Content uploaded by Sam Wouthuyzen
Author content
All content in this area was uploaded by Sam Wouthuyzen on Apr 12, 2021
Content may be subject to copyright.
Practical mapping methods of seagrass beds by
satellite remote sensing and ground truthing
journal or
publication title Coastal marine science
volume 43
number 1
page range 1-25
year 2020
URL http://doi.org/10.15083/00079480
Coastal Marine Science 43(1): 1–25, 2020
1
Practical mapping methods of seagrass
beds by satellite remote sensing and
ground truthing
Teruhisa K1, 2, Mazlan H3, Nurjannah N4, Thidarat N5,
Anchana P6, Milica S7, Tong Phuoc Hoang8, Pham Minh9,
Cao Van10, Sam 11, Sophany 12, Aidy M. 13,
Nurul Nadiah 14, Genki 15, Tatsuyuki 16 and Ken-ichi 17
1 Atmosphere and Ocean Research Institute, the University of Tokyo
5–1–5 Kashiwanoha, Kashiwa, Chiba 288–8564, Japan. E-mail: komatsu@aori.u-tokyo.ac.jp
2 Japan Fisheries Resource Conservation Association,
1–1 Akashicho, Chuoku, Tokyo 104–0044, Japan. E-mail: komatsu@aori.u-tokyo.ac.jp
3 Research Institute for Sustainable Environment, University of Technology Malaysia
81310 UTM Johor Bahru, Malaysia. E-mail: mazlanhashim@utm.my
4 Marine Science Department, Hasanuddin University
Jl.Perintis Kemerdekaan km.10, Makassar, 95245, Indonesia. E-mail: nurj_din@yahoo.com
5 Institute of Marine Science, Burapha University
Bangsaen, Chon Buri, 20131 Thailand. E-mail: sargassum2005@gmail.com
6 Faculty of Science, Prince of Songkla University, Hat Yai, Songkhla 90112, Thailand. E-mail: anchana.p@psu.ac.th
7 Faculty of Science, Prince of Songkla University, Hat Yai, Songkhla 90112, Thailand. E-mail: svesemenja@gmail.com
8 Institute of Oceanography, Vietnam Academy of Science and Technology
01 Cau Da, Nha Trang, Vietnam. E-mail: tongphuochoangson@gmail.com
9 Institute of Oceanography, Vietnam Academy of Science and Technology
01 Cau Da, Nha Trang, Vietnam. E-mail: phanminhthu@gmail.com
10 Institute of Marine Environment and Resources, Vietnam Academy of Science and Technology
246 Da Nang, Ngo Quyen, Hai Phong, Vietnam. E-mail: caoluongimer@gmail.com
11 Research Center for Oceanography, Indonesian Institute of Sciences (LIPI)
Jl. Pasir Putih I, Ancol Timur, Jakarta Utara, 11048, Indonesia. E-mail: swouthuyzen@yahoo.com
12 Faculty of Science, Royal University of Phnom Penh
Khan Toul Kork, Phnom Penh, 120000 Cambodia. E-mail: phauk.sophany@rupp.edu.kh
13 Institute of Oceanography (INOS), Universiti Malaysia Terengganu
21030 Kuala Terengganu, Terengganu, Malaysia. E-mail: aidy@umt.edu.my
14 Geoscience and Digital Earth Centre (INSTeG), 01–04–12, Ground Floor, Blok T06, Faculty of Built Environment and Surveying,
Universiti Teknologi Malaysia, 81310 Johor Bahru,
Johor Darul Ta’zim, Malaysia. E-mail: nurulnadiahy@utm.my
15 Northwest Pacic Action Plan (NOWPAP), Special Monitoring and Coastal Environmental Assessment Regional Activity Centre
(CEARAC), 5–5 Usijimashin-machi, Toyama 930–0856 Japan. E-mail: terauchi@npec.or.jp
16 Remote Sensing Technology Center of Japan
TOKYU REIT Toranomon Bldg., 3F, 3–17–1 Toranomon, MinatoKu, Tokyo, 105–0001, Japan. E-mail: sagawa_tatsuyuki@restec.or.jp
17 School of Marine Biosciences, Kitasato University, Kitasato
Minami-ku, Sagamihara, Kanagawa, 228–8555, Japan. E-mail: ken-ichi@kitasato-u.ac.jp
Received 16 March 2018; Accepted 1 November 2019
Abstract ―
-
-
-
-
Key words: -
Coastal Marine Science 43
2
1. Introduction
In coastal waters, seagrass and seaweed beds (Figs. 1
and 2) provide important ecological functions such as habi-
tats for animals (e.g. Coles et al. 1993, Dennison et al. 1993)
and plants, and stabilizing eects of environments such as
buering eect of currents and water movement, promotion
of sedimentation, absorption of nutrients, production of oxy-
gen, sediment retention, etc. (e.g. Ward et al. 1984, Jeudy de
Grissac and Boudouresque 1985, Komatsu 1989, 1996,
Komatsu and Yamano 2000). Therefore, they contribute to
marine biodiversity and also human society through ecologi-
cal services (e.g. Costanza et al. 1997). They have been de-
stroyed due to human impacts such as direct ones such as
bottom trawl and reclamation and indirect ones such as pol-
lution through aquaculture and urban and industrial wastewa-
ters (Komatsu 1997).
In the Seto Inland Sea, sh culture such as yellowtail
has been developed since 1960s. In Asia, aquaculture (e.g.
shrimp and marine sh farming) has also been developed
since 1970–80s (e.g. Gujja and Finger-Stich 1996, Huitric et
al. 2002). Aquacultures of sh such as yellowtail and sea
breams have been fed with a large quantity of baits which are
excreted as feces and urine into water. Since cultured sh do
not consume all fed bait, leftovers and feces become organic
load to the seabed. These organic matters are decomposed by
aerobic bacteria consuming oxygen in seawater. Nutrients
produced from organic matters with urine promote an in-
crease in phytoplankton and generate harmful algal blooms.
In such a phytoplankton rich environment, transparency is
decreased. Eventually, lower depth limits of seagrass beds
become shallower and seagrass beds are shrunk. On the sea
bed, sediments become anaerobic condition and muddy due
to accumulation of organic matters on the seabed and con-
sumption of oxygen by aerobic bacteria decomposing or-
ganic matters. While repeats of sh culture for long time,
feces, urine and leftovers of sh deteriorate water quality and
bottom sediment quality. It is called as auto-pollution. Con-
sequently, sandy bottom is changed to soft muddy bottom
due to increase in thickness of anaerobic layer. The soft
muddy bottom prevents seagrasses from rooting because
they are easily taken o from the bottom by the waves and
Fig. 1. Photo showing Zostera caulescens Miki in Funakoshi
Bay, Iwate Prefecture, Japan.
Fig. 2. Photos showing forests of brown algae, Stephanocystis hakodatensis (Yendo) Draisma, Ballesteros, Rousseau et Thibaut (left
photo), and Saccharina japonica (Areschoug) Lane, Mayes, Druehl et Saunders var. religiosa (Miyabe) Yotsukura, Kawashima, Kawai,
Abe et Druehl (right photo) o Kamoenai, west coast of Hokkaido, Japan.
Komatsu T. et al.: Seagrass mapping by remote sensing
3
currents. This is also true for waters where shrimp culture is
developed because discharge from shrimp ponds causes eu-
trophication and lead to destruction of seagrass and seaweed
beds due to water quality deterioration (e.g. Dierberg and Ki-
attisimkul 1996, Naylor et al. 2000) although seagrass and
seaweed harbor growth-inhibiting bacteria against dinoa-
gellates forming red tide (Onishi et al. 2014).
Seagrass and seaweed beds, however, have to be con-
served for marine biological diversity and sustainable devel-
opment of sheries and society. For maintaining sound sea-
grass and seaweed beds, it is necessary to map and monitor
their present spatial distributions (Komatsu et al. 2001,
2012), and to establish databases and information networks
to share and disseminate their data to manage these beds
through monitoring them (e.g. Kirkman 1996, Komatsu et al.
2002b). We can also analyze archived satellite images to
know past seagrass and seaweed distributions. Satellite re-
mote sensing studies on seagrass beds have been conducted
since mid-1980s with increase in spatial resolution of satel-
lite imagery such as LANDSAT TM with 30 m (e.g. Haddad
and Harris 1985) and SPOT 1 with 20 m (e.g. Fredj et al.
1990). Recently, spatial resolution of satellite imagery has
been improved to several meters suitable for reliable seagrass
mapping. In this article, we introduce a practical satellite re-
mote sensing method specialized for mapping seagrass beds,
which also applicable to seaweed beds, including validation
of satellite images by ground truthing.
2. Optics for Remote Sensing
Light is electromagnetic radiation, of which the true
units are W m−2 sr−1 μm−1, consisting of wide range of wave
lengths. Although visible and near-infrared bands are gener-
ally available in satellite images, only visible bands can pen-
etrate into the sea deeper than ultra-violet and near-infrared
which are easily absorbed by the surface thin layer (e.g. Lee
et al. 2013). Thus, visible bands are used for mapping habi-
tats in coastal waters under the sea. In the ocean, a blue band
reaches the deepest and a red band is attenuated in the shal-
lowest depths among blue, red and green bands (Fig. 3).
Remote sensing analysis with a satellite imagery utilizes
radiance of the visible bands reected by the sea beds re-
ceived by an optical onboard sensor. The radiance is dened
as the radiant ux emitted, reected, transmitted or received
by a given surface, per unit solid angle per unit projected
area. The passage of visible band radiation from the sun to
the satellite sensor is shown as a simplied schematic dia-
gram in Fig. 4. The visible band radiation from the sun
passes two layers: atmosphere and water. In both layers, a
part of light is absorbed and scattered by molecules and par-
ticulates from the sun to the sea bottom and from the sea bot-
tom to the satellite. A part of light is also reected by the sea
surface. Radiances of visible bands recorded as digital num-
bers of pixels by the satellite sensor include bottom reec-
tance depending on substrate types (Fig. 4). Reectance is a
ratio of leaving to incident radiation of a certain spectral win-
dow on surface of substrate, of which reectance depends on
the bottom surface substrate types such as sand, seagrass and
seaweeds. Thus, dierence in reectance of visible bands on
the bottom surface under the shallow sea can be used to clas-
sify substrate types. In general, an optical sensor mounted on
a satellite detects three bands of blue, green and red colors,
while spectral distribution of each band depends on the opti-
cal sensor. Blue band is important to detect bottom types in
deeper depth.
Figure 5 shows reectance of sand and one species of
Mediterranean seagrass, Posidonia oceanica L., from ultra-
violet to infrared measured with a spectrometer (FieldSpec
Pro, Analytical Spectral Devices Inc., USA) of which the in-
stantaneous eld of view was 25 degrees. We put samples
Fig. 3. Patterns of vertical penetration of light between ultravi-
olet and infrared into the water (courtesy of Prof. Tom Morris of
Fullerton College).
Fig. 4. Schematic view of passage of electromagnetic radia-
tion from the Sun to the satellite sensor through atmospheric
layer, sea water column and sea bottom surface.
Coastal Marine Science 43
4
into a basin with a diameter of 30 cm, of which inside and
outside were painted in black not to enter the light from sur-
rounding environment into the basin. We measured radiances
of the samples and a white disk around noon in a ne day
without clouds from a wave length of 350 nm to 2500 nm at
one nm intervals. Reectance of a sample was calculated by
dividing radiance of the sample with that of the white disk at
each wave length. We can nd the dierences in reectance
of visible bands between them. Since sand reects from short
wave length to long wave length, its color is white. On the
other hand, the seagrass reects green band around 550 nm.
Then its color is green. If dierences in reectance of sub-
strate types exist corresponding to spectral windows of satel-
lite optical sensor, it is possible to distinguish substrate types.
3. Satellite Images and Software
3.1 Spatial, spectral and radiometric resolutions of
satellite images
A multiband satellite image includes multiband data
which are generally composed of blue, green, red and near-
infrared layers and recently adds a coastal band and red edge.
Each layer consists of raster cells, pixels. One pixel corre-
sponds to an area with intensities of upwelling radiation of
the bands through the area on the Earth surface. For mapping
coastal habitats, it is better that the spatial resolution of satel-
lite images is ner than a habitat dimension to avoid a mixed
pixel called as a “mixel” where dierent objects are close to-
gether in one pixel. For example, if two objects of brown and
green exist in a pixel, its colour components of both objects
result in a brown-green mixed pixel, which is hard to ana-
lyze. To map a habitat, the most important thing is that plural
pixels more than ve cover a patch of a habitat for detecting
it. To overcome the mixel problem, it is necessary that a spa-
tial resolution of satellite image must be quite ner than an
area of a habitat. In generally, seagrass and seaweed beds are
distributed in a horizontal scale of several meters to hundred
or thousand meters. ALOS AVNIR-2, Geo Eye-1, IKONOS,
LANDSAT ETM+, LANDSAT 8 OLI, Pleiades-1/2, Quick
Bird-2, Sentinel-2, SPOT 6/7 and World View 2/3 have the
spatial resolutions described in Table 1. They can detect ma-
rine habitats on the shallow sea bottom greater than their spa-
tial resolutions. Images of GeoEye-1, IKONOS, Planet, Ple-
iades-1/2, Quick Bird-2 and World View 2/3/4 have very
high spatial resolutions of pixel less than 4 m (Table 1).
Radiometric resolution is an ability of sensor to discrim-
inate small dierences in the magnitude of radiation within a
ground area that corresponds to a single raster cell. When a
bit depth (number of data bits per pixel) of images that a sen-
sor records is great, its radiometric resolution is high. A sen-
sor of satellite records the intensity of electromagnetic radia-
tion from each spot viewed on the Earth’s surface as a digital
number (DN) for each spectral band of a pixel. The exact
range of DN that a sensor utilizes depends on its radiometric
resolution. The Geo Eye-1, IKONOS, QuickBird-2 and
World View 2/3/4 sensors, for example, have 11 bits (0–
2047) per band per pixel and LANDSAT 8 OLI, Planet, Ple-
iades-1/2 and Sentinel-2 have 12 bits (0–4094), while ALOS
AVNIR 2 and LANDSAT ETM+ have 8 bits (0–255) (Table
1). Thus, the rst two groups of satellites can distinguish
small dierences in the magnitude of radiation among sub-
strate types.
Spectral resolution is the ability of a sensor to detect
small dierences in wavelength. A panchromatic sensor sen-
sitive to a broad range of wavelengths has generally higher
spatial resolution than those of multibands (Table 1). An ob-
ject that reects a lot of energy in the blue area of the visible
band would be indistinguishable in a panchromatic photo
from an object that reected the same amount of energy in
other visible bands. A sensing system with a higher spectral
resolution would make it easier to distinguish the two objects
apart. ALOS AVNIR 2, Geo Eye-1, IKONOS, Planet, Pleia-
des 1/2, Quick Bird-2, Sentinel-2A/B and SPOT 6/7 have
multiband sensors measuring blue, green, red and infrared
bands while SPOT XS does green, red and infrared bands.
LANDSAT 8 OLI has coastal (new deep blue), blue, green,
red and infrared bands. WorldView-2/3/4 have coastal, blue,
green, yellow, red and infrared bands. Sensors with high
spectral resolution are a hyper spectral sensor. Compact air-
borne hyper spectral bands (CASI) have been often used for
mapping coral reef ecosystems. This system is eective with
a high spatial and spectral resolutions for coastal mapping.
However, their cost is expensive including cost of survey
with an airplane. A large quantity of data obtained by CASI
requires heavy processing of data.
Fig. 5. Reflectance level with reference to wavelength for each
bottom feature (Mahares). Values (bold lines) are shown as the
mean (±standard deviation represented by broken lines). For
each bottom feature, n=5 (source: Sagawa et al. 2010).
Komatsu T. et al.: Seagrass mapping by remote sensing
5
Table 1. Representative satellite multispectral sensors, their spatial resolutions, swath width, band spectral ranges of sensors and
panchromatic bands, average revisit days and dynamic ranges. Pan and Multi are panchromatic band and multiband, respectively.
Satellite Spatial resolution
(m)
Swath width
(km)
Multi and panchromatic
bands (nm)
Average revisit and
dynamic range
ALOS PRISM Pan 2.5 70 NIR 760–890 46 days
AVNIR-2 Multi 10 Blue 420–500
(from 2006 to 2011) Green 520–600 8 bit/pixel
Red 610–690
NIR 760–890
ALOS-3 PRISM-2
(launched in 2020)
Pan 0.8 70 Coastal 400–450 35 days
Multi 3.2 Blue 450–500
Green 520–600 PRSIM-2
Red 610–690 11 bit/pixel
Red edge 690–740
NIR 760–890
Pan 520–770
Dove (Planet) Multi 3.7 24 Blue 455–515 Everyday
(since 2014) Green 500–590
Red 590–670 12 bit/pixel
NIR 780–860
GeoEye-1 Pan 0.41 15.2 Blue 450–510 3 days
(since 2008) Multi 1.64 Green 520–580
Red 655–690 11 bit/pixel
NIR 780–920
Pan 450–900
IKONOS Pan 1 11.3 Blue 450–530 3 days
(from 1999 to 2015) Multi 4 Green 520–610
Red 640–720 11 bit/pixel
NIR 760–860
Pan 450–900
LANDSAT ETM+ Pan 15 180 Blue 450–520 16 days
(since 1999) Multi 30 Green 530–610 8bit/pixel
Red 630–690
NIR 780–900
Pan 520–900
LANDSAT 8 OLI Pan 15 180 New Deep
Blue
433–453 16 days
(since 2013) Multi 30 Blue 450–515
Green 525–600 12 bit/pixel
Red 630–680
NIR 845–885
Pan 500–860
Pan 0.5 20 Blue 450–530 4 days
Pleiades-1A Multi 2.8 Green 510–590
(since 2012) Red 620–700 12 bit/pixel
Pleiades-1B NIR 775–915
(since 2013) Pan 480–820
QuickBird-2 Pan 0.61 16.5 Blue 450–520 3.5 days
(from 2001 to 2015) Multi 2.4 Green 520–600
Red 630–690 11 bit/pixel
NIR 760–900
Pan 450–900
Coastal Marine Science 43
6
3.2 Availability of satellite images and software
Non-commercial satellite images of LANDSAT TM
have been archived since 1982. Downloading digital data of
LANDSAT 1–5 TM, LANDSAT 7 ETM+ and LANDSAT 8
OLI can be done via internet from Landsat Data Access of
United States Geological Survey (https://www.usgs.gov/
land-resources/nli/landsat/landsat-data-access?qt-science_
support_page_related_con=0#qt-science_support_page_re-
lated_con). Among the sites, the USGS Global Visualization
Viewer (GloVis) (https://glovis.usgs.gov/) or EarthExplorer
(https://earthexplorer.usgs.gov/) are easy to download
LANDSAT images free of charge. Due to problem of sensor,
the present LANDSAT 7 ETM+ hasn’t supplied any good
images for remote sensing since 2002. In 2006, non-commer-
cial satellite, Advance Land Observation Satellite (ALOS),
launched by the Japan Aerospace Exploration Agency
(JAXA) has a multispectral sensor, AVNIR-2, with 10 m spa-
tial resolution and a panchromatic sensor, PRISM, with 2.5 m
spatial resolution. Image data of Sentinel-2A/B with a spatial
resolution of 10 m for visible bands are also freely available
via internet (https://scihub.copernicus.eu/dhus/#/home).
These sensors that have spatially more precise than those of
LANDSAT images permit us to map coastal areas with vari-
ous ecosystems. While ALOS has been out of service since
April 2011 because of electric power depletion, archives of
ALOS AVNIR-2 from 2006 to 2011 are available. NASA
launched LANDSAT 8 that is a successor of LANDSAT 7 in
2013. LANDSAT 8 OLI (Operational Land Imager) has
higher radiometric and spectral resolutions and more visible
bands, which allow us to map coastal habitats by analyzing
their images. Commercial satellite images are also available.
GeoEye-1, IKONOS, Planet, Pleiades-1A/B, QuickBird-2,
SPOT and WorldView 2/3/4 have multiband images with
very high spatial resolutions (Table 1). JAXA schedules
launch of ALOS-3 in 2020.
There is a lot of commercial software for remote sensing
such as ENVI, ERDAS Imagine, TNTmips, etc. Although
their academic prices are reasonable, they are still expensive.
Free software for remote sensing and GIS is also available via
internet such as GRASS (http://grass.fbk.eu/index.php) and
Multispec (https://engineering.purdue.edu/˜biehl/MultiSpec/),
Sentinel-2 tool box (http://step.esa.int/main/toolboxes/sentinel-
2-toolbox/). QGIS (https://www.qgis.org/en/site/) is free GIS
software that can be used for analyzing satellite images with
a free open source plugin for QGIS. They are software to an-
alyze satellite images as eective as commercial software.
4. Ground-truthing of Sea Bottom
Ground truth is to obtain a coverage data of location in
situ corresponding to a pixel on satellite image in order to
Satellite Spatial resolution
(m)
Swath width
(km)
Multi and panchromatic
bands (nm)
Average revisit and
dynamic range
Sentinel-2A Multi 10 290 Blue 448–546 5 days (2A and 2B)
(since 2015) Green 538–583
Red 646–684 12 bit/pixel
NIR 763–908
Sentinel-2B Blue 443–541
(since 2017) Green 536–582
Red 646–685
NIR 767–900
Pan 1.5 60 Blue 455–525 26 days
Spot-6 Multi 8 Green 530–590
(since 2012) Red 625–695 12 bit/pixel
SPOT-7 NIR 760–890
(since 2014) Pan 455–745
WorldView-2 WV-2 Coastal 400–450 WV-2
(since 2009) Pan 0.46 16 Blue 450–510 3.7 days
WorldView-3/4 Mu1ti 1.85 Green 510–580 11 bit/pixel
Yellow 585–625
WorldView-3/4 WV-3/4 Red 630–690 WV-3
(since 2014/2016) Pan 0.34 13 Red Edge 705–745 4.5 days
Multi 1.38 NIR1 770–895 11 bit/pixel
NIR2 860–1040
Pan 450–800
Table 1. Continued.
Komatsu T. et al.: Seagrass mapping by remote sensing
7
verify contents of the pixel on the image on land. While, in
the sea, the words “sea truth” are used instead of ground
truth on land, it is better to use ground truth for verifying
bottom surface covers on the sea bottom. The classication
of the satellite image into groups of bottom surface cover
needs ground truth data as training data at supervised classi-
cation. After the classication of the image, ground truth
data were used for determining an accuracy of the classica-
tion.
Ground-truthing is conducted on site, performing bot-
tom observations and measurements of substrates covering
pixels on the remotely sensed digital image. The observa-
tions require highly accurate GPS to plot substrates on the
geographic coordinates. For geometric correction of image,
we also measure some typical locations that we can identify
on the satellite image with the highly accurate GPS. These
locations are called as ground control points (GCPs) used for
geometric correction. Software for remote sensing provides a
function for the geometric correction with position data of
GCPs as mentioned later.
Ground truth data are indispensable for a supervised
classication of an image or a decision tree classication.
When data of bottom cover types with location are available,
they can allocate attributes of pixels corresponding to their
locations on an image. The spectral characteristics of pixels
of the image corresponding to bottom covers on these sites
are used for decision rules for classifying the other pixels of
the image. In most cases, we divide ground truth data into
two groups: one for training and the other for classication
success. The latter data is to make an error matrix to evaluate
the accuracy of the classication.
Mapping methods of seagrass beds in situ are classied
into two categories. One is a direct observation through
walking, diving, grabbing, camera or video by researchers.
The other is an indirect method using a remote sensing appa-
ratus.
4.1 Determining positions of ground-truthing sites in
situ
It is very important to determine geographical positions
of ground-truthing sites as precise as possible because they
are used for classications and/or evaluation of classication
of pixels into bottom substrate types. The Global Positioning
System (GPS) became available in 1980s. It permits us to de-
termine a geographical position of a ground-truthing site in
real time. The accuracy of GPS has been ameliorated from
36 m (95% condence level) to 6 m (95% condence level)
since 2000 because of the removal of Selective Availability
(SA) from GPS (stopping the intentional degradation of the
GPS signals) on 2 May 2000. In the world, 34 countries have
already installed Dierential GPS (D-GPS) radio beacon net-
works in territory of each country, and more are considering
the adoption of this navigation standard. The improvement of
the basic GPS signal through elimination of SA may allow
the D-GPS radio beacons to transmit fewer error corrections
and more accurate localization. Accuracy of D-GPS is about
several decadal centimeters. However, D-GPS was abolished
in Japan on 1 March 2019 because the SBAS and QZSS
mentioned-below have been available (Japan Coast Guard
2017).
Another system to ameliorate precision of positions de-
termined with GPS is a satellite-based augmentation system
(SBAS), such as European Geostationary Navigation Over-
lay Service (EGNOS), MTSAT Satellite-based Augmentation
System of Japan (MSAS) and Wide Area Augmentation Sys-
tem (WAAS) of USA that are complement existing global
navigation satellite systems (GNSS) (Table 2). The SBAS
concept is based on GNSS measurements by accurately-lo-
cated reference stations deployed across an entire continent
(Fig. 6). The GNSS errors are then transferred to a comput-
Table 2. Satellite-based augmentation systems (SBASs) provided by a region or countries.
Country or region Name of satellite system
Europe European Geostationary Navigation Overlay Service (EGNOS)
USA Wide Area Augmentation System (WAAS)
Japan Multi-functional Satellite Augmentation System (MSAS)
India GPS and GEO Augmented Navigation (GAGAN)
Russia System for Dierential Corrections and Monitoring (SDCM)
Fig. 6. Flowchart for understanding the order of each compo-
nent for analyzing a satellite image and detecting seagrass distri-
bution. Numbers in parentheses indicate chapters or sections.
Coastal Marine Science 43
8
ing center, which calculate dierential corrections and integ-
rity messages which are then broadcasted over the continent
using geostationary satellites as an augmentation or overlay
of the original GNSS message. Several countries or a region
have implemented their own satellite-based augmentation
system as mentioned above. The SBAS can augment preci-
sion of positions within 1 m to 0.5 m.
On 11 September 2010, JAXA launched the Quasi-
Zenith Satellite System (QZSS) called “Michibiki”. It is a
proposed three-satellite regional time transfer system and
Satellite Based Augmentation System for the Global
Positioning System, which can be receivable in Asia and
Oceanica from 60ºE to 150ºW and from 60ºS to 60ºN
including Japan. Now four Michibiki launched by JAXA y
along an orbit of eight shapes between Japan and Australia.
Michibiki system runs from April 2018. It is estimated that
the errors to determine positions are within 1 m. Some recent
GPS can detect signals from Michibiki. Commercial
products with QZSS function including GPSs, drive
recorders, digital cameras etc. are listed and regularly
updated at the web site of Cabinet Oce of Japanese
Government (https://qzss.go.jp/en/usage/products/list.html).
According to this site, it is recommended to use the GPSs
which can receive Michibiki, SBAS, GLObal’naya
NAvigatsionnaya Sputnikovaya Sistema (GLONASS) in
Russia and BeiDou Navigation Satellite System in China
because more satellites give more precise positions.
4.2 Direct methods
Walking and diving
When bottom depths are less than 1 m or tidal ats
emerged from the sea surface, we can walk on the bottom to
observe bottom substrates with GPS. It is very easy to obtain
ground truth data. On the other hand, it is needed to dive to
observe the bottom when the sea level becomes high or bot-
tom depth is deeper than 1 m. Diving belongs to direct meth-
ods and is very sure for detecting bottom substrates. A merit
of diving is to identify bottom covers, especially species
compositions and densities of seagrass cover. However, it is
laborious and not ecient to take data at many points. It is
noted that the area observed by a diver at a point is within a
spatial scale of several meters. If the bottom depth is not
deep, a diver can use a GPS attached to a buoy attached to
the diver on the sea surface to determine its position. Calvo
et al. (1993) accurately located the upper limit of P. oceanica
meadow using a geodimeter when divers conducted ground
truthing. When the bottom depth is deeper than several me-
ters and another person measures positions of a diver from
the boat with GPS, the error of positions becomes greater.
This method cannot be applied to turbid water areas.
Manta tow
Manta tow is a simple method that a diver tracked by a
boat takes continuous pictures as shown in Fig. 8 (e.g. Chan-
cerelle et al. 2008, Miller et al. 2009). When a small boat is
available under warm water temperature, this method is very
useful to take ground truth data at places where water clarity
is high. A diver towed with a boat takes pictures with a digi-
tal camera from the sea surface at intervals of several sec-
onds. If the time of camera is synchronized with a GPS, we
can map bottom pictures on a geographical chart. The GPS
put in a plastic bag is xed on the head of diver not to sub-
merge it under the sea for receiving GPS signals.
Underwater video observation and towing
Underwater video observation is sometimes called as a
drop camera observation. It is easier than diving observation.
Researchers can observe bottom features with image on a
display monitored with a camera lowered from a boat and re-
corded images to a video recorder or a computer with infor-
mation through a microphone, time and date (Fig. 9). It is
better that the person who lowers the camera has GPS on his
pocket to identify positions of camera. Time of recorded im-
ages and GPS give position by synchronizing time of both
GPS and the recorder or computer.
Towing underwater video camera system has been de-
Fig. 7. Area covered by dierent SBAS systems in the world (source: http://sxbluegps.com/technology/sbas-made-easy/).
Komatsu T. et al.: Seagrass mapping by remote sensing
9
veloped (Norris et al. 1997). The camera is mounted in a
‘down-looking’ orientation on a towsh, which was deployed
directly o the stern of the vessel using the cargo boom. Re-
cently, we can buy cost-eective and high-performance un-
derwater camera such as Lumix (Panasonic Co.) and TG se-
ries (Olympus Co.) with function of interval photography
and underwater video cameras such as Hero (GoPro Inc.) and
Action Cam (Sony Co.). However, the method using under-
water camera or video are sensitive to turbidity.
Camera or video observation from a certain altitude above
the sea surface
Prof. Ken-ichi Hayashizaki, one of co-authors made
ground truth surveys in Philippines. He developed a system
using a digital camera or video protruding from the boat
abeam, supported by a bar mounted on a bamboo pole at a
height of 3 m from the boat deck (Hayashizaki and Ogawa
2011) (Fig. 10). Pictures were obtained by interval photo-
graphing. Positions of pictures were obtained with GPS by
synchronizing time of camera with that of GPS. Since GoPro
and Action Cam can be remotely operated with a controller
through Bluetooth, it is possible to verify whether images are
captured or not in real time in situ. When cameras and videos
include GPS with sub-meter precision inside, it is no need to
set a GPS on the top of pole or on the camera or video. Areas
of seagrass distributions can be estimated from the pictures
by using a length of the boat on the pictures.
Recent advances in unoccupied aircraft systems (UAS
or drones), coupled with their increased availability, present
a solution of ground truth survey for satellite remote sensing.
Since UAS can cover a study site with ultra-high resolution
(<5 cm) imagery allowing visual validation, Gray et al.
(2018) compared data of habitat distributions obtained by
UAS with those obtained by ground truthing in estuarine en-
Fig. 8. A schematic view of manta tow. Yellow circles, and pink and white rectangles are buoys, GPS in a container and underwater
camera or video, respectively. Buoys are attached to rope at 6 m intervals based on English et al. (1997).
Fig. 9. Photos showing video camera system for sea-truthing in Akkeshi Bay, east Hokkaido. Underwater video camera is lowered
from the boat (left photo). An observer monitors bottom features by the display in the cabin (right photo).
Fig. 10. Schematic diagram of a system using a digital cam-
era or video protruding from the boat abeam, supported by a bar
mounted on a bamboo pole at a height of 3 m from the boat
deck (left panel) and four continual pictures obtained by the sys-
tem (right panel).
Coastal Marine Science 43
10
vironments. They conclude that “Considering the similarity
in accuracy between UAS and eld-based assessments, our
conclusion is strongly in favor of validation with UAS where
feasible.” UAS increase sample validation points across a
larger proportion of the study site, while requiring less time
and less intrusion than ground truthing on the study area.
Drones with a high performance are available at a reasonable
price. When positions of pictures can be obtained at a sub-
meter precision, seagrass distributions and other bottom cov-
ers are used as ground truth data. Metashape (Agisoft Co.) of
which previous name was PhotoScan is a stand-alone soft-
ware product performs photogrammetric processing (mosaic)
of digital images and generates 3D spatial data to be used in
GIS applications (https://www.agisoft.com/). This software is
sold at a reasonable price.
4.3 Indirect methods
Echosounder
It is well known that acoustic backscattering informa-
tion can be used to infer seaoor physical and biological
properties (e.g. Hashimoto and Nishimura 1953a, Parnum
and Gavrilov 2011, Tecchiato et al. 2015). However, the
acoustic scatterings of seagrass and seaweed are poorly un-
derstood, compared with rock and sediment (De Falco et al.
2010). The backscattering strength of seagrass beds is usu-
ally higher than that of sandy or muddy bottoms, possibly
due to gas bubbles inside the foliage, sheath and shoot stem
structure (Komatsu and Tatsukawa 1998, Lyons and Abra-
ham 1999, Sabol et al. 2002, Riegl et al. 2005, Parnum 2007,
Wilson and Dunton 2009, De Falco et al. 2010, Tecchiato et
al. 2015).
The echosounders have advantages not only to continu-
ously measure biomass distributions and bottom topogra-
phies, but also to be used at a low cost and easy treatment
(Komatsu et al. 2002a). This method has been applied to sev-
eral studies in phanerogam beds in lakes (Duarte 1987), Zos-
tera marina L. beds (Hatakeyama and Maniwa 1978, Kom-
atsu and Tatsukawa 1998) (Fig. 11), P. oceanica meadows
(Colantoni et al. 1982, Rey and Diaz del Rio 1989) and
brown seaweed such as kelp (Hashimoto and Nishimura
1953b) and Sargassum species (Kitoh 1983).
Colantoni et al. (1982) tried to use a low frequency
echosounder (3.5 kHz); it proved to be rather ineective to
discriminate the acoustic character between P. oceanica bed
and the bottom. Although the high-resolution continuous
seismic reection (3.5 kHz) could distinguish the P. oceanica
and others (Rey and Diaz del Rio 1989), long wavelength of
ultrasonic brings worse vertical resolution. Echosounders
with an ultrasonic wave of 200 kHz is more appropriate for
detecting seagrass beds (Hatakeyama and Maniwa 1978,
Komatsu and Tatsukawa 1998).
The echosounder can scan seagrass beds when traveling
at about 1.0–1.5 m s−1 (2–3 knots). It is possible to investi-
gate 37 km per day when a ship with an echosounder travels
at 1 m s−1 (2 knots) for ten hours (Komatsu and Tatsukawa
1998). In this way, the echosounder is a very useful appara-
tus to map seagrass beds.
Hatakeyama and Maniwa (1978) used the echosounder
for mapping a Zostera bed, but they calculated only an index
of biomass: sum of canopy heights by unit sector along tran-
sects scanned by the echosounder. Since it is necessary to es-
timate seagrass or seaweed biomass for a quantitative com-
prehension of their ecosystems, Komatsu and Tatsukawa
(1998) proposed a simple converting method from the shad-
ing grades of seagrass on echograms to above-ground bio-
mass based on quadrat samplings (Fig. 12). From these echo-
grams, we can extract locations of seagrass and others for
satellite remote sensing.
The position of the lower bottom depth limit of seagrass
beds is related to the light extinction coecient inuencing
the minimum degree of light required for growth of seagrass
(Duarte 1991). Thus, it can be used for an indicator of water
quality. In France, the lower bottom depth limit of P. ocean-
ica was monitored by placing concrete markers (Meinesz and
Laurent 1978, Meinesz 1997, Boudouresque et al. 2000). In
this case, obtained results are very precise, but the observed
area is limited. The echosounder can be used to dene the
vertical distribution of seagrass bed and the lower bottom
depth limit of seagrass beds by correcting depths measured
by the echosounder to the sea level. Of course, it is necessary
to correct sea level depending on the time based on an obser-
vatory of the tide or tide table. Monitoring seagrass by the
echosounder is useful for detecting lower bottom depth limit
of seagrass beds not precisely but roughly in a wide area.
When these two types of monitoring are coupled, they com-
plement each other to obtain lower bottom depth limits.
Komatsu and Tatsukawa (1998) claried that the canopy
height was nearly proportional to the maximum blade length.
By cropping blades of seagrass, the height of seagrass cano-
pies on the echo-traces can be used as an indicator of the
maximum blade length of seagrass when the current speeds
Fig. 11. Raw records of echo trace along a transact in Ajino
Bay, Japan cited from Komatsu and Tatsukawa (1998). A depth
of 0 m is the sea surface, which is not standardized to the depth
relative to the mean sea level.
Komatsu T. et al.: Seagrass mapping by remote sensing
11
were not greatly dierent over the beds. Tanaka and Tanaka
(1985) also reported a similar proportional relation between
the canopy height and maximum frond lengths of Sargassum
species.
Sidescan Sonar
One acoustic method to map seagrass beds using a side-
scan sonar, which is more ecient than that of the ground
surveys, has been developed since 1970s in the Mediterra-
nean Sea. It scanned sea bottom at a width ranging 50–500 m
and could distinguish seagrass bed distributions and the oth-
ers successfully (Newton and Stefanon 1975, Meinesz et al.
1981, Lefèvre et al. 1984, Gloux 1984, Ramos and Ramos-
Espla 1989, Pasqualini et al. 1998). Figs. 13 and 14 show a
towing apparatus of sidescan sonar and the distribution map
of Z. caulescens in Koajiro Bay in Sagami Bay obtained by
the side-scan sonar, respectively. The patch structures are
clearly depicted. However, it is dicult for this method to
measure heights of plants except those just beneath the trans-
ducer along a transect.
Sagawa et al. (2008) proposed use of sidescan sonar
image as ground-truth data. They surveyed seagrass beds and
examined accuracy of results obtained from sidescan sonar
image with data obtained from drop camera observations.
They veried that maps of seagrass distributions surveyed
with a sidescan sonar corresponded to horizontal distribu-
tions of seagrass beds with drop camera observations and
proposed to use central areas (not border areas) with and
without seagrass beds as ground truthing locations because
central areas of habitats with horizontal spatial scale enough
broader than that of more than several satellite image pixels
consist of no mixed cells of habitats above-mentioned.
Multibeam sonar
Multibeam sonar (MBS) is one of the most eective
acoustic tools for mapping seagrass, because it can survey
with a broad swath to create a three-dimensional (3D) image
of the seagrass meadow (Komatsu et al. 2003a).
The study by Komatsu et al. (2003a) succeeded in map-
ping seagrass beds using bottom topography measured by a
MBS for the rst time. They were able to distinguish be-
tween echoes from seagrass and echoes from the actual sea-
bed. They removed echoes from the seagrass to create a sea
bottom without seagrass and subtract the sea bottom topogra-
phy without seagrass from the sea bottom topography with
seagrass to estimate seagrass distributions. They also esti-
mated biomass coupling quadrat sampling of seagrass with
volume and area of seagrass estimated by the MBS that
owering shoots and vegetative shoots occupied, respec-
tively. In a dierent study, Komatsu et al. (2004) mapped P.
oceanica in the Gulf of Gabes, Tunisia facing the Mediterra-
nean Sea. They were able to dierentiate the seagrass mead-
ows of P. oceanica and sand beds at depths ranging from 20
to 30 m using MBS backscattering information. These studies
required extensive post-processing, and multibeam sonar op-
eration needed to be carried out by highly trained personnel
due to the complexities associated with calibration and back-
scattering processes. In addition, the acquisition data volume
Fig. 12. Echogram (upper panels) and blade length distribution (lower panels) obtained by a quadrat sampling of 0.5 x 0.5 m at Sta-
tions 11–17 (Komatsu and Tatsukawa 1998). The mark “*” indicates transformed data from quadrat sampling of 1x1 m to that of 0.5 x
0.5 m due to small quantity of seagrass shoots. Grades of echo traces of seagrass beds were shown at the lower part of the vertical line
representing the station by the following characters: G1: Grade 1 (no seagrass); G2: Grade 2 (sparse seagrass); G3: Grade 3 (intermedi-
ate dense seagrass); G4: Grade 4 (dense seagrass). Blades lengths are shown as vertical lines from the smallest one to the largest one
in order in each Figure at the stations (lower panels).
Coastal Marine Science 43
12
by a MBS is larger than that by echosounder and sidescan
sonar processors (Anderson et al. 2008). Di Maida et al.
(2011) investigated the dierentiation of seagrass meadows
from sandy beds using bathymetry data obtained with the
MBS. They demonstrated that the MBS was capable of dis-
criminating between P. oceanica meadows and sand sub-
strate via decision-tree post-processing with use of the stan-
dard deviation of beam depths or beam depth range within
the grids of the bathymetry map. The grids were classied
into two categories, based on the presence or absence of
high-density seagrass. Recent studies using a MBS have ex-
plored the possibility of discriminating between seagrass and
macroalgae, to estimate abundance (dense or sparse) from
backscattering measurements (Tecchiato et al. 2015). How-
ever, their methodology is highly complicated, and their re-
sults remain inconclusive although discrimination between
the presence and absence of meadows was within acceptable
limits (78% accuracy in Tecchiato et al. (2015)).
The processing and analysis of backscattering strength
data from the MBS is still in the development stages. There
is no standard approach for MBS application (Parnum and
Gavrilov 2011). However, bathymetry data processing has
advanced to the point of survey area mapping in real time
(Lurton 2002). Hamana and Komatsu (2016) developed a
method to map seagrass meadows and estimate their relative
abundance on sand beds by using only bathymetry data
which simplify post-process analysis. This system can be
used to map seagrass meadows with long shoots growing in
the sublittoral zone in real time with use of a hydrographical
survey software with a function of 95% condence level of
bottom detections.
4.4 Seagrass cover and standing crop
Recently, satellite multiband images with high radiomet-
ric and/or spatial resolutions are obtained as explained
above. Using these data, researchers have studied to create
maps of seagrass cover and/or seagrass standing crop (e.g.
Hashim et al. 2014). Estimation of seagrass covers or stand-
ing crop require ground truthing data. This section explains
how to obtain ground-truthing data for this purpose.
Seagrass cover
In general, seagrass studies use a destructive sampling
with a quadrat to examine species compositions, shoot densi-
ties, above- and below-ground biomasses, leaf lengths, leaf
Fig. 13. Photo showing towing transducer of side-scan sonar
(Komatsu et al. 2003b).
Fig. 14. Map showing horizontal distribution of Zostera caulescens encircled with a white line surveyed by side-scan sonar (Komatsu
et al. 2003b).
Komatsu T. et al.: Seagrass mapping by remote sensing
13
area index and so on. The main disadvantage of the destruc-
tive sampling is that it takes much time to take samples in
situ and analyze samples in a laboratory. Remote sensing
studies require many ground truthing data. Thus, non-de-
structive sampling using visual assessment techniques are
applied to collect data for remote sensing studies. Percentage
cover of seagrass is estimated using a quadrat (e.g. McKen-
zie 2003). This method is to take digital pictures of seagrass
on a quadrat of 0.5×0.5 m in situ and classify seagrass covers
into percentage cover by using standardized pictures of sea-
grass covers. This method is quite repeatable when currents
are weak and seagrass blades have a vertical orientation.
However, the method becomes potentially unreliable when
current strength increases and forces the seagrass canopy into
a progressively horizontal (attened) plane (Mumby et al.
1997a).
Seagrass standing crop
Mumby et al. (1997a) propose an alternative method to
estimate seagrass standing crop based on Mellors (1991). An
intensive 3-day training period was undertaken prior to a sur-
vey work. After a broad reconnaissance survey, a provisional
biomass scale was established following the methods of Mel-
lors (1991). A quadrat of 0.5×0.5 m is placed in an area with
the lowest discernible biomass and was given the category 1.
The next quadrat is placed in seagrass which was both dens-
est and possessed greatest blade length. This was assigned a
category of 6. Quadrats for categories 2, 3, 4 and 5 are
placed by estimating a linear interpolation between catego-
ries 1 and 6. Three divers conduct surveys using the tech-
nique by haphazardly throwing quadrats and comparing cate-
gories. Over 100 quadrats are cross-compared in this fashion
until inter-observer agreement is highly consistent (complete
agreement on approximately 95% of occasions). At this
point, an adequate 1–6 scale is deemed to have been estab-
lished and 4 quadrats of each category are excavated for cali-
bration purposes. The quadrats are located haphazardly (i.e.
pseudo-randomly) and encompassed a range of seagrass spe-
cies composition rates if they are more than one species.
Harvested seagrass is washed in fresh water and sorted
to remove detritus and sediment. Each sample is divided by
species and then sub-sampled for biomass categories 4–6 if it
takes time to examine all seagrass in a quadrat of these cate-
gories. Sub-sampling is not necessary for most of the sam-
ples from categories 1–3. Epiphytes are removed from sea-
grass blades using either 5% citric acid or vinegar. Samples
are oven-dried at 80°C for 48 h and weighed to the nearest
0.1 g using an electronic balance. Epiphyte-free total dry
weights are calculated for each quadrat. It is important not to
confuse percent cover or density with standing crop estima-
tion. Although density and biomass arc closely related, the
determination of standing crop also takes blade length and
the relative dry weight of each species into account. From a
practical perspective, the assessment is carried out by consid-
ering the entire 3-dimensional standing crop within the quad-
rat (i.e. a volume of seagrass above the sediment). The high-
est standing crop categories of 5 and 6 dier mainly in blade
length rather than density. They recommend making a photo-
graphic record of the calibration quadrats which can be lami-
nated and taken underwater for guidance.
The seagrass standing crop is plotted with ordinal scale
(categories 1–6) for calibration of the scale. Mumby et al.
(1997a) obtained the regression line by transforming stand-
ing crop data with a modied square root function with a
good coecient of determination. Using this regression, vi-
sual assessment surveys on seagrass can bring standing crop
data.
5. Preparation for Processing Satellite
Images
5.1 Geometric correction
Satellite sensors project three-dimensional surface of the
earth to a plane. Satellite data were generally geocoded with
WGS 84 coordinate system. Therefore, it is necessary to ad-
just a spherical surface to a horizontal plane. In most cases,
remote sensing uses Universal Transversal Mercator (UTM)
projection. In the plane of UTM, x axis and y axis represent
east and north directions. UTM system has a zone number
depending on longitude of an area by every six degrees from
zone No. 1 between 180°W and 174°W to zone No. 60 be-
tween 174°E to 180°W. However, geometrical correction is
needed to t the image to the UTM coordinate system. This
correction is based on more than six GCPs whose longitude
and latitude have already been precisely obtained. In some
cases, it is necessary to measure longitude and latitude of
some discriminative points on the ground with a GPS with
Fig. 15. Calibration of ordinal scale for estimating seagrass
standing crop (Mumby et al. 1997a). The mean 95% confidence
intervals of the mean and range are shown from calibration data
of each category. The variance within actual dry weights has
been stabilized using a modified square-root transformation
(x’=√(x+3/8)). Coecient of determination, r2=0.94 (n=103), ac-
tual standing crop (g m−2). (source: Mumby et al. 1997a).
Coastal Marine Science 43
14
precision higher than sub-meter. If no GCPs are available,
we select some alternative points on the map as GSPs. We
relate GCPs with the corresponding points on the satellite
image with software for remote sensing. This operation is
called as geometric correction. It is recommended that GCPs
are scattered and also placed at edges and corners of struc-
tural object or roads for analysis.
5.2 Digital number to radiance
Raw digital number (DN) values recorded by a sensor
are proportional to upwelling electromagnetic radiation. The
majority of image processing has been based on DN values
in which actual spectral radiances are not of interest (e.g.
when classifying a single satellite image). However, there are
problems with this approach. The spectral signature of a hab-
itat expressed as DN values is not transferable because the
values are image specic under viewing geometry of the sat-
ellite when the image was pictured, the location of the sun,
specic weather conditions, etc. We cannot compare the val-
ues among the images taken at dierent time (e.g. seasons
and years) by dierent satellite sensors and on the area of
study larger than a single scene. Thus, it is necessary to con-
vert the DN values to spectral units that are universal among
dierent satellite images. If we obtain the spectral signature
of substrate types, we can compare “spectral libraries”—i.e.
libraries of spectral signatures containing lists of habitats and
their reectance. Converting the DN values to spectral units,
we can refer calibration equations depending on satellite sen-
sors that are included in the image data. In an example of
IKONOS, the following conversion equation is generally
used (Taylor 2005):
Li=Di/Ci, (1)
where L is the radiance at the sensor aperture (mW cm−2 sr−1),
C is the in-band radiance calibration coecient (cm2 sr mW−1)
and D is the DN value. In the equation the subscript i repre-
sents spectral band i. Spectral radiances can be obtained from
the calibration equation (1). USGS also provides the follow-
ing equation for converting DN to radiance of LANDSAT 8
OLI (USGS 2014).
Li=MLiDi+ALi, (2)
where MLi and ALi are band-specic multiplicative rescaling
factor and band-specic additive rescaling factor of band i
from the metadata, respectively. These radiances are those at
the top of atmosphere (TOA). DN of LANDSAT 8 OLI can
be also converted to TOA reectance as the following equa-
tion (USGS 2014):
( ) ( )
,
cos sin
ρρρρ
ρλ
θθ
＋＋
＝ ＝
ii i ii i
i
SZ SE
MDA MDA
(3)
where ρλi is TOA reectance of band i. Mρi and Aρi are band-
specic multiplicative rescaling factor and band-specic ad-
ditive rescaling factor of band i from the metadata, respec-
tively. θSZ and θSE are local sun elevation angle of the scene
center in degrees provided in the metadata and local solar ze-
nith angle, respectively.
5.3 Masking land areas and deep waters
It is necessary to exclude the land from satellite images
for classication of coastal habitats to avoid miss-classica-
tion of seagrass beds. Near-infrared bands are easily ab-
sorbed by the sea surface while reected by the land surface.
Thus, we use DN, reectance or radiance of an infrared band
on the image to discriminate the sea from the land. If ponds
or lakes are distributed on land, we need to manually mask
them.
Deep waters are also excluded with red or green bands
because they are absorbed with shallow water column. The
deeper water areas indicate constantly lower values of their
DN, reectance and radiances. Thus, deep water areas can be
classied with certain threshold values of red or green bands.
At the same time, it is needed to pay attention not to exclude
seagrass and seaweed beds because seagrass and seaweed
beds also show their lower values.
5.4. Atmospheric correction
The spectral radiances are those measured at the satellite
sensor. Figure 4 shows that electromagnetic radiation ob-
served by the satellite sensor has already passed through the
Earth’s atmosphere twice (sun to target and target to sensor).
During this passage, the radiation is aected by absorption
which reduces its intensity and scattering which alters its di-
rection. Absorption occurs when electromagnetic radiation
interacts with gases such as water vapor, carbon dioxide and
ozone. The electromagnetic radiation is scattered when it hits
both gas molecules and airborne particulate matter (aero-
Fig. 16. LANDSAT 8 OLI images of true color (upper left),
band 5 (upper right) and mask (lower) on Ajino Bay, Japan. The
mask was produced from pixels with DN values of near-infra red
band less than 7000.
Komatsu T. et al.: Seagrass mapping by remote sensing
15
sols). Scattering will redirect incident electromagnetic radia-
tion and deect the reected radiation from its path. Scatter-
ing also creates the adjacency eect in which the radiance re-
corded for a given pixel partly incorporates the scattered ra-
diance from neighboring pixels. In order to make a meaning-
ful measure of radiance at the Earth’s surface, the atmo-
spheric interferences must be removed from the data. This
process is called atmospheric correction. There are some so-
phisticated models of atmospheric corrections such as 5s and
6s models (e.g. Tanre et al. 1990, Vermote et al. 1997), etc.
After the atmospheric correction, light signals from the earth
surface become clearer.
Spectral radiance will depend on the degree of illumina-
tion of the object (i.e. the irradiance). Then, the spectral radi-
ance of features on the ground obtained by the atmospheric
correction is usually converted to reectance, which is called
surface reectance meaning the reectance of the surface of
the Earth. Be careful that the surface reectance is not that
on the sea bottom.
5.5. Water column correction
In coastal waters, Lyzenga’s model has been often used
for water column correction because this model’s simplicity
and eectiveness. In the scope of radiometric correction,
each pixel value within the image (DN value) is converted
into a radiance value as mentioned above. From an optical
perspective, bottom type can be identied by its reectance.
According to Lyzenga (1978), the relationship between the
radiance level recorded by an optical sensor and bottom re-
ectance is expressed by the following equation:
Li=Lsi+airiexp (−KigZ), (4)
where L is the radiance (mW cm−2 sr−1) same as in the equa-
tions (1) or (2), Ls is the radiance recorded over deep water
(external reection from the water surface and scattering in
the atmosphere), a is a constant which includes the solar irra-
diance, the transmittance of the atmosphere and the water
surface, and the reduction of the radiance due to refraction at
the water surface, r is the bottom surface reectance, K is the
eective attenuation coecient of the water (m−1), g is a geo-
metric factor to account for the path length through the water
and Z is the water depth (m). The subscript i means band i of
image. The value of g can be geometrically calculated from
the sun and satellite zenith angles at the moment when the
satellite image was taken. Bottom dierences are mirrored
by variations in L, as r changes according to the bottom type.
A radiometric correction index is required for estimating r.
We introduce two types of simple radiometric correction
suitable for coastal mapping: depth-invariant index proposed
by Lyzenga (1981) and bottom reectance index proposed by
Sagawa et al. (2010).
Depth-invariant index
In order to remove light scattering and absorption eects
within both the atmosphere and the water body, Lyzenga
(1981) suggested the calculation of Depth-invariant index.
This index is expressed as follows:
22
ln( ) ln( ) ,
j i si i i sj
ij
ij
K LL K LL
Index
KK
－－ －
＝＋
(5)
where L, Ls and K are the same as in equation (4), this time
with the i and j subscripts corresponding to two dierent
bands of satellite image. Equation (5) is derived from equa-
tion (4) and refers simultaneously to two bands (bands i and
j). For calculating this index, ratios of attenuation coe-
cients between bands are necessary. These coecients are
derived from ground truth data collected for a sandy bottom
along a bottom depth gradient (Lyzenga 1981). Using ground
truth data, we can plot satellite data against bottom depths
for sandy bottom type. The regression curve of Lyzenga’s
model is then obtained. When ground truth data are not avail-
able, we estimate sandy bottom and bottom depth from the
sea chart if it exists. When the sea chart is not available, we
estimate Kj/Ki as follows. Equation (5) can be transformed to
equation (6) through dividing numerator and denominator
with Ki:
22
/ ln( ) ln( )
,
1/
j i i si j sj
ij
ji
KK L L L L
Index
KK
－－ －
＝＋
(6)
where ln(Li−Lsi) and ln(Lj−Lsj) are dependent variables of
bottom depths from equation (4). If we can obtain reec-
tances of band i and j at the same pixels on a satellite image,
ln(Li−Lsi) and ln(Lj−Lsj) on pixels interpreted as the sand bot-
tom are plotted them on horizontal axis and vertical axis, re-
spectively. The regression line of points between ln(Li−Lsi)
and ln(Lj−Lsj) gives Kj/Ki as its slope. If bottom depth data
are available, it is possible to bottom depth versus ln(Li−Lsi)
or ln(Lj−Lsj). Sagawa et al. (2010) obtained very good nega-
tive exponential correlation between bottom depth and Li−Lsi
or Lj−Lsj in Mahares, Golf of Gabes, Tunisia (Fig. 17). This
shows that the regression line between ln(Li−Lsi) and ln(Lj−
Lsj) becomes straight.
Bottom reectance index
In order to improve mapping accuracy, Sagawa et al.
(2010) proposed an alternative reectance index (Bottom Re-
ectance index: BR index) expressed by the following equa-
tion:
()
,
exp( )
i si
i
i
LL
BR index K gZ
－
＝
－
(7)
where L, Ls, K, g and Z are the same as equation (4).
To calculate this BR index, it is needed to combine bot-
tom depth data, Z, with attenuation coecient, K. We use
each band attenuation coecient same as for the Depth-in-
Coastal Marine Science 43
16
variant index. Concerning bottom depth data, the bathymetry
map or sea chart supplied by local government or hydro-
graphical institutions is referred. It may be reasonable to take
advantage of these data, as they are easily available and rep-
resent generally accurate input. Once the numerator in equa-
tion (7) was replaced by airiexp(−KigZ) (from equation (4))
and the equation rearranged, the index becomes the follow-
ing equation including bottom reectance:
BR Indexi=airi (8)
where a and r are the same as in equation (4) and i corre-
sponds to a satellite image band i. Clearly, this index is di-
rectly related to bottom reectance. As a result, we named it
‘reectance index’. This index enables us to compare not
only the dierence in reectance ratios between bands but
also the dierence in absolute reectance for each band of
satellite image.
According to Lyzenga (1978), equation (4) should not
be applied to very shallow areas, as the model ignores inter-
nal reection eects occurring at the water surface. Thus, it
is better to apply both radiometric corrections exclusively to
areas deeper than about 1–2 m. On the other hand, we need
not underwater but atmospheric radiometric correction when
there are seagrasses exposed to the sea surface.
Sagawa et al. (2010) applied both radiometric correc-
tions of Depth-invariant index and Bottom Reectance index
to IKONOS image spotting seagrass beds o Mahares in
Golf of Gabes, south Tunisia, facing Mediterranean Sea, ac-
quired on 2 October 2005. In this area, P. oceanica is the
most abundant and common species and it mainly occurs on
a sandy bottom. Figure 18 shows results of supervised classi-
cation applied to Depth-invariant index and Bottom Reec-
tance index to classify sand and seagrass. In these waters
where attenuation coecients of blue band and green band
were 0.088−1 m and 0.093−1 m, respectively, water type was
Jerlov Water Type II–III (Jerlov 1976) suggesting turbid wa-
ters (Fig. 17). Overall accuracy of the former was 54%
meaning random classication between two bottom types.
On the other hand, that of the latter was over 90%. When the
bottom depth distribution is available, radiometric correction
Fig. 17. Relation between bottom depths and radiance levels
of blue (upper panel) and green (lower panel) on the sand bed in
Mahares, Golf of Gabes, Tunisia (source: Sagawa et al. 2010).
Fig. 18. Maps derived from satellite image analysis (Mahares). Black areas, described as ‘0–2 meters (unclassified area)’ in the leg-
ends, represent the data which were not included in the analysis. The maps are obtained by applying a radiometric correction based on
(a) the traditional Depth-invariant index (left map); (b) the Bottom Reflectance index (right map). (source: Sagawa et al. 2010).
Komatsu T. et al.: Seagrass mapping by remote sensing
17
using Bottom Reectance index is very practical to map sea-
grass beds under turbid waters rather than Depth-invariant
index. When water is clear, radiometric correction using
Depth-invariant index can also obtain good classication re-
sults. This is true because some studies showed good results
applying Depth-invariant index for mapping of seagrass beds
in Caribbean Sea (Mumby et al. 1998) or Mediterranean Sea
(Pasqualini et al. 2005, Belluco et al. 2006, Fornes et al.
2006) where transparencies are very high. Mumby and Ed-
wards (2000) conclude that Depth-invariant index for map-
ping coastal habitats is applicable to only waters with high
transparency belonging to Jerlov Water Type I–II (Jerlov
1976), which is roughly equivalent to waters with Secchi
depth of 50 m.
6. Image Classification
6.1 Pixel-based classication
In coastal habitat mapping, pixel-based classication
consisting of supervised or unsupervised classication is
generally applied to multiband satellite data after radiometric
correction mentioned above. Most studies on classications
for seagrass mapping have been conducted using a pixel-
based analysis of satellite multiband images. They used ei-
ther a supervised classication or unsupervised classication.
These pixel-based procedures analyze the spectral properties
of every pixel within the region of interest. Ground truth data
concerning each bottom type distributions are required.
These data must be prepared through ground-truthing such as
diving, observation by a lowered camera from the ship, side
scan sonar measurements, etc. as mentioned above and de-
scribed by Komatsu et al. (2003b). If ground truth data are
not available, unsupervised classication is useful. Unsuper-
vised classication groups pixels into some categories with
similar bottom reectance through a statistical standard such
as ISODATA etc. Using the unsupervised classication
method, the software is instructed to create the number of
clusters and the number of iterations that are inputs before
processing while attempting to meet a predetermined thresh-
old value within the number of iterations. By calculating and
plotting the cluster statistics, we can determine what number
of clusters (or classes) to use. Usually, we determine the
number of clusters more than classes that we need. After pro-
cessing, some clusters are merged to one cluster when they
represent a suitable class. Based on unsupervised classica-
tion, we can eectively conduct ground truth survey or eld
survey.
Mumby et al. (1999) stated that the number of samples
(ground truth data) necessary for one class with a supervised
classication is 30 ground truth data corresponding to 30
pixels of image. Calculation of classication accuracy needs
50 ground truth data (50 pixels of image). Thus, a total of 80
ground truth data are needed for one class with a supervised
classication. When we classify pixels into three classes, we
need 240 ground truth data corresponding to 240 pixels of
image. This is an ideal case. Since we do not have enough
time to take these samples, we use 10 to 30 samples for one
class for supervised classication (Green et al. 2000) and 20
to 30 samples for one class for accuracy evaluation.
The classication of supervised classication is based
on the spectral signature dened in the training set. The digi-
tal image classication software determines each class on
what it resembles most in the training set. Supervised classi-
cation is based on the idea that a user can select sample pix-
els in an image that are representative of each specic class
and then direct the image processing software to use these
training pixels as references for the classication of all other
pixels in the image. Training pixels are based on the ground
truth data above-mentioned. Supervised classication algo-
rithms frequently used are maximum likelihood and mini-
mum-distance classication. Maximum Likelihood method
assumes that the statistics for each class in each band are
normally distributed and calculates the probability that a
given pixel belongs to a specic class. Each pixel is assigned
to the class that has the highest probability (the maximum
likelihood). Minimum Distance method uses the mean vec-
tors for each class and calculates the Euclidean distance from
each unknown pixel to the mean vector for each class. The
pixels are classied to the nearest class.
Recently, machine learning and deep learning have been
applied to classication of seagrass beds from satellite
image. Machine Learning is a technique of parsing data,
learn from that data and then apply what they have learned to
make an informed decision (Anon. 2019). Deep learning is
actually a subset of machine learning. The main dierence
between deep and machine learning is, machine learning
models become better progressively but the model still needs
some guidance (Anon. 2019). Machine learning uses a model
such as Support Vector Machine, Random Forest, Logistic
Regression etc. to acquire a relation between inputs and out-
puts with training data sets while deep learning uses Neural
Network. If a machine learning model returns an inaccurate
prediction, then the programmer needs to x that problem
explicitly, but in the case of deep learning, the model does it
by himself. For example, Traganos and Reinartz (2018) ap-
plied three dierent methods to multispectral Sentinel-2 im-
agery for classifying four-class habitats (rocky algae, sand,
Cymodocea nodosa and P. oceanica) in the eastern Mediter-
ranean Sea. The machine-learning Random Forest and Sup-
port Vector Machine (SVM) methods obtained better overall
accuracies than Maximum Likelihood with and without ra-
diometric corrections. Islam et al. (2018) used deep learning
methods of Convolutional Neural Network and deep Capsule
Network to classify seagrass beds from 8 bands of World-
View-2 images in shallow waters in Florida State. Although
Coastal Marine Science 43
18
they did not apply any radiometric corrections to the images,
they obtained good classication results. It seems that inu-
ence of water column on classication might be corrected
through deep learning. It is necessary to increase studies on
machine learning and deep learning methods to classify sea-
grass beds from satellite imagery by applying them to dier-
ent areas.
6.2 Object-oriented classication
Recently, higher resolution satellite images such as
WorldView2/3/4 are available with reasonable prices for ar-
chived images. It is possible for classications to take the
spatial or contextual information related to pixels into ac-
count. Thus, the idea to classify objects stems from pixel-
based classications to object-oriented (or object-based)
classications based on groups of pixels with based on their
internal homogeneity and spectral separability at multi-scale
levels. The object-oriented analysis classies objects instead
of single pixels by multi-scale segmentation of pixels, which
lead to the extraction of spectrally and internally homoge-
neous units at a particular scale. It includes image segmenta-
tion to identify image objects and classication of the identi-
ed image objects. Objects form a hierarchical and scale-de-
pendent structure. This means that any object, in contrast to a
pixel, has not only neighbors but also sub-objects and super-
objects at dierent scales. Groups of pixels, due to their hier-
archical structure, are able to include many attributes which
can describe objects’ intrinsic characteristics (using physical
features like color, texture, and shape), typological character-
istics (relations to other objects, sub-objects and super-ob-
jects) and context. Representative softwares of object-ori-
ented approach are ArcGIS 10 (Esri), ENVI EX (Exelis VIS)
and eCognition (Trimble), which are based on edge to iden-
tify image objects and on FNEA (fractal net evolution ap-
proach) and multi-resolution segmentation, respectively
(Xiaohe et al. 2014).
The object-oriented method was applied to seagrass
mapping with images obtained from airborne digital cameras
by Lathrop et al. (2006). Following Robbins and Bell (1994)
approach, they set habitat structure at three dierent levels:
(a) meadow - a spatially continuous area of seagrass beds of
varying percent of cover composition; (b) bed—a spatially
continuous area of overall similar percent of cover composi-
tion; and (c) patch—a small discrete clump of seagrass- or
gap—an area within a seagrass bed not occupied by plants.
Lathrop et al. (2006) developed a hierarchical classica-
tion scheme to multiband images in a shallow (mean depth of
1.5 m at mean lower-low water) back-bay lagoonal type of
estuary on New Jersey’s Atlantic coast taken by an airborne
digital camera. Two GeoTi image products were created: a
true color imagery set, and an infrared imagery set, both at a
1-meter ground cell resolution and 8-bit radiometric resolu-
tion. They broadened three dierent levels of seagrass to six
levels: Level 1 of land and water, Level 2 of deep water/
channels (>1.5 to 2 m depth) and shallow water (<1.5 to 2 m
depth) from water of Level 1, Level 3 of sand/mud ats
(<1.5 to 2 m depth) and macrophyte from shallow water of
Level 2, Level 4 of the macroalgae and seagrass meadows
from macrophyte of Level 4, Level 5 of dense, moderate and
sparse seagrass beds from seagrass meadows of Level 4, and
Level 6 of seagrass patch and gap (bare bottom) of three sea-
grass levels of Level 5. Level 1 was classied with near-in-
frared band. Level 2 was classied a simple membership rule
based on the bathymetry layer. Level 6 was obtained by seg-
menting the whole area to ne scale objects using a scale pa-
rameter of 10. The image segmentation was then coarsened
to merge areas of like classes using a scale parameter of 15
(for Level 5) and 30 (for Level 4). The Level 4 or 5 image
objects were visually interpreted, and manual encoded as to
the appropriate bottom type based on the analyst judgment. If
the eld reference data was available, the analyst can consult
them. Thus, level 4 to 6 are dependent of scales of image
segmentation decided by scale parameters arbitrarily given.
The size of the object is associated with its physical and bio-
logical spatial structure. A homogeneity criterion for smooth-
ness and compactness of segmentation is based on a local
variance of pixels to be grouped, given by a parameter. The
seagrass density data for the 245 eld reference points were
categorized into four seagrass density classes (absent, sparse,
moderate and dense), compared with the same location from
the classication result.
Urbański et al. (2010) analyzed Quickbird data for map-
ping seagrass beds in sandy shoal habitat in the southern Bal-
tic Sea. The spatial resolution of the panchromatic band and
multi-spectral bands are 0.6 m and 2.5 m, respectively. Their
number and size at the particular segmentation level depend
on a scale parameter. The segmentation process aims to re-
tain objects of strong spectral and shape homogeneity. They
segmented each panchromatic image for the following scale
parameters: 400, 300, 200, 100, 50, and 20, which constitute
six levels of objects as mentioned above. The homogeneity
criterion was set to 0.9 for colour, 0.1 for shape, and 0.5 both
for smoothness and compactness. They selected level of sea-
grass meadows, beds and patch/gap with scale parameters of
200, 20 and 10, respectively. They classied segments of
seagrass bed level into ve zones with a seagrass index (SGI)
consisting of band 1, band 2 and standard deviation of band 2
of segment with a scale parameter of 20. Within each zone in
the object layer with a scale parameter of 20 (patch/gap
level), approximately 30 objects presenting vegetation cover
(Submerged Aquatic Vegetation: SAV) and approx. 30 ob-
jects without vegetation cover (not SAV) were selected by
manual sampling. In order to perform accuracy assessments,
samples from the patch/gap level were selected once more
but for the scale parameter of 10. The samples were assigned
to SAV or not SAV classes. They stated “Atypical, poten-
Komatsu T. et al.: Seagrass mapping by remote sensing
19
tially dicult to classify objects were purposely chosen”. A
problem of object-oriented classication is to set parameters
not objectively but subjectively and also needs manual
choosing of segments. In many cases, seagrasses grow patch-
ily. Thus, it isn’t problem that seagrass maps obtained by a
pixel-based classication applying satellite images with a
high spatial resolution show their distributions as salt-and-
pepper.
7. Validation of Accuracy
One of the most important themes for seagrass mapping
with remote sensing is an accuracy evaluation of classica-
tion. In general, accuracies of classication are evaluated
with an error matrix (confusion matrix or contingency table)
(Mumby and Green 2000). The accuracies are judged with a
user accuracy, a producer accuracy, an overall accuracy and a
tau coecient (Ma and Redmond 1995).
7.1 Error Matrix (Pixels and percent)
The error is calculated by comparing the class of each
ground truth pixel with the corresponding class in the classi-
cation image. Each column of the error matrix represents a
ground truth class and the values in the column correspond to
the classication image’s labeling of the ground truth pixels.
Table 3 shows the class distribution in pixels and percentage
for each ground truth class.
The user accuracy is a measure indicating the probabil-
ity that a pixel is Class A given that the classier has labeled
the image pixel into Class A. User accuracies are shown in
the rows of the error matrix. For example, in Table 3, sea-
grass pixels obtained with a ground survey are 50 of which
39 and 11 are correctly and incorrectly classied, respec-
tively. The percentages of the number of pixels correctly and
incorrectly classied into the seagrass class were 78% and
22% (Table 3), respectively, corresponding to a user accu-
racy and an error of commission.
The producer accuracy is a measure indicating the prob-
ability that the classier has labeled an image pixel into
Class A given that the ground truth is Class A. For example,
in Table 3, seagrass class of ground truth has a total of 48
pixels of which 39 and 9 pixels are correctly and incorrectly
classied, respectively. The percentages of the number of
seagrass pixels obtained with the ground survey classied
correctly and incorrectly are 81.3% and 18.8%, respectively,
corresponding to a producer accuracy and an error of omis-
sion (Table 1).
The overall accuracy is calculated by summing the num-
ber of pixels classied correctly and dividing by the total
number of pixels. The pixels correctly classied are found
along the diagonal of the error matrix table which lists the
number of pixels that were classied into the correct ground
truth classes. The total number of pixels is the sum of all the
pixels in all the ground truth classes. For example, in Table
3, the pixel counts of diagonal components consist of 39 pix-
els of seagrass and 51 pixels of sand, which are correctly
classied pixels. The overall accuracy (81.8%) is obtained by
dividing the correctly classied pixels number (39+51) by
the total number of ground truth pixels (110).
The overall accuracy is the overall degree of agreement
in the matrix. Generally, accuracies of classication of sur-
face covers of coastal sea bottom are lower than those of
land (e.g. Mumby et al. 1998). Mumby et al. (1999) stated
that a reasonable accuracy is between 60 and 80% for coarse
descriptive resolution such as corals/seagrasses and man-
groves/non-mangroves by using satellite images such as
LANDSAT TM or pansharpened image of LANDSAT TM
with SPOT. In any cases, overall accuracy is more than about
90% to monitor temporal changes in spatial distributions of
bottom covers using a remote sensing (Mumby and Green
2000).
7.2 Tau coecient
It is a reasonable way to describe the overall accuracy of
a map but does not account for the component of accuracy
resulting from chance alone. A chance component of accu-
racy exists because even a random assignment of pixels to
habitat classes would include some correct assignments.
The Tau coecient, T(e), is another measure of the ac-
curacy of the classication to exclude a chance component
and is expressed as the following equation:
Table 3. Example of error matrix of seagrass mapping. Shaded cells represent diagonal components that are correctly classified.
Numbers and those in parentheses are pixels and percentages of pixels, respectively.
Satellite image
classification data
Ground truth
Seagrass Sand Row total User accuracy
Seagrass 39 (35.5) 11 (10) 50 (45.5) 39/50 (78)
Sand 9 (8.2) 51 (46.3) 60 (54.5) 51/60 (85)
Column total 48 62 110 (100)
Producer accuracy 39/48 (81.2) 51/62 (82.3)
Overall accuracy 90/110 (81.8)
Tau coecient 0.636
Coastal Marine Science 43
20
(a) (e)
(e) ,
1 (e)
Pr Pr
T
Pr
－
＝
－
(9)
where Pr(a) and Pr(e) are the relative observed agreement
among classes and hypothetical probability of chance agree-
ment, respectively. For example, in Table 3, Pr(a) corre-
sponds to the overall accuracy, 0.818. Pr(e) derived from two
classes is 0.5. Then, we can obtain T(e) as 0.636 by dividing
(0.818–0.50) with (1.0–0.5). The Tau coecient ranges be-
tween −1.0 and 1.0. When the Tau coecient is −1.0 and 1.0,
classication is of perfect discrepancy and agreement, re-
spectively. When the Tau coecient is between 0.41 and
0.60, classication is of moderate agreement. When the Tau
coecient is between 0.61 and 0.80, classication is of good
agreement. When the Tau coecient is over 0.80, classica-
tion is of nearly prefect agreement.
The variance of Tau, σ2, is calculated as the following
equation (Ma and Redmond 1995):
2
2
(a) (1 (a))
,
(1 (e ))
Pr Pr
n Pr
－
＝
－
(10)
where n is the number of samples. Condence interval (CI)
is then calculated for each Tau coecient at the 95% con-
dence level (1−α), using the following form:
95%CI=T(e)±Za/2(σ2)0.5, (11)
where Z is a standard normal distribution with the lower
bound of α/2. Using Table 3, we obtain 95%CI as 0.636
±0.001. The coefcient’s distribution approximates to nor-
mality and Z-tests can be performed to examine dierences
between matrices (Ma and Redmond 1995). When two dif-
ferent classication methods (method 1 and method 2) are
applied, Z-tests are conducted to verify whether signicant
dierence in Tau coecients (T1 and T2) between results ob-
tained by method 1 and method 2 exists or not using the fol-
lowing equations:
12
22
12
,
σσ
－
＝
－
TT
Z
(12)
where σ2 is the variance of the Tau coecient, calculated
from equation (10). We can examine whether Tau coe-
cients have a 95% probability of being dierent or not.
8. Coverage and Biomass Estimation of
Seagrass
Some recent studies have documented methods for map-
ping seagrass species, cover and biomass properties from sat-
ellite images and eld data (e.g. Urbański et al. 2010, Baum-
stark et al. 2013). Phinn et al. (2008) compared accuracies of
seagrass percentage cover classication (1–10%, 10–40%,
40–70% and 70–100%) in the shallow subtidal areas of the
Eastern Banks, Moreton Bay, Australia, among images of
compact airborne spectrographic imager with a pixel size of
4 m×4 m, a radiometric resolution of 14 bit and 16 bands
(CASI), QuickBird-2 with a pixel size of 2.4 m×2.4 m, a ra-
diometric resolution of 11 bit and 4 bands, and LANDSAT 5
TM with a pixel size of 30 m × 30 m, a radiometric resolution
of 8 bit and 4 bands. The airborne hyper-spectral image data
returned high accuracies across all cover levels from very
high (70–100%) to very sparse (0–10%). In contrast, both
QuickBird-2 and LANDSAT 5 TM were unable to dierenti-
ate moderate to low (10–40%) and sparse (0–10%) levels of
seagrass cover. In the LANDSAT 5 TM image data, this was
caused by a function of its relatively large pixel size, limited
radiometric resolution and broad spectral bands. Phinn et al.
(2008) stated that the band placement and slightly lower ra-
diometric resolution of QuickBird-2 than the CASI data re-
duced its ability to detect the small reectance dierences
between seagrass cover.
Most of researches on seagrass biomass use relation be-
tween signals of multibands from bottom substrates and sea-
grass covers or standing crops of seagrass. Seagrass covers
and standing crops of seagrass are obtained by quadrat sam-
pling of seagrass through eld surveys. Mumby et al.
(1997b) used the Depth-invariant index (DII) as signals.
They applied principal component analyses to combine mul-
tiple DIIs into a single regressor (the rst principal compo-
nent). They obtained ground-truthing data by simple, precise,
non-destructive and quick method for measuring seagrass
standing explained in Section 4.4.
Above-ground biomass of seagrass has been estimated
by remote sensing. Most of studies have converted pixel val-
ues of reectance to above-ground biomass based on a re-
gression between reectance and biomass which have been
obtained by quadrat sampling in situ. Mumby et al. (1997b)
compared performances of above-ground biomass estimation
with use of regressions between eld measurements of stand-
ing crop and DIIs of LANDSAT TM, SPOT XS and CASI
images on seagrasses in the tropical Western Atlantic. They
found that predictions were high coecients of determina-
tion: 0.74, 0.79 and 0.81, respectively.
Hashim et al. (2014) used LANDSAT 8 OLI image to
map seagrass biomass in Johor Straight. After they classied
seagrass beds, they mapped above-ground biomass of classi-
ed seagrass area based on a regression between BR Index
and above-ground biomass of quadrat sampling of 0.5 ×
0.5 m. They obtained good regression between them with a
coecient of determination of 0.93 (r2) (see Fig. 13).
For estimating above-ground biomass of seagrasses, it is
necessary to consider their seasonal changes. When we com-
pare biomass distributions among dierent locations and pe-
riods, we need to take luxurious and scanty growth seasons
of seagrasses, interannual changes of timing of their growth
and interannual changes in their distributions into account.
Komatsu T. et al.: Seagrass mapping by remote sensing
21
Another problem is whether quadrat sampling of seagrass
biomass in only 0.5×0.5 m area represents biomass in a pixel
or not. For example, one pixel of LANDSAT TM is 30 ×
30 m. This pixel size includes not only quadrat sampling po-
sition but also other bottom covers, which produce a mixed
pixel. If seagrass beds are not homogeneous, quadrat samples
are not representative of the pixel because of its broader area.
If we use QuickBird-2 or CASI, we encounter georeferenc-
ing problems of quadrat samples in situ and pixels of an
image. In general, positions of quadrat sampling are deter-
mined with GPS including an error. Satellite and CASI im-
ages also include error in pixel positions. It is possible that a
position of quadrat sampling is not identical to a pixel on sat-
ellite or CASI images. To solve this georeferencing problem,
it is needed to use more accurate GPS system having a QZSS
system and also to conduct a quadrat sampling of a macro-
phyte at places where seagrass is homogeneously and
broadly distributed. In this case, image obtained by a side-
scan sonar can be used to evaluate a growth condition indi-
cated by Sagawa et al. (2008).
9. Summary
Most of remote sensing studies on seagrass beds aim to
map not only spatial distributions at a certain time but also
temporal changes in spatial distributions from time to time.
To conduct a reliable assessment of changes in seagrass ex-
tent and cover over time, the data sets compared should be
based on specic requirements as shown in Table 4 (Roelf-
sema et al. 2013). For assessing natural variability of sea-
grass distribution (intra- and inter-annual variations), it is im-
portant that eld sampling data and location, and date of re-
motely sensed data sets are considered.
We introduced simple and practical methods to map sea-
grass beds with satellite remote sensing. It is stressed that ra-
diometric correction is very important for satellite remote
sensing to correctly map bottom types. In tropical waters,
water transparency is usually very high. A simple method of
Lyzenga’s DII is very useful. On the other hand, in temperate
waters with low transparency, the Bottom Reectance index
(BRI) method proposed by Sagawa et al. (2010) is better
than the DII method to map seagrass beds by satellite remote
sensing when bottom topography data are available.
At this moment, LANDSAT 8 OLI and Sentinel-2 with
30 m and 10 m spatial resolution are the only freely available
non-commercial satellite images provided by NASA and
ESA, respectively and can be applied to coastal habitat map-
ping without cost. In scal 2020, JAXA will launch ALOS-3
that has higher spatial resolution of 4 × 4 m than LANDSAT
8 OLI and Sentinel-2. When this satellite is successfully
launched in 2020, seagrass mapping in not only Asia and the
Northwest Pacic but also the other regions will be advanced
enormously under the condition that Japanese Government
provides such images in free to everybody for contributing
Sustainable Development Goal 14.
Acknowledgements
This review was based on the results from several workshops
of Ocean Remote Sensing Project of Coastal Habitat Mapping
(ORSP-CHM) since 2010 under the IOC Sub-Commission for the
Western Pacic (WESTPAC), the Intergovernmental Oceanographic
Commission of the United Nations Educational, Scientic and Cul-
Table 4. Ideal requirements to conduct a reliable assessment of changes in seagrass extent and horizontal projected percentage sea-
grass cover over time. *The error that could occur when requirement is not adhered to (Source: Roelfsema et al. 2013).
Data set requirement Error type Example of impact error (DS1=data set 1, DS2=Data set 2)
Georeferenced Position shift Detected changes (false positive)
Near identical spatial extent Missing data If an area is not mapped for DS1 compared to DS2 due to
missing data, it could be detected as a change
Identical mapping categories Incomparable mapping categories Qualitative versus quantitative categories
Identical mapping scale Variation in level of detail
Small patches of seagrass are mapped in one and not in other
Reproducible mapping method Methodological error DS1 based on manual digitization, DS2 based on pixel-
based image classification
Seasonal sampling Natural variation DS1 in winter and DS2 in summer
Similar tidal stage and
water clarity
Aects ability to detect seagrass Satellite image for DS1 was derived at high tide with turbid
water, and for DS2 with low tide and clear water. Seagrass
could be mapped in deeper water for DS2.
Replicate field sampling Variation in calibration or valida-
tion
DS1 field data based on limited point-based sampling, DS2
based on detailed transect sampling for same area
Sampling accuracy Decreased map quality DS1 has high accuracy versus DS2 with low accuracy result-
ing in low reliability
Coastal Marine Science 43
22
tural Organization (IOC/UNESCO). The original manuscript was
presented for the workshop on case studies on seagrass mapping in
the selected sea areas in the NOWPAP region organized by the Ac-
tion Plan for the Protection, Management and Development of the
Marine and Coastal Environment of the Northwest Pacic Region
(NOWPAP) as a part of the Regional Seas Programme of the United
Nations Environment Programme (UNEP) in August 2015. The au-
thors thank to Ministry of Education, Sports, Culture and Science of
Japan for providing Japan Fund in Trust for “Promoting the Aware-
ness on Coastal Marine Environmental Changes and its Impact
(2011–2013)”, “Bolstering Knowledge and Institutional Capacity
for Ensuring Marine Biodiversity and Seafood Security (2014–
2017)” and “Accelerating the transfer of marine technology for the
conservation and sustainable use of oceans, seas and marine re-
sources (2018–2021)” to organize the workshops. These workshops
were also supported by Japan Society for the Promotion of Science-
Asian Core Program on Establishment of Research and Education
Network on Coastal Marine Science in. Southeast Asia (JSPS-
COMSEA) led by Prof. Shuhei Nishida of the University of Tokyo.
The authors appreciate Prof. Tom Morris of Fullerton College, Cali-
fornia, USA who kindly provided Figure 3 to this article. Finally,
they acknowledge to Mr. Wenxi Zhu, Head of IOC/WESTPAC Of-
ce, Prof. Emeritus Yasuwo Fukuyo of the University of Tokyo, for-
mer Vice Chairperson of WESTPAC and Prof. Emeritus Tetsuo
Yanagi of Kyushu University, former leader of JSPS project for their
support to have the workshops and their encouragements to conduct
the ORSP-CHM since 2010.
References
Anderson, J. T., van Holliday, D., Kloser, R., Reid, D. G. and Simard,
Y. 2008. Acoustic seabed classication: Current practice and
future directions. ICES J. Mar. Sci. J. Cons. 65: 1004–1011.
Anon. 2019. Articial intelligence vs Machine Learning vs Deep
Learning. https://www.geeksforgeeks.org/articial-intelli-
gence-vs-machine-learning-vs-deep-learning/ accessed on 1
May 2019.
Baumstark, R., Dixon, B., Carlson, P., Palandro, D. and Kolasa, K.
2013. Alternative spatially enhanced integrative techniques for
mapping seagrass in Florida’s marine ecosystem. Int. J. Re-
mote Sens. 34: 1248–1264.
Belluco, E., Camuo, M., Ferrari, S., Modenese, L., Silvestri, S.,
Marani, A. and Marani, M. 2006. Mapping salt-marsh vegeta-
tion
by multispectral and hyperspectral remote sensing. Re-
mote Sens. Environ. 105: 54–67.
Boudouresque, C.-F., Charbonel, E., Meinesz, A., Pergent, G., Per-
gent-Martini, C., Cadiou, G., Bertrandy, M. C., Foret, P., Raga-
zzi, M. and Rico-Raimondino, V. 2000. A monitoring network
based on the seagrass Posidonia oceanica in the northwestern
Mediterranean Sea. Biol. Mar. Mediterr. 7: 328–31.
Calvo, S., Frada Osterano, C. and Abbadessa, P. 1993. The suitabil-
ity of a topographical instrument for an integrated approach to
the cartography of Posidonia oceanica meadows. Oceanol.
Acta 16: 273–278.
Chancerelle, Y., Maurin, R., Poujade, S. and Vieux, C. 2008. Photo
tow: a new method for estimating coral reef status and changes
at large spatial scale. A coral bleaching event case study. J. Soc.
Océanis., 126–127. http://journals.openedition.org/jso/3312;
DOI:10.4000/jso.3312
Colantoni, P., Gallignani, P., Fresi, E. and Cinelli, F. 1982. Patterns of
Posidonia oceanica (L.) Delile beds around the Island of Ischia
(Gulf of Naples) and in adjacent waters. P.S.Z.N.: Mar. Ecol. 3:
53–74.
Coles, R. G., Lee Long, W. J., Watson, R. A. and Derbyshire, K. J.
1993. Distribution of seagrasses, and their sh and penaeid
prawn communities, in Cairns Harbour, a tropical estuary,
northern Queensland, Australia. Austral. J. Mar. Fresh. Res. 44:
193–210.
Costanza, R., d’Arge, R., de Groot, R., Farber, S., Grasso, M., Han-
non, B., Limburg, K., Naeem, S., O’Neill, R. V., Paruelo, J.,
Raskin, R. G., Sutton, P. and van den Belt, M. 1997. The value
of the world’s ecosystem services and natural capital. Nature
387: 253–260.
De Falco, G., Tonielli, R., di Martino, G., Innangi, S., Simeone, S.
and Parnum, I. M. 2010. Relationships between multibeam
backscatter, sediment grain size and Posidonia oceanica sea-
grass distribution. Cont. Shelf Res. 30: 1941–1950.
Dennison, W. C., Orth, R. J., Moore, K. A., Stevenson, J. C., Carter,
V., Kollar, S, Bergstrom, P. W. and Batiuk, R. A. 1993. Assess-
ing water quality with submerged aquatic vegetation: Habitat
requirements as barometers of Chesapeake Bay health. BioSci.
43: 86–94.
Di Maida, G., Tomasello, A., Luzzu, F., Scannavino, A., Pirrotta,
M., Orestano, C. and Calvo, S. 2011. Discriminating between
Posidonia oceanica meadows and sand substratum using mul-
tibeam sonar. ICES J. Mar. Sci. 68: 12–19.
Dierberg, F. E. and Kiattisimkul, W. 1996. Issues, impacts, and im-
plications of shrimp aquaculture in Thailand. Environ. Manage.
20: 649–666.
Duarte, C. M. 1987. Use of echosounder tracings to estimate the
above ground biomass of submerged plant in lakes. Canad. J.
Fisher. Aquat. Sci. 44: 732–735.
Duarte, C. M. 1991. Seagrass depth limits. Aquat. Bot. 40: 363–377.
English, S., Wilkinson, C. and Baker, V. (eds.) 1997. Survey manual
for tropical marine resources, 2nd edition. Australian Institute
of Marine Sciences, Townsville, Australia.
Fredj, G., Meinardi, M., Pierrot, S. and Roy, P. 1990. Cartographie
par le satellite SPOT 1 de communautés benthiques littorales
en Méditerranée occidentale. Bull. Inst. Oceanogr. Monaco 6:
71–85.
Fornes, A., Basterretxea, G., Orla, A., Jordi, A., Alvarez, A. and
Tintore, J. 2006. Mapping Posidonia oceanica from IKONOS.
ISPRS J. Photogram. Remote Sens. 60: 315–322.
Gloux, B. 1984. Méthode acoustiques et informatiques appliquées à
la cartographie rapide et détaillée des herbiers. In International
Workshop on Posidonia oceanica beds. Boudouresque, C.-F.,
Jeudy de Grissac, A. and Olivier, J. (eds.), pp. 45–48, GIS
Posidonie Press, Marseille.
Gray, P. C., Ridge, J. T., Poulin, S. K., Seymour, A. C., Schwantes, A.
M., Swenson, J. J. and Johnston, D. W. 2018. Integrating drone
imagery into high resolution satellite remote sensing assessments
of estuarine environments. Remote Sens. 10(8): 1257. https://
www.mdpi.com/2072-4292/10/8/1257/pdf
Green, E., Clark, C. and Edwards, A. 2000. Image classication and
habitat mapping. In Remote sensing handbook for tropical
coastal management. Edwards, A. J. (ed.), pp. 155–174,
Komatsu T. et al.: Seagrass mapping by remote sensing
23
UNESCO, Paris.
Gujja, B. and Finger-Stich, A. 1996. What price prawn? Shrimp
aquaculture’s impact in Asia. Environ. 38: 12–39.
Haddad, K. D. and Harris, B. A. 1985. Use of remote sensing to access
estuarine habitats. In Proceeding of the Fourth Symposium on
Coastal and Ocean management “Coastal Zone ‘85”, pp. 662–
675, ASCE, Baltimore, July 30-August 1. https://repository.lib
rary.noaa.gov/view/noaa/9221/noaa_9221_DS1.pdf
Hamana, M. and Komatsu, T. 2016. Real-time classication of sea-
grass meadows on at bottom with bathymetric data measured
by a narrow multibeam sonar system. Remote Sens. 8: 96. doi:
10.3390/rs8020096
Hashim, M., Misbari, S., Yahya, N. N., Ahmad, S., Reba, M. N. and
Komatsu, T. 2014. An approach for quantication of submerged
seagrass biomass in shallow turbid coastal waters. In Geosci-
ence and Remote Sensing Symposium (IGARSS), pp. 4439–
4442, 2014 IEEE International, IEEE.
Hashimoto, T. and Nishimura, M. 1953a. Sounding of articial
schools, rocks and sea plants elds by sh nder. Tech. Rep.
Fishing Boat Lab. 4: 138–142. (in Japanese).
Hashimoto, T. and Nishimura, M. 1953b. Study on detection of Jap-
anese tang eld by the ultrasonic sh-nder. Tech. Rep. Fishing
Boat Lab. 5: 187–194. (in Japanese).
Hatakeyama, Y. and Maniwa, Y. 1978. On the investigation of sea-
weed distribution by utilizing the sh nder. Tech. Rep. Fish-
ing Boat Lab. 73: 155–168. (in Japanese).
Hayashizaki, K. and Ogawa, H. 2011. Underwater observation on
seagrass communities. In Seagrasses: Resource status and
trends in Indonesia, Japan, Malaysia, Thailand and Vietnam.
Ogawa, H., Japar Sidik, B. and Muta Harar, Z. (eds.), pp. 132–
152, Seizsando-Shoten, Tokyo.
Huitric, M., Folke, C. and Kautsky, N. 2002. Development and gov-
ernment policies of the shrimp farming industry in Thailand in
relation to mangrove ecosystems. Ecol. Econo. 40: 441–455.
Islam, K. A., Pérez, D., Hill, V., Schaeer, B., Zimmerman, R. and
Li, J. 2018. Seagrass detection in coastal water through deep
capsule networks. In Chinese Conference on Pattern Recognition
a
nd Computer Vision (PRCV), pp. 320–331, Springer, Cham.
Japan Coast Guard. 2017. About the abolition of dierential GPS.
Japan Coast Guard, Tokyo. (in Japanese). https://www.kaiho.
mlit.go.jp/soshiki/koutsuu/dgps/haishi.pdf
Jerlov, N. G. 1976. Marine Optics. Elsevier Scientic Pub., New York.
Jeudy de Grissac, A. and Boudouresque, C.-F. 1985. Rôle des herbiers
de Phanérogames marines dans les mouvements des sédiments
côtiers: les herbiers à Posidonia oceanica. In Acte du colloque
pluridisciplinaire franco-japonais d’Océanographie Fiscicule 1.
Ceccaldi, H. J. and Champalbert, G. (eds.), pp. 143–151, So-
ciété Franco-japonaise d’Océanographie, Marseille.
Kirkman, H. 1996. Baseline and monitoring methods for seagrass
meadows. J. Environ. Manag. 47(2): 191–201.
Kitoh, H. 1983. Seaweed beds research by a scientic echosounder.
Seikai Regional Fish. Res. Inst. News 43: 2–4. (in Japanese).
Komatsu, T. 1989. Day-night reversion in the horizontal distributions
of dissolved oxygen content and pH in a Sargassum forest. J.
Oceanogr. Soc. Jpn. 45: 106–115.
Komatsu, T. 1996. Inuence of a Zostera bed on the spatial distribu-
tion of water ow over a broad geographical area, In Seagrass
biology: Proceedings of an international workshop Rottnest Is-
land, Western Australia, 25–29 January 1996. Kuo, J, Phillips,
RC, Walker DI. and Kirkman, H. (eds.), pp. 111–116, Faculty
of Science, The University of Western Australia, Nedlands.
Komatsu, T. 1997. Long-term changes in the Zostera bed area in the
Seto Inland Sea (Japan), especially along the coast of the
Okayama Prefecture. Oceanol. Acta 20: 209–216.
Komatsu, T. and Tatsukawa, T. 1998. Mapping of Zostera marina L.
beds in Ajino Bay, Seto Inland Sea, Japan, by using echo-sounder
and global positioning systems. J. Recherche Océanogr. 23:
39–46.
Komatsu, T. and Yamano, H. 2000. Inuence of seagrass vegetation
on bottom topography and sediment distribution on a small
spatial scale in the Dravuni Island Lagoon, Fiji. Biol. Mar.
Mediter. 7: 243–246.
Komatsu, T., Ben Mustapha, K., Shibata, K., Hantani, K., Ohmura,
T., Sammari, C., Igarashi, C. and El Abed, A. 2004. Mapping
Posidonia meadows on Messioua Bank o Zarzis, Tunisia,
using multi-beam sonar and GIS. In GIS/spatial analyses in
sheries and aquatic sciences. Nishida, T., Kaiola, P. J., Hol-
lingworth, C. E. (eds.), pp. 83–100, Fishery-Aquatic GIS Re-
search Group, Kawagoe, Saitama, Japan.
Komatsu, T., Igarashi, C., Tatsukawa, K., Nakaoka, M., Hiraishi, T.
and Taira, A. 2002a. Mapping of seagrass and seaweed beds
using hydro-acoustic methods. Fish. Sci. 68: sup. I, 580–583.
Komatsu, T., Igarashi, C., Tatsukawa, K., Sultana, S., Matsuoka, Y.
and Harada, S. 2003a. Use of multi-beam sonar to map seagrass
beds in Otsuchi Bay, on the Sanriku Coast of Japan. Aquat.
Living Resor. 16: 223–230.
Komatsu, T., Ishida, K., Iizumi, H., Okamoto, M. and Belsher, T.
2001. Utilization of ALOS data for mapping coastal ecosystem
and managing sheries activity, In Proceedings 1st ALOS PI
Workshop, March 28–March 30, 2001, pp. 384–386, NASDA,
Tokyo.
Komatsu, T., Sagawa, T., Sawayama, S., Tanoue, H., Mohri, A. and
Sakanishi, Y. 2012. Mapping is a key for sustainable develop-
ment of coastal waters: Examples of seagrass beds and aqua-
culture facilities in Japan with use of ALOS images. In Sus-
tainable development. Ghenai, C. C. (ed.), pp. 145–160, ISBN
979-953-307-649-1. In Tech Publishing Co., Rijeka, Croatia.
Komatsu, T., Tatsukawa, K., Ishida, K., Igarashi, C., Sultana, S.,
Takahashi, M. and Matsuoka, Y. 2002b. Development of meth-
ods mapping coastal zone along Sanriku Coast using satellite
imagery and acoustic survey. In Conserving our coastal envi-
ronment: Man and the ocean. UNU-ORI-Iwate Symposium
Marine Ecology and Environment, pp. 33–47, United Nations
University, Tokyo.
Komatsu, T., Mikami, A., Sultana, S., Ishida, K., Hiraishi, T. and
Tatsukawa, K. 2003b. Hydro-acoustic methods as a practical
tool for cartography of seagrass beds. Otsuchi Mar. Sci. 28:
72–79.
Lathrop, R. G. Montesano, P. and Haag, S. 2006. A multi-scale seg-
mentation approach to mapping seagrass habitats using. Photo-
gram. Eng. Remote Sens. 72: 665–675.
Lee, Z., Hu, C., Shang, S., Du, K., Lewis, M., Arnone, R. and
Brewin, R. 2013. Penetration of UV-visible solar radiation in
the global oceans: Insights from ocean color remote sensing. J
Geophys. Res: Ocean. 118: 4241–4255.
Lefèvre, J. R., Meinesz, A. and Gloux, B. 1984. Premières données sur
Coastal Marine Science 43
24
la comparaison de trois méthôdes de cartographie des biocéno-
ses marines. Rapp. P.-V. Réun. Comm. Int. Explor. Sci. Mer
Mediterr. 29: 209–211.
Lurton, X. 2002. An introduction to underwater acoustics: Principles
and applications. Springer, Berlin.
Lyons, A. P. and Abraham, D. A. 1999. Statistical characterization of
high-frequency shallow-water seaoor backscatter. J. Acoust.
Soc. Am. 106: 1307–1315.
Lyzenga, D. R. 1978. Passive remote-sensing techniques for mapping
water depth and bottom features. Appl. Optics 17: 379–383.
Lyzenga, D. R. 1981. Remote sensing of bottom reectance and water
attenuation parameters in shallow water using aircraft and
Landsat data. Int. J. Remote Sens. 10: 53–69.
Ma, Z. and Redmond, R. L. 1995. Tau coecients for accuracy as-
sessment of classication of remote sensing data. Photogram.
Eng. Remote Sens 61: 435–439.
McKenzie, L. J. 2003. Guidelines for the rapid assessment and
mapping of tropical seagrass habitats. QFS, NFC, Cairns, 46 pp.
https://www.researchgate.net/prole/Len_McKenzie/publication/
268423725_Guidelines_for_the_rapid_assessment_of_seagrass_
habitats_in_the_western_Pacic/links/573aa8e308aea45ee83f
976b/Guidelines-for-the-rapid-assessment-of-seagrass-
habitats-in-the-western-Pacic.pdf
Meinesz, A. 1997. Balisage de la limite inférieure de l’herbier de
Posidonia oceanica en rade de Villefranche-sur-Mer (Alpes-
Maritimes, France). Rapp. P.-V. Réun. Comm. Int. Explor. Sci.
Mer Mediterr. 24: 143–144.
Meinesz, A. and Laurent, R. 1978. Cartographie et état de la limite
inférieure de l’herbier de Posidonia oceanica dans les Alpes-
maritimes (France) —Campagne Poseidon 1976—. Bot. Mar.
21: 513–526.
Meinesz, A., Cuvelier, M. and Laurent, R. 1981. Méthode récentes
de
cartographie et de surveillance des herviers de Phanéro-
games marines. Vie Milieu 31: 27–34.
Mellors, J. E. 1991. An evaluation of a rapid visual technique for es-
timating seagrass biomass. Aquat. Bot. 42(1): 67–73.
Miller, I. R., Jonker, M. and Coleman, G. 2009. Crown-of-thorns
starsh and coral surveys using the manta tow and SCUBA
search techniques. Australian Institute of Marine Science.
Townsville, Australia.
Mumby, P. J. and Edwards, A. 2000. Water column correction tech-
niques. In Remote sensing handbook for tropical coastal man-
agement. Edwards, A. J. (ed.), pp. 121–128, UNESCO, Paris.
Mumby, P. J. and Green, E. P. 2000. Field survey: building the link
between image and reality. In Remote sensing handbook for
tropical coastal management. Edwards, A. J. (ed.), pp. 57–66,
UNESCO, Paris.
Mumby, P. J., Clark, C. D., Green, E. P. and Edwards, A. J. 1998.
Benets of water column correction and contextual editing for
mapping coral reefs. Int. J. Remote Sens. 19: 203–210.
Mumby, P. J., Edwards, A. J., Green, E. P., Anderson, C. W., Ellis,
A. C. and Clark, C. D. 1997a. A visual assessment technique
for estimating seagrass standing crop. Aquatic Conservation:
Mar. Fresh. Ecosys. 7(3): 239–251.
Mumby, P. J., Green, E. P., Edwards, A. J. and Clark, C. D. 1997b.
Measurement of seagrass standing crop using satellite and digital
airborne remote sensing. Mar. Ecol. Prog. Ser. 159: 51–60.
Mumby, P. J., Green, E. P., Edwards, A. J. and Clark, C. D. 1999.
The cost-eectiveness of remote sensing for tropical coastal
resources assessment and management. J. Environ. Manag. 55:
157–166.
Naylor, R. L., Goldburg, R. J., Primavera, J. H., Kautsky, N., Bev-
eridge, M. C., Clay, J., Folke, C., Lubchenco, J., Mooney H. and
Troell, M. 2000. Eect of aquaculture on world sh supplies.
Nature 405(6790): 1017–1024.
Newton, R. S. and Stefanon, A. 1975. Application of side scan sonar
in marine biology. Mar. Biol. 31: 287–291.
Norris, J. G., Wyllie-Echeverria, S., Mumford, T., Bauley, A. and
Turner, T. 1997. Estimating basal area coverage of subtidal
seagrass beds using underwater videography. Aquat. Bot. 58:
269–287.
Onishi, Y., Mohri, Y., Tuji A., Ohgi, K., Yamaguchi, A. and Imai, I.
2014. The seagrass Zostera marina harbors growth-inhibiting
bacteria against the toxic dinoagellate Alexandrium tamarense.
Fish. Sci. 80: 353–362.
Parnum, I. M. 2007. Benthic habitat mapping using multibeam
sonar
system. PhD Thesis, Department of Imaging and Applied
Phys-
ics, Curtin University of Technology, Perth, Australia.
Parnum, I. M. and Gavrilov, A. N. 2011. High-frequency multibeam
echo-sounder measurements of seaoor backscatter in shallow
water: Part 1-Data acquisition and processing. Underw. Tech-
nol. 30: 3–12.
Pasqualini, V., Pergent-Martini, C., Clabaut, P. and Pergent, G.
1998. Mapping of Posidonia oceanica using aerial photo-
graphs and side scan sonar: Application o Island of Corsica
(France). Estuar. Coast. Shelf Sci. 47: 359–367.
Pasqualini, V., Pergent-Martini, C., Pergent, G., Agreil, M., Skoufas,
G., Sourbes, L. and Tsirika, A. 2005. Use of SPOT 5 for map-
ping seagrasses: An application to Posidonia oceanica. Remote
Sens. Environ. 94: 39–45.
Phinn, S., Roelfsema, C., Dekker, A., Brando, V. and Anstee, J. 2008.
Mapping seagrass species, cover and biomass in shallow waters:
An assessment of satellite multi-spectral and airborne hyper-
spectral imaging systems in Moreton Bay (Australia). Remote
Sens. Environ. 112: 3413–3425.
Ramos, M. A. and Ramos-Espla, A. 1989. Utilization of acoustic
methods in the cartography of the Posidonia oceanica bed in
the bay of Alicante (SE, Spain). Posidonia Newsletter 2: 17–19.
Riegl, B., Moyer, R. P., Morris, L., Virnstein, R. and Dodge, R. E.
2005. Determination of the distribution of shallow-water sea-
grass and drift algae communities with acoustic seaoor dis-
crimination. Rev. Biol. Trop. 53: 165–174.
Rey, J. and Diaz del Rio, V. 1989. Cartographia de los fondos marinos
de la Bachia de Palma (Baleares, Espana): Distribution de las
praderas vegetales y sedimentos superciales. In International
Workshop on Posidonia beds 2 GIS. Boudouresque, C.-F.,
Meinesz, A., Fresi, E. and Gravez, V. (eds.), pp. 29–41, Posidonie
Press, Marseille.
Robbins, B. D. and Bell, S. S. 1994.. Seagrass landscapes: A terres-
trial approach to the marine subtidal environment. Trend. Ecol
Evol. 9: 301–304.
Roelfsema, C., Kovacs, E. M., Saunders, M. I., Phinn, S., Lyons, M.
and Maxwell, P. 2013. Challenges of remote sensing for quan-
tifying changes in large complex seagrass environments. Estuar.
Coast. Shelf Sci. 133: 161–171.
Sabol, B. M., Melton, R. E., Chamberlain, R., Doering, P. and Haunert,
Komatsu T. et al.: Seagrass mapping by remote sensing
25
K. 2002. Evaluation of a digital echo sounder system for detec-
tion of submersed aquatic vegetation. Estuar. 25: 133–141.
Sagawa, T., Mikami, A., Komatsu, T., Kosaka, N., Kosako, A., Miya
-
zaki, S. and Takahashi, M. 2008. Mapping seagrass beds using
IKONOS satellite image and side scan sonar measurements: A
Japanese case study. Int. J. Remote Sens. 29: 281–291.
Sagawa, T., Boisnier, E. Komatsu, T., Ben Mustapha, K., Hattour,
A., Kosaka, N. and Miyazaki, S. 2010. Using bottom surface
reectance to map coastal marine areas: A new application
method for Lyzenga’s model. Int. J. Remote Sens. 31: 3051–
3064.
Tanaka, M. and Tanaka, K. 1985. On the estimation of the abundance
of seaweed in coastal area I, estimation by echosounder. Fish.
Civil Eng. 21: 17–23. (in Japanese).
Tanre, D., Deroo, C., Duhaut, P., Herman, M., Morcrette, J. J., Perbos,
J. and Deschamps, P. Y. 1990. Description of a computer code
to simulate the satellite signal in the solar spectrum: 5S code.
Int. J. Remote Sens. 11: 659–668.
Taylor, M. 2005. IKONOS planetary reectance and mean solar
exoatmospheric irradiance, IKONOS planetary reectance QSOL
Rev.1. Available online at: http://www.geoeye.com/products/
imagery/ikonos/spectral.htm (accessed 15 January 2016).
Tecchiato, S., Collins, L., Parnum, I. M. and Stevens, A. 2015. The
inuence of geomorphology and sedimentary processes on
benthic habitat distribution and littoral sediment dynamics:
Geraldton, Western Australia. Mar. Geol. 359: 148–162.
Traganos, D. and Reinartz, P. 2018. Mapping Mediterranean sea-
grasses with Sentinel-2 imagery. Mar. Pollut. Bull. 134: 197–
209.
Urbański, J. A., Mazur, A. and Janas, U. 2010. Object-oriented
classication of Quickbird data for mapping seagrass spatial
structure. Int. J. Oceanogr. Hydrobiol. 38: 27–43.
USGS. 2014. Using the USGS Landsat 8 Product. Available online at:
http://landsat.usgs.gov/Landsat8_Using_Product.php (accessed
15 December 2015).
Vermote, E., Tanre, D., Deuze, J. L., Herman, M. and Morcrette, J.
J. 1997. Second simulation of the satellite signal in the solar
spectrum, 6S: An overview, IEEE T. Geosci. Remote Sens. 35:
675–686.
Ward, L. G., Kemp, W. M. and Boyton, W. R. 1984. The inuence
of waves and seagrass communities on suspended particulates
in an estuarine embayment. Mar. Geol. 59: 85–103.
Wilson, P. S. and Dunton, K. H. 2009. Laboratory investigation of
the acoustic response of seagrass tissue in the frequency band
0.5–2.5 kHz. J. Acoust. Soc. Am. 125: 1951–1959.
Xiaohe, Z., Liang, Z., Jixian, Z. and Huiyong, S. 2014. An object-
oriented classication method of high resolution imagery
based on improved AdaTree. IOP Conference Series: Earth and
Environmental Science, 17, 012212. doi:10.1088/1755-1315/
17/1/012212
