masking of land and sea ice. An algorithm for

suppressing sidelobes from strongly reflecting objects

such as large ships was found very useful for

determining ship lengths and breadths more

accurately. As result good comparison to AIS ground

truth numbers was found, however, with a large

offset of around 30m due to the corresponding

resolution of the S1 data. The standard deviation

between ship lengths from AIS and satellite data was

241

of similar size 24m and surprisingly similar to earlier

TerraSAR-X [5] analyses although this data has much

better resolution. In comparison Sentinel-2

multispectral analysis could determine ship lengths

and breadths much better [12].

For classification a change detection algorithm

proved very useful for identifying stationary objects

such as sea turbines, islands, piers, etc. by comparing

object positions from different satellite overpass. This

simplified the classification of the remaining changing

objects such as ships and icebergs. However, it was

found that for the S1 SAR data as opposed to S2

multispectral data [13], the spatial information such

as area, length and width was not very useful for

classifying ships in Greenland when icebergs are

abundant because they can have very different form

and sizes and the resolution is limited. Instead, the

average and cross radar polarization backscatter were

significantly larger for large ships and allowed for

correct classification of large ships vs. icebergs using a

simple k-nearest-neighbor method. However, smaller

ships and wakes proved very difficult to separate

from icebergs.

Neural networks show promising results for

discriminating smaller ships from icebergs [14-15].

Correlating to AIS data will be important for better

determination of true/false positive/negative alarms,

and finding dark ships. The detection of ships vs.

icebergs in all weather day and night SAR data will be

useful in ice infested arctic seas for surveillance,

monitoring navigation, rescue service, etc.

REFERENCES

[1] ESA Copenernicus Program, Sentinel Scientific Data

Hub. https://sentinel.esa.int/web/sentinel/user-guides/

sentinel-1-sar/document-library/-/asset_publisher/

1dO7RF5fJMbd/content/sentinel-1-product-definition

[2] C-CORE. “Summary of previous research in iceberg

and ship detection and discrimination in SAR,” DRDC

report no: R-13-060-1098, 2013.

[3] G. Saur, M. Teutsch, “SAR signature analysis for

TerraSAR-X-based ship monitoring,” Image and Signal

Processing for Remote sensing XVI. Proc. of SPIE Vol.

7830, pp. 78301O-1, 2010.

[4] C. Bentes, D. Velotto, S. Lehner, “Analysis of ship size

detectability over different TerraSAR-X modes,”

IGARSS IEEE, pp. 5137, 2014.

[5] S. Brusch, S. Lehner, T. Fritz, M. Soccorsi, A. Soloviev,

B.v. Schie, “Ship surveillance with TerraSAR-X,” IEEE

Transactions on Geoscience and Remote Sensing, 2010.

[6] S. Brusch, S. Lehner, E. Schwarz, T. Fritz, “Near real

time ship detection experiments,” Proc. SeaSAR 2010,

Frascati. ESA SP-679, 2010.

[7] D. Velotto, M. Migliaccio, S. Lehner, “Dual-

Polarimetric TerraSAR-X SAR data for target at sea

observation,” IEEE Geoscience and Remote Sensing

Letters, Vol. 10, p. 1114, 2013.

[8] C. Brekke, D.J. Weydahl, Ø. Helleren, R. Olsen, “Ship

traffic monitoring using multipolarisation satellite

SAR images combined with AIS reports,” Proc. of the

7th European Conference on Synthetic Aperture Radar

(EUSAR), Friedrichshafen, Germany, 2–5 June, 2008.

[9] C. Santamaria, H. Greidanus, M. Fournier, T. Eriksen,

M. Vespe, M. Alvarez, V.F. Arguedas, C. Delaney, P.

Argentieri, “Sentinel-1 contribution to monitoring

maritime activity in the Arctic,” Proc. Living Planet

Symposium 2016, Prague, ESA SP-740, 2016.

[10] M. Stasolla, H. Greidanus, “The exploitation of

Sentinel-1 images for vessel size estimation,” Remote

Sens. Lett., vol. 7, no. 12, pp. 1219-1228, 2016

[11] G. Hajduch, N. Longepe, J. Hanonneau, J.Y. Le Bras,

“Progress in automatic ship detection and

classification,” Proc. SEASAR 2012, Tromso, Norway,

2012, ESA SP-709 2013, id. 19

[12] H. Heiselberg, “A direct and fast methodology for ship

recognition in Sentinel-2 multispectral imagery by

supervised classification,” Remote Sens., vol. 8, pp.

1033, 2016. https://doi.org/10.3390/rs8121033

[13] P. Heiselberg, H. Heiselberg, “Ship-Iceberg

discrimination in Sentinel-2 multispectral imagery,”

Remote Sens., vol. 9(11), pp. 1156, 2017.

https://doi.org/10.3390/rs9111156

[14] C. Bentes, A. Frost, D. Velotto, B. Tings, “Ship-

Iceberg discrimination with convolutional neural

networks in high resolution SAR images,” Proc. of

EUSAR, 2016.

[15] Iceberg Classifier Challenge 2018 in Machine Learning,

https://www.kaggle.com/c/statoil-iceberg-classifier-

challenge