Remote Sensing of Environment: Contents Lists Available at

Remote Sensing of Environment 216 (2018) 598–614
Contents lists available at ScienceDirect
Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse
Coral reef applications of Sentinel-2: Coverage, characteristics, bathymetry T

and benthic mapping with comparison to Landsat 8
⁎
John D. Hedleya, , Chris Roelfsemab, Vittorio Brandoc, Claudia Giardinoc, Tiit Kutserd,
Stuart Phinnb, Peter J. Mumbye, Omar Barrilerof, Jean Laporteg, Benjamin Koetzh
a
Numerical Optics Ltd., Tiverton, UK
b
Remote Sensing Research Centre, School of Earth and Environmental Sciences, University of Queensland, Brisbane, Australia
c
National Research Council of Italy, Institute for Electromagnetic Sensing of the Environment, CNR-IREA, Via Bassini 15, 20133 Milano, Italy
d
Estonian Marine Institute, University of Tartu, Tallinn, Estonia
e
Marine Spatial Ecology Lab, School of Biological Sciences, University of Queensland, Brisbane, Australia
f
CS Systèmes d'Information, 5 rue Brindejonc des Moulinais, 31500 Toulouse, France
g
ARGANS Ltd., Plymouth Science Park, Plymouth, UK
h
European Space Agency, Earth Observation - Department Science, Applications and Climate, Frascati, Italy
A R T I C LE I N FO A B S T R A C T
Keywords: The Sentinel-2A and 2B Multi-Spectral Instrument (MSI) offers a specification of potential value toward a
Coral reefs number of objectives in remote sensing of coral reefs. Coral reefs represent a unique challenge for remote
Sentinel-2 sensing, being highly heterogeneous at metre scales and occurring at variable depths and water clarity regimes.
Landsat However, conservation initiatives, such as the United Nations Sustainable Development Goals, add urgency to
Bathymetry
the need for the large scale environmental monitoring information that remote sensing can provide. In the quest
Habitat mapping
Sun glint
to meet this challenge a range of satellite instruments have been leveraged, from Landsat to high spatial re-
solution sensors such as WorldView-2, toward objectives such as: mapping of bottom types, bathymetry, change
detection, and detection of coral bleaching events. Sentinel-2A and 2B offer a new paradigm of available in-
struments, with a 5-day revisit, 10 m multispectral spatial resolution and freely available data. Pre-launch si-
mulation analyses by several of the authors suggested Sentinel-2 would have good performance for reef appli-
cations, in this paper we follow up on this study by reviewing the potential based on the substantial archive of
actual data now available.
First we determine to what extent the World's reefs are covered by Sentinel-2, since the mission requirements
do not by default include all reefs. Secondly we review how a 5-day revisit translates to a usable acquisition rate
of clear images, given that cloud and surface glint are common confounding factors. The usable acquisition rate
is the real determinant of the objectives to which the data can be applied. Finally we apply current processing
algorithms to Sentinel-2 data of several sites over the Great Barrier Reef, including physics-based bathymetry
inversion and object-orientated benthic mapping. Landsat 8 OLI is most comparable current sensor to Sentinel-2
MSI, so direct comparisons and the possibilities for data synthesis are explored.
Our findings confirm that Sentinel-2 has excellent performance for meeting several essential coral reef sci-
entific and monitoring objectives. Taking into account cloud and sun glint, the usable acquisition rate for a large
proportion of reefs is likely to be around 20 clear images a year on average, giving a new potential for evaluation
of short time-scale disturbances and impacts. The spatial resolution of 10 m is a key threshold for delineating
benthic features of interest such as coral structures, and there is evidence from image and field data that
bleaching is detectable. Radiometrically Sentinel-2 data can support good results in physics-based methods, such
as bathymetric mapping, comparable to Landsat 8 and WorldView-2. In addition the large scale acquisition area,
provided by the 290 km wide swath, offers advantages over high spatial resolution imagery for mapping at
multi-reef scales.
Sentinel-2 data can be immediately leveraged with existing methods, to provide a new level of reef mon-
itoring information compared to that previously available by remote sensing. Combined with Landsat 8 and the
historical Landsat archive, the data collected today will be invaluable for decades or even centuries to come. In
this context, the main downside of the Sentinel-2 mission is that approximately 12% of the World's reefs cur-
rently lie outside the acquisition plan and are not imaged. Surprisingly, for a European initiative, coral reefs in
⁎
Corresponding author.
E-mail address: j.d.hedley@numopt.com (J.D. Hedley).
https://doi.org/10.1016/j.rse.2018.07.014
Received 13 October 2017; Received in revised form 30 April 2018; Accepted 10 July 2018
Available online 27 July 2018
0034-4257/ © 2018 Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/BY-NC-ND/4.0/).
J.D. Hedley et al. Remote Sensing of Environment 216 (2018) 598–614
European governed territories are among the worst served globally. These omissions, approximately only 1/
200th of the currently imaged area, limit the global scope which otherwise would be one of Sentinel-2's greatest
strengths.
1. Introduction spatial resolution offers substantial benefits, and this kind of data is
widely used for one-off or occasional reef mapping exercises at local
Coral reefs are currently under threat worldwide from numerous scales (Yamano, 2013). However, the expense of commercial imagery,
environmental and anthropogenic stresses (Hoegh-Guldberg et al., relative infrequency of acquisition, in particular of images clear of
2007; Pandolfi et al., 2003), and effective management and policy cloud and surface glint, make it unsuitable for systematic repeated
formation requires monitoring across all scales from local to global. monitoring or large scale change detection. With respect to scientific
While not as detailed as in-situ surveys, remote sensing provides a va- missions from space agencies, coral reefs are somewhere between a
luable complementary monitoring approach, in particular at large land and an ocean colour application and are not specifically priori-
scales (> 10's km2) where manual surveys would be prohibitively ex- tised, so the available satellite instruments are typically a compromise
pensive and impractical (Hedley et al., 2016). Key mapping objectives toward requirements. The Landsat series has seen some use in coral reef
are the abundance, distribution and health of living flora and fauna, context: freely available data facilitates the potential for global appli-
also known as the benthos, and include benthic type (e.g. coral and cations such as the Millennium Coral Reef Habitat Mapping Project
algae), benthic change detection and coral bleaching. These objectives (Andréfouët et al., 2004) and time series analysis (Palandro et al.,
require data within specific spatial and temporal ranges (Fig. 1). Geo- 2008). Landsat 8 in particular, with 12-bit digitisation and the addi-
physical parameters, such as geomorphic zonation and bathymetry, are tional blue band at 443 nm which offers high depth of penetration in
essential inputs for the growing number of ecosystem models used in a clear waters (Table 1), has good capability for techniques based on
management context (e.g. Hock et al., 2014). radiative transfer models (Giardino et al., 2016). The 16-day revisit
Over the last forty years the number of satellite instruments avail- ensures a reasonable chance of clear acquisitions several times a year,
able for coral reef applications has grown, starting with the early dependent on location. Capturing a 185 km wide swath means larger
Landsat and SPOT sensors (Smith et al., 1975; Bour, 1988) and in- areas can be processed without the challenges presented when mosai-
creasing rapidly in recent years with space agency platforms and cing images acquired under different conditions. While restricted by
commercial offerings such as the WorldView series of satellites. Sen- 30 m spatial resolution Landsat 8 data benefit from being the con-
tinel-2 satellites A and B, part of the Copernicus programme headed by tinuation of a long term dataset stretching back to the early 1980’s, a
the European Commission in partnership with the European Space requirement for analysis of long term processes (El-Askary et al., 2014).
Agency (ESA), are the most recent addition, and with the Multi Spectral Other short term missions with hyperspectral characteristics such as
Imager (MSI) instrument provide a 5-day revisit, 10 m pixels in visible Hyperion or HICO have also been tested (Kutser et al., 2006; Garcia
bands, and freely available data: specifications which cover a number of et al., 2014a) but without a long term mission commitment these ap-
reef monitoring requirements (Fig. 1). Further, Sentinel-2 is a European plications remain scientific investigations. Upcoming missions such as
mission and the physical area and biological, cultural and economic HyspIRI or EnMAP (Devred et al., 2013; Guanter et al., 2015) may re-
values of tropical reefs occurring in European territories are substantial. invigorate hyperspectral reef applications but currently the options for
France is the fourth nation when listed by reef area, after Australia, applied benthic mapping of coral reefs by satellite instruments remain
Indonesia and the Philippines (UNEP-WCMC, 2010). The United dominated high spatial resolution commercial offerings and Landsat 8.
Kingdom is 11th in the same list and has over half as many reefs as the Sentinel 2 differs significantly to the previously discussed instru-
United States. The European Union is committed to management and ments in both spatial and temporal resolution. The spatial resolution of
conservation of coral reefs through being a signatory to a number of
international initiatives that specifically mention coral reefs: The
United Nations Environment Assembly, The Johannesburg Declaration,
Convention on Biological Diversity, and The UNESCO World Heritage
Convention. In particular, a number of targets under the United Nations
Sustainable Development Goal 14 (SDG 14) (oceans, seas and marine
resources, https://sustainabledevelopment.un.org/sdg14) require re-
gion-scale information on the extent, health, resilience, and sustain-
ability of coastal ecosystems, of which coral reefs are a key priority. The
impetus for European initiatives such as Copernicus to put coral reefs
central to their efforts is therefore unquestionable and the opportunity
to leverage Sentinel-2 toward this aim is of great interest.
The current options for satellite based reef mapping are dominated
by high spatial resolution multispectral instruments (pixels ≤ 5 m) and
moderate resolution (pixels 10–50 m) e.g. Landsat 8 (30 m multi-
spectral pixels). This is now changing with instruments such as
Sentinel-2 and other initiatives such as the next generation of CubeSat
imaging systems. While hyperspectral data is considered advantageous
(Mumby et al., 1997; Hochberg and Atkinson, 2003), the drive with
commercial instruments has primarily been toward very high spatial
resolution multispectral data (e.g. Worldview-4, Pleiades, pixel
size ≤ 2 m, Planet Labs constellation, 3–5 m). Coral reefs are highly
heterogeneous at scales of a few meters or less (Fig. 1), differing benthic
types such as corals, seagrasses or macroalgae typically occur at dif- Fig. 1. The relationship between the spatial and temporal characteristics of
ferent spatial scales and so the optimal minimum spatial resolution may Sentinel-2 acquisitions and the requirements for coral reef remote sensing ob-
differ dependent on the objective and site (Phinn et al., 2010). High jectives.
599
Table 1 MODIS cloud climatology data set (Mercury et al., 2012). Finally, two
Band specifications of Sentinel-2A MSI and Landsat 8 OLI. Information from: frequently-used and current mapping techniques, physics-based
https://earth.esa.int/web/sentinel/user-guides/sentinel-2-msi/resolutions/ bathymetry estimation and object-orientated bottom classification, are
radiometric and https://landsat.gsfc.nasa.gov/landsat-8/landsat-8-bands. applied to Sentinel 2 images of sites on the GBR. The bathymetric
Sentinel-2A MSI Landsat 8 OLI mapping is compared directly to Landsat 8 in a parallel analysis. For the
object-orientated benthic mapping both the process and results are
Band Pixel size Wavelength range Band Pixel size Wavelength range compared to previous mapping efforts using Landsat 8 and other sen-
(m) (nm) (m) (nm)
sors. Some of the methods applied here are the same or similar to al-
gorithms that will be provided in ESA's SNAP toolbox as an output of
01 60 430–457 1 30 433–453 the Sen2Coral project (http://step.esa.int/main/toolboxes/snap/). In
02 10 448–546 2 30 450–515 summary we present four analyses:
03 10 538–583 3 30 525–600
04 10 646–684 4 30 630–680
05 20 694–713 1) The global status of coverage of coral reefs by Sentinel-2 acquisi-
06 20 731–749 tions.
07 20 769–797 2) How a 5-day revisit translates into ‘usable acquisition frequency’.
08 10 763–908
3) Performance of physics-based bathymetry method for Sentinel-2 vs.
8A 20 848–881 5 30 845–885
09 60 932–958
Landsat 8.
10 60 1336–1411 9 30 1360–1390 4) Practicalities and results of object-orientated benthic classification
11 20 1542–1685 6 30 1560–1660 using Sentinel-2 data.
12 20 2081–2323 7 30 2100–2300
8 15 500–680
The paper concludes with a summary of the benefits, and caveats, of
the application of Sentinel-2 MSI data toward coral reef management
10 m in four visible and near infra-red (NIR) bands lies directly between objectives.
the ~2 m pixel high resolution sensors and the Landsat 8 multispectral
bands at 30 m (Table 1). This spatial regime has not been widely used 2. Methods
on reefs, although similar resolutions are available in the SPOT series
(10 m and 6 m), and Rapid-Eye (5 m), uptake of these data for reef 2.1. Global coverage estimation
applications is far less than that of the commercial high resolution of-
ferings. The 5-day revisit of the Sentinel-2A and 2B instrument pair, The original Sentinel-2 mission requirements guaranteed only land
combined with freely available data, is unprecedented. This offers many masses > 100 km2 in area, islands in the European Union and
potential benefits: from simply increased likelihood of finding imagery areas < 20 km from the coast (SUHET 2013). However in practice the
clear of cloud and sun glint; to high temporal resolution time series for acquisition area is larger and has been modified several times, for ex-
change detection; or to new multiple-image analysis techniques. The ample the whole of the GBR is now routinely imaged. ESA does not
imaging swath of Sentinel-2 at 290 km is wider than that of Landsat provide an official on-going acquisition plan, therefore to assess coral
8185 km, which carries the advantage of a larger simultaneous image reef coverage it was assumed that the combined weekly Sentinel-2A
area but with caveat of including a wider range of view angles (Roy acquisition plans published by ESA in the 12 months prior to September
et al., 2017). On the downside, the current Sentinel-2 mission re- 2017 were representative. The acquisition area was reconciled against
quirements do not ensure global coverage of all reefs (nominally in- the UNEP-WCMC Global Distribution of Coral Reefs dataset (United
cludes coastal areas within 20 km of landmasses > 100 km2) and this Nations Environment Programme - World Conservation Monitoring
potentially undermines the advantage of free data at large scales. Ad- Centre, UNEP-WCMC, 2010). This dataset was compiled from a number
ditionally, unlike Landsat 8, the useful bands of Sentinel 2 are at mixed of sources by UNEP-WCMC and includes data from the Millennium
resolutions, 60 m for the equivalent of Landsat 8’s deep blue band Coral Reef Mapping Project and the World Atlas of Coral Reefs
(Table 1). Pre-launch sensitivity analysis on simulated Sentinel 2 data (Spalding et al., 2001). The area of individual reefs lying inside or
(Hedley et al., 2012) indicated that good performance could be ex- outside the acquisition area was quantified for each of the 85 sover-
pected, but that mixed spatial resolutions may introduce some artefacts. eignties listed in the UNEP-WCMC dataset. To quantify the status of
In simulations spectral reflectances at 10 m resolution over dark benthic global coverage an analysis based on the Sentinel-2 tiling grid was
features such as corals had corresponding Band 1 values that were too conducted (SUHET, 2015), in which a list of all 110 km × 110 km
high, being the average over a larger area (60 m), subsequently in a Sentinel-2 tiles that contained reefs was made, and the number of those
bathymetry analyses these pixels were erroneously estimated as deep tiles that were wholly or partially missed from the acquisitions were
water (Hedley et al., 2012). The similarity of Sentinel 2 band specifi- noted. A summary of coverage is provided in the results and discussion,
cation to Landsat 8 (Table 1) does however offer good possibilities for a report on the full analysis is supplied as supplementary information.
data synthesis, and may facilitate interpretation of Sentinel 2 data
within the context of the Landsat historical archive if the relative per- 2.2. Usable acquisition frequency (sun glint and cloud)
formance of the sensors can be characterised.
This paper is a follow-up of the pre-launch sensitivity analysis of Both cloud cover and surface glint are serious confounding factors
Hedley et al. (2012). After 18 months of operation of Sentinel-2A the for coral reef applications of remotely sensed imagery. Understanding
imagery archive is now sufficient to review and assess the current the practical value of a 5-day revisit time requires estimating the like-
capability and future potential for coral reef applications. The first lihood that each image will be clear of cloud and sun-glint. The ‘usable
question addressed is to what extent are the World's reefs covered by acquisition frequency’ is what is relevant for time series or change
Sentinel 2 acquisitions, and where are the significant omissions, if any. detection algorithms.
Secondly, while the 5-day revisit is an impressive specification, the With respect to cloud, first a local scale analysis was conducted
question of relevance to time series and change detection is how many using images of Heron Reef (23.442° S, 151.915° E) on the GBR. A total
usable images this translates to, given the possibility of cloud and water of 69 Sentinel-2A and Landsat 8 images acquired over a 12 month
surface glint (sun glint). These factors are assessed from a series of period were visually assessed as being ‘clear’, ‘minor cloud’ or ‘unu-
images on the Great Barrier Reef (GBR) and with reference to a global sable’, where ‘unusable’ approximately corresponded to > 20% cloud
cover over reef areas. Results were compared to a global seasonal cloud
600
cover climatology derived from 10 years of MODIS data (Mercury et al., Surface glint is dependent on solar zenith angle, view zenith angle,
2012). Secondly, the Mercury et al. (2012) dataset was applied in global relative azimuth and sea surface state (Kay et al., 2009). The brightest
context by determining the percentage of global reef area under dif- glint occurs when the view and solar zenith angles, and relative azi-
ferent regimes of cloud free acquisitions, using the UNEP-WCMC global muth, are such that the view is close to the forward direct reflectance
reef map. I.e. to show how much of the World's reefs are cloud free 50% direction. Increased surface roughness widens the glint ‘spot’ and in-
of the time, etc. crease the chance of glint outside the direct reflectance direction.
While correction algorithms for glint exist (Kay et al., 2009) images Sentinel-2 incident view angles vary with position in the swath with a
with minimal glint are preferred since strong correction is imperfect maximum of around 12° at the edges (Roy et al., 2017). Some locations,
and introduces noise. Glint correction of Sentinel-2 data is potentially such as part of Heron Reef, are located at swath edges in the orbit
more challenging than with higher spatial resolution imagery because overlap area and so are imaged on two overpasses and benefit from
10 m pixels may contain a combination of glint at different scales: 1) a double acquisitions. However these locations have the largest view
component that varies pixel-to-pixel due to wave slopes spaced > 10 m zenith angles. Typically in one overpass the solar-view geometry will be
apart; 2) a more constant sub-pixel component arising from surface in a forward reflectance direction (prone to glint) and in the other a
waves smaller than 10 m. Regression based corrections (Hedley et al., reverse direction (less prone to glint). Glint is also substantially de-
2005) fail to capture sub-pixel glint whereas statistical correction pendent on solar zenith angle which varies with season.
methods (Kay et al., 2009) cannot handle pixel-to-pixel variation. The To estimate the frequency and conditions of glint occurrence on
frequent acquisitions of Sentinel-2 imply the best solution is simply to typical GBR sites cloud screened Sentinel-2A images of Heron Reef and
select images with minimal glint, so the question of interest is how Lizard Island (14.660° S, 145.459° E) were selected from an 18 month
frequently and under what conditions such images occur. period from January 2016 to August 2017. Unlike Heron Reef, Lizard
Fig. 2. Various sites on the Great Barrier Reef as referred to in this paper: (a) Lizard Island, (b) northern part of Hall-Thompson Reef, (c) SSCMR (Southern Section
Cairns GBR Management Region), (d) Adelaide Reef and (e) Heron and Wistari Reefs. Images are RGB composites of bands 4, 3, and 2 of Sentinel-2A images.
601
Island is located closer to the centre of a swath with typical view zenith 2.4. Bathymetry and benthic mapping: validation data
angles of ~3°. Glint was estimated as the top-of-atmosphere reflectance
at 660 nm (band 4 in Sentinel-2, Table 1) over deep water areas (in- The bathymetry dataset for Lizard Island is described in detail in
terpretation of the atmospheric contribution is discussed in the Results Hamylton et al. (2015) and consists of approximately 30,000 echo
section). This resulted in a total of 31 data points of solar-view geo- sound data points collected in December 2011 and tide corrected to a
metry and top of atmosphere reflectance. These data are discussed later datum of mean sea level (MSL). The bathymetry data for SSCMR site
to provide insight on how surface glint may limit the practical usable was collected in 2014 using a Garmin GPS echo sounder 550C and si-
revisit time. milarly corrected to MSL, it consisted of approximately 11,000 data
points.
Benthic cover data at the SSCMR site was derived from georefer-
2.3. Bathymetry and benthic mapping: study sites and imagery enced benthic field photographs collected on dive and snorkel transects
in the study area in January and May 2017. A detailed methods de-
For assessing the capability for bathymetry and benthic mapping scription is provided in (Phinn et al., 2012; Roelfsema and Phinn,
with Sentinel-2 and Landsat 8 images were analysed from two sites on 2010). Photos with 1 m2 footprint were taken of benthos at 2–3 m in-
the Great Barrier Reef (GBR), Australia. The first site was used for tervals along transects while a GPS at the surface was tracking the
bathymetry only: Lizard Island (14.660° S, 145.459° E) was chosen position of the diver or snorkeler. Photos were assigned coordinates by
since it has a long record of scientific study and characterisation, in- synchronisation of the GPS and camera through time. Diving transects
cluding a substantial bathymetry dataset (Hamylton et al., 2015). The were approximately 500 m in length, conducted at depths > 5 m on the
Lizard Island site is fairly small, the reef lies within an area of reef slope and were directed to represent the main aspects of the reefs.
7 km × 7 km, the Sentinel 2 image subsets analysed here being Snorkel transects were 500–1000 m in length at 1–3 m depth and
1044 × 1062 pixels in the 10 m bands. The entire archive of Sentinel- transects were directed to cross the different geomorphic zones on the
2A and Landsat 8 to the start of 2015 was reviewed to find the best reef top. Benthic composition for each individual photo was derived
available images in terms of cloud cover, surface glint and atmospheric through automated photo analysis using CORALnet, with 3% of photos
clarity. Two Sentinel-2A images and two Landsat 8 images were se- manually processed for calibration and validation of the machine
lected with dates of 2 July 2016, 30 August 2016, 24 July 2015 and 11 learning algorithm (Beijbom et al., 2015; González-Rivero et al., 2016).
August 2016 respectively (Fig. 2a). The main aim of the analysis at this
site was to compare bathymetry results between Sentinel-2A and 2.5. Atmospheric and glint correction
Landsat 8 using the same methods, and to assess repeatability by using
two images from each instrument. The physics-based inversion method for bathymetry described
For the second site the primary focus was benthic mapping. This site below required bottom of atmosphere spectral reflectance Rrs(λ) as
was larger, approximately 25 km × 28 km, and included nine reefs input, so a simple atmospheric correction (Hamylton et al., 2015) was
from further south on the GBR, in the area bounded by Ellison Reef, applied to the Landsat 8 and Sentinel 2 scenes. The correction is based
Gilbey Reef and Farquharson Reef (centred around 17.692° S 146.485° on look-up tables for atmospheric reflectance and transmission gener-
E, Fig. 2b). This site was chosen as it coincided with recent field data of ated by libRadtran (Emde et al., 2016) with a maritime 99% relative
bathymetry and benthic composition collected in the context of the humidity aerosol model as described by Antoine and Morel (1999). The
Great Barrier Reef Habitat Mapping Project (Roelfsema et al., 2017, look-up tables are parameterised on solar-view geometry with the only
2018, https://www.rsrc.org.au/gbrcommon). Henceforth this site is two free parameters being aerosol optical thickness: τ(550), 0 to 0.83;
referred to as ‘SSCMR’ (Southern Section Cairns GBR Management and wind speed: u10, 0 to 10 ms−1. The wind speed translates to an
Region). The main purpose of the analysis at the SSCMR site was to assumed spatially homogenous glint via the Cox-Munk equations under
perform object orientated benthic mapping using Sentinel-2 data, to certain solar-view geometries (Kay et al., 2009), therefore the main
gain insight to both the required process and results in relation to the effect of these two parameters is to contribute a component to the at-
same methods using other sensors on GBR (Roelfsema et al., 2013, mospheric reflectance which includes a spatially homogenous compo-
2018). It was expected that the results would also be relevant for nent of sea-surface reflectance. The aerosol component is typically
simpler mapping methods such as per-pixel classifications based on spectrally blue and the glint component is spectrally flat. Aerosol
digital numbers (Andréfouët et al., 2003), which remain popular. thickness and wind speed were assumed constant over the image area
Bathymetric analysis is an intermediate step in the object-orientated and were estimated by taking one or more clear deep water areas and
benthic mapping and this was performed for both Sentinel-2 and determining which values of τ(550) and u10 enabled the bathymetric
Landsat 8 data at the SSCMR site. model described below to correctly estimate deep water. The atmo-
For all analyses the first five bands were used from both sensors, spheric correction is therefore not independent of the bathymetry
covering wavelengths from 433 nm to 713 nm for Sentinel 2 and analysis, the approach is closer to that of Kutser et al. (2006) where
430 nm to 880 nm with Landsat 8 (Table 1). In some cases additional atmospheric parameters were included in the inversion, except here the
bands were used for glint correction (see below). For Sentinel 2 the atmospheric parameters were constrained to be constant over a scene.
bands at 60 m (B1) and 20 m (B5) resolution were resampled to 10 m As discussed in the last section, the spatial resolution of Sentinel 2
without interpolation. The panchromatic band at 15 m resolution from images and the solar-view geometry at acquisition is such that pixel-to-
Landsat 8 (band 8, Table 1) was not used due to its spectral width. pixel surface glint is often visible in the 10 m bands. Although the
Image geo-locations were used as supplied in the level 1 data. clearest possible images were selected for this study close inspection
Examination of image time series indicated level 1C Sentinel-2 images indicated that all three Sentinel 2A scenes would benefit from per-pixel
were typically geo-located within 2 pixels of each other (20 m) which is glint correction. The correction was applied according to Hedley et al.
within the stated quality requirements for absolute geo-location (ESA, (2005), using a far red band (NIR or SWIR) to estimate and correct the
2017). Improvement of geo-location through the production of the glint in visible wavelength bands. Due to the differing band resolutions
Global Reference Image (GRI) is an ongoing ESA activity (Clerc, 2017) three steps were required: Band 1 (60 m) was corrected using Band 9,
and so was not an assessment objective within the scope of this paper. Bands 2, 3, and 4 (10 m) were corrected using Band 8, and Band 5
The potential consequences of geo-location error are considered in the (20 m) was corrected using Band 7. The correction procedure was ap-
interpretation of the results. plied to Band 1 just for consistency; in practice it had very little effect
because pixel-to-pixel glint was absent at the 60 m scale. The Landsat 8
scenes did not have the per-pixel glint correction applied because pixel-
602
to-pixel glint was not visible. Note the glint correction of the Sentinel- relative spectral response (RSR) functions for application to imagery.
2A scenes was applied before the atmospheric correction described This approach is in theory more accurate than convolving to band RSRs
above. Note also the glint correction methodology was the same as will first and performing calculations in ‘band space’, since the bands can be
be made available in the ESA's SNAP toolbox an output of the Sen2Coral wide and may encompass spectral features that combine non-linearly
project. (e.g. Sentinel-2A Band 2 is 470–524 nm, at full width half maximum).
Pahlevan et al. (2014) published a set of vicarious calibration ad- However, in the development of this work it was deduced that there
justments for Landsat 8 which can be beneficial for physics-based was an issue with the published Sentinel-2A RSR for Band 2. In short,
shallow water mapping (Giardino et al., 2016). However, when tested initially a box function was used and gave good results, when the
on the imagery and methods used here, these coefficients were not of published RSR was implemented the results became substantially
clear benefit. A similar analysis for Sentinel-2 has only just become worse. Also, the published RSR for Sentinel-2A Band 2 had an unusual
available at the time of manuscript submission (Pahlevan et al., 2017), almost triangular shape whereas for Sentinel-2B the Band 2 RSR ap-
so the analysis presented here does not include the application of any peared more typical and closer to a box function. Enquiries to the
vicarious calibration coefficients for Landsat 8 or Sentinel-2. Sentinel technical support team confirmed that there was a known issue
with the published Sentinel-2A Band 2 RSR. This has recently been
2.6. Bathymetry mapping stated in the September 2017 Data Quality Report (Clerc, 2017) with
the advice to use the Sentinel-2B RSR, and this is what was done. It
Bathymetric maps were produced using the physics based inversion solved the issue and is the basis of all the results presented here. The
method described in Hedley et al. (2009), the same approach as used in most recent update is that since the initial submission of this paper an
the sensitivity analysis on simulated Sentinel-2 data (Hedley et al., updated Sentinel-2A RSR has been published by ESA and is indeed
2012). The method is also similar to the SWAM (Shallow Water semi- much more similar to the Sentinel-2B RSR in Bands 1 and 2 (March
Analytical Model) implementation that will be released as part of ESA's 2018 Data Quality Report, Clerc, 2018).
SNAP toolbox. The basis of the method is a forward model of above
water spectral remote sensing reflectance Rrs(λ) based on the equations 2.8. Benthic mapping
of Lee et al. (1998) of the form:
For the Sentinel 2 benthic mapping analysis at the SSCMR site an
Rrs (λ ) ≈ f (P , G, X , H , e1 , e2 , m, λ ) (1)
image dated 22 September 2015 was used, being completely cloud free
where P, G, X, and H respectively represent the water column phyto- for the nine reefs within the area. The image was used to produce a
plankton, coloured dissolved organic matter (CDOM), particulate bathymetry map, by the method described above, and also a water
backscatter and depth. The bottom reflectance is a linear mix of two column corrected bottom reflectance image, i.e. ρ(λ) in the five image
endmember reflectance spectra drawn from a set of size ne, and indexed bands, for input to the object-orientated classifier. The estimate of
by e1 and e2 such that 0 ≤ e1, e2 < ne. The mix fraction of e1 vs. e2 is bottom reflectance in each pixel was given by re-arranging the forward
given by m which ranges from 0 to 1. Here the same six endmember model equation (Eq. 3 in Hedley et al., 2009) to give ρ(λ) as a function
spectra were used as in the previous sensitivity analysis, and included of Rrs(λ) once P, G, X and H have been estimated from the inversion
reflectances of coral, sand, algae and seagrass (Fig. 2 from Hedley et al., procedure. While the forward model contains an explicit estimate of
2012, excluding bleached coral). At each pixel in the image a look-up- ρ(λ) from the linear mix of the endmembers (parameters e1, e2 and m)
table inversion procedure (ALUT, Hedley et al., 2009) was applied to this is not used directly as a bottom reflectance output because being a
find the input values to the model which produced the closest spectral linear mix of a finite set of endmembers it is dimensionally constrained.
match to image reflectance. The look-up-table approach requires the Object-based image analysis (OBIA) through Trimble eCognition 9.3
parameters to lie in bounded ranges, here the limits were: phyto- software was used to map geomorphic zones and benthic cover type by
plankton (P) ranged from 0 to ~2 mg m−3; CDOM (G), absorption at adaptation of a protocol developed previously and described in
440 nm, a(440), ranged from 0 to 0.2 m−1; Particulate backscatter at Roelfsema et al. (2013). The first part of the OBIA consisted of seg-
440 nm (X) bbp(440) ranged from 0 to 0.1 m−1; and depth estimations mentation of the image into groups of pixels with similar character-
were limited as being between 0 and 30 m. istics, e.g. colour or texture, or a physical property such as water depth.
To provide per-pixel bathymetry uncertainty estimates, each pixel This was followed by labelling of segments using a membership rule set
was inverted 20 times with a random spectral error term added to the that determined which class each segment was identified as (Blaschke,
pixel reflectance derived from the covariance matrix over a deep water 2010). Some OBIA parameters required adjustment to correspond to the
area (Hedley et al., 2010; Garcia et al., 2014b). This error term, the 10 m pixel size: this is part of the protocol as for example in Roelfsema
environmental noise equivalent radiance, NEΔRrs(λ), (Brando et al., et al. (2013, 2018) the OBIA was applied on high spatial resolution
2009) notionally includes all sources of pixel-to-pixel variation from the imagery varying from 2 to 4 m pixels, whereas for Roelfsema et al.
water surface upwards. The forward model had no terms to accom- (2018) pan sharpened Landsat 8 OLI imagery was used at 15 m pixel
modate such variations so they are effectively noise. Typically this noise scale.
term is dominated by variable water surface reflectance (even after The following attributes were incorporated into OBIA: Sentinel-2
glint correction) but may also contain atmospheric fluctuations and derived bathymetry, slope (calculated from bathymetry) and Sentinel-2
instrument noise. The best estimate for the bathymetry at each pixel reflectances derived at the water surface and bottom. For each geo-
was taken as the mean over the 20 inversions and the 90% confidence morphic zone category, a rule set was developed to assign a label to
intervals for bathymetry were taken as the range after discarding the each segment based on a set of biophysical attributes such as water
highest and the lowest estimate. Tide correction for the time of image depth (e.g. MSL depth > 3 m implies reef top), colour (e.g. brightest
acquisition was applied using the OTPS2 tide model (Egbert and colour implies sand), slope derived from water depth (> 10° implies
Erofeeva, 2002) assuming that a single tide estimation near the centre slope region) and neighbourhood relationships, e.g. reef slope is ad-
of the scene was representative over the entire area. All results are jacent to reef crest. Geomorphic zones were mapped down to 15 m
presented relative to mean sea level (MSL). depth using the at-surface reflectance image, water depth and slope.
Dominant benthic cover type was mapped for the reef top, here defined
2.7. Sentinel-2A relative spectral response function as area above −3 m (MSL) which includes reef crest, outer and inner
reef flat and shallow lagoon. Wherefore, membership rules were created
The atmospheric correction values and reflectance model (Eq. 1) based on expert knowledge of reefs, field data, bottom reflectance
were evaluated at 2 nm intervals and then convolved by the sensor band image and rules from previous OBIA studies (Joyce et al., 2002; Phinn
603
et al., 2012; Roelfsema et al., 2013, 2018). Membership rules to assign Table 2
dominant benthic cover type labels to segments were based on the Ten least-served sovereignties ordered by physical area of reef that has not been
brightness of the segments, band ratios, segment location within each of imaged by Sentinel-2A once in the 12 months to September 2017.
the geomorphic zones. These rules varied between geomorphic zones Sovereignty Total reef area % of reef area Reef area not
dependent upon the type of relationship and/or the threshold value for imaged imaged
a dominant benthic cover type. Dominant benthic cover type labels
1 France 8296 km2 63% 3051 km2
assigned included coral/algae, rock, rubble and sand.
2 Micronesia 3192 km2 7% 2962 km2
3 United States 4738 km2 46% 2562 km2
4 United Kingdom 2937 km2 27% 2152 km2
3. Results and discussion 5 Kiribati 1973 km2 5% 1873 km2
6 Seychelles 1560 km2 17% 1289 km2
3.1. Global coral reef coverage by Sentinel-2 7 Tuvalu 879 km2 4% 847 km2
8 India 2036 km2 60% 812 km2
9 Fiji 3404 km2 86% 483 km2
In total 2293 tiles are required to cover the World's reefs (Fig. 3a) of 10 New Zealand 367 km2 14% 316 km2
which 275 were wholly omitted in the acquisitions in the 12 month
period to September 2017. A small number of reefs were missed in tiles
that are partially imaged but this is a comparatively minor factor. Table 2 lists the countries that have the largest coral reef areas that
Therefore ~12% of the tiles required for global reef coverage were are not currently imaged. Interestingly, although Sentinel-2 is a Eur-
missed, which is only ~1/200th of the total effort of 56,686 tiles that opean mission, and France and the United Kingdom have substantial
were imaged at least once in that period. In terms of reef area missed combined reef area (~11,000 km2), European territories are among the
this also translates to 12% of the total global reef area, according to the worst served in terms of coral reef coverage (Table 2). Globally, around
UNEP-WCMC dataset. Here ‘reef area’ means the polygonal areas in the 30% of the coral reef area missed by Sentinel-2 is in European overseas
UNEP-WCMC dataset, which is primarily shallow reef areas that could territories (~5200 km2). Other poorly served territories are Micronesia,
be mapped by remote sensing (UNEP-WCMC et al., 2010). In many Kiribati, and Tuvalu with < 7% of their reef area imaged. The Phoenix
places coverage was excellent, since coral reefs are frequently asso- Island group in Kiribati has the status of the largest UNESCO World
ciated with islands sufficiently large to be included in the Sentinel-2 Heritage site in the world, but its reefs are entirely missed. Many of the
land acquisition plan. Many countries had 100%, or close to, coverage areas which are missed are those in remote and expensive to access
of their reef areas, including the three countries with the largest reef locations, the places where satellite remote sensing offers the greatest
area: Australia, Indonesia and the Philippines (with approximately potential (Hedley et al., 2016).
32,000 km2, 20,000 km2 and 12,000 km2 reef area respectively). The By way of specific example Fig. 3b to d illustrate the coverage of the
majority of missed tiles are over small islands in the Pacific (Fig. 3a).
Fig. 3. (a) Global coverage of reefs by Sentinel-2A over the 12 month period to September 2017: red indicates tiles containing reefs, dark blue is acquired area. (b–d)
Example detailed coverage of reefs in UK territories: black dots are reefs, red indicates tiles including reefs in UK territories, dark blue is acquired area. (b) shows
omission of Bermuda, (c) Chagos Islands, (d) Pitcairn Islands. (For interpretation of the references to colour in this figure legend, the reader is referred to the web
version of this article.)
604
reefs of the United Kingdom (UK), see the supplementary information reflectances, and in particular the upper bound for wind speed of
for similar plots for all 85 sovereignties listed in UNEP-WCMC dataset. 5 ms−1 and aerosol optical thickness of 0.3 (Fig. 5). The model and data
While every reef area probably has special status by some criteria, it is are in agreement that minimal glint is virtually guaranteed for solar
worth noting the areas missed in Fig. 3 include Bermuda which has the zenith angles above 40°, whereas for angles below 30° glint may or may
northern-most coral reefs in the Atlantic (Spalding et al., 2001); the not be present, dependent on sea state and site location in the swath.
Chagos Archipelago (one of the world's largest marine reserves) and the The images of Heron Reef on the west edge of the swath where relative
Pitcairn Islands, a biodiversity hotspot at which 79 new species were azimuth is in the back-scatter direction were generally glint free. For
recently discovered (Friedlander et al., 2014). While interest in the those on the east edge the relative azimuth varied between 30° and 60°
potential use of Sentinel-2 for coral reefs is likely to be high, this will be from direct forward reflectance and glint was in some cases very high
mitigated if areas of special interest are omitted from the acquisition (Fig. 5). Glint at Lizard Island, close to the swath centre and with ap-
plan. proximately nadir view (~3°), was dependent on solar zenith angle and
relative azimuth which co-varied (Fig. 5). Relative azimuth was in some
3.2. Usable image acquisition rate (cloud and sunglint) cases exactly in the forward reflectance direction and this gave rise to
the highest glint data points. Both images used for the bathymetry test
On average 35% of 69 acquisitions over Heron Reef in the GBR were had solar zenith angles > 30° and relative azimuths > 60° from direct
clear of cloud but as expected there was seasonal variation, with no reflectance direction. Note the Lizard Island outlier in Fig. 5c had
clear images in March to April period (Fig. 4a). Other months varied substantial atmospheric haze visible over the land so this was an aerosol
between 30% and 60% cloud free images but there is likely some sta- effect.
tistical noise in the variation due to the small sample size (n = 69). The Overall, with respect the sites on GBR, regardless of position in the
MODIS analysis of Mercury et al. (2012) matched the average cloud swath, glint was minimised for solar zenith angles > 40°, i.e. the local
cover, giving 25% to 40% cloud cover with the same seasonal tendency winter with low sun positions. This is an additional seasonal con-
but less pronounced (Fig. 4a). If minor cloud cover (approx. < 20% tribution to the usable image acquisition rate, but at least corresponds
cloud over reef areas) is acceptable then on average 51% of acquisitions to a relatively cloud free period at these sites (Fig. 4a). While some glint
were usable. free images were found for all positions in the swath the west edge is
Globally, the Mercury et al. (2012) dataset implies that the occur- certainly favourable (Fig. 5). Using a glint threshold of 0.015 as ‘ac-
rence of cloud cover is at least 50% for almost all reefs (~95% by area), ceptable’ estimates the usable images due to glint as anywhere between
when averaged over a year (Fig. 4b). The most common annual pro- 100% (Heron west) to 42% (five out of twelve, Lizard). Combined with
portion of cloud free acquisition for all reef scenes globally was the cloud free estimates from above gives a range of usable image ac-
20–30%, so Heron Reef seems quite typical in this respect. Therefore quisition rates from 10-day revisit (minimal cloud, no glint) giving >
due to cloud alone, a 5-day revisit more realistically will produce a 10 30 clear images a year, to a worst case of three images a year (cloudy
to 15 day on-average usable image acquisition rate with respect to site, frequent glint). This worst case figure is likely overly pessimistic
Heron Reef. This figure would also seem a reasonable estimate at global due to the large-scale perspective of the Mercury et al. (2012) cloud
scales. Although dependent on season and location, > 50% cloud-free climatology, and would only apply in a very limited number of loca-
images annually is unlikely to be achieved anywhere (effectively a 10- tions. Usable image rate will be seasonally variable and this may affect
day revisit, Fig. 4b), but equally unlikely is < 1 in 10 images cloud free specific objectives, on GBR bleaching is most likely in the summer
(approximately six cloud free acquisitions a year). period from February to April which unfortunately coincides with
Solar zenith angle had a strong relation with surface glint in the periods of increased cloud and tendency for glint (Figs. 4, 5). While
Sentinel-2A data over Heron Reef and Lizard Island, assuming glint is these estimates are very approximate they at least contextualise what a
the major contributor to top of atmosphere reflectance at 660 nm over 5-day revisit may translate to locally. A ball-park figure of 20 clear
deep water (Fig. 5). To help interpret these results, the atmospheric images a year is certainly a reasonable a priori expectation.
model used in the atmospheric correction (Section 2.5) was used to
model top of atmosphere nadir reflectance over a ‘black ocean’ (in- 3.3. Bathymetry mapping
cludes glint but no subsurface reflectance). Reflectances were generated
at 600 nm, as function of solar zenith angle and for two extreme All four bathymetric maps of Lizard Island, produced with either
treatments of wind speed, u10, of 0 and 5 ms−1 paired with aerosol Sentinel-2A data or Landsat 8, were qualitatively similar at the scale of
optical thicknesses τ(550) of 0 and 0.3 respectively (Fig. 5). The sa- the imagery (i.e. kilometre scales). Fig. 6b illustrates one example: at
tellite data corresponded very well to the range of the modelled this scale the visible difference in the other plots (not shown) was
Fig. 4. (a) Frequency of cloud free Sentinel-

2A images over Heron Reef in a 12 month
period. Cloud free was judged subjectively
as images that by visual inspection did not
contain significant cloud. (b) Histogram of
amount of time reefs are cloud free ac-
cording to MODIS climatology analysis
Mercury et al. (2012), in terms of relative
reef area.
605
Fig. 5. Sun glint, estimated as top of atmosphere reflectance at 660 nm, as a function of solar zenith angle, assessed from (a, b) deep water areas in Sentinel-2A
imagery near Heron Reef and Lizard Island, GBR. (c) Dotted lines show modelled reflectance according to the Cox-Munk equations for (lower line) a wind speed of
0 ms−1 (flat surface) and aerosol optical thicknesses (AOT, at 550 nm) of zero and (upper line) wind speed 5 ms−1 and AOT of 0.3, both with a marine 99% relative
humidity aerosol model. Grey area is the possible reflectance between the model extremes. Data points come from three positions in a swath, west edge and east edge
for Heron Reef and approximately central for Lizard Island. Example images are (a) low glint, swath west edge, (b) high glint, swath east edge.
primarily in the maximum depth attributed to deep water areas. In With respect to systematic errors, the Sentinel-2 results showed a
areas where the bottom is not visible the attributed depth is very sen- slight tendency to over-estimate depth by a metre, at and beyond
sitive to the slightest variation in image reflectances. Regression plots depths of 10 m (Fig. 7a, b), but more noticeably both Landsat 8 results
against the in-situ data (Fig. 7) indicate that depths were well estimated exhibited a slight curved deviation from the 1:1 line (Fig. 7c, d). This
to around 15 m, and an r2 value of 0.89 achieved with one Sentinel 2 deviation could be indicative of a ‘calibration’ issue between the water
image (Fig. 7a) is equal to that obtained in a WorldView 2 analysis with column model (Eq. (1)) and the atmospherically corrected data, in that
same in-situ data (Hamylton et al., 2015). Results are possibly even the image data in some bands may not match the spectra produced from
better than those predicted by the simulated image analysis of Hedley the forward model. The inversion used bands 1 to 5 of Landsat 8
et al. (2012). In that paper the equivalent analysis gave an r2 value of (Table 1). Different bands penetrate the water to different depths, so if
0.93 but visually the spread of points in Fig. 7a appears closer to the 1:1 the image data and model are not radiometrically aligned in all bands
line. then this can manifest as a curve in the bathymetry data. Introducing
Fig. 6. (a) Location of in-situ bathymetry ground truth points at Lizard Island, (b) example bathymetric map produced from Sentinel-2A image of 21 July 2016.
606
Fig. 7. Bathymetric results at Lizard Island for the two Sentinel-2A images and the two Landsat 8 images. Red lines show 90% confidence intervals from the
uncertainty analysis averaged in 1 m bins. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
the Pahlevan Landsat 8 aquatic calibration coefficients, which give a distributed around Lizard Island (Fig. 6a), geo-location errors probably
maximum adjustment of just 3.4% in Band 1 (Pahlevan et al., 2014), did not introduce systematic depth errors. The higher spatial resolution
did not improve the result but rather introduced a systematic under- of Sentinel-2 visibly affected the structure of the regression plots, since
estimation. With respect to the Sentinel-2 results, without using the data was not grouped into 30 m pixels at it is for Landsat 8 (Fig. 7), but
corrected Band 2 spectral response function as described in the Methods overall the spread of points was not appreciably less. This suggests that
(Section 2.7) a larger systematic error was introduced. Therefore it is the bathymetric image and the in situ data cannot be reconciled at a
clear that good correspondence to the 1:1 line in all the plots of Fig. 7 resolution less than around 30 m. This can also be demonstrated in the
was highly dependent on the details of the calibration and atmospheric bathymetry results produced for the benthic mapping exercise for the
correction to within a few per cent. SSCMR area (Fig. 8). The initial results from Sentinel-2 exhibited quite
The uncertainty bounds were slightly better for Sentinel-2 than wide spread of deviations from the 1:1 line (Fig. 8a). In this case the
Landsat 8 (Fig. 7). The propensity for very shallow image derived image geolocation was optimised by manual adjustment of up to two
depths from Landsat 8 (i.e. uncertainty bounds reaching zero for depths pixels in all directions (20 m) to maximise the r2 value, an adjustment of
at around 10 m) was likely due to the fact that the Sentinel-2 images one pixel proved maximal but improved the r2 only by a tiny fraction.
were glint corrected but Landsat 8 was not. These low depth features Applying a filter to the bathymetry image, in which each pixel became
are probably due to surface glint or white caps. While glint correction the mean of the 3 × 3 pixel window around it, reduced the spread in
cannot effectively correct white-caps it will modify the reflectance so the results and substantially improved the r2 value from 0.64 to 0.77
that those pixels are not interpreted as shallow (Hedley et al., 2005). (Fig. 8b). This implies either that the bathymetry image contains spatial
Image spatial resolution and geo-location of the in situ data in the noise at the scale of 10 m, or that the in situ data is spatially noisy and
imagery could also be a potential source of spread or systematic errors, not precisely geo-located at that scale, or both.
particularly if the in situ data were primarily located on slopes with a The image simulation analysis of Hedley et al. (2012) suggested that
certain orientation. Imagery geo-location was used as supplied in the combining the 60 m resolution band with the 10 m bands for a physics-
level 1 data with no further corrections, but since the in situ data was based analysis may introduce anomalies due to non-physical spectra
607
Fig. 8. Bathymetric results for SSCMR area from Sentinel-2A imagery, (a) direct results (b) results after a 3 × 3 pixel mean filter was applied to image output. Red
lines show 90% confidence intervals from the uncertainty analysis averaged in 1 m bins. (For interpretation of the references to colour in this figure legend, the reader
is referred to the web version of this article.)
over heterogeneous regions. Visual inspection of the bathymetry results inverted; therefore objective conclusions about the calibration of the
(Fig. 9c) showed no evidence of serious errors due to this factor, but sensors or the accuracy of the atmospheric correction cannot be made.
there were slight discontinuities present at the 60 m pixel boundaries. Since the only verification available for the physics-based inversion was
The previous sensitivity analysis was based on re-scaling of 1 m re- the in situ bathymetry data intermediate results cannot be assessed. The
solution airborne imagery and therefore pixels in the simulated image results presented here required the best imagery, but with Sentinel-2
were much more sharply delineated than in real imagery, with no numerous images are available to make this choice, and that alone
consideration of the instrument point spread function (PSF) (Radoux provides a substantial advantage. Even with the clearest images, due to
et al., 2016). Past experience with high spatial resolution imagery the spatial resolution of Sentinel-2, per-pixel glint correction is bene-
(pixels ~2 m) indicates that spatial filtering frequently improves ficial but requires further development to reduce artefacts.
bathymetric results (unpublished data). An important general point is
that the spatial resolution of a source image may not be the optimal 3.4. Benthic mapping
spatial resolution for a given product produced from that image. Sen-
tinel-2 is clearly able to resolve finer scale bathymetric features than As expected, Sentinel-2 was able to resolve benthic features at
Landsat 8 (Fig. 9c, d) but the optimal spatial resolution of the bathy- smaller scales than Landsat 8. In the bottom reflectance map (Fig. 9e, f)
metry is likely < 10 m, and dependent on depth since deeper estimates features such as coral structures on the reef slope, which were indistinct
are noisier (Fig. 10c). Likewise, incorporating the 15 m pixel panchro- blurs in the Landsat 8 image (upper part of Fig. 9f), were clearly deli-
matic Landsat 8 band into the bathymetry inversion may have some neated in the Sentinel-2 map (Fig. 9e). Equally Sentinel-2 revealed a
influence on the effective spatial resolution, although this would be dense population of patch reefs (bommies, < 500 m2) in places where
difficult to interpret since the depth estimates will be highly reliant on Landsat 8 presented just a rough texture (lower left part of Fig. 9e, f).
the differential attenuation of light in the 30 m multispectral bands. The clarity in these features was translated to the object orientated
Another issue with the SSCMR bathymetry map was that the analyses of geomorphic zones and benthic type (Fig. 10). The outline of
Sentinel-2A image was acquired at the time of a relatively low tide of mapped features such as patch reefs or edges between sand and rock,
−0.5 m (MSL) according to OTPS2. This lead to a slight mis-correction were delineated with more detail than is typically achieved by Landsat
of the shallowest areas by the deglint procedure because the water- 8 (Roelfsema et al., 2018) and represented more closely what was
leaving radiance in the NIR band could be non-zero. Consequentially visible in the field or in high spatial resolution imagery such as
the shallowest areas on the reef crest are estimated as slightly too deep, Quickbird or Worldview 2 (Roelfsema et al., 2018). The benthic com-
as can be seen in Fig. 9c vs. 9d. This was not apparent in the scatter position map for the eight reefs had overall accuracy of 49% using 6
plots (Fig. 8) since there was no in situ depth data in depths < 5 m. The categories which was higher than the previously conducted Landsat
Lizard analyses with tides of 0.22 m and − 0.19 m MSL seemed un- study with 34% (Roelfsema et al., 2018), but lower than studies on
affected by this issue (Fig. 7). The Sentinel-2 results overall were im- single reefs using high spatial resolution imagery. Similar improve-
proved by the deglint procedure so this indicates that further work on ments in accuracy can be expected for simpler per-pixel classification
the application of this step is required. approaches (Andréfouët et al., 2003) since spatial delineation of
Overall, Sentinel-2 appears equally capable of bathymetric mapping benthic patches is equally important for those methods. Object based
as Landsat 8, if not potentially better, both from a radiometric and approaches can also use texture information to improve classification
spatial point of view. Within the capacity to relate in situ data to (Benfield et al., 2007; Saul and Purkis, 2015; Wahidin et al., 2015). So
imagery it can produce results comparable to commercial sensors such an additional advantage in the smaller pixel size of Sentinel-2 in com-
as WorldView 2 (Fig. 7a vs. Hamylton et al., 2015). However physics- parison to Landsat 8 was that textures that are naturally present in reef
based inversion requires precise atmospheric corrections and calibra- bottom features were better visible in the Sentinel-2 imagery (Fig. 9).
tion (Dekker et al., 2011; Goodman et al., 2008). The methods used Previous studies mapping geomorphic zonation using OBIA of
here involved an atmospheric correction that effectively aligns the Landsat data have used reflectance values only (Leon and Woodroffe,
imagery to the deep water reflectance produced by the model to be 2011), whereas this study included image derived attributes depth and
608
Fig. 9. Examples comparing spatial resolutions over reefs in the SSCMR region, specifically the northern part of Hall-Thompson Reef. (a, b) RGB composite from
original images using bands 2, 3 and 4 of both sensors, (c, d) bathymetric map, (e, f) bottom reflectance RGB composite based on bands 1, 2 and 3.
slope, which increase the discriminatory ability and repeatability of the The Sentinel-2 MSI swath is 290 km wide (larger than Landsat 8’s
mapping process. The derived depth and slope attributes from Sentinel- 185 km) so large areas are captured under the same environmental
2 are more robust than from Landsat 8 because, for example, a reef conditions, albeit with some variation in view angle (Roy et al., 2017).
slope from 0 to 10 m depth may stretch horizontally for 50 to 100 m, At approximately 20 km wide the eight reefs in the SSCMR site was a
which is just three Landsat 8 30 m pixels (six 15 m pan-sharpened small subset of the available acquired area on that date. The eight reefs
pixels) or ten Sentinel-2 10 m multispectral pixels. Regions of slope are being within the same image allows OBIA mapping using same rule set,
more reliably identifiable when characterised by more pixels, and this which is a strong advantage over high spatial resolution sensors as it
also applies to other geomorphological features that can be useful dis- improves consistency. To perform the same analysis using high spatial
criminators for benthic mapping. However, the smaller pixel size in resolution imagery it is likely that eight individual scenes would be
comparison to Landsat 8 also leads to small scale pixel-to-pixel noise in required, each captured under varying environmental conditions (e.g.
the imagery, especially in deeper areas of the bathymetry and bottom tides, clouds, sea surface state, solar-view geometry, water clarity) and
reflectance layers (Fig. 9). This needs to be compensated for when each requiring adjustment of rule sets (Phinn et al., 2012; Roelfsema
applying the OBIA rule sets. The benthic mapping approaches applied et al., 2013; Saul and Purkis, 2015). Validation would also then require
in this study did not use wave exposure parameters derived from water in situ data within each image, representing a substantially larger field
depth as was done in the previous Landsat based study (Roelfsema work effort, whereas in this study only three out of the eight reefs were
et al., 2018). It is expected that this is another aspect where the gain in visited in the field. Therefore at larger scales (> 10 km) there are dis-
pixel resolution will propagate to an improved water depth and ex- advantages to high spatial resolution imagery and Sentinel-2's large
posure model and therefore to habitat maps. area acquisition may be a practical advantage.
609
Fig. 10. (a, b) Geomorphological zone and (c, d) benthic classification maps of the SSCMR region produced by object orientated analysis of Sentinel-2 data. Panels (b,
d) show zoomed in area of (a, c).
3.5. Scope for remote sensing objectives on coral reefs 3.5.1. Geomorphic zone mapping
Geomorphic zones are usefully mapped at scales hundreds to
The results and previous discussion highlights several distinct thousands of meters and delineate regions at sizes of 100 s m (Figs. 1,
strengths of Sentinel-2 for reef applications, and some caveats. Table 3 10). While this zonation can be captured by the spatial resolution of
summarises these findings in the context of specific objectives relevant both Sentinel-2 and Landsat 8, Sentinel-2 can provide a finer delinea-
to reef management. This section expands on the table with a brief tion of the edges of zones. Further, geo-physical parameters such as reef
discussion on each topic. slope regions can be more robustly identified with Sentinel-2 due to
610
Table 3
Summary of the benefits and caveats in the application of Sentinel-2 imagery towards various coral reef remote sensing objectives.
Objective Comment
Geomorphic zone mapping • For large scale zones (> 100 m) Landsat 8 and Sentinel-2 equally applicable.
• Sentinel-2 provides better delineation and more robust identification of zones.
Benthic mapping • Annual or biannual maps can be created from Landsat 8 or Sentinel-2.
• Sentinel-2 permits mapping details that are more representative of the heterogeneous environment of reefs.
• Large 2
area (> 1000 km ) acquisitions of Sentinel-2 are an advantage over high spatial resolution sensors for multi-reef mapping.
Bathymetry • Performance of Sentinel-2 within imagery spatial scale can be as good as Landsat 8 or high resolution sensors.
• Repeatability, atmospheric and water surface corrections, remain an issue that requires more work, for Sentinel-2 and Landsat 8.
Benthic change detection • Change detection benefits from readily available multiple images from Sentinel-2 and Landsat 8 to characterise variability baseline.
• Cloud and surface glint limit number of usable images as a function of location and season, for both Sentinel 2 and Landsat 8.
Bleaching detection • High revisit rates of Sentinel-2 and availability of imagery make bleaching detection a good prospect, Landsat-8 offers additional data.
• AUnderstanding
known bleaching event appears to be visible in some Sentinel-2 imagery, further work is needed to develop methods.
• High revisit ratesglobal patterns is limited by lack of global reef coverage by Sentinel-2.
Science and understanding • Long term missionofcommitment
Sentinel-2 and availability of imagery likely to foster new methods and insights.
• many decades. of Sentinel-2 and continuity with previous Landsat series data gives future potential for time series analyses over
Global scale applications • Sentinel-2 data and current methods could produce valuable new global scale information resource.
• Limited by lack of global reef coverage by Sentinel-2.
being covered by larger groups of pixels, and this can be beneficial for 3.5.2. Benthic mapping
benthic mapping and other derived products such as hydrodynamic The improved spatial resolution of Sentinel-2 over Landsat 8 is
models. realised in features which occur at scales of around 10–100 m. These
features can be better identified or more clearly mapped, such as the
Fig. 11. (a) Bleaching as visible in Sentinel-2A imagery at Adelaide reef in February 2017, in comparison to June 2016. The two lower images show a zoomed-in area
as represented by a white rectangle in top left panel. (b) Bottom reflectance at ~560 nm, derived as described in Section 2.8, from a time series of Sentinel-2A images
over one of the bleached areas. The brightening anomaly was present in two successive images (red circle). Coral mortality at the site was verified by field surveys
before and after February 2017, these data will be presented elsewhere (Roelfsema, unpublished data). (For interpretation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)
611
coral patches visible in Fig. 9e and Fig. 10. Beyond simply higher resource can be challenging.
solution maps, object orientated techniques may facilitate increasing
the number of classes that can be mapped over Landsat 8 due to tex- 3.5.5. Bleaching detection
tures revealed at smaller scales, for example the smaller coral heads Bleaching detection requires a high frequency of clear images, the
that became visible in Fig. 9e are highly characteristic. Realising the full onset of coral bleaching stretches over 1–4 week period, and the actual
benefit of 10 m pixel imaging requires further study on ecological bleached white period lasts up to 4–6 weeks after which recovery or
spatial complexity at that scale (Lim et al., 2009). Higher spatial re- colonisation by algae returns the corals to dark reflectance. A 10 to 20-
solution sensors (pixels ≤ 2 m) may still be favoured for local scale day usable image revisit should be adequate for bleaching studies and is
mapping in areas less than ~100 km2 but the cost effectiveness and likely achieved over many reefs, such as in the GBR area. With respect
relative likelihood of finding clear images means Sentinel-2 will be to spatial resolution and detectability, detection of bleaching by remote
increasingly considered as an option. Sentinel-2 is likely to find sub- sensing has been demonstrated in 4 m multispectral pixel imagery
stantial use in developing countries and by NGOs with constrained (Elvidge et al., 2004; Rowlands et al., 2008), and Yamano and Tamura
budgets. One advantage over high resolution sensors is that the large (2004) suggested a 25% sub-pixel bleaching area was required for de-
acquisition area of Sentinel-2 facilitates more consistent mapping at tection in 30 m Landsat TM pixels. A review of Sentinel-2A imagery
multi-reef scales. over GBR indicates that the 2017 bleaching event was visible at least in
some locations (Fig. 11).
3.5.3. Bathymetry
Bathymetry mapping results with Sentinel-2 can be excellent in 3.5.6. Science and understanding
comparison with both Landsat 8 and high spatial resolution sensors. At swath edges some reef locations benefit from double acquisitions,
The spatial resolution of captured features is as expected, and the 60 m e.g. Wistari Reef to the west of Heron Island (Fig. 5). Although glint can
band is useful and appears to only induce small artefacts in otherwise be an issue at the eastern swath edge some images will be clear
noisy areas. The limiting factor in quantitative assessment of spatial (Fig. 5c). These high frequency imaged locations will be an excellent
features is in relating in situ data to the bathymetric maps. As with all resource for science applications, for example: understanding baseline
physics-based methods, challenges remain with atmospheric and water- variation for change detection algorithms; providing comparable views
surface corrections. These are well known issues for all sensors under different solar-view geometries (Roy et al., 2017); or simply
(Goodman et al., 2008) but the spatial and spectral characteristics of providing substantial datasets for algorithm testing and vicarious cali-
Sentinel-2 suggest work on unique solutions are required, this work is bration adjustments (Pahlevan et al., 2014, 2017). In general this fre-
ongoing through ESA initiatives and the scientific community (e.g. see quency of images and spatial resolution over reefs has not been pre-
atmospheric correction session at the EO Open Science Conference, viously available, so there is an exciting opportunity to look for
September 2017, http://eoopenscience.esa.int/page_session23.php). processes at spatial and temporal scales not previously observed.
3.5.4. Benthic change detection 3.5.7. Global scale applications

Change detection benefits greatly from multiple images to under- Reefs are under threat globally (Pandolfi et al., 2003; Wolff et al.,
stand baseline variation, since apparent changes between any two 2015) and recent initiatives are increasing the availability of large-scale
images could be due to irrelevant causes such as tide, sea surface state, in situ data collection, including automated processing of video images
water constituents. The usable acquisition rate of Sentinel-2 is variable for coral community composition (González-Rivero et al., 2014). The
due to location and season, but a minimal expectation would be several availability of such field data together with Sentinel-2 would make for
good images a year. Coupled with improved spatial resolution over an immediate and useful application: updating the global reef map of
Landsat 8, which enables clearer identification of features, Sentinel-2 the Landsat-based Millennium Coral Reef mapping project (Andréfouët
will not only be a strong choice for change detection applications but is et al., 2004). That map was produced largely by visual interpretation;
likely to catalyse scientific developments in multi-image analysis on the application of current methods as presented in this paper could
reefs. Currently commercial price structures and licences mean the cost produce not just an incremental higher resolution update to the global
of commercial high resolution imagery is prohibitive for multi-image reef map, but rather a fundamentally more comprehensive resource
analysis, so Sentinel-2 is a unique proposition in this application. This including geomorphic zones, benthic classification and bathymetric
may change in the next five years as the traditional high spatial re- maps. These would form the baseline not only for cataloguing and
solution satellite companies move to service-fee access models to match monitoring global reefs, but a resource for ecological and biophysical
the next generation of imaging cube-sat constellations. The latter now models, long term change detection and understanding global scale
provides accessible and useable very high spatial resolution, multi- patterns in reef ecology.
spectral global coverage, although reef mapping applications are only The main drawback of Sentinel-2 is the current lack of global reef
just being developed (Asner et al., 2017). Change detection requires coverage, this affects not only global-scale applications but many re-
precise image alignment, and for Sentinel-2 this is ongoing work mote territories where the cost-effectiveness of remote sensing has the
through the development of the Global Reference Image (Clerc, 2017) most to offer. Hopefully as the community develops coral reef appli-
with a stated target image-image registration of 0.3 of a pixel (pre- cations of Sentinel-2 ESA and the European Commission may review the
sumably only over land, ESA, 2017). Sensor performance consistency is acquisition area. There may be mission constraints on what can be
also beneficial for change detection, and both Sentinel-2 and Landsat 8 achieved: in some cases extending the coverage over missed reefs may
benefit from substantial cal-val efforts from space agencies and the simply be a case of extending the acquisition time within a swath (e.g.
scientific community (Gascon et al., 2017; Pahlevan et al., 2014, 2017). Chagos Islands, Fig. 3c), but for the major omissions in the Pacific ad-
Another option to increase time series data is synthesis with Landsat ditional swath segments may be required (Fig. 3d). Nevertheless, that
8. The methods described in this paper are equally applicable to the major omissions of an ESA mission are within European territories
Sentinel-2 and Landsat 8, and at the Landsat 8 spatial scale, produced (Table 2) would seem a priority to address. The required additional
similar results (Figs. 7, 9). In suitable applications Sentinel-2 and imaging effort amounts to ~1/200th of the current acquisitions, and
Landsat 8 data could be used side-by-side, both increasing the effective would facilitate progress on European obligations to reef conservation
data acquisition rate but also providing a path to integrate Sentinel-2 under various environmental initiatives. Further, time is of the essence
into the legacy Landsat series data set and long term change detection since long term data continuity is essential for change detection in the
analyses (Palandro et al., 2008). However, further work on this is context of long term stressors such as climate change. The omitted reefs
needed since alignment of data and products even from the same image have already missed out on over 18 months of acquisitions, which in
612
some cases just require swaths to be extended by the equivalent of a References

handful of tiles (Fig. 3c).
Andréfouët, S., Kramer, P., Torres-Pulliza, D., Joyce, K.E., Hochberg, E.J., Garza-Pérez, R.,
et al., 2003. Multi-site evaluation of IKONOS for classification of tropical coral reef
environments. Remote Sens. Environ. 88, 128–143.
4. Conclusions Andréfouët, S., Muller-Karger, F.E., Robinson, J.A., Kranenburg, C.J., Torres-Pulliza, D.,
Spraggins, S., Murch, B., 2004. Global assessment of modern coral reef extent and
Sentinel-2 has the capability to deliver essential scientific, mon- diversity for regional science and management applications: a view from space. In:
Proceedings of the 10th International Coral Reef Symposium, Okinawa, Japan, 28
itoring and management-ready information for coral reef applications. June–2 July 2004.
This capability is a step up from Landsat 8 due to spatial resolution and Antoine, D., Morel, A., 1999. A multiple scattering algorithm for atmospheric correction
revisit time, but Sentinel-2 also performs well in comparison high re- of remotely sensed ocean colour (MERIS instrument): principle and implementation
for atmospheres carrying various aerosols including absorbing ones. Int. J. Remote
solution imagery and offers advantages in terms of cost effectiveness Sens. 20, 1875–1916.
and large scale coverage. Existing methods, as demonstrated here, can Asner, G.P., Martin, R.E., Mascaro, J., 2017. Coral reef atoll assessment in the South China
immediately enable effective use of Sentinel-2 by the reef science Sea using planet dove satellites. Remote Sens. Ecol. Conserv. 3, 57–65.
Beijbom, O., Edmunds, P.J., Roelfsema, C., Smith, J., Kline, D.I., Neal, B.P., Dunlap, M.J.,
community and management agencies globally, to address their es-
Moriarty, V., Fan, T.-Y., Tan, C.-J., Chan, S., Treibitz, T., Gamst, A., Mitchell, B.G.,
sential environmental information requirements – this also applies to Kriegman, D., 2015. Towards automated annotation of benthic survey images:
hydrographic agencies, defence applications and private industry. The variability of human experts and operational modes of automation. PLoS One 10,
revisit time and cost-effectiveness of imagery is unparalleled, and offers e0130312.
Benfield, S.L., Guzman, H.M., Mair, J.M., Young, J.A.T., 2007. Mapping the distribution
many advantages: from simply being able to select the clearest images of coral reefs and associated sublittoral habitats in Pacific Panama: a comparison of
to opportunities for fundamentally new methods and insights. In the optical satellite sensors and classification methodologies. Int. J. Remote Sens. 28,
near future new methods are certain to be developed to meet the op- 5047–5070.
Blaschke, T., 2010. Object based image analysis for remote sensing. J. Photogramm.
portunity that Sentinel-2 presents, and this can only further add to the Remote Sens. 65, 2–16.
use of Sentinel-2 in coral reef applications. Synergy with Landsat 8, and Bour, W., 1988. SPOT images for a coral reef mapping in New Caledonia. A fruitful ap-
future and past Landsat missions, also adds substantial value to proach for classic and new topics. In: Proceedings of the 6th International Coral Reef
Symposium, Australia, 1988. vol. 2. pp. 445–448.
Sentinel-2 data. Brando, V.E., Anstee, J.M., Wettle, M., Dekker, A.G., Phinn, S.R., Roelfsema, C., 2009. A
That nearly 90% of the World's reefs are imaged by what was en- physics based retrieval and quality assessment of bathymetry from suboptimal hy-
visaged as a land mission is certainly a positive outcome, but the perspectral data. Remote Sens. Environ. 113, 755–770.
Clerc, S., 2017. S2 MPC data quality report. In: S2-PDGS-MPC-DQR Issue 19 (06/09/
missing 12% is likely to be a disappointment to many potential reef 2017). European Space Agency. https://earth.esa.int/web/sentinel/user-guides/
users. Under the current acquisition plan, many regions will not benefit sentinel-2-msi/document-library.
from Sentinel-2 at all, and lack of global coverage is limiting for the Clerc, S., 2018. S2 MPC data quality report. In: S2-PDGS-MPC-DQR Issue 25 (06/03/
2018). European Space Agency. https://earth.esa.int/web/sentinel/user-guides/
global-scale science and management applications for which Sentinel-2
sentinel-2-msi/document-library.
would otherwise be a very strong proposition. It is our hope that there is Dekker, A.G., Phinn, S.R., Anstee, J., Bissett, P., Brando, V.E., Casey, B., Fearns, P.,
the possibility for ESA and the European Commission to review and Hedley, J., Klonowski, W., Lee, Z.P., Lynch, M., 2011. Intercomparison of shallow
extend the Sentinel-2 acquisition plan to cover more, if not all, reefs. water bathymetry, hydro-optics, and benthos mapping techniques in Australian and
Caribbean coastal environments. Limnol. Oceanogr. Methods 9, 396–425.
Full coral reef coverage would remove the only obstacle for immediate Devred, E., Turpie, K.R., Moses, W., Klemas, V.V., Moisan, T., Babin, M., Toro-Farmer, G.,
uptake of Sentinel-2 in providing global-scale environmental informa- Forget, M.-H., Young-Heon, J., 2013. Future retrievals of water column bio-optical
tion required to meet the European Union's conservation obligations properties using the hyperspectral infrared imager (HyspIRI). Remote Sens. 5,
6812–6837. https://doi.org/10.3390/rs5126812.
under numerous initiatives, including the UN Sustainable Development Egbert, G.D., Erofeeva, S.Y., 2002. Efficient inverse modeling of barotropic ocean tides. J.
Goals. In particular, given the surprising and substantial omission of Atmos. Ocean. Technol. 19, 183204. https://doi.org/10.1175/1520-0426(2002)
coverage of reefs in European territories, an extended acquisition plan 019<0183:EIMOBO>2.0.CO;2.
El-Askary, H., Abd El-Mawla, S.H., Li, J., El-Hattab, M.M., El-Raey, M., 2014. Change
would be very much in line with the aims of the Copernicus programme detection of coral reef habitat using Landsat-5 TM, Landsat 7 ETM+ and Landsat 8
to support Earth Observation in a European context. The true value of OLI data in the Red Sea (Hurghada, Egypt). Int. J. Remote Sens. 35, 2327–2346.
100% global coverage of reefs would be fully realised in the decades Elvidge, C.D., Dietz, J.B., Berkelmans, R., Andréfouët, S., Skirving, W., Strong, A.E.,
Tuttle, B.T., 2004. Satellite observation of Keppel Islands (Great Barrier Reef) 2002
and even centuries to come, when, inevitably, future conservation ef-
coral bleaching using IKONOS data. Coral Reefs 23, 123–132.
forts seek to look back in time to understand long term trajectories at Emde, C., Buras-Schnell, R., Kylling, A., Mayer, B., Gasteiger, J., Hamann, U., Kylling, J.,
global scales. Richter, B., Pause, C., Dowling, T., Bugliaro, L., 2016. The libRadtran software
package for radiative transfer calculations (version 2.0.1). Geosci. Model Dev. 9,
1647–1672. https://doi.org/10.5194/gmd-9-1647-2016.
ESA, 2017. Sentinel-2 MSI Technical Guide. https://earth.esa.int/web/sentinel/technical-
Acknowledgements guides/sentinel-2-msi.
Friedlander, A.M., Caselle, J.E., Ballesteros, E., Brown, E.K., Turchik, A., Sala, E., 2014.
The real bounty: marine biodiversity in the Pitcairn Islands. PLoS One 9 (6), e100142.
The work presented here was part funded by the European Space https://doi.org/10.1371/journal.pone.0100142.
Agency project Sen2Coral as part of the SEOM programme (S2 for Land Garcia, R., Fearns, P., McKinna, L.I.W., 2014a. Detecting trend and seasonal changes in
and Water, Coral Reefs). The global coral reef map was used with kind bathymetry derived from HICO imagery: a case study of Shark Bay, Western
Australia. Remote Sens. Environ. 147, 186–205.
permission of UNEP-WCMC. The bathymetry data from Lizard Island Garcia, R.A., McKinna, L.I.W., Hedley, J.D., Fearns, P.R.C.S., 2014b. Improving the op-
was provided by Sarah Hamylton and Rob Beaman. The Great Barrier timization solution for a semi-analytical shallow water inversion model in the pre-
Reef Foundation funded development work, field work and analysis sence of spectrally correlated noise. Limnol. Oceanogr. Methods 12, 651–669.
https://doi.org/10.1016/j.rse.2014.03.010.
related to the benthic mapping exercise. OBIA image analysis was Gascon, F., Bouzinac, C., Thépaut, O., Jung, M., Francesconi, B., Louis, J., et al., 2017.
performed with eCognition and software support was provided by Copernicus sentinel-2A calibration and products validation status. Remote Sens. 9,
Trimble. Fieldwork and field data analysis could not have been done 584. https://doi.org/10.3390/rs9060584.
Giardino, C., Bresciani, M., Fava, F., Matta, E., Brando, V.E., Colombo, R., 2016. Mapping
without the help of volunteer field teams and the charter vessel MV
submerged habitats and mangroves of Lampi Island Marine National Park (Myanmar)
Kalinda. from in situ and satellite observations. Remote Sens. 8, 2. https://doi.org/10.3390/
rs8010002.
González-Rivero, M., Bongaerts, P., Beijbom, O., Pizarro, O., Friedman, A., Rodriguez-
Ramirez, A., Upcroft, B., Laffoley, D., Kline, D., Vevers, R., et al., 2014. The Catlin
Appendix A. Supplementary data seaview survey—kilometre-scale seascape assessment, and monitoring of coral reef
ecosystems. Aquat. Conserv. Mar. Freshwat. Ecosyst. 24, 184–198.
Supplementary data to this article can be found online at https:// González-Rivero, M., Beijbom, O., Rodriguez-Ramirez, A., Holtrop, T., González-Marrero,
Y., Ganase, A., Roelfsema, C., Phinn, S., Hoegh-Guldberg, O., 2016. Scaling up
doi.org/10.1016/j.rse.2018.07.014.
613
ecological measurements of coral reefs using semi-automated field image collection demonstrations and validations. Remote Sens. Environ. 201, 47–56.
and analysis. Remote Sens. 8, 30. Palandro, D.A., Andréfouët, S., Hu, C., Hallock, P., Muller-Karger, F., Dustan, P.,
Goodman, J.A., Lee, Z.P., Ustin, S.L., 2008. Influence of atmospheric and sea-surface Callahan, M.K., Kranenburg, C., Beaver, C.R., 2008. Quantification of two decades of
corrections on retrieval of bottom depth and reflectance using a semi-analytical shallow-water coral reef habitat decline in the Florida Keys National Marine
model: a case study in Kaneohe Bay, Hawaii. Appl. Opt. 47, F1–F11. Sanctuary using Landsat data (1984–2002). Remote Sens. Environ. 112, 3388–3399.
Guanter, L., Kauffmann, H., Segl, K., Foerster, S., Rogass, C., et al., 2015. The EnMAP https://doi.org/10.1016/j.rse.2008.02.015.
spacebourne imaging spectroscopy mission for Earth observation. Remote Sens. 7, Pandolfi, J.M., Bradbury, R.H., Sala, E., Hughes, T.P., Bjorndal, K.A., Cooke, R.G.,
8830–8857. https://doi.org/10.3390/rs70708830. McArdle, D., McClenachan, L., Newman, M.J.H., Paredes, G., et al., 2003. Global
Hamylton, S.M., Hedley, J.D., Beaman, R.J., 2015. Derivation of high-resolution bathy- trajectories of the long-term decline of coral reef ecosystems. Science 301, 955–958.
metry from multispectral satellite imagery: a comparison of empirical and optimi- https://doi.org/10.1126/science.1085706.
sation methods through geographical error analysis. Remote Sens. 7, 16257–16273. Phinn, S.R., Roelfsema, C.M., Stumpf, R., 2010. Remote sensing: discerning the promise
https://doi.org/10.3390/rs71215829. from the reality. In: Longstaff, B.J., Carruthers, T.J.B., Dennison, W.C., Lookingbill,
Hedley, J.D., Harborne, A.R., Mumby, P.J., 2005. Simple and robust removal of sun glint T.R., Hawkey, J.M., Thomas, J.E., Wicks, E.C., Woerner, J. (Eds.), Integrating And
for mapping shallow-water benthos. Int. J. Remote Sens. 26, 2107–2112. Applying Science: A Handbook For Effective Coastal Ecosystem Assessment. IAN
Hedley, J.D., Roelfsema, C., Phinn, S.R., 2009. Efficient radiative transfer model inversion Press, Cambridge, Maryland.
for remote sensing applications. Remote Sens. Environ. 113, 2527–2532. Phinn, S.R., Roelfsema, C.M., Mumby, P.J., 2012. Multi-scale, object-based image analysis
Hedley, J.D., Roelfsema, C., Phinn, S., 2010. Propagating uncertainty through a shallow for mapping geomorphic and ecological zones on coral reefs. Int. J. Remote Sens. 33,
water mapping algorithm based on radiative transfer model inversion. In: 3768–3797.
Proceedings of the Ocean Optics XX Conference, Anchorage, AK, USA, 27 Radoux, J., Chomé, G., Jacques, D.C., Waldner, F., Bellemans, N., Matton, N., Lamrche,
September–1 October 2010. C., D'Andrimont, R., Defourney, P., 2016. Sentinel-2's potential for sub-pixel land-
Hedley, J.D., Roelfsema, C., Koetz, B., Phinn, S., 2012. Capability of the sentinel 2 mission scape feature detection. Remote Sens. 8, 488. https://doi.org/10.3390/rs8060488.
for tropical coral reef mapping and coral bleaching detection. Remote Sens. Environ. Roelfsema, C., Phinn, S., 2010. Integrating field data with high spatial resolution multi-
120, 145–155. https://doi.org/10.1016/j.rse.2011.06.028. spectral satellite imagery for calibration and validation of coral reef benthic com-
Hedley, J.D., Roelfsema, C.M., Chollett, I., Harborne, A.R., Heron, S.F., Weeks, S., munity maps. J. Appl. Remote. Sens. 4, 043527. https://doi.org/10.1117/1.3430107.
Skirving, W.J., Strong, A.E., Eakin, C.M., Christensen, T.R.L., Ticzon, V., Bejerano, S., Roelfsema, C.M., Phinn, S.R., Jupiter, S., Comley, J., Albert, S., 2013. Mapping coral reefs
Mumby, P.J., 2016. Remote sensing of coral reefs for monitoring and management: a at reef to reef-system scales, 10s–1000s km2, using object-based image analysis. Int. J.
review. Remote Sens. 8, 118–157. Remote Sens. 34, 6367–6388.
Hochberg, E.J., Atkinson, M.J., 2003. Capabilities of remote sensors to classify coral, Roelfsema, C.M., Kovacs, E.M., Phinn, S.R., 2017. Georeferenced Photographs of Benthic
algae, and sand as pure and mixed spectra. Remote Sens. Environ. 85, 174–189. Photoquadrats Acquired Along 160 Transects Distributed Over 23 Reefs in the Cairns
Hock, K., Wolff, N.H., Condie, S.A., Anthony, K.R.N., Mumby, P.J., Paynter, Q., 2014. to Cooktown Region of the Great Barrier Reef. January and April/May, 2017.
Connectivity networks reveal the risks of crown-of-thorns starfish outbreaks on the PANGAEA. https://doi.org/10.1594/PANGAEA.877578.
Great Barrier Reef. J. Appl. Ecol. 51, 1188–1196. Roelfsema, C., Kovacs, E., Ortiz, J.C., Phinn, S., Mumby, P., Callaghan, D., Ronan, M.,
Hoegh-Guldberg, O., Mumby, P.J., Hooten, A.J., Steneck, R.S., Greenfield, P., Gomez, E., Wolf, N., Hamylton, S., Wettle, M., 2018. Coral reef habitat mapping: a combination
et al., 2007. Coral reefs under rapid climate change and ocean acidification. Science of object-based image analysis and ecological modelling. Remote Sens. Environ. 208,
318, 1737–1742. 27–41. https://doi.org/10.1016/j.rse.2018.02.005.
Joyce, K., Phinn, S., Roelfsema, C., Neil, D., Dennison, W., 2002. Mapping the southern Rowlands, G.P., Purkis, S.J., Riegl, B.M., 2008. The 2005 coral-bleaching event, Roatan
Great Barrier Reef using Landsat ETM and the Reef Check classification scheme. In: (Honduras): use of pseudo-invariant features (PIFs) in satellite assessments. J. Spat.
11th Australasian Remote Sensing & Photogrammetry Conference. Causal Sci. 53, 99–112.
Publications, pp. 239–250. Roy, D.P., Li, J., Zhang, H.K., Yan, L., Huang, H., Li, Z., 2017. Examination of sentinel-2A
Kay, S., Hedley, J.D., Lavender, S., 2009. Sun glint correction of high and low spatial multi-spectral instrument (MSI) reflectance anisotropy and the suitability of a general
resolution images of aquatic scenes: a review of methods for visible and near-infrared method to normalize MSI reflectance to nadir BRDF adjusted reflectance. Remote
wavelengths. Remote Sens. 1, 697–730. Sens. Environ. 199, 25–38.
Kutser, T., Miller, I.R., Jupp, D.L.B., 2006. Mapping coral reef benthic substrates using Saul, S., Purkis, S., 2015. Semi-automated object-based classification of coral reef habitat
hyperspectral space-borne images and spectral libraries. Estuar. Coast. Shelf Sci. 70, using discrete choice models. Remote Sens. 7, 15810.
449–460. Smith, V.E., Rogers, R.H., Reed, L.E., 1975. Automated mapping and inventory of Great
Lee, Z., Carder, K.L., Mobley, C.D., Steward, R.G., Patch, J.S., 1998. Hyperspectral remote Barrier Reef zonation with Landsat data. In: Ocean 75 Conference Record. Institute of
sensing for shallow waters: 1. A semianalytical model. Appl. Opt. 37, 6329–6338. Electrical and Electronics Engineers, Inc, New York.
Leon, J., Woodroffe, C.D., 2011. Improving the synoptic mapping of coral reef geomor- Spalding, M.D., Ravilious, C., Green, E.P., 2001. World Atlas of Coral Reefs. UNEP-
phology using object-based image analysis. Int. J. Geogr. Inf. Sci. 25, 949–969. WCMC/University of California Press, London.
Lim, A., Hedley, J.D., Ledrew, E., Mumby, P.J., Roelfsema, C., 2009. The effects of eco- SUHET, 2015. Sentinel-2 User Handbook, Issue 1 Revision 2. European Space Agency.
logically determined spatial complexity on the classification accuracy of simulated https://earth.esa.int/web/sentinel/user-guides/sentinel-2-msi.
coral reef images. Remote Sens. Environ. 113, 965–978. UNEP-WCMC, WorldFish Centre, WRI, TNC, 2010. Global Distribution of Warm-water
Mercury, M., Green, R., Hook, S., Oaida, B., Wu, W., Gunderson, A., Chodas, M., 2012. Coral reefs, Compiled From Multiple Sources Including the Millennium Coral Reef
Global cloud cover for assessment of optical satellite observation opportunities: a Mapping Project. Version 1.3. http://data.unep-wcmc.org/datasets/1.
HyspIRI case study. Remote Sens. Environ. 126, 62–71. https://doi.org/10.1016/j. Wahidin, N., Siregar, V.P., Nababan, B., Jaya, I., Wouthuyzen, S., 2015. Object-based
rse.2012.08.007. image analysis for coral reef benthic habitat mapping with several classification al-
Mumby, P.J., Green, E.P., Edwards, A.J., Clark, C.D., 1997. Coral reef habitat-mapping: gorithms. Procedia Environ. Sci. 24 (Supplement C), 222–227.
how much detail can remote sensing provide? Mar. Biol. 130, 193–202. https://doi. Wolff, N.H., Donner, S.D., Cao, L., Iglesias-Prieto, R., Sale, P.F., Mumby, P.J., 2015.
org/10.1007/s002270050238. Global inequities between polluters and the polluted: climate change impacts on coral
Pahlevan, N., Lee, Z., Wei, J., Schaaf, C.B., Schott, J.R., Berk, A., 2014. On-orbit radio- reefs. Glob. Chang. Biol. 21, 3982–3994.
metric characterization of OLI (Landsat-8) for applications in aquatic remote sensing. Yamano, H., 2013. Multispectral applications. In: Goodman, J.A., Purkis, S.J., Phinn, S.R.
Remote Sens. Environ. 154, 272–284. (Eds.), Coral Reef Remote Sensing. vol. 3. Springer, Dordrecht, pp. 51–78.
Pahlevan, N., Sarkar, S., Franz, B.A., Balasubramanian, S.V., He, J., 2017. Sentinel-2 Yamano, H., Tamura, M., 2004. Detection limits of coral reef bleaching by satellite remote
MultiSpectral instrument (MSI) data processing for aquatic science applications: sensing: simulation and data analysis. Remote Sens. Environ. 90, 86–103.
614

Remote Sensing of Environment: Contents Lists Available at

Uploaded by

Copyright:

Available Formats

Remote Sensing of Environment: Contents Lists Available at

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Remote Sensing of Environment: Contents Lists Available at

Uploaded by

Copyright:

Available Formats

Remote Sensing of Environment 216 (2018) 598–614

Contents lists available at ScienceDirect

Remote Sensing of Environment

Coral reef applications of Sentinel-2: Coverage, characteristics, bathymetry T

Fig. 4. (a) Frequency of cloud free Sentinel-

3.5.4. Benthic change detection 3.5.7. Global scale applications

some cases just require swaths to be extended by the equivalent of a References

You might also like