Available via license: CC BY 4.0
Content may be subject to copyright.
remote sensing
Article
A Direct and Fast Methodology for Ship Recognition
in Sentinel-2 Multispectral Imagery
Henning Heiselberg 1, 2, *
1National Space Institute, Technical University of Denmark, 2800 Kongens Lyngby, Denmark
2Joint Research & Test Centre, Danish Defense Acquisition & Logistics Org., Lautrupbjerg 1-5,
2750 Ballerup, Denmark
Academic Editors: Clement Atzberger and Prasad S. Thenkabail
Received: 22 September 2016; Accepted: 14 December 2016; Published: 19 December 2016
Abstract:
The European Space Agency satellite Sentinel-2 provides multispectral images with pixel
sizes down to 10 m. This high resolution allows for ship detection and recognition by determining
a number of important ship parameters. We are able to show how a ship position, its heading,
length and breadth can be determined down to a subpixel resolution. If the ship is moving, its
velocity can also be determined from its Kelvin waves. The 13 spectrally different visual and infrared
images taken using multispectral imagery (MSI) are “fingerprints” that allow for the recognition
and identification of ships. Furthermore, the multispectral image profiles along the ship allow for
discrimination between the ship, its turbulent wakes, and the Kelvin waves, such that the ship’s
length and breadth can be determined more accurately even when sailing. The ship’s parameters are
determined by using satellite imagery taken from several ships, which are then compared to known
values from the automatic identification system. The agreement is on the order of the pixel resolution
or better.
Keywords:
Sentinel-2; multispectral; ship; recognition; identification; turbulent wake; Kelvin waves
1. Introduction
Marine situation awareness is of vital importance for monitoring, the control of piracy, smuggling,
fisheries, irregular migration, trespassing, spying, traffic safety, shipwrecks, and also for the
environment (oil or chemical dumping), etc. In cases of cooperative transponder systems, e.g.,
the automatic identification system (AIS), vehicle monitoring system (VMS) or long-range identification
and tracking system (LRIT), they may not be transmitting deliberately or accidentally, they are able to
be jammed, spoofed, and sometimes experience erroneous returns or are simply turned off. AIS satellite
coverage at high latitudes is sparse which means that other means of non-cooperative surveillance
systems are required such as, for example, satellite or airborne systems.
The Sentinel satellites under the Copernicus program provide excellent and freely available
multispectral [
1
] and Synthetic Aperture Radar (SAR) imagery with resolutions down to 10 m.
The orbital period is 10 days for the Sentinel-2 (S2) satellite A, with the image strips overlapping at an
given point on the Earth, and the typical revisit period is five days in Europe, being more frequent in
the Arctic. When the S2-B is launched, the revisit time for most of Europe will be two or three days.
Unless it is obscured by clouds, the multispectral and SAR satellite imagery will provide frequent
coverage of the Earth that increases with latitude, and this condition is particularly useful for arctic
surveillance [
2
,
3
], ship detection [
4
–
9
], oil spills [
10
], crops and trees [
11
], as well as hyperspectral
imaging [12].
This article focuses on S2 MSI ship detection, ship recognition, and ship identification. Detection
is relatively easy due to the high sensitivity and the generally dark sea background. Recognition is
based on high-resolution images, which allow for an accurate and robust determination of a ship’s
Remote Sens. 2016,8, 1033; doi:10.3390/rs8121033 www.mdpi.com/journal/remotesensing
Remote Sens. 2016,8, 1033 2 of 11
position, and its length, breadth and form, heading and velocity. Finally, identification is possible from
the 13 MSI spectral bands. The accuracy and confidence of all of this is reliant on the ship’s spectral
reflectances and the ship’s size. The analysis and discussion of the accuracy of the recognition and
identification from a ship’s S2 MSI data return are the results of the S2 MSI detection, recognition and
identification process.
The paper is organized such that the S2 data is described, followed by how it is then analyzed
within the ship model. The complication of wakes and how they can be included in the modeling is
also discussed. The ship model and analysis method used is described in the Appendix A. Results for
ship parameters are given for a number of ships. Ship lengths and breadths have been compared to
reported values by using AIS ground truth numbers and, finally, a discussion, summary and outlook
are also offered.
2. Satellite Data and Ship Modeling
2.1. Sentinel-2 Data and Analysis
S2 carries the wide-swath, high-resolution, multispectral imager (MSI) with 13 spectral bands
with 10, 20 and 60 m resolution [
3
]. We will mainly analyze the four bands with the 10 m resolution,
namely B2 (blue), B3 (green), B4 (red) and B8 (near-infrared). The S2 images as shown here are almost
cloud-free Level 1C recorded on 23 August 2016 (shown in Figure 1) and 5 September 2016; both
were taken over Skagen, Denmark. Ships from other satellite MSI images have also been analyzed
which confirm and support the conclusions of this article but they have not been included for reasons
of brevity.
Remote Sens. 2016, 8, 1033 2 of 11
Detection is relatively easy due to the high sensitivity and the generally dark sea background.
Recognition is based on high-resolution images, which allow for an accurate and robust
determination of a ship’s position, and its length, breadth and form, heading and velocity. Finally,
identification is possible from the 13 MSI spectral bands. The accuracy and confidence of all of this is
reliant on the ship’s spectral reflectances and the ship’s size. The analysis and discussion of the
accuracy of the recognition and identification from a ship’s S2 MSI data return are the results of the
S2 MSI detection, recognition and identification process.
The paper is organized such that the S2 data is described, followed by how it is then analyzed
within the ship model. The complication of wakes and how they can be included in the modeling is
also discussed. The ship model and analysis method used is described in the Appendix. Results for
ship parameters are given for a number of ships. Ship lengths and breadths have been compared to
reported values by using AIS ground truth numbers and, finally, a discussion, summary and outlook
are also offered.
2. Satellite Data and Ship Modeling
2.1. Sentinel-2 Data and Analysis
S2 carries the wide-swath, high-resolution, multispectral imager (MSI) with 13 spectral bands
with 10, 20 and 60 m resolution [3]. We will mainly analyze the four bands with the 10 m resolution,
namely B2 (blue), B3 (green), B4 (red) and B8 (near-infrared). The S2 images as shown here are almost
cloud-free Level 1C recorded on 23 August 2016 (shown in Figure 1) and 5 September 2016; both were
taken over Skagen, Denmark. Ships from other satellite MSI images have also been analyzed which
confirm and support the conclusions of this article but they have not been included for reasons
of brevity.
Figure 1. Sentinel-2A image tile VNK from 23 August 2016 at 10:30 a.m. UTC showing Skagen—the
northernmost tip of Denmark. The image is RGB contrast-enhanced. A number of container and
tanker ships (see Table 1) are moored just east of Skagen in the tranquil sea of Kattegat, a number of
which are waiting for bulk fuel. Close to the coast a few clouds can be seen, with their shadows being
cast northward.
Figure 1.
Sentinel-2A image tile VNK from 23 August 2016 at 10:30 a.m. UTC showing Skagen—the
northernmost tip of Denmark. The image is RGB contrast-enhanced. A number of container and
tanker ships (see Table 1) are moored just east of Skagen in the tranquil sea of Kattegat, a number of
which are waiting for bulk fuel. Close to the coast a few clouds can be seen, with their shadows being
cast northward.
Remote Sens. 2016,8, 1033 3 of 11
A number of methods do exist for tracking, searching and signature analysis in visual and infrared
imagery. We have developed dedicated Matlab software that enables us to search for, detect, and track
objects such as ships at sea or aircraft in the sky. We restrict the search to open sea without clouds
where objects such as ships are to be found. A smaller box is selected around each object as shown
in Figure 2, which then constitutes a single ship image segment. The background is determined and
subtracted in each band. Each ship image is then subsequently analyzed as described in Appendix A.
2.2. Ship Model and Parameters
The analysis is based on a basic ship model as shown in Figure 2. The ship’s signature is
established with an approximation made by the rectangular form of length Land breadth B, and
rotated by heading angle
θ
. In a more precise analysis we also make a determination of the bow and
Kelvin waves, and remove the turbulent wake effect.
Remote Sens. 2016, 8, 1033 3 of 11
A number of methods do exist for tracking, searching and signature analysis in visual and
infrared imagery. We have developed bespoke and dedicated Matlab software that enables us to
search for, detect, and track objects such as ships at sea or aircraft in the sky. We restrict the search to
open sea without clouds where objects such as ships are to be found. A smaller box is selected around
each object as shown in Figure 2, which then constitutes a single ship image segment. The background
is determined and subtracted in each band. Each ship image is then subsequently analyzed as
described in Appendix A.
2.2. Ship Model and Parameters
The analysis is based on a basic ship model as shown in Figure 2. The ship’s signature is
established with an approximation made by the rectangular form of length L and breadth B, and
rotated by heading angle θ. In a more precise analysis we also make a determination of the bow and
Kelvin waves, and remove the turbulent wake effect.
Figure 2.The RGB ship image box of NS Burgas (see Table 1, also present in Figure 1). Multiplying the
pixels (i,j) by l = 10 m gives the ship coordinates (x,y). The axes show the ship axis coordinate system
(x’,y’). Insert shows a horizontal visual image of the ship (with a substantially higher pixel resolution).
The ship images contains reflectances Ic(i,j) for each of the multispectral bands (c). The pixel
coordinates (i,j) are the (x,y) coordinates in units of the pixel resolution l, which is l = 10 m for the
blue, green, red and near-infrared bands c = B2, B3, B4, and B8, respectively, but l = 20 m for the bands
c = B5, B6, B7, B8a, B11, B12, and l = 60 m for c = B1, B9, B10. The total ship reflectivity in spectral band
c is:
=
(,
)
, . (1)
Once these as a background have been subtracted, the sum can run over all of the pixels on the
ship image. However, due to wakes and clutter we apply a minimum threshold between the
Figure 2.
The RGB ship image box of NS Burgas (see Table 1, also present in Figure 1). Multiplying the
pixels (i,j) by l = 10 m gives the ship coordinates (x,y). The axes show the ship axis coordinate system
(x’,y’). Insert shows a horizontal visual image of the ship (with a substantially higher pixel resolution).
The ship images contains reflectances I
c
(i,j) for each of the multispectral bands (c). The pixel
coordinates (i,j) are the (x,y) coordinates in units of the pixel resolution l, which is l= 10 m for the blue,
green, red and near-infrared bands c= B2, B3, B4, and B8, respectively, but l= 20 m for the bands c= B5,
B6, B7, B8a, B11, B12, and l= 60 m for c= B1, B9, B10. The total ship reflectivity in spectral band c is:
Ic=
>
∑
i,j
Ic(i,j)(1)
Once these as a background have been subtracted, the sum can run over all of the pixels on
the ship image. However, due to wakes and clutter we apply a minimum threshold between the
Remote Sens. 2016,8, 1033 4 of 11
background and ship reflectances, and only pixels with reflectances above this threshold are included.
This is indicated by the >index above the sum. For all ships, one threshold is chosen well above the
sea reflectance but well below typical ship reflectances. The ship parameters are found to decrease
only slowly with increasing threshold for a wide region of threshold values.
If we alternatively just sum up the number of pixels with reflectances above the threshold,
Ac=l2
>
∑
i,j
1 (2)
we ideally get the ship area A
c
= B
·
L, but generally it will depend on the threshold, spectral band,
image clutter and pixel resolution. The two methods used are based on the reflectance and pixel
counting and both are very useful, but we shall use (1) with a proper choice of threshold.
As described in Appendix A, the corresponding sums in Equations (1) and (2) of the reflectances
weighted with pixel numbers iand jgive the ship center-of-mass coordinates (
x
,
y
), when normalized
with the total reflectance. Likewise, weighting with the moments i
2
,j
2
, and i
·
jgives the ship’s
covariances, from which the ship length (L), breadth (B) and orientation (
θ
) can be determined (see the
formulas in Appendix A). From even higher moments (e.g., i
3
,j
3
,i
·
j
2
, etc.) one can also estimate
asymmetries such as the ship’s bow and other details of the ship’s form. However, such higher
moments are increasingly sensitive to clutter and wakes, and are threshold-dependent. As described
in Appendix A, the model is similar to a principal component analysis but calculates the ship
parameters directly.
Instead we find it useful, once the ship’s heading angle θhas been determined and the ship axis
is known, to transform to the ship axis coordinate system as shown in Figures 2and 3. As described in
Appendix A, one is then able to calculate the ship’s breadth along the ship’s axis, B(x’). A narrower
ship’s bow can be estimated from B(x’); consequently, from this its breadth and length can be
determined much more precisely.
2.3. Turbulent Wakes and Kelvin Waves
A ship moving relative to the sea creates a turbulent wake with the breadth increasing as a function
of the distance (r) behind the ship. In Reference [13] it is parameterized at long distances as
W(r)≈1.9 B4
5r1
5(3)
On the other hand, in Reference [
14
] the power scaling r
1/7
fits the turbulent wake of two warships
better. The turbulent wakes are often very long and this is very useful for detecting and tracking ships
with visual and infrared satellite images [
13
]. However, the turbulent wakes appear to be dependent
on the ship type, breadth, length and speed, and therefore are not a very good determinations to use
for the accuracy of a ship’s breadth.
The wakes also reflect solar light which is very difficult to separate from the ship’s reflectance [
15
].
The inclusion of turbulent wakes can also dramatically increase the ship’s length if included.
The multispectral differences between the wake and ship, in particular in the S2 near-infrared band
c = 8, can, however, can be exploited to distinguish between the ship and its wake, as shown below.
A sailing ship also creates Kelvin waves bounded by cusp-lines separated by an angle of
±
arcsin(1/3) =
±
19.47
◦
on each side of the referenced ship, as in Figure 3. The Kelvin wave length is
related to the ship speed Vas [16]
λ=2πV2
g(4)
where g= 9.98 m/s2is the gravitational acceleration at the surface of the Earth.
The cusps of the Kelvin waves lead to a larger apparent ship length and breadth. The amplitude
of the Kelvin waves increases sternwards and will, as shown in Figure 5b below, appear as oscillations
Remote Sens. 2016,8, 1033 5 of 11
in the breadth along the ship axis B(x’). Hereby the Kelvin waves can effectively be removed and the
wavelength determined.
Remote Sens. 2016, 8, 1033 5 of 11
Figure 3. Sketch of Kelvin waves and turbulent wake from a sailing ship.
3. Results
The S2 multispectral images are analyzed by using dedicated and bespoke developed Matlab
software for this purpose. The processing time is mainly for searching for ships, which can take
seconds depending on the complexity and size of the mega- to giga-pixel 12 bit images. For each of
the detected ships, as shown in Figure 1, a box is then drawn around the ship, and the reflectances,
center of mass system coordinates and covariances are calculated in the described ship model as
described in Appendix A. The calculation of the ship parameters is extremely simple and fast.
3.1. Multispectral Signatures
A typical medium-sized ship L = 100 m and B = 20 m appears in ca. 20 pixels for each of the
high-resolution bands c = B2, B3, B4, B8, as well as in fewer pixels for the other bands. The collected
multispectral ship signature data as shown in Figure 1 contains just over 100 data points, which can
be exploited as ship “fingerprints” for identification. As expected, the sea is dark and mainly reflects
in the blue band c = 2. Ships, wakes and clouds have high radiances for all spectral bands c = 2, 3, 4;
however, ships often reflect a color different from white as, e.g., seen in Figure 2. More interestingly,
ships and land tend to reflect or emit more infrared light than wakes. Therefore, the high-resolution
near-infrared band c = B8 is particularly useful for discriminating between the ship and wake, which
then allows for a more accurate determination of the ship’s length.
Clouds are seen in Figure 1 to cast shadows displaced ca. 570 m northward due to the solar
elevation. In Skagen, at a 57°43′ latitude on 23 August, two months or 360°·2/12 = 60° after summer
solstice, the solar elevation above the horizon is 57.7–23.44cos(60°) = 46°. The cloud altitude is
therefore 570 m/tan(46°) = 550 m. The cloud shadows are also slightly displaced westward as seen
from the sun-synchronous S2. Ship reflectances can also be affected by shadows, especially in the
winter months at higher latitudes. In fact, a darker reflectance can be seen in Figure 2 at the ship’s
starboard stern (northward), which is most likely to be a shadow cast by the ship bridge.
3.2. Ship’s Total Reflectance, Position, Heading, Length, and Breadth
The ship parameters for large- and medium-sized ships in the two Skagen datasets including
Figure 1 are shown in Table 1. The total ship reflectance is panchromatic, i.e., summed over RGB
bands c = B2 + B3 + B4 and ship pixels. The calibration factor for standard radiation units W/sr may
be calculated. Due to currents along the coast, most moored ships head southwest and south in the
two datasets. Ship ellipticities are in most cases ε ≥ 0.90, and high ellipticity is a good ship classifier.
Figure 3. Sketch of Kelvin waves and turbulent wake from a sailing ship.
3. Results
The S2 multispectral images are analyzed by using dedicated and bespoke developed Matlab
software for this purpose. The processing time is mainly for searching for ships, which can take
seconds depending on the complexity and size of the mega- to giga-pixel 12 bit images. For each of the
detected ships, as shown in Figure 1, a box is then drawn around the ship, and the reflectances, center
of mass system coordinates and covariances are calculated in the described ship model as described in
Appendix A. The calculation of the ship parameters is extremely simple and fast.
3.1. Multispectral Signatures
A typical medium-sized ship L= 100 m and B= 20 m appears in ca. 20 pixels for each of the
high-resolution bands c = B2, B3, B4, B8, as well as in fewer pixels for the other bands. The collected
multispectral ship signature data as shown in Figure 1contains just over 100 data points, which can
be exploited as ship “fingerprints” for identification. As expected, the sea is dark and mainly reflects
in the blue band c = 2. Ships, wakes and clouds have high radiances for all spectral bands c = 2, 3, 4;
however, ships often reflect a color different from white as, e.g., seen in Figure 2. More interestingly,
ships and land tend to reflect or emit more infrared light than wakes. Therefore, the high-resolution
near-infrared band c = B8 is particularly useful for discriminating between the ship and wake, which
then allows for a more accurate determination of the ship’s length.
Clouds are seen in Figure 1to cast shadows displaced ca. 570 m northward due to the solar
elevation. In Skagen, at a 57
◦
43
0
latitude on 23 August, two months or 360
◦·
2/12 = 60
◦
after summer
solstice, the solar elevation above the horizon is 57.7–23.44cos(60
◦
) = 46
◦
. The cloud altitude is therefore
570 m/tan(46
◦
) = 550 m. The cloud shadows are also slightly displaced westward as seen from the
sun-synchronous S2. Ship reflectances can also be affected by shadows, especially in the winter months
at higher latitudes. In fact, a darker reflectance can be seen in Figure 2at the ship’s starboard stern
(northward), which is most likely to be a shadow cast by the ship bridge.
3.2. Ship’s Total Reflectance, Position, Heading, Length, and Breadth
The ship parameters for large- and medium-sized ships in the two Skagen datasets including
Figure 1are shown in Table 1. The total ship reflectance is panchromatic, i.e., summed over RGB
Remote Sens. 2016,8, 1033 6 of 11
bands c = B2 + B3 + B4 and ship pixels. The calibration factor for standard radiation units W/sr may
be calculated. Due to currents along the coast, most moored ships head southwest and south in the
two datasets. Ship ellipticities are in most cases ε≥0.90, and high ellipticity is a good ship classifier.
Table 1.
Ship parameters for the ships in S2 Figure 1: reflectance (I), heading angle (
θ
), ellipticity (
ε
),
length (L) and breadth (B), see Appendix A. Numbers in parentheses are lengths and breadths found
from reported values in the AIS ship data.
Ship I2+3+4 θ ε L (m) B (m)
NS Burgas 26,370 33◦0.913 240 (275) 51 (48)
Eagle Barents 32,976 21◦0.912 252 (276) 55 (46)
GijonKnutsen 9939 17◦0.950 187 (183) 30 (27)
MarmaraMariner 2827 4◦0.968 130 (129) 17 (17)
Trade Navigator 2738 11◦0.963 125 (118) 17 (16)
Afines Sky 8013 19◦0.949 152 (162) 24 (23)
Skaw Provider 818 28◦0.966 110 (95) 15 (15)
Loireborg 2997 33◦0.955 113 (122) 17 (14)
Solstraum 2880 23◦0.899 89 (94) 21 (18)
Fjellstraum 2134 1◦0.950 94 (100) 15 (16)
BW Yangtze 14,530 33◦0.951 204 (229) 32 (32)
StenFjell 4343 −84◦0.921 131 (149) 27 (24)
SCL Basilia 4130 −97◦0.947 127 (140) 21 (22)
Karen Knutsen 20,809 −91◦0.914 256 (274) 54 (50)
Edith Kirk 6208 −100◦0.939 171 (183) 30 (27)
Grumant 5256 −104◦0.966 147 (181) 19 (23)
ChampionTrader 4223 −85◦0.975 185 (189) 21 (30)
Wilson Mersin 1412 −74◦0.942 84 (107) 15 (15)
Voorneborg 4590 −69◦0.944 116 (132) 20 (16)
Coral Monactis 1765 −69◦0.885 74 (95) 18 (15)
AtlanticaHav 358 −50◦0.845 53 (82) 15 (11)
Coral Obelia 741 −49◦0.930 83 (93) 16 (15)
Coral Pearl 3201 −60◦0.934 105 (115) 19 (19)
HHL Amur 8078 25◦0.920 128 (138) 26 (21)
HDW Herkules 1615 −88◦0.819 53 (54) 17 (10)
Elly Kynde 165 0◦1 17 (19) ≈5 (5)
Gottskar 434 0◦1 17 (21) ≈5 (6)
Frank Maiken 1081 18◦0.485 26 (18) 16 (6)
Haukur 15366 −7◦0.878 95 (75) 24 (13)
Ritz Dueodde 1525 28◦0.904 32 (15) 7 (5)
Torland 114,076 −1◦0.930 183 (140) 35 (22)
Sea Endurance 24350 11◦0.731 90 (110) 35 (18)
Bow Triumph 211,227 2◦0.852 162 (183) 46 (32)
1Sailing ship and includes ship wake. 2Includes adjacent fueling ships.
The ship lengths and breadths as referenced in Table 1are plotted in Figure 4, which are generally
in good agreement between those found in the ship model using S2 data and the ground truth numbers
from AIS. A few special cases disagree for obvious reasons. Three ships are sailing and the generated
wake extends the ship length. Two ships have fueling ships docked alongside them which extends
their breadth. By excluding these ships we are then able to calculate the standard deviation for the
lengths
σ
(L
sat −
L
GT
) = 16 m and breadths
σ
(B
sat −
B
GT
) = 3.8 m between the S2 satellite data and the
AIS ground truth. The S2 ship lengths are, on average, shorter than the ground truth lengths, which
we attribute to the ship’s bow and stern. Both extend the ship’s length. We are able to find a best fit
when the ship’s length is extended with half of the ship’s breadth. The resulting standard deviation is
reduced to
σ
(L
sat
+ B/2
−
L
GT
) = 10 m. Geometrically the extension depends on the bow and stern
angles with respect to the ship’s axis. Bow lengths around half the breadth are compatible with most
ship constructions. The corrected values and resulting 10 m standard deviation in ship length are
representative as they are the same for the two datasets separately.
Remote Sens. 2016,8, 1033 7 of 11
Remote Sens. 2016, 8, 1033 7 of 11
(a) (b)
Figure4. (a) Ship lengths and (b) breadths. The ground truth lengths and breadths are from AIS ship
records. The satellite lengths and breadths are the calculated values as used in the model given in
Table 1. Blue lines indicate agreement whereas dashed line includes bow corrections as described in
text.
Ships longer than ca. 30 m are seen in six pixels which is a minimum requirement in order to
determine the six ship parameters: total reflectance, position, heading, length and breadth (Ic, (̅,),
θ, B, L). The limited pixel resolution adds its own uncertainty, and the results become more
dependent on the threshold. For example, in Table 1 the two small sail boats are detected in two pixels
only, the ship’s breadth is simply an assumed half a pixel. Small ships can only be detected in S2
images. For large- and medium-sized ships longer than ca. 30 m, we find that the ship model is robust
and accurate for determining ship parameters.
3.3. Wake Removal and Ship Speed
In Figure 5a, a S2 panchromatic image of a sailing ship is shown. In Figure 5b, the corresponding
ship breadth B(x’) (see Appendix A) is shown along the ship axis coordinate x’. One observes the ship
breadth and wake width to be around 6̇·l = 60 m, with an increasing oscillation added on due to
Kelvin wave cusps. A Fourier transform yields the wave number, which in turn gives the Kelvin
wave length λ ≈ 5.5·l = 55 m. According to Equation (4), the ship speed is V = 9.3 m/s = 18 kts.
(a) (b)
Figure 5. (a) Ship with Kelvin waves and turbulent wake; (b) Corresponding ship breadth B(x’) along
ship axis x’ including wake oscillations.
Figure 4.
(
a
) Ship lengths and (
b
) breadths. The ground truth lengths and breadths are from AIS ship
records. The satellite lengths and breadths are the calculated values as used in the model given in
Table 1. Blue lines indicate agreement whereas dashed line includes bow corrections as described
in text.
Ships longer than ca. 30 m are seen in six pixels which is a minimum requirement in order to
determine the six ship parameters: total reflectance, position, heading, length and breadth (I
c
, (
x
,
y
),
θ
,
B, L). The limited pixel resolution adds its own uncertainty, and the results become more dependent
on the threshold. For example, in Table 1the two small sail boats are detected in two pixels only,
the ship’s breadth is simply an assumed half a pixel. Small ships can only be detected in S2 images.
For large- and medium-sized ships longer than ca. 30 m, we find that the ship model is robust and
accurate for determining ship parameters.
3.3. Wake Removal and Ship Speed
In Figure 5a, a S2 panchromatic image of a sailing ship is shown. In Figure 5b, the corresponding
ship breadth B(x’) (see Appendix A) is shown along the ship axis coordinate x’. One observes the ship
breadth and wake width to be around 6
·
l = 60 m, with an increasing oscillation added on due to Kelvin
wave cusps. A Fourier transform yields the wave number, which in turn gives the Kelvin wave length
λ≈5.5·l=55m. According to Equation (4), the ship speed is V = 9.3 m/s = 18 kts.
Remote Sens. 2016, 8, 1033 7 of 11
(a) (b)
Figure4. (a) Ship lengths and (b) breadths. The ground truth lengths and breadths are from AIS ship
records. The satellite lengths and breadths are the calculated values as used in the model given in
Table 1. Blue lines indicate agreement whereas dashed line includes bow corrections as described in
text.
Ships longer than ca. 30 m are seen in six pixels which is a minimum requirement in order to
determine the six ship parameters: total reflectance, position, heading, length and breadth (Ic, (̅,),
θ, B, L). The limited pixel resolution adds its own uncertainty, and the results become more
dependent on the threshold. For example, in Table 1 the two small sail boats are detected in two pixels
only, the ship’s breadth is simply an assumed half a pixel. Small ships can only be detected in S2
images. For large- and medium-sized ships longer than ca. 30 m, we find that the ship model is robust
and accurate for determining ship parameters.
3.3. Wake Removal and Ship Speed
In Figure 5a, a S2 panchromatic image of a sailing ship is shown. In Figure 5b, the corresponding
ship breadth B(x’) (see Appendix A) is shown along the ship axis coordinate x’. One observes the ship
breadth and wake width to be around 6̇·l = 60 m, with an increasing oscillation added on due to
Kelvin wave cusps. A Fourier transform yields the wave number, which in turn gives the Kelvin
wave length λ ≈ 5.5·l = 55 m. According to Equation (4), the ship speed is V = 9.3 m/s = 18 kts.
(a) (b)
Figure 5. (a) Ship with Kelvin waves and turbulent wake; (b) Corresponding ship breadth B(x’) along
ship axis x’ including wake oscillations.
Figure 5.
(
a
) Ship with Kelvin waves and turbulent wake; (
b
) Corresponding ship breadth B(x’) along
ship axis x’ including wake oscillations.
Remote Sens. 2016,8, 1033 8 of 11
The ship length cannot be accurately determined from the panchromatic plot in Figure 5a, because
the turbulent wake and Kelvin waves contribute to the ship breadth B(x) at a substantial length behind
the ship. We find that the wake is reduced in the high-resolution near-infrared B8 and we also find
that shorter ship lengths are better in agreement with ground truth lengths.
The ship’s bow can also be estimated from B(x). It increases from 0 to 6
·
l in around three pixels,
and therefore the bow length is
≈
3
·
l = 30 m, which is also half of the ship’s breadth. Therefore, half of
this bow length should be added to the ship’s length Lin order to obtain the ground truth length.
Ship maneuvers do not affect the ship’s breadth B(x’) because the ship heading is automatically
corrected for in the ship’s axis coordinate system. Strong side winds may cause asymmetries in wakes
and waves but do not change the Kelvin wavelengths.
4. Discussion
Ship detection, recognition and identification in S2 MSI has been analyzed. The images allow
for the detection of even small ships. Recognition and determination of ship parameters require that
the ship extend over several pixels, which for S2 data requires medium-sized ships above ca. 30 m.
The ship parameters of position, length, breadth, and heading can all be reliably determined as can the
ship’s speed when Kelvin waves are found. Length estimations can be improved by determining the
bow length as well as correcting for shadows. The ship lengths deviate from reported AIS numbers
with a standard deviation of 16 m without bow corrections, but 10 m for ship lengths with bow
corrections. The standard deviation is 3.8 m for ship breadths. These deviations are on the order of or
less than the S2 pixel resolution. The high resolution and the 13 multispectral bands’ information are
able to identify most medium- and large-sized ships.
The detection of sailing ships is simpler, due to the long turbulent wakes and Kelvin waves behind
the ship; however, their reflectances do complicate the determination of ship parameters. By exploiting
the multispectral information in the reflectances, in particular the smaller near-infrared reflectance of
wakes as compared to most ships, one can separate ships from wakes and determine the ship lengths
more accurately. For medium- and large-sized ships it is useful to determine the ship axis and plot the
breadth along this axis. Thereby one can separate the Kelvin waves and determine the ship breadth,
bow length, Kelvin wavelength and speed more accurately.
In Reference [
16
], the ship lengths for 53 vessels were determined from Geo-Eye-1 satellite images
with a higher resolution l = 2 m. Their segmentation method differs from the ship model employed
here specifically on wake removal. Their optimized method results in a standard deviation between the
satellite and ground truth ship lengths of 26 m. This may seem large considering the high resolution
but most of the ships are seafaring and wake removal seems difficult.
5. Conclusions and Outlook
S2 MSI is very useful for the detection, recognition of ship parameters larger than ca. 30 m, and
detailed identification of ships. If the ship reflectances in the 13 multispectral bands have been recorded
for all ship pixels and stored in a database, this will enable the reflectances to serve as “fingerprints”
for later identification. When Sentinel-2B is launched, the revisit time will be a few days in Europe and
almost daily in the Arctic. S2 MSI will greatly improve the marine situational awareness, especially for
non-cooperative ships—weather permitting. For example, in September 2016 Crystal Serenity was the
first cruise ship that risked sailing along the Northwest Passage which is infested with unchartered
reefs and titanic icebergs. The ship is non-cooperative since satellite AIS coverage is very limited at
these latitudes.
An obvious extension of these first results for ship classification from MSI is the discrimination of
ice floes in the Arctic and to develop methods for automatic ship searches in an Arctic environment.
In future work we will apply all 13 multispectral bands using pansharpening or hypersharpening
techniques [
17
] for the bands with lower resolution in order to improve the classification, discrimination
and multispectral identification of ships and threats from icebergs.
Remote Sens. 2016,8, 1033 9 of 11
Additionally, a comparison to Sentinel-1 SAR imagery will provide complementary and
weather-independent information but with a lower resolution (except in the rare and narrow swath SM
GRD full resolution mode). Also SAR is unable to detect low dielectric glass fiber boats. The synergy
of S1 and S2 imagery should be investigated for daily searching and tracking.
Acknowledgments: We are grateful for many useful comments from the referees.
Conflicts of Interest: The author declare no conflict of interest.
Appendix A. Determination of Ship Parameters
The total ship reflectance I
c
in band c is given by the sum in Equation (1). Inserting factors of i
and jin this sum, and normalizing it with the total reflectance, we obtain the ship cms coordinates:
x=l
Ic
>
∑
i,j
i·Ic(i,j)(A1)
y=l
Ic
>
∑
i,j
j·Ic(i,j)(A2)
Likewise, inserting factors i2,j2and i·j, the ship covariances are:
σ2
xx =l2
Ic
>
∑
i,j
i2·Ic(i,j)−x2(A3)
σ2
yy =l2
Ic
>
∑
i,j
j2·Ic(i,j)−y2(A4)
σ2
xy =l2
Ic
>
∑
i,j
i·j·Ic(i,j)−x·y(A5)
It is convenient to translate the coordinates into the ship cms coordinates (x’,y’) rotated by the
ship heading angle θsuch that the ship is aligned along the x´-axis (see Figure 2):
x−x=x0cos(θ)−y0sin(θ)(A6)
y−y=x0sin(θ)+y0cos(θ)(A7)
At the first approximation, we assume that the ship bow is short. The ship is then a simple
rectangle of length Land breadth B, and the ship extends from:
−
L/2 < x’ < L/2 and
−
B/2 < y’ < B/2.
The corresponding ship covariances in the (x’,y’) coordinate systems are:
σ2
x0x0=L2/
12,
σ2
y0y0=B2/
12
and
σ2
x0y0=
0, when the ship is large compared to the pixel resolution. The covariances
σ2
xx
,
σ2
yy
and
σ2
xy
are now found by replacing the pixel coordinates i = x/l and j = y/l in Equations (A3)–(A5) as given
in Equations (A1) and (A2). We obtain
σ2
xx =1
12 (L2cos (θ)2+B2sin (θ)2)(A8)
σ2
yy =1
12 (B2cos (θ)2+L2sin (θ)2)(A9)
σ2
xy =1
12 (L2−B2)sin(θ)cos(θ)(A10)
Remote Sens. 2016,8, 1033 10 of 11
It is now straightforward to determine the three ship parameters L,Band
θ
from the three
covariances σxx,σyy and σxy . Firstly, the ship heading angle is given by
tan(2θ)=2σ2
xy
σ2
xx −σ2
yy
(A11)
Next we introduce the ship ellipticity
e=L2−B2/L2+B2
. It can be expressed and
calculated in terms of the covariances as
e=L2−B2
L2+B2=
σ2
xx −σ2
yy
σ2
xx +σ2
yy
1
cos(2θ)(A12)
The ship ellipticity is usually between
e≈
0.89–0.96, corresponding to the length/breadth
ratio
L/B=√(1+e)/(1−e)≈4−7
, and can be used for classifying elongated ship-like objects.
It discriminates ships from icebergs, ice floes, wind turbines and other less-elongated objects.
The ship breadth and length squared can now be conveniently expressed as:
B2=6(1−e)(σ2
xx +σ2
yy)(A13)
L2=6(1+e)(σ2
xx +σ2
yy)(A14)
The model described here is similar to principal component analysis in 2D (2DPCA).
The symmetry axes are aligned with the principal eigenvectors, and Land Bare related to the principal
eigenvalues. The model described is simpler, faster and able to calculate the ship parameters directly.
The 2DPCA analysis is more elaborate and complicated and it can provide more eigenvalues which,
in our case, has not been necessary, given we only have ground truth values for ship lengths and
breadths from AIS.
In most of the S2 images, the background is dark sea, and consequently a threshold is able to be
set between sea and ship reflectances. Moving ships, however, create turbulent wakes and Kelvin
waves, which do have an effect on the calculated reflectance. These can effectively be removed by
calculating the ship breadth B(x’) along the ship axis x’. For simplicity, let us assume that we have
aligned the xand x’ axes (i.e., θ= 0). In that case, the variance in the y’ (or y) direction is
σ(i)2=l2
Ic
>
∑
j
j2·Ic(i,j)−y2(A15)
as function of the pixel i= x/l along the ship axis. The corresponding ship breadth is (cf. Figure 5b)
Bx0=√12 ·σ(i)(A16)
References
1.
ESA Sentinel-2 Delivers First Images. Available online: http://www.esa.int/Our_Activities/Observing_
the_Earth/Copernicus/Sentinel-2/Sentinel-2_delivers_first_images (accessed on 7 January 2016).
2.
Krogager, E.; Heiselberg, H.; Møller, J.G.; von Platen, S. Fusion of SAR and EO imagery for Arctic surveillance.
In Proceedings of the NATO IST-SET-128 Specialist Meeting, Norfolk, VA, USA, 4–5 May 2015.
3.
Brekke, C.; Weydahl, D.J.; Helleren, Ø.; Olsen, R. Ship traffic monitoring using multipolarisation satellite SAR
images combined with AIS reports. In Proceedings of the 7th European Conference on Synthetic Aperture
Radar (EUSAR), Friedrichshafen, Germany, 2–5 June 2008.
4.
Daniel, B.; Schaum, A.; Allman, E.; Leathers, R.; Downes, T. Automatic ship detection from commercial
multispectral satellite imagery. Proc. SPIE 8743 2013. [CrossRef]
Remote Sens. 2016,8, 1033 11 of 11
5.
Burgess, D.W. Automatic ship detection in satellite multispectral imagery. Photogramm. Eng. Remote Sens.
1993,59, 229–237.
6.
Zhu, C.; Zhou, H.; Wang, R.; Guo, J. A novel hierarchical method of ship detection from spaceborne optical
image based on shape and texture features. IEEE Trans. Geosci. Remote Sens.
2010
,48, 3446–3456. [CrossRef]
7.
Corbane, C.; Marre, F.; Petit, M. Using SPOT-5 HRG data in panchromatic mode for operational detection of
small ships in tropical area. Sensors 2008,8, 2959–2973. [CrossRef] [PubMed]
8.
Corbane, C.; Najman, L.; Pecoul, E.; Demagistri, L.; Petit, M. A complete processing chain for ship detection
using optical satellite imagery. Int. J. Remote Sens. 2010,31, 5837–5854. [CrossRef]
9.
Tang, J.; Deng, C.; Huang, G.-B.; Zhao, B. Compressed-domain ship detection on spaceborne optical image
using deep neural network and extreme learning machine. IEEE Trans. Geosci. Remote Sens.
2015
,53,
1174–1185. [CrossRef]
10. Gade, M.; Hühnerfuss, H.; Korenowski, G. Marine Surface Films; Springer: Heidelberg, Germany, 2006.
11.
Immitzer, M.; Vuolo, F.; Atzberger, C. First experience with Sentinel-2 data for crop and tree species
classifcations in Central Europe. Remote Sens. 2016,8, 166. [CrossRef]
12. Eismann, M.T. Hyperspectral Remote Sensing. SPIE 2012,PM210, 748.
13.
Lapierre, F.D.; Borghgraef, A.; Vandewal, M. Statistical real-time model for performance prediction of ship
detection from microsatellite electro-optical imagers. EURASIP J. Adv. Signal Process.
2009
,2010, 1–15.
[CrossRef]
14.
Golbraikh, E.; Eidelman, A.; Soloviev, A. On the helical behavior of turbulence in the ship wake. J. Hydrodyn.
Ser. B 2013,25, 83–90. [CrossRef]
15.
Bouma, H.; Dekker, R.J.; Schoemaker, R.M.; Mohamoud, A.A. Segmentation and Wake Removal of Seafaring
Vessels in Optical Satellite Images. Proc. SPIE 2013,8897. [CrossRef]
16. Thomson, W. On ship waves. Proc. Inst. Mech. Eng. 1887,38, 409–434. [CrossRef]
17.
Selva, M.; Aiazzi, B.; Butera, F.; Chiarantini, L.; Baronti, S. Hyper-sharpening: A first approach on SIM-GA
data. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2015,8, 3008–3024. [CrossRef]
©
2016 by the author; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC-BY) license (http://creativecommons.org/licenses/by/4.0/).
