remote sensing

Article
Brazilian Mangrove Status: Three Decades of Satellite
Data Analysis
Cesar Diniz 1,2, * , Luiz Cortinhas 1 , Gilberto Nerino 1 , Jhonatan Rodrigues 1 , Luís Sadeck 1 ,
Marcos Adami 2,3,4 and Pedro Walfir M. Souza-Filho 2,5
1 Solved—Solutions in Geoinformation, Belém 66075-750, Brazil; [email protected] (L.C.);
[email protected] (G.N.); [email protected] (J.R.);
[email protected] (L.S.)
2 Geoscience Institute, Federal University of Pará, Belém 66075-110, Brazil; [email protected] (M.A.);
[email protected] (P.W.M.S.-F.)
3 Regional Center of the Amazon, National Institute for Space Research (INPE), São Paulo 12227-010, Brazil
4 Global Land Analysis and Discovery (GLAD) laboratory, Department of Geographical Sciences, University
of Maryland, College Park, MD 20742, USA
5 Instituto Tecnológico Vale (ITV), Belém 66055-090, Brazil
* Correspondence: [email protected]; Tel.: +55-91-98239-1597




Received: 25 January 2019; Accepted: 22 March 2019; Published: 4 April 2019


Abstract: Since the 1980s, mangrove cover mapping has become a common scientific task. However,
the systematic and continuous identification of vegetation cover, whether on a global or regional
scale, demands large storage and processing capacities. This manuscript presents a Google Earth
Engine (GEE)-managed pipeline to compute the annual status of Brazilian mangroves from 1985 to
2018, along with a new spectral index, the Modular Mangrove Recognition Index (MMRI), which has
been specifically designed to better discriminate mangrove forests from the surrounding vegetation.
If compared separately, the periods from 1985 to 1998 and 1999 to 2018 show distinct mangrove area
trends. The first period, from 1985 to 1998, shows an upward trend, which seems to be related more
to the uneven distribution of Landsat data than to a regeneration of Brazilian mangroves. In the
second period, from 1999 to 2018, a trend of mangrove area loss was registered, reaching up to
2% of the mangrove forest. On a regional scale, ~85% of Brazil’s mangrove cover is in the states
of Maranhão, Pará, Amapá and Bahia. In terms of persistence, ~75% of the Brazilian mangroves
remained unchanged for two decades or more.

Keywords: mangroves; machine learning; Google Earth Engine; spectral indices; Brazil; Landsat

1. Introduction
Following global change scenarios, coastal areas are exposed to a wide range of environmental
hazards, including sea-level rise and its associated effects. At the same time, coastal areas are
more densely populated than the hinterland and exhibit higher rates of population growth and
urbanisation [1], hosting almost half of the planet’s population [2]. Nevertheless, coastal areas comprise
only 20% of the Earth’s land area [3].
The Brazilian coastal zone (BCZ) shows the same pattern. Extending approximately 9200 km, this
dynamic landscape of quick physical and socio-economic changes is home to ~18% of the country’s
population, along which 16 out of 28 metropolitan regions lie [4]. The Brazilian coastal zone presents a
very diverse suite of coastal environments that evolved during the Quaternary in response to changes
in climate and sea-level changes, showing an interaction between different sediment supplies and
a geologic heritage that dates back to the breakup of South America and Africa [5]. The Brazilian
mangrove systems are among this diverse suite of coastal environments.

Remote Sens. 2019, 11, 808; doi:10.3390/rs11070808 www.mdpi.com/journal/remotesensing

Remote Sens. 2019, 11, 808 2 of 19

Globally, mangrove forests are distributed in tropical and subtropical intertidal regions between
approximately 30◦ N and 30◦ S [6]. In 2000, mangrove forests represented a total area of 137,760 km2 ,
distributed in 118 countries and making up ~1% of the tropical forests in the world [7]. Mangrove
forests are an evergreen type of vegetation typically distributed from the mean sea level to the highest
spring tide [8] and grow in extreme environmental conditions such as extreme tides, high salinity, high
temperatures and muddy anaerobic soils [9].
Mangrove systems play an essential role in human sustainability, providing a wide range of
ecosystem services, including nutrient cycling, soil formation and wood production. They also
provide fish spawning grounds and carbon (C) storage [10–12], being one of the most productive and
biologically complex ecosystems on earth [13]. Mangroves and coastal wetlands sequester carbon at an
annual rate two to four times greater than that of mature tropical forests and store three to five times
more carbon per equivalent area than do tropical forests [10]. Despite its importance, this environment
is still highly threatened due to population growth and urbanisation processes.
Since the 1980s, mapping and change detection in mangrove areas at the global scale have been
carried out [7,11,14–16]. However, there are few studies in the current literature that include the
systematic and continuous identification of mangroves and associated changes, whether on the global
or regional scale. In Brazil, the first national mangrove map was published in 1991 [17], based on
airborne real aperture radar data collected from 1972 to 1975. At that time, the national mangrove
area was ~13,800 km2 . In the same period, Schaeffer-Novelli et al. [18] described the variability in the
mangrove ecosystems along the Brazilian coast. In the last two decades, several papers were published
that focused on regional-scale mangrove mapping and short temporal windows, for example, [19–23]
concentrated on the northern coast, [24,25] focused on the north-eastern coast and [26,27] focused on
the south-eastern coast of Brazil.
In 2010, national-scale mangrove maps, based on 2009 Landsat-5 data, were again published and
the Brazilian mangrove area reported was ~11,143 km2 [28]. Global-scale mangrove maps published
in the last two decades estimated Brazil’s mangrove area to be ~9627 km2 [7,16]. More recently, the
Brazilian Environmental Ministry released the Brazilian Mangrove Atlas, based on 2013 Landsat-8
data, which estimated the national mangrove area to be near 13,989 km2 [29]. However, the current
literature lacks long time series analyses that allow an exhaustive and systematic understanding of
Brazilian mangrove coverage dynamics.
Regardless of the terrestrial cover to be identified, any systematic and exhaustive identification of
patterns, including vegetation patterns, requires large storage and processing capacities. These two
requirements have only recently been circumvented with the advent of cloud computing platforms,
such as Google Earth Engine (GEE) [30] and Amazon Web Services (AWS) [31], combining several
petabytes of orbital and geospatial data with statistical analysis resources on the planetary scale.
Moreover, the integration of remotely sensed time series data in such platforms minimizes one of
the major problems inherent to land use and land cover mapping: the persistent cloud cover over some
areas of the planet [32]. Intertropical coastal zones are no exception to this characteristic. Coastal zones
are severely affected by cloudy conditions due to their proximity to oceans and their position. The GEE
platform provides fast filtering and sorting capabilities, which greatly facilitates searching through
millions of individual images and pixels to select data that meet specific spatial, temporal, spectral
or other criteria [30]. However, there are few spectral mechanisms specifically designed to support
mangrove identification and to distinguish mangroves from surrounding vegetation. Traditionally,
mangrove detection uses classical vegetation indices, such as the Normalized Difference Vegetation
Index (NDVI) [33] and the Normalized Difference Water Index (NDWI) [34,35], visual interpretation,
supervised classification, unsupervised classification and microwave imagery [7,16,23,28,36–41].
This paper presents the annual Brazilian mangrove cover status as part of a continental-scale
analysis from 1985 to 2018 based on Landsat time series data and GEE cloud computing capabilities.
To facilitate the recognition, mapping and monitoring of mangrove forests from surrounding vegetation,
Remote Sens. 2019, 11, 808 3 of 19

this study presents and verifies the robustness of a new spectral vegetation index called the MMRI, the
Modular Mangrove Recognition Index.
Remote Sens. 2018, 10, x FOR PEER REVIEW 3 of 19
2. Materials and Methods
2. Materials and Methods
Data processing and analysis occurred inside the GEE platform, as shown in Figure 1. All raster
data andData
theirprocessing and analysis
sub-products occurred
were derived inside
from thethe GEE platform,
United as shown in
States Geological Figure(USGS)
Survey 1. All raster
Landsat
data and their sub-products were derived from the United States Geological Survey (USGS) Landsat
Collection 1 Tier 1 Top of Atmosphere (TOA) data, which include Level-1 Precision Terrain
Collection 1 Tier 1 Top of Atmosphere (TOA) data, which include Level-1 Precision Terrain (L1TP)
(L1TP) processed data that have well-characterized radiometric values across the different Landsat
processed data that have well-characterized radiometric values across the different Landsat sensors
sensors [42–44].
[42–44].

Figure
Figure 1. Data-flow
1. Data-flow diagram.All
diagram. Allprocessing
processing and
and analysis
analysisoccur
occurinside thethe
inside Google Earth
Google Engine
Earth (GEE)(GEE)
Engine
platform. Steps related to image processing are in blue. The steps in green are related to the
platform. Steps related to image processing are in blue. The steps in green are related to the sample sample
design. Classification procedures are in yellow. The concordance assessment phase is in red and,
design. Classification procedures are in yellow. The concordance assessment phase is in red and, finally,
finally, the data availability is in salmon. BQA and SRTM denotes Band Quality Assessment and
the data availability is in salmon. BQA and SRTM denotes Band Quality Assessment and Shuttle Radar
Shuttle Radar Topography Mission, respectively.
Topography Mission, respectively.
For each year, Landsat TOA data were used to produce annual cloud-free composites, ranging
For each year, Landsat TOA data were used to produce annual cloud-free composites, ranging
from the 1 January to the 31 December. The cloud/shadow removal script takes advantage of the
fromquality
the 1 January
assessment to (QA)
the 31bandDecember.
and the GEEThe median
cloud/shadow removal
reducer. When script
used, QA takes
valuesadvantage
can improve of the
quality
data integrity by indicating which pixels might be affected by artefacts or subject to cloud data
assessment (QA) band and the GEE median reducer. When used, QA values can improve
integrity by indicating
contamination [45]. which pixels might
In conjunction, be affected
GEE can by artefacts
be instructed to pick or
thesubject
mediantopixelcloud contamination
value in a stack of [45].
In conjunction,
images. By doingGEE can be instructed
so, the to pick
engine rejects thethat
values median
are toopixel value
bright in clouds)
(e.g., a stack or of too
images.
dark By doing
(e.g.,
shadows)
so, the engineand picks
rejects the median
values pixel
that are toovalue in each
bright (e.g.,band overortime.
clouds) too dark (e.g., shadows) and picks the
Subsequently,
median pixel value in the
eachannual
band mosaics
over time.were sub-set to include only areas where mangrove forests
are likely to occur (e.g., low-lying
Subsequently, the annual mosaics were coastal areas and intertidal
sub-set zones)
to include andareas
only to exclude
where vast areas where
mangrove forests
mangrove forests are not expected to exist (e.g., highlands, areas distant from the shore and open
are likely to occur (e.g., low-lying coastal areas and intertidal zones) and to exclude vast areas where
ocean areas). Sub-setting allows the reduction of processing time and decreases the diversity of
mangrove forests are not expected to exist (e.g., highlands, areas distant from the shore and open ocean
flooded vegetation types, because non-coastal areas were excluded, thereby improving the overall
areas). Sub-setting allows the reduction of processing time and decreases the diversity of flooded
accuracy of the final maps.
vegetation Duetypes,
to itsbecause
length and non-coastal areas characteristics
biogeographical were excluded, thereby
[5,18], improving
the Brazilian the region
coastal overallwas
accuracy
split of
the final maps.
into six (6) distinct sites, as shown in Figure 2.
DueTheto itsfollowing
length and biogeographical
spectral indices were characteristics
calculated from[5,18],extracted
the Brazilian coastal
spectral region was
attributes: the split
into Normalized
six (6) distinct sites, asVegetation
Difference shown in Figure 2.
Index (NDVI), the Enhanced Vegetation Index (EVI) [46], the
Normalized
The following Difference
spectralWater
indicesIndex
were(NDWI),
calculatedthefrom
Modified Normalized
extracted spectralDifference
attributes:Water Index
the Normalized
(MNDWI) [47], the Normalized Difference Soil Index (NDSI) [48] and
Difference Vegetation Index (NDVI), the Enhanced Vegetation Index (EVI) [46], the Normalized the proposed MMRI. The
MMRI was inspired by the NDDI, the Normalized Difference Drought
Difference Water Index (NDWI), the Modified Normalized Difference Water Index (MNDWI) [47],Index [49], which is given by the
the following:
Normalized Difference Soil Index (NDSI) [48] and the proposed MMRI. The MMRI was inspired by
the NDDI, the Normalized Difference Drought 𝑁𝐷𝑉𝐼 − 𝑁𝐷𝑊𝐼
𝑁𝐷𝐷𝐼 = Index [49], which is given by the following: (1)
𝑁𝐷𝑉𝐼 + 𝑁𝐷𝑊𝐼
Remote Sens. 2019, 11, 808 4 of 19

NDV I − NDW I
NDDI = (1)
NDV I + NDW I
Remote Sens. 2018, 10, x FOR PEER REVIEW 4 of 19
Since the values of both the NDVI and the NDWI are between −1 and 1, the direct application of
the NDDISince
would theresult
valuesinofan undefined
both the NDVImathematical
and the NDWIlimit, −∞ to +∞.
are between −1 andThus,
1, thethe MMRI
direct was based
application of on
the NDDI
a slightly would
different result in an that
formulation undefined mathematical
uses the limit, −∞of
MNDWI instead tothe
+∞. NDWI
Thus, theandMMRI was basedof
the modulus oneach
indexawithin
slightlyadifferent
normalizedformulation thatstructure.
difference uses the MNDWI instead
Therefore, the of the NDWI
MMRI and the modulus
is a combination of each
of two classical
index within a normalized difference structure. Therefore, the MMRI is a combination
indices, a vegetation and a water index, which enhances the mangrove cover contrast. Its equation is of two classical
indices, a vegetation and a water index, which enhances the mangrove cover contrast. Its equation is
given by the following:
given by the following: | MNDW I | − | NDV I |
MMRI = |𝑀𝑁𝐷𝑊𝐼| |𝑁𝐷𝑉𝐼| (2)
| MNDW I | −
𝑀𝑀𝑅𝐼 =
+ | NDV I | (2)
|𝑀𝑁𝐷𝑊𝐼| + |𝑁𝐷𝑉𝐼|
In the sample acquisition process, the global mangrove cover data [7] was buffered (50 km) and
set as the In the sample
training acquisition
boundary. Then,process, the global
over each annualmangrove cover
composite butdata [7]inside
solely was buffered (50 km) training
the narrowed and
set as the training boundary. Then, over each annual composite but solely inside the narrowed
region, a K-means cluster analysis was run, resulting in a refined sampling area for the mangrove and
training region, a K-means cluster analysis was run, resulting in a refined sampling area for the
non-mangrove classes.
mangrove and non-mangrove classes.

FigureFigure 2. The
2. The Brazilian
Brazilian coastalregion
coastal region was
was split
splitinto
intosixsix
(6)(6)
distinct sectors.
distinct Site 1—Amapá;
sectors. Site 2— Site
Site 1—Amapá;
Marajó Island; Site 3—Pará/Maranhão; Site 4—Piauí/Bahia; Site 5—Espírito Santo/São
2—Marajó Island; Site 3—Pará/Maranhão; Site 4—Piauí/Bahia; Site 5—Espírito Santo/São Paulo; Paulo; and Site and
6—Paraná/Laguna (SC). The north and south boundaries of the Brazilian mangroves are in green. The
Site 6—Paraná/Laguna (SC). The north and south boundaries of the Brazilian mangroves are in green.
black grids represent the WRS-2 Path and Row footprints. The following acronyms represent the
The black grids represent the WRS-2 Path and Row footprints. The following acronyms represent the
Brazilian coastal states: AL (Alagoas), AP (Amapá), BA (Bahia), CE (Ceará), ES (Espírito Santo), MA
Brazilian coastal states: AL (Alagoas), AP (Amapá), BA (Bahia), CE (Ceará), ES (Espírito Santo), MA
(Maranhão), PA (Pará), PB (Paraíba), PE (Pernambuco), PI (Piauí), PR (Paraná), RJ (Rio de Janeiro),
(Maranhão), PA (Pará), PB (Paraíba), PE (Pernambuco), PI (Piauí), PR (Paraná), RJ (Rio de Janeiro), RN
RN (Rio Grande do Norte), SC (Santa Catarina), SE (Sergipe) and SP (São Paulo).
(Rio Grande do Norte), SC (Santa Catarina), SE (Sergipe) and SP (São Paulo).
Having set the refined training region, ~1000 stratified random samples were distributed per
Having set the refined
class, mangrove training region,
and non-mangrove, ~1000per
per sector stratified random
year. Once samples
collected, were distributed
the samples per class,
were statistically
mangrove and non-mangrove, per sector per year. Once collected, the samples were statistically
filtered (with the 80th percentile function) and visually inspected to remove inadequate training filtered
(withsamples.
the 80thStratified
percentile function)
sampling andand visually
statistical inspected
filtering to removetoinadequate
are necessary training samples.
address imbalanced class
problems [50], allowing the removal of outliers from the sample bag. The presence of imbalanced
Remote Sens. 2019, 11, 808 5 of 19

Stratified sampling and statistical filtering are necessary to address imbalanced class problems [50],
allowing the removal of outliers from the sample bag. The presence of imbalanced classes within the
coastal region
Remote Sens.is a probable
2018, scenario,
10, x FOR PEER REVIEWas water surface samples may, in general, greatly surpass 5 of other
19
rare class occurrences (e.g., mangrove cover).
classes within
Among the coastal
supervised region is a probable
and unsupervised scenario,
methods, GEEas water surface
presents samples
more than may, in general,
15 different classifiers.
greatly surpass other rare class occurrences (e.g., mangrove cover).
However, in the last 5 years, nearly 15,000 papers based on random forests (RFs) classifying a variety
Among supervised and unsupervised methods, GEE presents more than 15 different classifiers.
of land use/land cover classes were produced. Thus, due to its apparent robustness, the RF algorithm
However, in the last 5 years, nearly 15,000 papers based on random forests (RFs) classifying a variety
was selected here as the
of land use/land coverclassification
classes weremethod to Thus,
produced. categorize
due tothe Brazilianrobustness,
its apparent coastal zone
theinto two distinct
RF algorithm
classes: mangrove (Mg) and non-mangrove (N-Mg). This entire process was then repeated
was selected here as the classification method to categorize the Brazilian coastal zone into two distinctfor each
year from 1985 to 2018. Table 1 shows the classifier attributes and classification parameters.
classes: mangrove (Mg) and non-mangrove (N-Mg). This entire process was then repeated for each In total,
ten classification
year from 1985 attributes
to 2018. were
Table used.
1 shows the classifier attributes and classification parameters. In total,
ten classification attributes were used.
Table 1. Classifier attributes and classification parameters. In total, ten (10) distinct attributes were
used. Table
NIR, stands for near-infrared;
1. Classifier attributes and SWIR1 and SWIR2,
classification short-wave
parameters. In total,infrareds 1 and 2.
ten (10) distinct attributes were
used. NIR, stands for near-infrared; SWIR1 and SWIR2, short-wave infrareds 1 and 2.
Parameters Values
Parameters
Classifier Values
Random Forest
Classifier
Trees Random Forest 100
Trees
Samples ~1000 per100class per sector per year
Samples
Attributes ~1000
10 (Green, Red, NIR,per class per
SWIR1, sector
SWIR2, per year
NDVI, EVI, MNDWI, NDSI, MMRI)
Classes
Attributes 2 (Mangrove
10 (Green, Red, NIR, SWIR1, and Non-Mangrove)
SWIR2, NDVI, EVI, MNDWI, NDSI, MMRI)
Classes 2 (Mangrove and Non-Mangrove)
Due to the pixel-based nature of the classification method and the very long temporal series, a
Due to the pixel-based nature of the classification method and the very long temporal series, a
chain of post-classification filters was applied. The chain starts by filling in possible no-data values.
chain of post-classification filters was applied. The chain starts by filling in possible no-data values.
In a long time series of severely cloud-affected regions, such as tropical coastal zones, it is expected
In a long time series of severely cloud-affected regions, such as tropical coastal zones, it is expected
that no-data values may populate some of the resultant median composite pixels. In this filter,
that no-data values may populate some of the resultant median composite pixels. In this filter, no-
no-datadata values(“gaps”)
values (“gaps”)arearetheoretically
theoretically not allowedand
not allowed andare
arereplaced
replacedbyby the
the temporally
temporally nearest
nearest valid
valid
classification. In this
classification. procedure,
In this procedure,if no
if no“future”
“future”validvalidposition
position isis available,
available,then
thenthe the no-data
no-data value
value is is
replaced by itsbyprevious
replaced valid
its previous class.
valid UpUp
class. totothree
threeprior
prioryears
years can
can bebe used
usedtotofill
fillininpersistent
persistent no-data
no-data
positions. Therefore,
positions. gapsgaps
Therefore, should onlyonly
should existexist
if a ifgiven pixel
a given hashas
pixel been
beenpermanently
permanentlyclassified
classifiedasasno-data
no-
data throughout the entire temporal domain. To keep track of pixel temporal origins,
throughout the entire temporal domain. To keep track of pixel temporal origins, a mask of years was a mask of years
built, was built, as
as shown inshown
Figurein3.Figure 3.

FigureFigure 3. Gap-filling mechanism. The next valid classification replaces existing no-data values. If no
3. Gap-filling mechanism. The next valid classification replaces existing no-data values. If no
“future” valid position is available, then the no-data value is replaced by its previous valid
“future” valid position is available, then the no-data value is replaced by its previous valid classification,
classification, based on up to a maximum of three (3) prior years. To keep track of pixel temporal
based on up to a maximum of three (3) prior years. To keep track of pixel temporal origins, a mask of
origins, a mask of years was built.
years was built.
Remote Sens. 2019, 11, 808 6 of 19
Remote Sens. 2018, 10, x FOR PEER REVIEW 6 of 19

After gap
After gapfilling,
filling, a temporal
a temporal filterexecuted.
filter was was executed. The filter
The temporal temporal filter uses
uses sequential sequential
classifications
classifications
in in a three-yearmoving
a three-year unidirectional unidirectional
windowmoving window
to identify to identify
temporally temporally
non-permitted non-permitted
transitions. Based
transitions. Based on a single generic rule (GR), the temporal filter inspects the central
on a single generic rule (GR), the temporal filter inspects the central position of three consecutive yearsposition of
three consecutive years (“ternary”), and if the extremities of the ternary are identical but
(“ternary”), and if the extremities of the ternary are identical but the centre position is not, then the the centre
positionpixel
central is not, then the central
is reclassified pixelitsistemporal
to match reclassified to match
neighbour its temporal
class, as shownneighbour
in Table 2.class, as shown
in Table 2.
Table 2. The temporal filter inspects the central position of three consecutive years, and in cases of
Table 2. extremities,
identical The temporal thefilter inspects
centre theiscentral
position position
reclassified of three
to match consecutiveT1,
its neighbour. years, andT3instand
T2 and casesfor
of
identical extremities,
positions one (1), twothe
(2)centre position
and three is reclassifiedGR
(3), respectively. to match
meansits neighbour.
“generic T1,
rule”, T2 and
while MgT3andstand for
N-Mg
positions mangrove
represent one (1), two
and(2)non-mangrove
and three (3), respectively.
pixels. GR means “generic rule”, while Mg and N-Mg
represent mangrove and non-mangrove pixels.
Input (Year) Output
Rule Input (Year) Output
Rule T1 T2 T3 T1 T2 T3
T1 T2 T3 T1 T2 T3
GR Mg N-Mg Mg Mg Mg Mg
GR GRN-Mg Mg N-Mg
Mg Mg
N-Mg MgN-Mg Mg N-Mg
Mg N-Mg
GR N-Mg Mg N-Mg N-Mg N-Mg N-Mg

Next,
Next, aa spatial
spatial filter
filter was
was applied.
applied. To To avoid
avoid unwanted
unwanted modifications
modifications to
to the
the edges
edges of
of the
the pixel
pixel
groups
groups (blobs), a spatial filter was built based on the “connectedPixelCount” function. Native to
(blobs), a spatial filter was built based on the “connectedPixelCount” function. Native to the
the
GEE platform, this function locates connected components (neighbours) that
GEE platform, this function locates connected components (neighbours) that share the same pixel share the same pixel
value.
value. Thus,
Thus, only
only pixels
pixels that
that do not share
do not share connections
connections toto aa predefined
predefined number
number of identical neighbours
of identical neighbours
are
are considered isolated, as shown in Figure 4. In this filter, at least ten connected pixels are needed
considered isolated, as shown in Figure 4. In this filter, at least ten connected pixels are needed toto
reach
reach the minimum connection value. Consequently, the minimum mapping unit is directly affected
the minimum connection value. Consequently, the minimum mapping unit is directly affected
by
by the
the spatial
spatial filter
filter applied,
applied, andand itit was
was defined
defined as
as 10
10 pixels
pixels (~1
(~1 ha).
ha).

Figure 4.4.The
Figure Thespatial
spatial filter
filter removes
removes pixels
pixels that
that do notdo notneighbours
share share neighbours of value.
of identical identical
Thevalue. The
minimum
minimum connection
connection value was 10 value was 10 pixels.
pixels.

The last step

step of
of the
thefilter
filterchain
chainisisthethefrequency
frequency filter,
filter, as as shown
shown in Figure
in Figure 5. This
5. This filterfilter
takestakes
into
into consideration
consideration the the mangrove
mangrove occurrence
occurrence frequency
frequency throughout
throughout theentire
the entiretime
time series.
series. Thus,
Thus, all
mangrove
mangrove occurrences
occurrenceswith withless
less than
than 10%
10% temporal
temporal persistence
persistence(3 (3 years
years or
or fewer
fewer out
out of 33) are filtered
out and
and incorporated
incorporated into into the
the non-mangrove
non-mangrove class. class. This mechanism contributes to reducing the the
temporal
temporal oscillation of the mangrove signal, decreasing decreasing the number of false positives and preserving
consolidated
consolidated mangrove
mangrove pixels.
pixels.
Finally,
Finally, contingency
contingency table
table analyses
analyses [51,52]
[51,52] were
were used
used toto compare
compare twotwo distinct
distinct products:
products: (1) thethe
MMRI spectral
spectral separability
separability andand (2)
(2) the
theBrazilian
Brazilianmangrove/non-mangrove
mangrove/non-mangrove map. map. The contingency
contingency
table cross-tabulates the Mg and N-Mg classes of the reference map and the classified products; thus,
a map-to-map
map-to-map analysis
analysis waswas performed.
performed. From From the contingency table, the following metrics were were
calculated: the overall agreement (OA); the mangrove positive disagreement proportion (MgPDp),
which
which refers
refers to
to the
the proportion
proportion of of mangrove
mangrove pixels
pixels that
that are
are exclusive
exclusive to
to the
the classified
classified maps;
maps; and and the
the
mangrove negative disagreement proportion (MgNDp), which is the proportion of mangrove pixels
exclusive
exclusive toto the
the reference
reference map.
map.
Remote Sens. 2019, 11, 808 7 of 19
Remote Sens. 2018, 10, x FOR PEER REVIEW 7 of 19

Figure 5. Red,
Figure 5. Red, yellow
yellow and
and green
green represent
represent mangrove
mangrove pixels
pixels with
with high
high (23
(23 or
or more years, yy ≥
more years, 23),
≥ 23),
average (between 11 and 22 years, 11 ≤ y ≤ 22) and low (ten years or less, y < 11) occurrence
average (between 11 and 22 years, 11 ≤ y ≤ 22) and low (ten years or less, y < 11) occurrence frequencies,
respectively. The top image
frequencies, respectively. shows
The mangrove
top image shows pixels beforepixels
mangrove applying theapplying
before frequency filter.
the The bottom
frequency filter.
image shows mangrove pixels after applying the frequency filter. The black boxes are centred on areas
The bottom image shows mangrove pixels after applying the frequency filter. The black boxes are
that have been significantly affected by the filter. Note that all mangrove occurrences with less than
centred on areas that have been significantly affected by the filter. Note that all mangrove occurrences
10% temporal persistence (3 years in 33 possible years) were filtered out.
with less than 10% temporal persistence (3 years in 33 possible years) were filtered out.

To examine the performance of the MMRI mangrove separability, a comparison of distinct

To examine the performance of the MMRI mangrove separability, a comparison of distinct
mangrove and classical vegetation indices was performed. Each map was constructed based on a
mangrove and classical vegetation indices was performed. Each map was constructed based on a
single-attribute classifier, in which only one of the following spectral indices per execution were
single-attribute classifier, in which only one of the following spectral indices per execution were
considered: Combined Mangrove Recognition Index (CMRI) [38,53], NDVI [33], NDWI [34] or MMRI.
considered: Combined Mangrove Recognition Index (CMRI) [38,53], NDVI [33], NDWI [34] or MMRI.
The CMRI [53] and the NDWI-NDVI difference [38] are technically the same, as both are simple
The CMRI [53] and the NDWI-NDVI difference [38] are technically the same, as both are simple
arithmetical differences between the NDVI and the NDWI, in which the order of the variables changes
arithmetical differences between the NDVI and the NDWI, in which the order of the variables
from “NDV I − NDW I” to “NDW I − NDV I”. This rearrangement imposes a histogram brightness
changes from “ 𝑁𝐷𝑉𝐼 − 𝑁𝐷𝑊𝐼 ” to “ 𝑁𝐷𝑊𝐼 − 𝑁𝐷𝑉𝐼 ”. This rearrangement imposes a histogram
reversal but does not amplify the separability between the mangrove and non-mangrove strata. Thus,
brightness reversal but does not amplify the separability between the mangrove and non-mangrove
for the sake of simplicity, only the NDVI [33], the NDWI [34], the CMRI [53] and the MMRI were
strata. Thus, for the sake of simplicity, only the NDVI [33], the NDWI [34], the CMRI [53] and the
further compared.
MMRI were further compared.
Each classification was trained based on the same set of samples, with a total of one thousand
Each classification was trained based on the same set of samples, with a total of one thousand
(1000) training samples collected, five hundred (500) per class. The training sample distribution obeyed
(1000) training samples collected, five hundred (500) per class. The training sample distribution
a stratified strategy, where the mangrove/non-mangrove data obtained from the Socioeconomic Data
obeyed a stratified strategy, where the mangrove/non-mangrove data obtained from the
and Applications Center (SEDAC) [54] served as stratifying reference. To assess the performance of
Socioeconomic Data and Applications Center (SEDAC) [54] served as stratifying reference. To assess
each index in distinguishing between mangrove and non-mangrove pixels and defining how distinct
the performance of each index in distinguishing between mangrove and non-mangrove pixels and
are the mangrove/non-mangrove strata, a contingency table analysis was performed and the kappa
defining how distinct are the mangrove/non-mangrove strata, a contingency table analysis was
coefficient (KC) and the Bhattacharya coefficient (BC) were computed.
performed and the kappa coefficient (KC) and the Bhattacharya coefficient (BC) were computed.
The testing dataset used to evaluate the performance of each index was independent of the
The testing dataset used to evaluate the performance of each index was independent of the
training dataset. The testing dataset was designed based on a new round of stratified random sampling
training dataset. The testing dataset was designed based on a new round of stratified random
sampling over the same reference [54]. This analysis took place along the Caucaia mangrove patch in
the metropolitan region of Fortaleza in the Northeast Region of Brazil.
Remote Sens. 2019, 11, 808 8 of 19

over the same reference [54]. This analysis took place along the Caucaia mangrove patch in the
metropolitan region of Fortaleza in the Northeast Region of Brazil.
The agreement level of the national-scale maps was derived from three distinct dataset sources:
the SEDAC [54] dataset, serving as a reference for the year 2000 mapping; the Global Mangrove Watch
(GMW) Project [16], providing the 2010 reference data; and the Chico Mendes Institute for Biodiversity
Conservation (ICMBio)—Brazilian Atlas of Mangroves [29], which was used to evaluate the 2013
mapping agreement. As before, all samples used to assess the national agreement measures were
entirely independent of the training dataset. A total of twenty thousand (20,000) samples per year
were used, with ten thousand (10,000) samples in each class. Finally, all the results, from the auxiliary
products to the final annual maps, were made publicly available in raster and vector formats on a Web
platform: www.solved.eco.br/mangroveplatform.

3. Results

3.1. Modular Mangrove Recognition Index (MMRI)

Spectral indices may be described as the mathematical combinations of bands used to indicate the
relative abundance of features of interest or to enhance the spectral separability between them. With
this in mind, the MMRIs were compared to already existing mangrove indices and classical vegetation
indices, such as the CMRI [53], the NDVI [33] and the NDWI [34].
Based on a set of modular operations and to present a normalized difference structure (normalized
ratio), the MMRI imposes an enhancement of the mangrove/non-mangrove contrast level while
simultaneously redefining the resultant mathematical limits to fit the usual −1 to 1 interval. Figure 6
shows the visual aspect of each index, as well as a contingency table from which the overall agreement,
the mangrove positive disagreement proportion, the mangrove negative disagreement proportion
and the kappa coefficient were calculated, as well as the Bhattacharya coefficient, which denotes the
spectral distance between mangrove and non-mangrove pixels.
Regardless of the aspect to be considered—whether visual characteristics, contingency table
attributes or the spectral distinction between mangrove and non-mangrove classes—the classification
based exclusively on the MMRI achieved more robust results than did those based solely on NDVI,
NDWI or the difference between these two. All the metrics, the overall agreement, the mangrove
positive disagreement proportion, the mangrove negative disagreement proportion and the kappa
coefficient were better or equivalent in the MMRI classification (Figure 6).
Regarding the agreement metrics, the OA and the KC, the values reported by the MMRI mapping
were ~1.5 and ~4 times higher, respectively, than the values reported by the remaining indices. Likewise,
the disagreement metrics were ~4 times lower in the MMRI results, which indicates a more significant
agreement between the MMRI classification and the reference used.
From the perspective of mangrove and non-mangrove spectral separability, the Bhattacharya
coefficient applied to the MMRI reached the value of 0.22, which is the lowest among all the calculated
coefficients (Figure 6). This suggests that the referred index provides greater distinction between the
mangrove and non-mangrove strata compared to the other evaluated indices. The BC is a spectral
distance metric that measures the overlap between two populations or samples, where a value of zero
represents no overlap and a value of one represents perfect overlap [55].
Remote
RemoteSens. 2018,11,
Sens.2019, 10,808
x FOR PEER REVIEW 9 of
10 of19
19

Figure6.6. The
Figure The first
firstrow
rowshows
showsthethevisual
visualaspects
aspectsof
ofeach
eachindex—NDVI,
index—NDVI, NDWI,
NDWI, CMRI
CMRI and
and MMRI.
MMRI. The
The second
second row
row shows
shows the
theclassification
classification results.
results. The
The bottom
bottom row
row
shows the
shows the contingency
contingency tables,
tables, where
where Mg
Mg stands
stands for
for the
themangrove
mangrove class
classand
andN-Mg
N-Mgstands
standsfor
forthe
thenon-mangrove
non-mangroveclass.
class. OA
OA stands
stands for
for overall
overall agreement,
agreement, MgPDp
MgPDp
indicatesmangrove
indicates mangrovepositive
positive disagreement
disagreement proportion,
proportion, MgNDp
MgNDp refers
refers to mangrove
to mangrove negative
negative disagreement
disagreement proportion,
proportion, KC means
KC means the coefficient
the kappa kappa coefficient and BC
and BC denotes
denotes the Bhattacharya
the Bhattacharya coefficient. coefficient.
Remote
Remote Sens. 2018, 11,
Sens. 2019, 10, 808
x FOR PEER REVIEW 10 of
11 of 19
19

3.2. Brazilian Mangrove Status

3.2. Brazilian Mangrove Status
For the first time in the scientific literature, Brazilian mangrove cover mapping was
For the first time in the scientific literature, Brazilian mangrove cover mapping was systematically
systematically and continuously carried out, producing maps and annual statistics ranging from 1985
and continuously carried out, producing maps and annual statistics ranging from 1985 to 2018, as
to 2018, as shown in Figure 7. Data of this nature allow a better understanding of the dynamics of
shown in Figure 7. Data of this nature allow a better understanding of the dynamics of Brazilian
Brazilian mangroves, updating the status of this type of coverage throughout the country over a span
mangroves, updating the status of this type of coverage throughout the country over a span of 33 years.
of 33 years.

Brazilian mangrove
Figure 7. Brazilian mangrove areas
areas from
from 1985
1985 to
to 2018.
2018. The x-axisrepresents
The x-axis representsthe
theyears
yearsofofthe
the study.
study.
The y-axis represents total area
area of Brazilian mangroves in ten thousand kilometers squared. The
of Brazilian mangroves in ten thousand kilometers squared.
numbers above each bar are the area values for that
that bar.
bar.

In Brazil, 2 in 1985 and ~9900 km2 in 2018, representing

In Brazil, mangroves
mangroves occupied
occupied areas areas of of ~9700
~9700 kmkm2 in 1985 and ~9900 km2 in 2018, representing
an area 2
an area increase
increase of of ~200
~200 km
km2.. ItIt is
is imperative,
imperative, however,
however, not not to
to rush
rush the
the interpretation
interpretation thatthat Brazilian
Brazilian
mangroves have experienced a regeneration movement over the past
mangroves have experienced a regeneration movement over the past three decades. The percentage three decades. The percentage
difference between
difference between 1985 1985 and
and 2018
2018 is is ~2%.
~2%.
Compared to previous mappings
Compared to previous mappings [16,29,54], [16,29,54], which
which refer
refer toto the
the years
years 2000,
2000, 2010
2010 and
and 2013,
2013,
respectively, the areas reported here are 4% more, 7% more and 28%
respectively, the areas reported here are 4% more, 7% more and 28% less, respectively. In absolute less, respectively. In absolute
terms, the
terms, the difference
difference for for the
the year
year 20002000 was
was ~400 km22,,with
~400 km withthe
thevalue
valuereported
reported by by the
the reference
reference reaching
reaching
~9600 km 2 and the value calculated here was ~10,010 km 2 . For the year 2010, the difference was
~9600 km and the value calculated here was ~10,010 km . For the year 2010, the difference was 800
2 2

8002,km 2 , with the reference indicating 10,9002km2 , and this manuscript yielding 10,100 2km2 . Finally,
km with the reference indicating 10,900 km , and this manuscript yielding 10,100 km . Finally, for
for 2013, 2 , with 13,989 km 2 according to the reference and ~10,020 km22
2013, thethe valuesvalues differed
differed byby ~3900
~3900 kmkm2, with 13,989 km2 according to the reference and ~10,020 km
calculated
calculated by us.by us.
The overall
The overall pattern
patternshows
showsananupward upward trend
trend(Figures
(Figures7 and
7 and8). However,
8). However, fromfrom
19851985
to March 1998,
to March
only the Landsat-5 satellite remained operational. In this period, for
1998, only the Landsat-5 satellite remained operational. In this period, for the BCZ, the averagethe BCZ, the average number of
images per year was ~500. In the last decade between 2008 and 2018, this
number of images per year was ~500. In the last decade between 2008 and 2018, this figure tripled to figure tripled to ~1500 images
per
~1500year,
imagesas shown in Figure
per year, 8. in Figure 8.
as shown
On a regional scale, the statesofofMaranhão,
On a regional scale, the states Maranhão,Pará, Pará,Amapá
Amapá and Bahia
and Bahiaareare
thethefederal units
federal withwith
units the
most extensive mangrove cover in the country, reaching ~4350 km 2 (~46% of the national coverage),
the most extensive mangrove cover in the country, reaching ~4350 km (~46% of the national 2

~2100 km2 ~2100

coverage), (~22%), km~830 km2 ~830
2 (~22%), (~9%)km and ~ 670and
2 (~9%) km2~ (~7%),
670 kmrespectively, in 2018. Together,
2 (~7%), respectively, the four
in 2018. Together,
states represent ~85% of Brazil’s mangrove cover, as shown
the four states represent ~85% of Brazil’s mangrove cover, as shown in Figure 9. in Figure 9.
Remote Sens. 2019, 11, 808 11 of 19
Remote Sens. 2018, 10, x FOR PEER REVIEW 12 of 19
Remote Sens. 2018, 10, x FOR PEER REVIEW 12 of 19

Figure
Figure 8. The top
8. The top image
imageshows
showsthe thegeneral
generalpattern
patternofoffluctuation
fluctuationinin
thethe mangrove
mangrove areas
areas from
from 1985
1985 to
Figure
to
2018. 8.The
2018.
The The
solidtop
solid image
line line shows the
represents
represents general
thethe pattern
variation
variation of fluctuation
in mangrove
in mangrove inover
areaarea
overthe mangrove
time.
time. TheThe areas
dotted
dotted from
line line 1985
showsshowsto
the
the general
2018. The
general trend.
solid
trend. lineThe
The barsbars
showshow
represents thethe
the distribution
variation
distribution of Landsat
inofmangrove
Landsat images
area
imagesover along
time.
along the
theThe
timetime
dotted series.
series. line L5 stands
L5 shows
stands the
for
for Landsat-5,
general trend. L7
The refers
bars to Landsat-7
show the and L8
distribution stands
of for
Landsat Landsat-8.
images The
along bottom
the time
Landsat-5, L7 refers to Landsat-7 and L8 stands for Landsat-8. The bottom left image shows the left image
series. L5 shows
stands the
for
mangrove
mangrove area
Landsat-5, L7 variation
arearefers
variation and
andtrend
to Landsat-7 line from
andline
trend 1985 to
L8 from
stands 1998.
for
1985 to The bottom
Landsat-8.
1998. The
The right image
bottom
bottom shows
left
right image
image theshows
mangrove
shows the
the
areas and
mangrove trends from
area variation
mangrove areas 1999
and trendsand to 2018.
from trend
1999line from 1985 to 1998. The bottom right image shows the
to 2018.
mangrove areas and trends from 1999 to 2018.
50 46.46
50 46.46
40
40
30
30 22.28
%%

22.28
Area

20
Area

20 8.87 7.10
10 8.87 7.10 2.14 1.95 1.88 1.79
10 1.51 1.36 1.26 1.05 0.69 0.65 0.57 0.45
0 2.14 1.95 1.88 1.79 1.51 1.36 1.26 1.05 0.69 0.65 0.57 0.45
0 MA PA AP BA SE PR SP CE PE RN PB RJ ES SC PI AL
MA PA AP BA SE PR SPFederal CE PE UnitsRN PB RJ ES SC PI AL
Federal Units
Figure 9. Brazilian mangrove area in percentage of total mangrove cover per state. The x-axis
Figure 9. Brazilian
9.
represents Brazilianmangrove
the coastal states’area
mangrove in percentage
area of total
in percentage
two-letter of mangrove
abbreviations, total
as in cover per
mangrove
Figure state.per
2.cover
The x-axis
Thestate.
y-axis represents
The x-axis
represents the
the coastal
represents states’
the two-letter
coastal abbreviations,
states’
mangrove area % of the year 2018. two-letter as in Figure
abbreviations, 2. The
as y-axis
in represents
Figure 2. The the mangrove
y-axis area
represents %the
of
the year 2018.
mangrove area % of the year 2018.
From the perspective of mangrove cover persistence, ~75% of the Brazilian mangroves remained
From thethe perspective ofof mangrove
mangrove cover
cover persistence,
persistence, ~75%
~75% of of the
the Brazilian
Brazilian mangroves
mangroves remained
remained
unchanged forperspective
From two decades or more, ~10% remained stable between one and two decades and ~15%
unchanged for
unchanged for two decades
decades or
or more,
more, ~10% remained stable
stable between
between one one and
and two
two decades
decades and
and ~15%
~15%
remained stabletwo for ten or fewer ~10%
years. remained
In this scenario, the state of Amapá is the state that shows the
remained
remained stable for
stable for ten or fewer
ten or fewer years.
years. In this scenario, the state of Amapá is the state that shows
lowest stability, at ~40%, followed by In this
the scenario,
states of thethe state of Amapá
south-eastern andis southern
the state that shows
regions of the
the
the lowest
lowest stability,
stability, at at ~40%,
~40%, followed
followed by by the
the states
states ofofthe
thesouth-eastern
south-easternand andsouthern
southernregions
regions of
of the
the
country, Espírito Santo, Rio de Janeiro, São Paulo and Santa Catarina, which all present levels of
country, Espírito 60%
stability between Santo, Rio
and de as
65%, Janeiro,
shownSão Paulo 10.
in Figure and Santa Catarina, which all present levels of
stability between 60% and 65%, as shown in Figure 10.
Remote Sens. 2019, 11, 808 12 of 19

country, Espírito Santo, Rio de Janeiro, São Paulo and Santa Catarina, which all present levels of
stability between 60% and 65%, as shown in Figure 10.
Remote Sens. 2018, 10, x FOR PEER REVIEW 13 of 19

Figure 10.
Figure Brazilian
10.Brazilian mangrove
mangrove cover
cover persistence
persistence at national
at the the national and regional
and regional scale.scale.
The topThe
bartop
showsbar
shows
the the overall
overall mangrove
mangrove persistence.
persistence. The bottom
The bottom graphgraph shows
shows thethe mangrove
mangrove persistenceper
persistence perstate,
state,
abbreviated as in Figure 2. The x-axis represents the state distributions, whereas the y-axis represents
abbreviated as in Figure 2. The x-axis represents the state distributions, whereas the y-axis represents
the mangrove
the mangrovecover
covertemporal
temporal persistence percentages
persistence (%).(%).
percentages BlackBlack
represents 20 years
represents 20 or moreorofmore
years stability,
of
stability, dark grey indicates stability between 10 and 20 years and light grey represents stabilitythan
dark grey indicates stability between 10 and 20 years and light grey represents stability for less for
10 years.
less than 10 years.

Since the 1980s, mangrove cover mapping at diverse scales has been carried out, although in
Since the 1980s, mangrove cover mapping at diverse scales has been carried out, although in a
a non-systematic way. Here, three sets of data—two global datasets [16,54] from 2000 and 2010,
non-systematic way. Here, three sets of data—two global datasets [16,54] from 2000 and 2010,
respectively, and one national dataset [29] from 2013—were used to test the concordance levels
respectively, and one national dataset [29] from 2013—were used to test the concordance levels
between the mapping presented in this article and the references cited above. Figure 11 below shows
between the mapping presented in this article and the references cited above. Figure 11 below shows
the contingency tables for the years 2000, 2010 and 2013.
the contingency tables for the years 2000, 2010 and 2013.
Compared to the year 2000 reference data [54], the mapping developed herein achieves an overall
Compared to the year 2000 reference data [54], the mapping developed herein achieves an
agreement of 87% (OA = 0.87) and a kappa coefficient of 74% (KC = 0.74), and presents a low proportion
overall agreement of 87% (OA = 0.87) and a kappa coefficient of 74% (KC = 0.74), and presents a low
of mangrove negative disagreement, MgNDp = 2% (0.02). There are, however, a large number of
proportion of mangrove negative disagreement, MgNDp = 2% (0.02). There are, however, a large
positive disagreements, namely, 2474 pixels or ~25% (0.25).
number of positive disagreements, namely, 2474 pixels or ~25% (0.25).
In relation to the 2010 reference [16], the mapping developed herein also achieves a great level
In relation to the 2010 reference [16], the mapping developed herein also achieves a great level
of overall agreement, with OA = 85% (0.85) followed by a kappa coefficient of 70%, KC = 0.70. The
of overall agreement, with OA = 85% (0.85) followed by a kappa coefficient of 70%, KC = 0.70. The
MgNDp reached ~1% (0.01). On the other hand, the proportion of positive disagreements was ~29%
MgNDp reached ~1% (0.01). On the other hand, the proportion of positive disagreements was ~29%
(0.29).
(0.29).
For the year 2013, based on the data produced by the Chico Mendes Institute for Biodiversity
For the year 2013, based on the data produced by the Chico Mendes Institute for Biodiversity
Conservation (ICMBio) [29], the smallest kappa coefficient and overall agreement were achieved, KC
Conservation (ICMBio) [29], the smallest kappa coefficient and overall agreement were achieved, KC
= 60% (0.60) and OA = 80% (0.80), and the highest negative disagreement proportion was reached,
= 60% (0.60) and OA = 80% (0.80), and the highest negative disagreement proportion was reached,
MgNDp = 40% (0.40). The proportion of positive disagreement was low, MgPDp = 1% (0.01).
MgNDp = 40% (0.40). The proportion of positive disagreement was low, MgPDp = 1% (0.01).
 Remote Sens. 2019, 11, 808 13 of 19
Remote Sens. 2018, 10, x FOR PEER REVIEW 14 of 19

Figure
Figure11.11.On
Onthe
theleft
leftare
arethe
the2000,
2000,2010
2010and
and2013
2013median
mediancomposites.
composites.InInthe
thecentral
centralportion,
portion, red
redrepresents
representsthe
themangroves
mangroves ofofthe
thereference
referencedata,
data,blue
bluedenotes
denotesthe
the
classifiedmangroves
classified mangrovesand andgreen
greenrepresents
represents the
the agreement
agreement between blues
blues and
and reds.
reds. The
Thecontingency
contingencytables
tables(right)
(right)show
showthe
theagreement
agreement levels
levelsbetween
betweenthethe
reference and
reference
classified
and data.
classified Values
data. Values on on
thethe
main diagonal
main diagonalareare
thethe
numbers
numbers of of
concordant
concordantpixels. OnOn
pixels. thethe
off-diagonal,
off-diagonal,those
thoseabove
aboveareare
positive
positive differences
differences and
and those
thosebelow
beloware
negative
are negative differences.
differences.OA OAstands
standsfor
foroverall
overallagreement,
agreement,MgPDp
MgPDp means
means the mangrove positive
positive disagreement
disagreementproportion,
proportion,MgNDp
MgNDprefers referstotothe
themangrove
mangrovenegative
negative
disagreementproportion
disagreement proportionand andKC KCdenotes
denotesthethekappa
kappacoefficient.
coefficient.
Remote Sens. 2019, 11, 808 14 of 19
Remote Sens. 2019, 11, x FOR PEER REVIEW 15 of 19

3.3. Filter-Chain
3.3. Filter-Chain Influence
Influence
In total,
In total, four
four different
different filters
filters may
may have
have influenced
influenced aa resultant
resultant classification: gap-fill, temporal,
classification: gap-fill, temporal,
spatial and frequency filters. The first one, gap-fill, may only add mangrove data
spatial and frequency filters. The first one, gap-fill, may only add mangrove data into a given year;into a given year; it
it
is an
is an exclusively
exclusively positive
positive filter
filter (PF).
(PF). The
The last
last two,
two, the
the spatial
spatial and
and frequency
frequency filters,
filters, will
will only
only remove
remove
mangrove pixels from a given year and are therefore negative filters (NF). The
mangrove pixels from a given year and are therefore negative filters (NF). The temporal filter is temporal filter is the
the
only filter that may add or remove mangrove classifications, depending on the
only filter that may add or remove mangrove classifications, depending on the pixel temporal pixel temporal trajectory.
Below, Figure
trajectory. 12 and
Below, Table123 and
Figure categorized
Table 3and aggregated
categorized andeach pixel, year
aggregated by year,
each pixel,according
year by year,to its
filtered or no-change aspect, which varied as follows: positive filter (mangrove
according to its filtered or no-change aspect, which varied as follows: positive filter (mangrove addition), negative
filter (mangrove
addition), negativeremoval) or unmodified
filter (mangrove (no change).
removal) or unmodified (no change).

Figure 12. On the left is the 1988 MMRI median composite: dark green represents mangroves, light
Figure 12. On the left is the 1988 MMRI median composite: dark green represents mangroves, light
green is non-mangroves and blue is water. In the centre is the 1988 Landsat-5 band 543 median
green is non-mangroves and blue is water. In the centre is the 1988 Landsat-5 band 543 median
composite. On the right, PF denotes a positive filter (mangrove addition), NF stands for a negative
composite. On the right, PF denotes a positive filter (mangrove addition), NF stands for a negative
filter (mangrove removal) and NC stands for no change.
filter (mangrove removal) and NC stands for no change.
Table 3. Annual filter-affected area, in percentages. PF% indicates a positive filter (mangrove addition),
Table 3. Annual
PN% stands filter-affected
for a negative area, in percentages.
filter (mangrove removal) and PF%
NC% indicates
denotes anaunfiltered
positive pixel
filter (no-change).
(mangrove
addition), PN% stands for a negative filter (mangrove removal) and NC% denotes an unfiltered pixel
Year 1985
(no-change). 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996
PF% 21 17 18 17 28 24 38 40 47 47 51 47
NF% Year 1985
23 18 198615 198714 1988 18 1989 1990
22 1991
13 1992131993 171994 199514 1996 09 11
NC% PF% 83
79 21 1782 18 83 17 7228 24
76 3862 40 60 47 5347 51
53 4749 53
Year NF% 1998
1997 23 18
1999 152000 14 200118 2002
22 13
2003 132004 17 2005
14 2006
09 11
2007 2008
PF%
NC%
27
79
26
8313
82 05
83 04
72 76
03
6204
60 04
53 05
53 49
03
5305 04
NF% Year 1997
10 20 1998
16 199912 2000 17
2001 2002
20 2003
18 2004182005 17
2006 2007
16 2008 17 19
NC% 73
PF% 7427 2687 13 95 05 9604 97
03 0496 04 96 05 9403 97
05 0495 96
Year NF% 2010
2009 10 20
2011 162012 12 201317 2014
20 18
2015 182016 17 2017
16 2018
17 19
PF% NC%
06 73
04 7405 87 10 95 0996 97
12 9607 96 07 94 0797 95
10 96
NF% Year 2009
13 18 2010
22 201122 2012 22
2013 2014
19 2015
19 2016232017 24
2018 10
NC% 94
PF% 9606 0495 05 91 10 9109 88
12 0793 07 93 07 9310 90
NF% 13 18 22 22 22 19 19 23 24 10
NC%percentage
The mean 94 96values
95 for 91 91 and88negative
positive 93 filters
93 were93 relatively
90 low: 17% and 18%,
respectively. On the other hand, the annual average values for the no-change pixels were quite high:
82%.The
Themean percentage
highest values
quantity of for filtering
positive positive happened
and negative filterswhereas
in 1995, were relatively low:
the lowest 17% and
appeared in 18%,
2006.
respectively. On the
The highest (24%) other
and hand,
lowest the
(9%) annual
values of average
negativevalues forappeared
filtering the no-change
in 2017pixels wererespectively.
and 1995, quite high:
82%.
FromThe
2000highest quantity
until 2018, the NCof percentages
positive filtering happened
were always in 1995,
above whereas the lowest appeared in
the average.
2006. The highest (24%) and lowest (9%) values of negative filtering appeared in 2017 and 1995,
respectively. From 2000 until 2018, the NC percentages were always above the average.
Remote Sens. 2019, 11, 808 15 of 19

4. Discussion
For the first time in the scientific literature, Brazilian mangrove cover has been systematically and
exhaustively mapped. Although data of this nature allow a better understanding of the dynamics of
mangrove coverage, it is necessary to understand the effects of the non-homogeneous distribution of
Landsat data throughout the time series to avoid superficial interpretations regarding the recorded
fluctuations in the country’s mangrove areas.
In scenarios of low data frequency, such as the period from 1985 to 1998, it is reasonable to observe
a degradation of the quality of “cloud-free composites” since these are based on the median behaviour
of pixels composing the orbital data stack. Therefore, the smaller the number of pixels for a given
position of the matrix, the higher the likelihood of including atmospheric and radiometric noise as
part of the composite, and thus, the decision of the classifier is less accurate in distinguishing between
the categories of mangrove and non-mangrove.
When compared separately, the periods from 1985 to 1998 and 1999 to 2018 show distinct trends,
as shown in Figures 8 and 9. The first period, from 1985 to 1998, shows an upward trend with a positive
slope, as shown by the equation y = 10.64x + 9827.7. Aside from environmental and human-related
changes that are present in this period, most of this pattern of growth seems to be related more to
the uneven distribution of Landsat data throughout the 33 years than to a regeneration movement of
Brazilian mangroves. From 1989 to 1998, the PF% values were greater than 20%; for five of those years,
the PF% values were greater than 40%.
The second period, from 1999 to 2018, when the PF% values are primarily smaller than 10% and
the NC% values are above 85%, a slightly negative slope is shown, presenting a trend of mangrove area
loss, which is evidenced by the equation y = −3.13x + 10070 and compatible with the global standard
expected for mangrove coverage in recent decades [56].
In absolute terms, the direct difference between the years 2000 and 2018 shows an area loss of
approximately 70 km2 : a difference of ~0.6%. However, it is interesting to note that there is a wave-like
pattern present in the variation in mangrove areas. Upon comparing the data from 2003, which has
the largest reported area of mangroves, and the 2017 data, which has the smallest quantified area,
the absolute difference is ~200 km2 , which represents an area loss of ~2%, a value almost three times
greater than that previously reported by the comparison between 2000 and 2018.
This pattern may be associated with intra-annual oceanographic and climatic variables (rainy and
dry seasons), which alter the prevailing tide and humidity conditions in the median composites [57–60].
On one-year intervals, especially over areas often covered by clouds such as the BCZ, the definition
of additional image-selection parameters (e.g., the tide condition and the dominant climatic pattern)
greatly reduces the number of images available, which makes spatial analysis over shorter time
frames impractical.
Regarding the concordance levels, when compared to the year 2000 data [54], an overall agreement
of 87% and a kappa coefficient of 74%, KC = 0.74, were reached. However, for this reference,
there is a large quantity of positive disagreements, which is reflected by the MgPDp of ~25%.
Nevertheless, most of these positive differences may be associated with unfiltered clouds and shadow
residues in the reference, which were included as part of the non-mangrove class and thus generated
positive disagreement.
In comparison to the 2010 data [16], there is significant overall agreement between the published
data and our mapping effort, reaching an overall agreement of 85% and a kappa coefficient of 70%.
However, the MgNDp is approximately 29%, which is likely to be associated with differences between
the edges of the mangrove classes (border errors) and to the inclusion of the coastal rivers’ dendritic
patterns as part of the mangrove reference class.
For 2013, the year with the greatest mangrove area difference between the reference and our map,
approximately 28% (~3900 km2 ), and the lowest overall agreement and kappa coefficient, 80% and
60%, respectively, the mangrove concept adopted by the reference was different. Unlike the previous
concepts, [29] focused on ecosystem identification rather than mangrove forests [16,54]. Thus, the
Remote Sens. 2019, 11, 808 16 of 19

omissions are quite high and the MgNDp reaches nearly 40%. Furthermore, the reference was created
through visual interpretation and, therefore, had a reduced need for the application of spatial filters.
As a result, the reported omissions are minorly associated with border differences and are mostly
attributed to the inclusion of “apicuns” (salt marshes) as part of the mangrove reference class.
Moreover, it is necessary to highlight the difficulties of implementing continental-scale mapping
in a systematic and exhaustive way. The uneven distribution of Landsat data, scarce prior to 1998
and much more frequent in the last two decades, imposes obstacles within the construction of annual
cloud-free composites because the quantity of orbital data is greatly unbalanced over time.
Increasing the composites’ temporal windows to three or more years may be important to sharply
select data that match specific climatic and oceanographic conditions. In addition, the use of multiple
satellite families (Landsat and Sentinel, for example) should also be considered as an alternative to
increase the frequency of observations.
Compared to previous mappings produced on an approximately annual timeframe, without
considering the prevailing tide and climatic conditions and focusing on mangrove vegetation mapping,
the results expressed herein are spatially coherent, reaching a kappa coefficient of 0.7 and an overall
agreement above 0.85, as seen in Figure 11.
Along with a broader temporal window, the restriction of the climatic pattern (rainy or dry), as
well as the selection of data in specific tide conditions, whether high or low, can reduce the wave-like
pattern associated with the reported mangrove areas. Furthermore, the addition of more detailed
land use/land cover classes seems to be essential to the ability to identify and separate natural coastal
changes from anthropogenic coastal changes, as a more detailed pixel trajectory would allow a more
accurate identification of human interference.
From the standpoint of MMRI robustness, compared to other indices, the MMRI seems to
provide a greater distinction between the mangrove and non-mangrove strata, which facilitates
the discrimination of such classes by automatic classifiers. The new index can be used in conjunction
with multiple families of satellites, digital elevation models (DEM) and microwave data and, whenever
possible, can incorporate information from the main domains and subdomains of remote sensing,
space, time, spectra and context.
The next steps of this research include the creation of a mechanism capable of identifying tidal
conditions from orbital data, the expansion of the analysis to the South American continent, and the
insertion of new classes of coastal environments such as “apicuns”, beaches and dunes, aquaculture
and urban areas, which would help separate human-related changes from natural coastal changes.

5. Conclusions
This manuscript detailed a GEE-managed pipeline to discriminate mangrove forests from
surrounding vegetation. The developed pipeline is scalable and suitable for large-scale mangrove
cover analyses, allowing systematic and continuous mapping of the Brazilian mangrove cover and
producing maps and annual statistics ranging from 1985 to 2018. When compared to previous
research at global and national scales, the data produced have high levels of spatial and statistical
agreement. All the data produced are available for download through the website http://www.solved.
eco.br/mangroveplatform and will be transferred to the MapBiomas project (www.mapbiomas.org) in
MapBiomas Collection 4.0.
The developed spectral index, MMRI, was demonstrated to be robust and helped to increase the
spectral separability between the mangrove and non-mangrove classes. The expansion of the temporal
window for the construction of mosaics, the restriction of the climatic pattern (rainy or dry) and the
ability to discern between the predominant tidal regimes, whether high or low, may help reduce the
area fluctuations associated with the coverage of mangroves.
Remote Sens. 2019, 11, 808 17 of 19

Author Contributions: Conceptualization, C.D. and L.C.; methodology, C.D. and L.C.; software, L.C., G.N. and
J.R.; validation, C.D., L.S. and M.A.; formal analysis, C.D.; investigation, C.D.; resources, P.W.M.S.-F.; data curation,
C.D., M.A. and P.W.M.S-F.; writing—original draft preparation, C.D.; writing—review and editing, C.D., M.A.
and P.W.M.S-F.; visualization, C.D and L.S.; supervision, P.W.M.S-F.; project administration, P.W.M.S-F.; funding
acquisition, P.W.M.S-F.
Funding: This research was funded by the Brazilian National Council for Scientific and Technological
Development (CNPq): 870005/1997-9 and the MapBiomas Project. The processing charges were covered by the
Dean of Research and Graduate Studies (PROPESP) from the Federal University of Para (UFPA).
Acknowledgments: The authors are grateful to Tasso Azevedo, general coordinator of the MapBiomas Project and
the System for Estimation of Green House Gases Emissions (SEEG) initiative, Carlos Souza Jr., senior researcher at
the Amazon Institute (IMAZON) and Washington Rocha, full professor at Feira de Santana State University (UEFS),
for their useful comments, suggestions and encouragement. The authors would like to thank the anonymous
reviewers and the editor for their helpful and constructive comments that greatly contributed to improving
this paper. Lastly, the authors acknowledge the Norwegian Government for the financial support given to the
Mapbiomas Project, via World Resources Institute (WRI), and Google for the development of the Google Earth
Engine platform.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Neumann, B.; Vafeidis, A.T.; Zimmermann, J.; Nicholls, R.J. Future Coastal Population Growth and Exposure
to Sea-Level Rise and Coastal Flooding—A Global Assessment. PLoS ONE 2015, 10, e0118571. [CrossRef]
2. Small, C.; Nicholls, R.J. A Global Analysis of Human Settlement in Coastal Zones. J. Coast. Res. 2003, 19,
584–599. [CrossRef]
3. Burke, L.; Kura, Y.; Kassem, K.; Revenga, C.; Spalding, M.; McAllister, D. Pilot Analysis of Global Ecosystems:
Coastal Ecosystems; World Recourses Institute: Washington, DC, USA, 2001.
4. Nicolodi, J.L.; Petermann, R.M. Potential vulnerability of the Brazilian coastal zone in its environmental,
social, and technological aspects. Panam. J. Aquat. Sci. 2010. [CrossRef]
5. Dominguez, J.M.L. The Coastal Zone of Brazil. In Geology and Geomorphology of Holocene Coastal Barriers of
Brazil; Springer: Berlin/Heidelberg, Germany, 2009; pp. 17–51. ISBN 978-3-540-44771-9.
6. Giri, C. Observation and Monitoring of Mangrove Forests Using Remote Sensing: Opportunities and
Challenges. Remote Sens. 2016, 8, 783. [CrossRef]
7. Giri, C.; Ochieng, E.; Tieszen, L.L.; Zhu, Z.; Singh, A.; Loveland, T.; Masek, J.; Duke, N. Status and distribution
of mangrove forests of the world using earth observation satellite data. Glob. Ecol. Biogeogr. 2011, 20, 154–159.
[CrossRef]
8. Alongi, D.M. The Energetics of Mangrove Forests, 1st ed.; Springer: Dordrecht, The Netherlands, 2009;
ISBN 978-1-4020-4270-6.
9. Alongi, D.M. Mangrove forests: Resilience, protection from tsunamis, and responses to global climate change.
Estuar. Coast. Shelf Sci. 2008, 76, 1–13. [CrossRef]
10. Murdiyarso, D.; Purbopuspito, J.; Kauffman, J.B.; Warren, M.W.; Sasmito, S.D.; Donato, D.C.; Manuri, S.;
Krisnawati, H.; Taberima, S.; Kurnianto, S. The potential of Indonesian mangrove forests for global climate
change mitigation. Nat. Clim. Chang. 2015, 5, 1089–1092. [CrossRef]
11. Saenger, P.; Hegerl, E.J.; Davie, J.D.S. Global Status of Mangrove Ecosystems; International Union for
Conservation of Nature and Natural Resources: Gland, Switzerland, 1983.
12. Alongi, D.M. Present state and future of the world’s mangrove forests. Environ. Conserv. 2002, 29, 331–349.
[CrossRef]
13. Donato, D.C.; Kauffman, J.B.; Murdiyarso, D.; Kurnianto, S.; Stidham, M.; Kanninen, M. Mangroves among
the most carbon-rich forests in the tropics. Nat. Geosci. 2011, 4, 293–297. [CrossRef]
14. Spalding, M.; Kainuma, M.; Collins, L. World Atlas of Mangroves; Taylor & Francis Group:
Abingdon-on-Thames, UK, 2010; ISBN 9781849776608.
15. Thomas, N.; Bunting, P.; Lucas, R.; Hardy, A.; Rosenqvist, A.; Fatoyinbo, T. Mapping Mangrove Extent and
Change: A Globally Applicable Approach. Remote Sens. 2018, 10, 1466. [CrossRef]
16. Bunting, P.; Rosenqvist, A.; Lucas, M.R.; Rebelo, L.-M.; Hilarides, L.; Thomas, N.; Hardy, A.; Itoh, T.;
Shimada, M.; Finlayson, M.C. The Global Mangrove Watch—A New 2010 Global Baseline of Mangrove
Extent. Remote Sens. 2018, 10, 1669. [CrossRef]
Remote Sens. 2019, 11, 808 18 of 19

17. Herz, R. Manguezais do Brasil; United States Pharmacopeia (USP): North Bethesda, MD, USA, 1991.
18. Schaeffer-Novelli, Y.; Cintrón-Molero, G.; Adaime, R.R.; de Camargo, T.M.; Cintron-Molero, G.;
de Camargo, T.M. Variability of Mangrove Ecosystems along the Brazilian Coast. Estuaries 1990, 13, 204–218.
[CrossRef]
19. Souza Filho, P.W.M. Costa de manguezais de macromaré da Amazônia: Cenários morfológicos, mapeamento
e quantificação de áreas usando dados de sensores remotos. Rev. Bras. Geofísica 2005, 23, 427–435. [CrossRef]
20. Walfir Martins Souza Filho, P.; Renato Paradella, W. Recognition of the main geobotanical features along the
Bragança mangrove coast (Brazilian Amazon Region) from Landsat TM and RADARSAT-1 data. Wetl. Ecol.
Manag. 2002, 10, 121–130. [CrossRef]
21. Souza Filho, P.W.M.; Paradella, W.R. Use of RADARSAT-1 fine mode and Landsat-5 TM selective principal
component analysis for geomorphological mapping in a macrotidal mangrove coast in the Amazon Region.
Can. J. Remote Sens. 2005, 31, 214–224. [CrossRef]
22. Rodrigues, S.W.P.; Souza-Filho, P.W.M. Use of Multi-Sensor Data to Identify and Map Tropical Coastal
Wetlands in the Amazon of Northern Brazil. Wetlands 2011, 31, 11–23. [CrossRef]
23. Nascimento Jr, W.R.; Souza-Filho, P.W.M.; Proisy, C.; Lucas, R.M.; Rosenqvist, A. Mapping changes in
the largest continuous Amazonian mangrove belt using object-based classification of multisensor satellite
imagery. Estuar. Coast. Shelf Sci. 2013, 117, 83–93. [CrossRef]
24. Santos, L.C.M.; Matos, H.R.; Schaeffer-Novelli, Y.; Cunha-Lignon, M.; Bitencourt, M.D.; Koedam, N.;
Dahdouh-Guebas, F. Anthropogenic activities on mangrove areas (São Francisco River Estuary, Brazil
Northeast): A GIS-based analysis of CBERS and SPOT images to aid in local management. Ocean Coast.
Manag. 2014, 89, 39–50. [CrossRef]
25. Queiroz, L.; Rossi, S.; Meireles, J.; Coelho, C. Shrimp aquaculture in the federal state of Ceará, 1970–2012:
Trends after mangrove forest privatization in Brazil. Ocean Coast. Manag. 2013, 73, 54–62. [CrossRef]
26. Rocha de Souza Pereira, F.; Kampel, M.; Cunha-Lignon, M. Mapping of mangrove forests on the southern
coast of São Paulo, Brazil, using synthetic aperture radar data from ALOS/PALSAR. Remote Sens. Lett. 2012,
3, 567–576. [CrossRef]
27. Pereira, F.; Kampel, M.; Cunha-Lignon, M. Mangrove vegetation structure in Southeast Brazil from phased
array L-band synthetic aperture radar data. J. Appl. Remote Sens. 2016, 10, 036021. [CrossRef]
28. Magris, R.A.; Barreto, R. Mapping and assessment of protection of mangrove habitats in Brazil. Panam. J.
Aquat. Sci. 2010, 5, 546–556.
29. ICMBio. Atlas dos Manguezais do Brasil, 1st ed.; ICMBio: Brasilia, Brazil, 2017; ISBN 978-85-61842-75-8.
30. Gorelick, N.; Hancher, M.; Dixon, M.; Ilyushchenko, S.; Thau, D.; Moore, R. Google Earth Engine:
Planetary-scale geospatial analysis for everyone. Remote Sens. Environ. 2017, 202, 18–27. [CrossRef]
31. Chen, X.; Huang, X.; Jiao, C.; Flanner, M.G.; Raeker, T.; Palen, B. Running climate model on a commercial
cloud computing environment: A case study using Community Earth System Model (CESM) on Amazon
AWS. Comput. Geosci. 2017, 98, 21–25. [CrossRef]
32. Zhu, Z.; Wang, S.; Woodcock, C.E. Improvement and expansion of the Fmask algorithm: Cloud, cloud
shadow, and snow detection for Landsats 4–7, 8, and Sentinel 2 images. Remote Sens. Environ. 2015, 159,
269–277. [CrossRef]
33. Tucker, C.J. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens.
Environ. 1979, 8, 127–150. [CrossRef]
34. McFeeters, S.K. The use of the Normalized Difference Water Index (NDWI) in the delineation of open water
features. Int. J. Remote Sens. 1996, 17, 1425–1432. [CrossRef]
35. Gao, B. NDWI—A normalized difference water index for remote sensing of vegetation liquid water from
space. Remote Sens. Environ. 1996, 58, 257–266. [CrossRef]
36. Tong, P.H.S.; Auda, Y.; Populus, J.; Aizpuru, M.; Al Habshi, A.; Blasco, F. Assessment from space of
mangroves evolution in the Mekong Delta, in relation to extensive shrimp farming. Int. J. Remote Sens. 2004,
25, 4795–4812. [CrossRef]
37. Fei, S.X.; Shan, C.H.; Hua, G.Z. Remote Sensing of Mangrove Wetlands Identification. Procedia Environ. Sci.
2011, 10, 2287–2293. [CrossRef]
38. Alsaaideh, B.; Al-Hanbali, A.; Tateishi, R.; Kobayashi, T.; Hoan, N.T. Mangrove Forests Mapping in the
Southern Part of Japan Using Landsat ETM+ with DEM. J. Geogr. Inf. Syst. 2013, 5, 9. [CrossRef]
Remote Sens. 2019, 11, 808 19 of 19

39. Nardin, W.; Locatelli, S.; Pasquarella, V.; Rulli, M.C.; Woodcock, C.E.; Fagherazzi, S. Dynamics of a fringe
mangrove forest detected by Landsat images in the Mekong River Delta, Vietnam. Earth Surf. Process.
Landforms 2016, 41, 2024–2037. [CrossRef]
40. Pham, D.T.; Yokoya, N.; Bui, T.D.; Yoshino, K.; Friess, A.D. Remote Sensing Approaches for Monitoring Mangrove
Species, Structure, and Biomass: Opportunities and Challenges. Remote Sens. 2019, 11, 230. [CrossRef]
41. Kuenzer, C.; Bluemel, A.; Gebhardt, S.; Quoc, T.V.; Dech, S. Remote Sensing of Mangrove Ecosystems: A
Review. Remote Sens. 2011, 3, 878–928. [CrossRef]
42. USGS Landsat. 8 (L8)Data Users Handbook; USGS Landsat: Sioux Falls, SD, USA, 2015.
43. Storey, J.; Choate, M.; Lee, K. Landsat 8 Operational Land Imager On-Orbit Geometric Calibration and
Performance. Remote Sens. 2014, 6, 11127–11152. [CrossRef]
44. Teillet, P.M.; Barker, J.L.; Markham, B.L.; Irish, R.R.; Fedosejevs, G.; Storey, J.C. Radiometric cross-calibration
of the Landsat-7 ETM+ and Landsat-5 TM sensors based on tandem data sets. Remote Sens. Environ. 2001, 78,
39–54. [CrossRef]
45. USGS Landsat. USGS Landsat Collection 1 Level 1 Product Definition; USGS Landsat: Sioux Falls, SD, USA,
2017; Volume 26.
46. Liu, H.Q.; Huete, A. Feedback based modification of the NDVI to minimize canopy background and
atmospheric noise. IEEE Trans. Geosci. Remote Sens. 1995, 33, 457–465. [CrossRef]
47. Xu, H. Modification of normalised difference water index (NDWI) to enhance open water features in remotely
sensed imagery. Int. J. Remote Sens. 2006, 27, 3025–3033. [CrossRef]
48. Rogers, A.S.; Kearney, M.S. Reducing signature variability in unmixing coastal marsh Thematic Mapper
scenes using spectral indices. Int. J. Remote Sens. 2004. [CrossRef]
49. Gu, Y.; Brown, J.F.; Verdin, J.P.; Wardlow, B. A five-year analysis of MODIS NDVI and NDWI for grassland
drought assessment over the central Great Plains of the United States. Geophys. Res. Lett. 2007, 34. [CrossRef]
50. Bogner, C.; Seo, B.; Rohner, D.; Reineking, B. Classification of rare land cover types: Distinguishing annual
and perennial crops in an agricultural catchment in South Korea. PLoS ONE 2018, 13, e0190476. [CrossRef]
[PubMed]
51. Olofsson, P.; Foody, G.M.; Herold, M.; Stehman, S.V.; Woodcock, C.E.; Wulder, M.A. Good practices for
estimating area and assessing accuracy of land change. Remote Sens. Environ. 2014, 148, 42–57. [CrossRef]
52. Pontius, R.G.; Santacruz, A. Quantity, exchange, and shift components of difference in a square contingency
table. Int. J. Remote Sens. 2014, 35, 7543–7554. [CrossRef]
53. Gupta, K.; Mukhopadhyay, A.; Giri, S.; Chanda, A.; Datta Majumdar, S.; Samanta, S.; Mitra, D.; Samal, R.N.;
Pattnaik, A.K.; Hazra, S. An index for discrimination of mangroves from non-mangroves using LANDSAT 8
OLI imagery. MethodsX 2018, 5, 1129–1139. [CrossRef] [PubMed]
54. Giri, C.; Ochieng, E.; Tieszen, L.L.; Zhu, Z.; Singh, A.; Loveland, T.; Masek, J.; Duke, N. Global Mangrove Forests
Distribution, 2000; NASA Socioeconomic Data and Applications Center (SEDAC): Palisades, NY, USA, 2013.
55. Ben Ayed, I.; Punithakumar, K.; Li, S. Distribution Matching with the Bhattacharyya Similarity: A Bound
Optimization Framework. IEEE Trans. Pattern Anal. Mach. Intell. 2015, 37, 1777–1791. [CrossRef] [PubMed]
56. Thomas, N.; Lucas, R.; Bunting, P.; Hardy, A.; Rosenqvist, A.; Simard, M. Distribution and drivers of global
mangrove forest change, 1996–2010. PLoS ONE 2017, 12, e0179302. [CrossRef]
57. Xia, Q.; Qin, C.-Z.; Li, H.; Huang, C.; Su, F.-Z. Mapping Mangrove Forests Based on Multi-Tidal
High-Resolution Satellite Imagery. Remote Sens. 2018, 10, 1343. [CrossRef]
58. Rogers, K.; Lymburner, L.; Salum, R.; Brooke, B.P.; Woodroffe, C.D. Mapping of mangrove extent and
zonation using high and low tide composites of Landsat data. Hydrobiologia 2017, 803, 49–68. [CrossRef]
59. Zhang, X.; Treitz, P.M.; Chen, D.; Quan, C.; Shi, L.; Li, X. Mapping mangrove forests using multi-tidal
remotely-sensed data and a decision-tree-based procedure. Int. J. Appl. Earth Obs. Geoinf. 2017, 62, 201–214.
[CrossRef]
60. Chen, B.; Xiao, X.; Li, X.; Pan, L.; Doughty, R.; Ma, J.; Dong, J.; Qin, Y.; Zhao, B.; Zhixiang, W.; et al. A mangrove
forest map of China in 2015: Analysis of time series Landsat 7/8 and Sentinel-1A imagery in Google Earth
Engine cloud computing platform. ISPRS J. Photogramm. Remote Sens. 2017, 131, 104–120. [CrossRef]

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).