remote sensing

Article
The Google Earth Engine Mangrove Mapping
Methodology (GEEMMM)
J. Maxwell M. Yancho 1 , Trevor Gareth Jones 1,2,3, *, Samir R. Gandhi 1,4 , Colin Ferster 5 ,
Alice Lin 1 and Leah Glass 1
1 Blue Ventures Conservation—Mezzanine, The Old Library, Trinity Road, St Jude’s, Bristol BS2 0NW, UK;
[email protected] (J.M.M.Y.); [email protected] (S.R.G.); [email protected] (A.L.);
[email protected] (L.G.)
2 Department of Forest Resources Management, University of British Columbia, Vancouver, BC V6T 1Z4, Canada
3 Terra Spatialists, The Blue House, 660 West 13th Avenue, Vancouver, BC V5Z 1N9, Canada
4 The Jolly Geographer, 7 Yorke Gate, Watford, Hertfordshire WD17 4NQ, UK
5 Department of Geography, University of Victoria, P.O. Box 1700 STN CSC, Victoria, BC V8W 2Y2, Canada;
[email protected]
* Correspondence: [email protected]




Received: 1 October 2020; Accepted: 9 November 2020; Published: 16 November 2020


Abstract: Mangroves are found globally throughout tropical and sub-tropical inter-tidal coastlines.
These highly biodiverse and carbon-dense ecosystems have multi-faceted value, providing critical
goods and services to millions living in coastal communities and making significant contributions
to global climate change mitigation through carbon sequestration and storage. Despite their many
values, mangrove loss continues to be widespread in many regions due primarily to anthropogenic
activities. Accessible, intuitive tools that enable coastal managers to map and monitor mangrove
cover are needed to stem this loss. Remotely sensed data have a proven record for successfully
mapping and monitoring mangroves, but conventional methods are limited by imagery availability,
computing resources and accessibility. In addition, the variable tidal levels in mangroves presents
a unique mapping challenge, particularly over geographically large extents. Here we present a
new tool—the Google Earth Engine Mangrove Mapping Methodology (GEEMMM)—an intuitive,
accessible and replicable approach which caters to a wide audience of non-specialist coastal managers
and decision makers. The GEEMMM was developed based on a thorough review and incorporation of
relevant mangrove remote sensing literature and harnesses the power of cloud computing including a
simplified image-based tidal calibration approach. We demonstrate the tool for all of coastal Myanmar
(Burma)—a global mangrove loss hotspot—including an assessment of multi-date mapping and
dynamics outputs and a comparison of GEEMMM results to existing studies. Results—including both
quantitative and qualitative accuracy assessments and comparisons to existing studies—indicate that
the GEEMMM provides an accessible approach to map and monitor mangrove ecosystems anywhere
within their global distribution.

Keywords: GEEMMM; mangroves; remote sensing; google earth engine; Myanmar; cloud computing;
digital earth

1. Introduction
Mangroves are a species of woody plants which comprise unique, halophytic communities in the
tropical and sub-tropical inter-tidal coastlines of the world [1]. When meeting accepted definitions
based on attributes including height, diameter and canopy closure, mangroves can qualify as forest [2].
Areas not qualifying as forest are peripheral parts of wider mangrove ecosystems, including expanses

Remote Sens. 2020, 12, 3758; doi:10.3390/rs12223758 www.mdpi.com/journal/remotesensing

Remote Sens. 2020, 12, 3758 2 of 35

dominated by submerged, dwarf or scrub, and fringe plants [1,3–5]. Mangrove ecosystems —both forest
and non-forest—are found in 102 countries and 21 territories [5]. The value of mangrove ecosystems
is multifaceted, including the provisioning of critical goods (e.g., fuel wood, fish, shellfish, medicine,
fiber, and timber) and services (e.g., shoreline stabilization, storm protection, and cultural, recreational
and tourism opportunities) to millions of people residing in coastal communities [6–9]. In addition,
mangrove ecosystems are incredibly biodiverse, providing habitat for numerous species, many of which
are rare, at-risk, or endangered [10–12]. Mangrove forests are also incredibly carbon-dense and meet or
exceed many of their terrestrial peers in sequestration and storage [13–15]. Increasingly, the conservation,
restoration and managed-use of mangrove ecosystems is being pursued through payments for ecosystem
services (PES) programs, including forest carbon initiatives (e.g., REDD+, Plan Vivo) [16,17].
Despite their multifaceted value, global mangrove loss is widespread. In the last two decades of
the 20th century the world lost an estimated 35% of mangrove forest cover [18]. While globally the
rate of loss has thus far slowed in the 21st century—an estimated 4% from 1996 to 2016—many parts
of the world, notably SE Asia, remain loss hotspots [19–21]. The primary driver of mangrove loss
is anthropogenic activities including aquaculture, agriculture, urban development, and unmanaged
harvest [22]. Accurate, reliable, contemporary, and easily updated information representing the extent
of mangrove ecosystems is required by decision makers and managers and to help countries pursue
and meet environmental targets (e.g., Millennium Development Goals and Ramsar Convention on
Wetlands of International Importance especially as Waterfowl Habitat) [23–25]. Remotely sensed data
have a well-established utility for mapping and monitoring the multi-date distribution of mangrove
ecosystems and quantifying change over time; however, the remote sensing of mangrove environments
has its own unique set of challenges which must be overcome to produce accurate results, including
the variable presence or absence of water associated with daily tidal fluctuations [23,26]. Fluctuating
tides can drastically influence the spectral properties of mangrove ecosystems making information on
tidal condition at time of image acquisition vital [27]. Many mangrove studies have ignored variable
tidal conditions, combining images ranging from low to high tide [23]. Recently, studies have used
image composites that include imagery acquired during selective tides (i.e., high and/or low); however,
these have covered limited areas (e.g., a single bay within a single Landsat scene) where reliable
local tidal stations or modeled tidal products are available, and have not evaluated dynamics [27–29].
Other studies demonstrated the potential to use remote sensing or models to calibrate tides across larger
areas; however, these approaches depend on substantial expertise to run specialized or customized
software and the models depend on high quality training data—which is not always available—making
them too complex and inaccessible for most potential users [30–33].
Beyond tidal considerations, conventional mapping techniques—while successful and
informative—remain limited by imagery availability, required computing resources, and necessary
technical expertise [34]. A single uncompressed Landsat 8 scene is larger than 1.6 gigabytes,
and applications using multiple scenes require computing resources that present a barrier to many
practitioners [35]. Emerging tools and technologies are ushering in a new era for land-cover
mapping and monitoring [26,36]. Cloud-based platforms, most notably Google Earth Engine (GEE),
provide unprecedented volumes of ready-to-use geospatial data, including the entire Landsat archive
(i.e., radiometrically and geometrically corrected), and tool and computing resources for rapid and
seamless processing [34]. GEE stores data and completes processing on numerous remote servers
(i.e., parallel processing), removing the need to download and process data on local stand-alone
computers. This eliminates many barriers related to the hardware and technical expertise required
for remote sensing. All that is required to use GEE is a computer capable of running a modern web
browser and an internet connection—for development, research, or educational purposes, access is
freely granted through Google, LLC (Limited Liability Corporation), by signing up through the GEE
Homepage. These advancements allow for developing and carrying out mapping methodologies over
unprecedented spatial extents with drastically increased speed (e.g., University of Maryland Global
Forest Dynamics), making advanced remote sensing applications accessible to considerably broader
Remote Sens. 2020, 12, 3758 3 of 35

audiences [34,37]. In addition, tools built for GEE and distributed over the Internet can facilitate
methodological repeatability while providing opportunities for adaptability and customization [38].
To date, several studies have explored and demonstrated the utility of GEE for mapping mangroves
yielding encouraging results and improvements over conventional methods [39–43]. While there is
clear utility for mapping and monitoring mangrove ecosystems using GEE, published methodologies
remain inaccessible to many would-be users. To replicate published methods requires an advanced
level of specialized expertise with remote sensing, geospatial processing techniques, and/or coding.
To date, no intuitive and accessible version of a mangrove mapping methodology within GEE has
been proposed which caters to a wider audience of non-specialist conservation managers and decision
makers. In addition, existing tools fail to fully capitalize on the wealth of local knowledge and
understanding often held by coastal managers. Lastly, no single methodology comprehensively
incorporates all of the best available options for mapping and monitoring mangrove ecosystems from
across existing published studies and includes a widely applicable approach toward tidal calibration.
Herein we present a comprehensive, intuitive, accessible, and replicable methodology encapsulated
in a new tool—the Google Earth Engine Mangrove Mapping Methodology (i.e., the GEEMMM).
The GEEMMM was designed to provide a ready-to-go methodology for non-expert practitioners
to map and monitor mangrove ecosystems, enabling them to combine their local knowledge with
GEE’s cloud computing capabilities. We developed the GEEMMM following a thorough review
of mangrove remote sensing literature and incorporating the best available practices. In addition,
our approach to tidal calibration operates completely within GEE based entirely on shoreline reflectance
(i.e., image-based). To demonstrate the tool, we present an example of multi-date, desk-based (i.e.,
involving no field work) mapping and change assessment for Myanmar (Burma)—a global loss
hotspot [19]. The GEEMMM—freely accessible to non-profit users—runs on detailed and well
commented code within the GEE environment and is adaptable to any mangrove area of interest.
GEEMMM outputs include multi-date classified maps, accuracies, and dynamic assessments. To set
the stage for trailing the GEEMMM for Myanmar and contextualizing the outputs, and similar to
methods detailed in Gandhi and Jones [19], all existing single- and multi-date mangrove maps for
Myanmar were inventoried, described, and compared, with an emphasis on existing information on
distribution and dynamics. We introduce the pilot area of interest (i.e., AOI), describe existing datasets,
overview the GEEMMM tool, and compare the results to existing datasets.

2. Materials and Methods

2.1. Google Earth Engine Mangrove Mapping Methodology (GEEMMM) Pilot AOI

2.1.1. Regional Context

The region encompassing south (S) Asia, southeast (SE) Asia, and Asia-Pacific is home to
approximately 46% of the world’s mangroves [5,44]. This region includes some of the world’s
most productive, oldest, and biodiverse mangrove forests [45]. Regional loss—the highest in the
world—is driven by conversion to aquaculture ponds (i.e., shrimp and fish farms), oil palm plantations
and rice paddies, coastal development, and over-extraction for wood [11,46–57]. Natural processes and
phenomena (e.g., rising ocean temperatures and sea-levels, severe tropical storms, and natural disasters)
also contribute to regional dynamics [48,53,56,58–66]. Notably, SE Asia is exceptionally biodiverse
containing 51 of the world’s 73 documented mangrove species, compared to 10 in the Americas and
Africa [5,67]. SE Asia alone contains an estimated 34% of the world’s mangroves [5,68]. Recent studies
show that mangrove areas in SE Asia are experiencing the highest prevalence of anthropogenic activity
in the world [68,69].
Remote Sens. 2020, 12, 3758 4 of 35

2.1.2. Myanmar—A Regional and Global Loss Hotspot

Located within SE Asia, the preliminary AOI for this pilot study is all of coastal Myanmar (Figure 1).
As confirmed by Gandhi and Jones [19], within SE Asia, mangrove loss is most notable in Myanmar,
making the country both a regional and global loss hotspot. Giri et al. [70] reported a 35% decrease
in mangrove extent from 1975–2005 whereas De Alban et al. [57] reported a 52% decrease from
1996–2016 [57,70]. According to De Alban et al. [57] and Estoque et al. [56], the primary anthropogenic
drivers of this loss include conversion to rice paddies, oil palm and rubber plantations, and increasingly
for aquaculture (e.g., shrimp, fish) [56,57]. Natural drivers include tsunamis triggered by seismic
activity, and tropical storms [68,71,72]. Within Myanmar, according to Giri et al. [70], Saah et al. [73],
Bunting et al. [74], De Alban et al. [57], and Clark Labs [75] sub-national loss hotspots include the
northwestern (NW) coastline, much of the Ayeyarwady peninsula, and a smaller area slightly east of
the Ayeyarwady peninsula (Figure 1).

2.1.3. Myanmar—Inventory, Summary and Acquisition of Existing Datasets

All national-level mangrove datasets providing single- or multi-date coverage for Myanmar up to
July 2020 were inventoried through an exhaustive online search and literature review. When available,
datasets were obtained from online repositories or through contacting authors. When not available,
datasets were described based on associated literature. All datasets were summarized based on
producer/organization/reference, single- vs. multi-date, temporal and spatial extent, availability,
imagery source(s), mapping methods, and whether discrete or continuous (Table 1).
Remote Sens. 2020, 12, 3758 5 of 35

Figure 1. The preliminary region of interest (ROI) for the GEEMMM pilot representing coastal
Myanmar; sub-national AOIs wherein qualitative accuracy assessments (QAAs) were untaken for
existing maps (i.e., Baseline QAA AOIs); sub-national AOIs wherein GEEMMM QAAs were undertaken
(i.e., GEEMMM QAA AOIs). Also shown are sub-national AOIs wherein classification reference areas
(CRAs) were derived (i.e., CRA AOIs) and the location of known mangrove loss hotspots based on
existing studies (i.e., Giri et al. [70], Saah et al. [73], GMW (Bunting et al. [74]), De Alban et al. [57],
Clark Labs [75].
Remote Sens. 2020, 12, 3758 6 of 35

Table 1. Inventory and summary of existing national-level mangrove datasets for Myanmar—July 2020.

Spatial Imagery Discrete/

Author(s) Year(s) Availability Methods
Extent/Resolution Source(s) Continuous
1975, 1990, 2000, 6 tsunami-affected Hybrid supervised/unsupervised
Giri et al. [70] Available from authors Landsat Discrete
2005 countries/30 m classification (ISODATA† clustering)
SERVIR-Mekong Downloadable from
Regional Land-Cover Greater Mekong SERVIR-Mekong website (at Supervised classification (Support
1987–2018 (V3) Landsat, MODIS Discrete
Monitoring region/30 m c. 120 m resolution; available Vector Machine; Random Forest)
System—Saah et al. [73] from authors at 30 m)
Global Mangrove
1996, 2007–2010, Downloadable from Ocean Jers-1, ALOS, Supervised classification (Random
Watch—Bunting et al. Global/25 m Discrete
2015–2016 Data Viewer ALOS-2, Landsat Forest); histogram thresholding [57]
[74]
Landsat, JERS-1, Supervised classification (Random
de Alban et al. [57] 1996, 2007, 2016 National/30 m Available from authors Discrete
ALOS, ALOS-2 Forest)
Unsupervised maximum likelihood
Stibig et al. [76] 1998–2000 S and SE Asia/1 km Downloadable from JRC SPOT-4 Discrete
classification
Bangladesh and Visual interpretation and supervised
Blasco et al. [77] 1999 Not available SPOT 1, 2, 3 Discrete
Myanmar/20 m classification
Downloadable from Clark Mahalanobis classifier; hybrid
Clark Labs [75] 1999, 2014, 2018 Multi-national/30 m Landsat Discrete
Labs website supervised/ISOCLUST ‡ clustering
World Atlas of
Downloadable from Ocean
Mangroves (WAM)— 2000–2007 Global/30 m Landsat Not disclosed Discrete
Data Viewer
Spalding et al. * [5]
Mangrove Forests of the
Downloadable from Ocean Hybrid supervised/unsupervised
World (MFW)— 2000 Global/30 m Landsat Discrete
Data Viewer classification (ISODATA † clustering)
Giri et al. [44]
2000–2012 data downloaded Masked Global Forest Change (GFC)
CGMFC-21—Hamilton
2000–2014 Global/30 m from CGMFC-21, 2013–2014 Landsat [47] maps using MFW [47]) to Continuous
and Casey [78]
data available from authors calculate dynamics
Masked GFC maps using MFW [56]
Richards and Friess [47] 2000, 2012 SE Asia/30 m Not available Landsat Continuous
to calculate loss
Unsupervised classification
Estoque et al. [56] 2000, 2014 National/30 m Not available Landsat Discrete
(ISODATA † clustering)
* WAM data over Myanmar from Ministry of Forestry’s Remote Sensing and GIS Section, derived from Landsat imagery 2000–2007. † Iterative Self-Organizing Data Analysis Techniques.
‡ Iterative Self-Organizing Clustering.
Remote Sens. 2020, 12, 3758 7 of 35

2.1.4. Myanmar—Comparison of Existing Datasets and Baseline QAA

Once inventoried, all known datasets were compared based on mapped classes, mangrove
distribution, accuracy, dynamics (when available), and known limitations. Where provided, mangrove
distributions and dynamics were extracted from publications and supporting materials. If not readily
apparent—and if the datasets were available—dynamics were calculated. Adding to the standard
reported metrics, the accuracy was further qualitatively assessed for all available datasets through
cross-checking in reference to high spatial resolution satellite imagery viewable through Google Earth
Pro (GEP) [79]. This secondary qualitative accuracy assessment—or QAA—first reported in Gandhi
and Jones [19], provides a more thorough understanding of existing mangrove datasets.
The QAA of existing maps (i.e., baseline QAA) was undertaken for the most recent entry in
each discrete dataset, when available. Datasets were acquired in both raster and vector format,
and in a range of coordinate systems, necessitating several pre-processing steps. For each baseline
QAA, three 100 × 100 km sub-national AOIs were selected across Myanmar: in the north (Rakhine),
in the center (Ayeyarwady Delta), and in the south (Tanintharyi) (Figure 1). Each baseline QAA AOI
was divided into 10 × 10 km boxes, and working from NW to SE, every sixth box was selected for
spot-checking, such that approximately 17% was systematically assessed. QAA of the Giri et al. [70]
dataset was already conducted [19]. For the remaining datasets, each spot-check entailed comparing
mangrove coverage to GEP imagery as close to the dataset’s capture date as possible. In some instances,
particularly in the southernmost Tanintharyi AOI, GEP imagery was partly/fully cloud-covered,
limiting the ability to conduct QAA (limitations also noted by Estoque et al. [56]). A single mangrove
class, representing the variability of canopy cover, height and stand structure in mangrove forests
(as used in GEEMMM pilot classifications and defined below) was qualitatively assessed within each
spot-check as either well-, under-, or over-represented. For each dataset, results help contextualize the
representation of mangrove distribution and dynamics.

2.2. The Google Earth Engine Mangrove Mapping Methodology (GEEMMM)

The GEEMMM is intended to facilitate the mapping and monitoring of mangrove ecosystems
anywhere in the world, without requiring a dedicated in-house geospatial expert. Intended users
need basic computer skills and an understanding of the key steps required for mapping mangroves,
but are not expected to hold advanced expertise in remote sensing, geospatial analysis, and/or coding.
The interactive tool is broken into three modules—Module 1: defines customized region of interest (ROI)
boundaries and generates multi-date imagery composites; Module 2: examines spectral separation
between target map classes and undertakes multi-date classifications and accuracy assessments;
Module 3: explores dynamics and offers an optional QAA. Each module is broken into thoroughly
commented and referenced sections, bringing the user through all steps while making reference to this
manuscript for full methodological details and context. Each module and the parameters used in this
pilot study are described below. Table 2 provides a summary of all GEEMMM user inputs and variable
selections for the Myanmar pilot study.

2.2.1. Module 1—Defining the ROI and Compositing Imagery

In the first step of Module 1: Section 1, the user must identify key datasets to be used in the
GEEMMM. The first user-defined dataset is a preliminary ROI. This is generated using the ‘drawing
tools’ function built into GEE and clips all subsequent user-defined datasets. The second user-defined
dataset is the known extent of mangroves which is used to calculate elevation and slope thresholds and
shoreline buffer distance. The user can select the baseline GEE data set representing global mangrove
distribution circa 2000 (i.e., Giri et al. [44]) or upload their own. The third user-defined dataset is
coastline, for which the user can select the baseline Large Scale International Boundary Polygons [80]
or upload their own. The fourth user-defined dataset concerns elevation which is required to generate
topographic masks (i.e., elevation and slope). The user can select the GEE JAXA-ALOS satellite radar
Remote Sens. 2020, 12, 3758 8 of 35

DSM (30 m) [81] or upload their own. For the Myanmar pilot, the preliminary ROI is shown in
Figure 1, the GEE global mangrove distribution circa 2000 was used for known mangrove extent,
the Global Administrative Boundaries database (GADM) Myanmar dataset (v3.6, www.gadm.org,
an external source) for coastline, and the GEE JAXA-ALOS Global PALSAR-2/PALSAR Yearly Mosaic
25 m land-cover data for elevation [81,82].

Table 2. Summary of GEEMMM user inputs and selected variables used in the Myanmar pilot.

GEEMM User Inputs.

Module Input Type Selected Variables
Preliminary ROI Dataset (vector) GUI Generated
Known Mangrove Extent Dataset (raster) Giri et al. [44]
Coast Line Dataset (vector) GADM—Myanmar [82]
Digital Surface Model Dataset (raster) JAXA-ALOS DSM (30 m) [81]
Contemporary Year(s) Date range (YYYY) 2014–2018
Module 1

Historic Year(s) Date range (YYYY) 2004–2008

Month(s) Date range (MM) 06–12
Cloud Cover Limit Integer (%) 30%
Cloud Cover Mask Variable Aggressive
Tidal Zone Numeric (m) 1500 m
Water Mask Variable Combined
Fringe Mangroves Boolean False (not included)
Topographic Mask Variable Uses Known Mangrove Extent [44]
Classification Reference Areas
Dataset (vector) See Table 4
(CRAs)
Module 2

Class Names Variable See Table 4

Class Numbers Integer Defined by Authors
Classification Algorithm Random Forest Random Forest [82]
Number of Trees Integer 200
Output Classification Maps Variable Hist. and Cont. Combined
Classification Reference Areas
Dataset (vector) See Table 4
(CRAs)
Module 3

Class Names Variable See Table 4

Class Numbers Integer Defined by Authors
Classification View Variable See Figure 7
Mangrove Class Number(s) Integer Defined by Authors

In the second step of Module 1: Section 1, the user defines input variables and sets how workflow
thresholds are calculated. Table 2 lists all of the user variables and user inputs for the GEEMMM
including those used in this pilot study. GEE provides unprecedented access to the Landsat catalog,
offering approximately 1.3 M scenes from 1984 to present [34]. While it is certainly advantageous
to have access to so many images, the choice of imagery based on parameters such as year(s) and
time of year(s) must be considered carefully. Two variables define contemporary and historic year(s)
of interest. There are two four-digit date inputs to bookend the historic and contemporary year
windows. If the user wishes to isolate a single historic or contemporary year the same is selected
for each book-end. Following the year(s) of interest, the month(s) of interest are selected. Seasonal
variations can affect terrestrial vegetation adjacent to mangroves, and atmospheric conditions can
change throughout the year, so the ability to target specific months is essential to generating optimal
image composites [83–85]. The user identifies the month(s)-of-interest using two book-end numbers
corresponding to the 12 months of the year; they may overlap the new year; e.g., “11” (Nov.) to
“2” (Feb.). Next, the allowable cloud cover limit, an integer between 0 and 100, is used to filter the
Landsat metadata [86]. Also related to cloud cover, the user decides whether to mask the imagery,
and to what extent, i.e., setting a mild cloud mask using the USGS-provided (United States Geological
Survey) quality band, or an aggressive cloud mask where pixels are excluded based on their ‘whiteness’
Remote Sens. 2020, 12, 3758 9 of 35

and a temperature band threshold [35]. For the sixth input, the approximate tidal zone—a numeric
input (in m) that represents the tidally active zone buffered inland from the coastline—is entered.
Approximate tidal zone helps isolate the portion of images subject to reflectance changes from tidal
variation, while reducing influence from other non-tidal variability. The default value is 1000 m. Next,
the user chooses how water is masked out of the imagery, either using a mask developed from the
water present in the contemporary imagery alone, or a combination mask based on pixels determined
to be water in both historic and contemporary imagery. A pixel is determined to be water if its
value was greater than the 0.09 modified normalized difference water index (i.e., MNDWI) threshold
established by Xu [87]. The modified normalized difference water index (MNDWI) was developed to
detect water pixels by calculating the normalized difference between the green and short-wave infrared
(e.g., Landsat 8 Operational Land Imager, 1.57–1.65 µm) bands, making it suitable for measuring the
amount of water present in an acquisition. Topographic thresholds are set to generate masks based on
elevation and slope. The user can either manually enter the elevation (m) and slope (%) thresholds,
or have them automatically calculated based on the 99th percentile values extracted from within the
known mangrove extent dataset. The user can further opt to search for inland-fringe mangroves,
which have been documented as far as 85 km inland [75,88]. If inland-fringe mangroves are targets
for the classification(s), the preliminary ROI is doubled for elevations lower than 5 m based on [89].
The last step in Module 1: Section 1 is the selection of spectral indices which the user would like to
calculate for each image composite. After the workflow begins, the user chooses which indices they
would like to calculate from a list of fourteen indices, including some which are mangrove-specific.
The complete list of indices included in the GEEMMM can be found in Table 3. The contemporary
and historic windows from which imagery was selected for the Myanmar pilot study were 2014–2018
and 2004–2008, respectively. The months of acquisition were limited to June through December,
corresponding with the wet season and the months directly following that time [90]. The imagery was
filtered using cloud cover information for each acquisition at a 30% threshold. All 14 spectral indices
were selected for calculation.
Module 1: Section 2 determines the finalized ROI for processing. Numerous studies have
demonstrated the utility of reducing the classification extent to the minimum required area—this
approach helps reduce spectral confusion with unnecessary scene components [44,91]. The preliminary
ROI is used to isolate a section of shoreline which is buffered at 5, 10, 15, 20, 25, 30, and 35 km intervals.
5 km intervals were used to ensure observable differences in buffer distances. 35 km was used as
a maximum extent based on observations in several countries, including Myanmar. These buffer
distances are used to calculate the area of known mangroves that falls within their respective bounds.
The user either selects their buffer distance preference from a drop-down menu containing values in
between, greater than, or less than the listed intervals.
In either case, the buffer distance is used to create the finalized ROI. This ROI is used to
select Landsat path/row tiles and generate image composites, clip composite imagery and masks
(i.e., elevation, slope, and water), define the classification and dynamics extent, and provide a visual aid
for optional QAAs. The finalized ROI used in the pilot study was based on a 23 km buffered shoreline
which represents the maximum observed distance between known mangrove extent (i.e., Giri et al. [44])
and Myanmar’s coastline.
Module 1: Section 3 generates the imagery composites required for multi-date classifications.
Given the daily dynamic nature of mangrove ecosystems—wherein tides inundate 2–3 times
per day on average—tidal conditions and the associated presence (or lack thereof) of water must be
considered—there are a growing number of mangrove detection indices which rely on the isolation
of high and low tide imagery [29,92,93]. The GEEMMM uses an image-based approach to calibrate
imagery based on high and low tide. For each available image, an MNDWI is generated and the land
is masked out using JAXA-ALOS Global PALSAR-2/PALSAR Yearly Mosaic 25 m land-cover data.
The MNDWI was selected as the key spectral index because it has been proven to be an improvement
over the normalized difference water index (NDWI), and was developed explicitly for detecting water
Remote Sens. 2020, 12, 3758 10 of 35

and non-water pixels [87]. The shoreline is buffered to the user-defined tidal zone value and mean
MNDWI is used to create a constant value band wherein the greater the MNDWI mean value, the more
water present within the tidal zone, corresponding to higher-tidal conditions. A second value band is
added to each available image by multiplying mean MNDWI by −1, isolating lower-tide conditions.
Clouds, if present and opted to be, are masked prior to the calculation of mean MNDWI using only
the pixel quality band or an aggressive approach where the three visible (red, greed, and blue) and
thermal bands mask based on digital number reflectance thresholds. Under the aggressive filter,
a pixel is considered to be a cloud if its visible spectrum bands digital number reflectance values are
greater than or equal to 1850, and the thermal band (brightness temperature, Kelvin) digital number is
less than or equal to 2955. For the Myanmar pilot, the aggressive cloud filter option was selected to
filter the imagery in an effort to remove low-altitude clouds which were not correctly classified by
the Landsat cloud detection algorithm. If/once clouds have been masked, all available images and
their corresponding tidal value bands are used to create best available pixel-based highest observable
tide (i.e., HOT) and lowest observable tide (i.e., LOT) composites. Composite generation works as
if all available images were stacked and organized by desired tidal condition. For example, as the
LOT composite is being generated, the imagery with the lowest observed tidal condition is placed on
top, and any missing pixels in that image, e.g., clouds masked, would be filled by the next best tidal
observation and so on until all the gaps are filled. This process takes place for both the contemporary
and historic data sets, resulting in a maximum of four composites (i.e., HOT and LOT contemporary,
HOT and LOT historic). Because tides are determined using value bands, it is possible that all of the
pixels for HOT and/or LOT composites within a particular area may be from one image (e.g., if no
clouds were present and that image represented best available tidal conditions). The GEEMMM
employs USGS surface reflectance Landsat products, which are readily available within GEE [35,94].
Module 1: Section 4 calculates the user selected indices from Section 1 of Module 1 (Table 3).
There are a growing number of Landsat-related spectral indices available, many of which relate directly
to mangroves such as the submerged mangrove recognition index (SMRI) and the modular mangrove
recognition index (MMRI) [29,43]. The GEEMMM provides the user with the option to select from
14 spectral indices, of which four are mangrove-specific. The selected indices are calculated for both
contemporary and historic HOT and LOT composites and added as potential classification inputs.
Figure 2 compares the appearance of a typical mangrove-dominated area in Myanmar across all of
the available mangrove-specific spectral indices (i.e., combined mangrove recognition index (CMRI),
MMRI, SMRI, MRI) in the GEEMMM [29,43,93,98].
In Module 1: Section 5 the classification extent is further reduced through masking.
In accordance with numerous mangrove mapping studies (e.g., Jones et al. [91], Thomas et al. [68],
and Weber et al. [105]), the GEEMMM incorporates cloud, water, slope, and elevation masks to produce
a finalized AOI. The cloud mask is generated and applied before composites are produced. The water
mask is calculated for each composite using the methodology established in Xu [87], where the
MNDWI layer for historic and contemporary LOT composites are generated and then a threshold
is applied. Pixels with a value greater than 0.09 are considered to be water and a binary mask is
produced. Depending on user selection, the water mask is finalized by either using just contemporary
or combining the historic and contemporary and selecting only pixels determined to be water in both
composites. This pilot study used the combined water mask. The two topographic masks are generated
through user-defined thresholds or automatically determined using the 99th percentile of elevation and
slope for known mangroves. The Myanmar pilot study used the known mangrove extent to generate
topographic masks based on elevation values > 39 m and slope values > 16%. Noting how minor
elevation is within mangrove ecosystems, the elevation threshold actually represents an approximate
combined elevation + canopy height past which mangroves are not found. The generated masks are
combined to create a binary, single unified final mask which is applied to all composites within the
finalized ROI.
Remote Sens. 2020, 12, 3758 11 of 35

Module outputs include: (1) HOT contemporary composite, (2) LOT contemporary composite,
(3) HOT historic composite, (4) LOT historic composite, (5) finalized ROI, and (6) Finalized Mask.

Table 3. List of all spectral indices available in the GEEMMM including mangrove-specific.

Index Abbreviation Calculation Citation

Simple Ratio SR NIR/Red Jordan [95]
Normalized Difference
NDVI (NIR − Red)/(NIR + Red) Tarpley et al. [96]
Vegetation Index
Normalized Difference
NDWI (Green − NIR)/(Green + NIR) Gao [97]
Water Index
Modified Normalized
MNDWI (Green − SWIR1)/(Green + SWIR1) Xu [87]
Difference Water Index
Combined Mangrove
CMRI * NDVI − NDWI Gupta et al. [98]
Recognition Index
Modular Mangrove
MMRI * (|MNDWI| − |NDVI|)/(|MNDWI| + |NDVI|) Diniz et al. [43]
Recognition Index
Soil-Adjusted Vegetation
SAVI 1.5*(NIR − Red)/(NIR + Red + 0.5) Huete [99]
Index
Optimized Soil-Adjusted
OSAVI (NIR − Red)/(NIR + Red + 0.16) Rondeaux et al. [100]
Vegetation Index
Enhanced Vegetation Index EVI 2.5*((NIR − red)/NIR + 6*Red − 7.5*Blue + 1)) Huete et al. [101]
Mangrove Recognition Index MRI * |GVI(l) − GVI(h)|*GVI(l)* (WI(l) + WI(h)) Zhang and Tian [93]
Submerged Mangrove (NDVI(l) − NDVI(h))* ((NIR(l) −
SMRI * Xia et al. [29]
Recognition Index NIR(h))/(NIR(h))
Land Surface Water Index LSWI (NIR − SWIR1)/(NIR + SWIR1) Chandrasekar et al. [102]
Normalized Difference
NDTI (MIR − SWIR2)/(MIR + SWIR2) Van Deventer et al. [103]
Tillage Index
Enhanced Built-up and √
EBBI (SWIR1 − NIR)/(10* (SWIR1 + LWIR)) As-syakur et al. [104]
Bareness Index
* denotes mangrove-specific spectral index.

Figure 2. The appearance of a typical mangrove-dominated area in Myanmar across all of the available
mangrove-specific spectral indices (i.e., CMRI, MMRI, SMRI, MRI) in the GEEMMM [91].
Remote Sens. 2020, 12, 3758 12 of 35

2.2.2. Module 2—Spectral Separability, Classifications and Accuracy Assessment

For Module 2: Section 1, user inputs address classification variables and settings. The user enters
the asset path for historic and contemporary classification reference areas (CRAs) (i.e., the user-defined
examples of target map classes) and identifies the unique column labels for class names and numeric
codes. Next, the user identifies whether CRAs are spatio-temporally invariant (i.e., each CRA represents
a class example in both contemporary and historic imagery). If the CRAs are not spatio-temporally
invariant, the spectral properties of the contemporary CRAs are extracted and used to define class
boundaries in the historic classification(s). For classification algorithm the single option is currently
random forest [106]. The user determines how many trees are employed. The final input determines
classification outputs. Users have the option to select outputs from either HOT or LOT composites for
contemporary and historic inputs (i.e., four possible outputs), and/or a combined classification where
HOT and LOT composites are merged to create single outputs (i.e., two more possible outputs), totaling
six possible classification outputs. Zhang and Tian [93] demonstrated the utility of using combined
HOT and LOT image composites as classification inputs. For the Myanmar pilot, 200 trees were
selected with outputs based on combined (i.e., HOT and LOT) historic and contemporary classifications
(i.e., two classifications).
In Module 2: Section 2, the user can examine correlation between potential spectral indices and
the spectral separability of CRAs across all potential classification inputs. The Pearson’s correlation is
calculated for each selected index to all others and these values are used to generate a correlation matrix
with values ranging from −1 to 1 [107]. A value of 1 means that the potential inputs have a perfect,
positive, linear correlation, and a value of −1 indicates that the indices have a perfect, negative, linear
correlation. Users are encouraged to select indices that are not highly correlated indicating that they
provide unique information. As a general rule, correlation coefficients with absolute values greater
than 0.7 are considered moderately to strongly correlated and thus present similar information [107].
Users are advised to consider that correlation coefficients are also impacted by the amount of variability
in the data, the shapes of distributions, and the presence of outliers among other factors [108].
The spectral separability between target map classes as represented by CRAs is explored through
the generation of three types of graphs. First, the user can view the spectral separability between each
target class and each Landsat band—the user has the option to view this output for each of the four
imagery composites. Box-and-whisker plots show the min, max, and inter-quartile range for each band
and each map class. The second set of graphs is similar to the first, except that spectral separability is
shown for individual indices across all of the target classes, showing only one index at a time. The final
graph shows spectral feature space, where the x and y axes are user selected bands or indices. For the
pilot study, and based on previously established precedents in Jones et al. [91], we included the visible,
NIR and SWIR Landsat bands. Based on the correlation matrices and further the spectral separability
they provided, the MNDWI, CMRI, MMRI, enhanced vegetation index (EVI), and Land Surface Water
Index (LSWI) indices were selected as additional classification inputs.
For piloting the GEEMMM in Myanmar, six classes were initially targeted, including, (1) closed-canopy
mangrove, (2) open-canopy mangrove, (3) terrestrial forest, (4) non-forest vegetation, (5) exposed/barren,
and (6) residual water. Table 4 provides class descriptions and an overview of how many CRAs were
digitized per class. CRAs can be derived within the GEE environment or externally. For this pilot,
90 × 90 m (i.e., 3 × 3 Landsat pixels) CRAs were derived externally. To ensure that internal class variability
was captured for each class and across the AOI, three sub-national AOIs were used to define CRAs
(Figure 1). CRAs were derived referring to finer spatial resolution satellite imagery viewable in Google
Earth Pro (Google, Mountain View, CA, USA), existing contemporary land-cover maps for Myanmar
(i.e., Giri et al. [44], Saah et al. [73], and De Alban et al. [57]), and expert interpretive knowledge gained
with mapping mangroves in other regions of the world. Two mangrove classes were defined to ensure
that the internal variability of mangrove forests based on stature, canopy cover and density was captured.
Figure 3 shows examples of all targeted classes in HOT, LOT, a key spectral index, and finer spatial
resolution imagery viewable in Google Earth Pro (Google, Mountain View, CA, USA) [79].
Remote Sens. 2020, 12, 3758 13 of 35

Table 4. Names and description of classes and numbers of classification reference areas (CRAs).
Also shown is how many CRAs were derived within each sub-national CRA AOI (Figure 1).

Contemporary Historic
Class Class Description
AOI 1 AOI 2 AOI 3 Total AOI 1 AOI 2 AOI 3 Total
Grass and/or shrubs dominate; some
Non-Forest exposed soil + scattered trees; canopy
10 8 7 25 3 7 0 10
Vegetation < 30% closed; active cropland, vegetation
appears green
Terrestrial Forested areas; canopy > 30% closed
10 8 7 25 1 9 0 10
Forest (includes plantations (e.g., palm))
Closed-Canopy
Tall, mature stands; canopy > 60% closed 12 16 9 37 9 1 0 10
Mangrove
Open-Canopy Short-medium stands; canopy 30–60%
6 3 2 11 0 10 0 10
Mangrove closed
Soil/sediment/sand dominates; includes
Exposed/Barren senesced/unhealthy (i.e., inactive) crops, 4 4 4 12 2 4 4 10
mudflats, recently deforested areas
Residual
Water areas missed from masking 4 3 3 10 3 4 4 11
Water
120 61

In Module 2: Section 3: once the user confirms their final choice of classification inputs and
target classes, classification—the process by which remotely sensed data is assigned land-cover
classes—can occur [109,110]. There are many established algorithms for classifying Landsat data
to produce maps of mangrove distribution, including classification and regression trees (CART),
support vector machines (SVM), unsupervised k-means, decision trees, and maximum likelihood
(ML) [28,29,42,111,112]. Many of these algorithms are available to use within the GEE environment;
however, random forests—also available in GEE—is well established and used to map mangroves
across the world, with distinct success within the GEE environment [27,41,43,106,113]. The inputs for
random forest include an imagery data set (i.e., selected Landsat bands and spectral indices), training
data (i.e., randomly selected 70% of CRAs), and a numeric parameter determining the number of
‘trees’ to be employed. For each classification the output is a single band raster with the same spatial
resolution as the input data (30 m), with each pixel assigned a map value based on target classes.
Following classification, the user can choose to merge map classes—this is particularly advantageous
in scenarios where initial map classes were used to capture variability, but for which confidence in class
boundaries may be lacking. For example, in the Myanmar pilot, we merged the two mangrove classes
(i.e., closed- and open-canopy) post-classification. This ensured capturing mangrove variability while
not having to draw a distinct boundary between these potentially overlapping classes in the final map.
Classification accuracy—defined as “a comparison of the derived product to ground condition”—is
not reported in numerous studies involving mangrove mapping [114,115]. Following classification and
optional class merging, in Module 2: Section 3, the GEEMMM automatically produces resubstitution
and error matrices for all output classifications [116]. The resubstitution matrices determine end
land-cover class for the CRAs used for training the classifier. The error matrices use 30% of CRAs
held back from classification to independently evaluate map accuracies. The overall accuracy is
reported using the error matrix ‘accuracy’ tool, found within the GEE library. Overall accuracy is
printed below both the error and resubstitution matrices. By reviewing the error matrices and visually
inspecting the output maps the user may wish to collapse/further collapse classes (e.g., if two classes
are very confused). If the user combines classes, they can opt to re-calculate accuracy, re-generating
resubstitution and error matrices. The final step for all users to exporting the classification maps to
their assets. Module 2 outputs include, (1) correlation and spectral separability graphs, (2) classified
maps, and (3) accuracy assessments.
Remote Sens. 2020, 12, 3758 14 of 35

Figure 3. The appearance of all targeted classes in highest observable tide (HOT), lowest observable
tide (LOT), key spectral indices, and fine spatial resolution satellite imagery viewable in Google Earth
Pro (Google, Mountain View, CA, USA) [79]. The HOT and LOT composites represent 432 (R: NIR,
G: red, B: green) or 453 (R: NIR, G: SWIR, B: red) false color. The spectral indices include enhanced
vegetation index (EVI—[101]), combined mangrove recognition index (CMRI—[98]) and modified
normalized difference water index (MNDWI—[87]).

2.2.3. Module 3—Dynamics and QAA

In Module 3: Section 1, the user indicates which classification(s) will be used to calculate dynamics
and/or assess optional QAA. If desired, the user can further clip classifications to a country’s boundary—
if pertinent—using the GEE Large Country Boundary Polygons, or by uploading an external dataset. For the
Remote Sens. 2020, 12, 3758 15 of 35

Myanmar pilot we further clipped using a uniquely uploaded boundary (GADM) and exclusive economic
zone (EEZ) from Marine Regions (v10 World EEZ,) [117]. For the QAA, the user enters CRA information
(e.g., asset path, class names, and unique class numbers).
In Module 3: Section 2, multi-date outputs are used to quantify dynamics. This is foundational
to understanding long-term trends and the effectiveness of conservation efforts. The user selects
which map class they would like to view, and loss, persistence, and gain (i.e., LPG) are calculated.
The automatically produced, self-masked layers are added to the GEE-GUI map interface. The resulting
area for each dynamic assessment is printed to the console, expressed in hectares. Building on the
inventory, description, acquisition and comparison of existing datasets, the dynamics resulting from
this GEEMMM pilot were also compared to published values.
Module 3: Section 3—building on the previously referenced methods detailed in Gandhi and
Jones [19]—facilitates an optional QAA. For this GEEMMM QAA, an interactive map is divided
into three linked maps (Figure 1). In each map, two sets of grids are automatically generated,
(1) 100 km by 100 km grids, and (2) within each of those cells, sub-divided 10 km by 10 km grids.
The 100 km × 100 km grid cells are randomly selected, retaining 50% of the grid cells that intersect the ROI.
In slight contrast to the baseline QAA described in Section 2.1.4, for the QAA tool in the GEEMMM, within
each selected grid cell, 20% of the sub-grid cells are selected. The tool works by cycling through the sub grid
cells, and giving the user the option to view simultaneously on linked maps showing Landsat composites
where the date can be changed at the user’s preference, the classifications produced in Module 2, and the
imagery used for the classifications generated in Module 1. The user then has the ability to record in the GUI
whether each map class is under-, well-, or over-represented, and record ‘free comments’ for each sub-cell.
Module 3 outputs include: automatically generated LPG as raster and—if performed—QAA grid
(for viewing outside of GEE). The user also has the option to export the QAA table (containing the under,
over, and well representation statistics, and the free comments) as a CSV (i.e., comma separated values) file
at any point during the QAA.

3. Results and Discussion

3.1. Myanmar—Comparison of Existing Datasets

Table 5 provides a comparison of all single- and multi-date datasets based on dataset/authors,
year, extent (ha), dynamics (ha and %), whether discrete or continuous, mapped classes, accuracy,
and known limitations. Figure 4 provides a comparison of all distributions across time across all
datasets. Results show that Myanmar’s mangrove distribution ranged from 851,452–1,323,300 ha circa
1975–1987 (i.e., historic) to 475,637–1,002,098 ha circa 2014–2018 (i.e., contemporary). Of the 11 existing
studies, only five provided quantitative accuracy assessments, with overall accuracies ranging from 76%
(i.e., Saah et al. [73]) to 97% (i.e., Estoque et al. [56]), mangrove producer’s accuracies ranging from 75%
(i.e., De Alban et al. [57]) to 93.1% (i.e., also De Alban et al. [57]), and mangrove user’s accuracies ranging
from 92.3% (i.e., De Alban et al. [57]) to 98.1% (i.e., Clark Labs [75]). Of the existing studies, eight provided
dynamics, including a loss of 300,091 ha/35.2% from 1975–2005 (Giri et al. [70]), 195,227 ha/16.3% from
1987–2018 (Saah et al. [73]), 43,208 ha/8.0% from 1996–2016 (Bunting et al. [74]), 694,600 ha/52.5% from
1996–2016 (De Alban et al. [57]), 76,465 ha/10.9% from 1999–2018 (Clark Labs [75]), 27,064 ha/9.7% from
2000–2014 (Hamilton and Casey [78]), 27,770 ha/5.5% from 2000–2012 (Richards and Friess [47]),
and 191,122 ha/28.7% from 2000–2014 (Estoque et al. [56]). Two reported specifically on sub-national
loss hotspots (i.e., De Alban et al. [57] and Estoque et al. [56]). According to De Alban et al. [57], Bago,
Mon, Yangon—the three states immediately to the east of the Ayeyarwady delta—suffered greatest
proportionate loss from 1996–2016 totaling more than 80% of their extents. In terms of absolute loss,
from 2000–2014, Estoque et al. [56] reported Rakhine as the state with the greatest loss (75,494 ha/39.5% of
Myanmar’s total loss), followed by Ayeyarwady experiencing 69,431 ha/36.3% of Myanmar’s total loss.
Remote Sens. 2020, 12, 3758 16 of 35

Table 5. Comparison of single- and multi-date datasets based on mapped classes, accuracy, mangrove distribution (ha), dynamics, and known limitations. Accuracy:
OA = overall accuracy; UA = user’s accuracy; PA = producer’s accuracy.

Dataset/ Mapped
Year Extent (ha) Dynamics (ha, %) Discrete/Continuous Accuracy Known Limitations
Author(s) Classes

1975 851,452 4 classes Semantic differences in class definitions, positional,

−300,091 Positional root mean square and classification errors. Mangrove patches smaller than
Giri et al. [70] Discrete including
−35.2% error of ± 1/2 pixel 1 ha not mapped likely reducing distribution figures.
2005 551,361 Mangrove

1987 1,197,325 Gap in 2012 data due to removal of ETM+ imagery

SERVIR-Mekong 21 classes following Landsat 7 Scan Line Corrector failure.
−195,227 including OA 76% (2016 map)
(Saah et al. [73]) Discrete 2012 primitives interpolated using Whittaker smoothing
−16.3% Mangrove
2018 1,002,098 algorithm. Bias in reference data toward more recent
past, due to availability of high-resolution imagery.

1996 537,428 Mangrove Fine-scale features commonly misclassified, e.g.,

GMW (Bunting −43,208 OA 95.3% (2010 baseline aquaculture features, riverine environments, and coastal
Discrete presence vs no
et al. [74]) −8.0% map) fringes. Minimum mapping unit of 1 ha suggested for
presence
2016 494,220 end user mapping.

1996 1,323,300 1996: OA 85.6%

De Alban et al. 10 classes Mangrove UA 92.3%
−694,600 including Mangrove PA 93.1% No significant limitations disclosed.
[57] Discrete
−52.5% Mangrove 2016: OA 89.2%
2016 628,700 Mangrove UA 97.5%
Mangrove PA 75.0%

7 classes OA 96.9%
1999 703,945 including Mangrove UA 98.11%
Clark Labs [75] −76,465 Mangrove Mangrove PA 93.04%
−10.9% Discrete No significant limitations disclosed.
33 classes 2014: OA 93.7%
2018 627,480 including Mangrove UA 94%
Mangrove Mangrove PA 92%
Limitations with use of ‘quick look’ data due to modest
8, including 6
technical performance. The authors state that
Blasco et al. [77] 1999 690,000 n/a Discrete mangrove Not disclosed
classification accuracy could be improved by 10% if
classes
NDVI and empirical thresholds were included.
Mangrove Small patches of mangrove (<0.09–0.27 ha) not well
MFW (Giri et al. Positional root mean square
2000 494,584 n/a Discrete presence vs no captured.
[44]) error of ±1/2 pixel
presence
Remote Sens. 2020, 12, 3758 17 of 35

Table 5. Cont.

Dataset/ Mapped
Year Extent (ha) Dynamics (ha, %) Discrete/Continuous Accuracy Known Limitations
Author(s) Classes

2000 279,260 Pixels containing just 0.01% forest canopy cover are
CGMFC-21 −27,064 Mangrove Positional root mean square included as mangrove falling well below commonly
Continuous
(Hamilton and −9.7% canopy cover error of ±1/2 pixel used minimum canopy cover definitions
Casey [78]) 2014 252,196 (e.g., [78,118,119]).

Richards and 2000 502,466 −27,770 Mangrove Positional root mean square Reported figures reflect rates of mangrove loss rather
Continuous
Friess [47] −5.5% deforestation error of ±1/2 pixel than net mangrove change, likely reducing areal figures.
2012 474,696
Estoque et al. 2000 666,759 −191,122 Mangrove 2000: OA 91%
Discrete No significant limitations disclosed.
[56] −28.7% presence vs no 2014: OA 97%
2014 475,637
presence
WAM (Spalding
2004 502,911 n/a Discrete Not disclosed Not disclosed No significant limitations disclosed.
et al. [5] *)

GEEMMM 2004–2008 995,411 6 classes

−352,752 including 2004–2008: OA 97.01%
(Yancho et al., Discrete Refer to Results and Discussion; Conclusion.
−35.4% combined 2014–2018: OA 96.08%
2020) 2014–2018 642,659
Mangrove.
* WAM data over Myanmar from Ministry of Forestry’s Remote Sensing and GIS Section, derived from Landsat imagery 2000–2007.
Remote Sens. 2020, 12, 3758 18 of 35

Figure 4. Comparison of distribution for all existing single- and multi-date mangrove distribution
maps for Myanmar, including results of GEEEMMM pilot.

Direct comparisons of existing datasets are challenging due to differences in temporal coverage,
methodologies, and imagery sources. Although most studies use optical imagery (typically
medium-resolution Landsat), some of the more recent studies combine optical with radar imagery
(e.g., Bunting et al. [74]; De Alban et al. [57]). Several different mapping techniques are employed,
while two of the datasets (Hamilton and Casey [78]; Richards and Friess [47]) calculate and present
continuous measures of mangrove canopy cover, rather than discrete (i.e., presence vs no presence).
Interpreting continuous datasets for areal mangrove extent is problematic as pixels containing just
0.01% canopy cover are included as mangrove falling well below commonly used minimum definitions
mangrove forest (e.g., 30%) [18,78,91].
Of the five datasets reporting, all achieve overall accuracies of >75%, with four >85% [56,57,74,75].
QAAs further identified Clark Labs [75] as mapping mangroves in Myanmar most consistently.
Mangroves were under-represented in the remaining five datasets assessed, particularly in Giri et al. [70]
and Saah et al. [73], but also in Bunting et al. [74] (Table 6).

Table 6. Results of QAA for available/acquired datasets (1 = best).

AOI 1— AOI 2— AOI 3— Overall

Rank Dataset Comments
Rakhine Ayeyarwady Tanintharyi Representation
Well- Well- Well- Mangrove very well-
1 Clark Labs [75] Well- represented
represented represented represented represented
Mangrove slightly
under- represented;
De Alban et al. Under- Under- Well- Under-
2 some confusion
[57] represented represented represented represented
between cropland and
mangrove
MFW Well- Under- Under- Under- Mangrove under-
3
(Giri et al. [44]) represented represented represented represented represented
GMW (Bunting Under- Under- Well- Under- Mangrove under-
4
et al. [74]) represented represented represented represented represented
Mangrove under-
Under- Under- Under- Under-
5 Giri et al. [70] represented,
represented represented represented represented
considerably in places
SERVIR- Mangrove under-
Under- Under- Under- Under-
6 Mekong represented,
represented represented represented represented
(Saah et al. [73]) considerably in places

Existing studies clearly establish that Myanmar has experienced consequential mangrove loss;
however, baseline distributions and dynamics (when available) are highly variable. These discrepancies
are likely attributed to the differences highlighted in Table 5. In addition, the definitions for mangroves
Remote Sens. 2020, 12, 3758 19 of 35

and surrounding land-cover classes and the actual examples used for classification (i.e., CRAs) likely
further account for differences. Only with agreed upon conventions for defining mangroves and
providing examples as CRAs can cross-study comparisons become standardized and optimized. Falling
short of this, discrepancies will remain common.

3.2. Results of the Google Earth Engine Mangrove Mapping Methodology (GEEMMM)

3.2.1. Module 1—Defining AOI and Compositing Imagery

As confirmed through qualitative yet systematic spot-checks, the imagery generated from the
Myanmar pilot reflects the selected inputs well—both historic and contemporary composites are mostly
cloud- and artifact-free, and clearly represent distinct HOT and LOT conditions (Figure 5). Figure 6
shows a national overview of the AOI including contemporary and historic HOT and LOT composites.
The challenges that have been identified can be attributed to the extent of the study area and trying to
capture a long, complex coastline in a series of contiguous composites. The most notable challenge relates
to seasonal variability observed primarily in areas where large clouds were masked out of one image
and the pixels selected to fill captured seasonally different land-cover conditions. Notably, this issue
was almost entirely associated with areas which undergo significant changes throughout the year, i.e.,
agricultural mosaics and non-forest vegetation. Even within the defined seasonal window, variability was
observed. Users are advised to select meaningful seasonal windows that restrict such variability while
still offering enough imagery to make optimal composites—this is constrained by the size of the AOI.

Figure 5. Examples of image composite outputs from the GEEMMM showing lowest observable tide
(LOT), panel (a) and highest observable tide (HOT), panel (b). The north oriented, false colour (R: NIR
G: SWIR B: Red) Landsat image is over Kaingthaung Island, Ayeyarwady Region, Myanmar.
Remote Sens. 2020, 12, 3758 20 of 35

Figure 6. National overview of image composite outputs from the GEEMMM showing highest
observable tide (HOT) and lowest observable tide (LOT) (false color composites, R: NIR, G: SWIR,
B: Red). The composites were further reduced in area using topographic and combined water masks.
(A) Contemporary HOT; (B) Contemporary LOT; (C) Historic HOT; (D) Historic LOT.
Remote Sens. 2020, 12, 3758 21 of 35

3.2.2. Module 2—Spectral Separability, Classifications and Accuracy Assessment

Based on correlation analysis of all available spectral indices, five (i.e., MNDWI, CMRI, MMRI,
EVI, and LSWI) stood out as not correlated (Appendix A) and were selected as classification inputs.
Using the spectral separability tools, all target classes as represented by CRAs were assessed across all
non-thermal (red, green, blue, NIR, SWIR1, and SWIR2) Landsat bands (Figure 7), and each selected index
was further evaluated to confirm that it provided additional separation for one or more classes (e.g., MMRI:
Figure 8). Results indicate that bands SWIR1 and SWIR2 were particularly helpful in separating non-forest
vegetation. Non-forest vegetation was most confused with other vegetation classes in the visible spectrum
and MNDWI. For terrestrial forest, NIR, MNDWI, and MMRI provided separability. In particular,
MNDWI provided good separation from mangroves; whereas within the visible spectrum and CMRI the
most confusion was noted, particularly with other vegetation classes. Closed-canopy mangrove was best
distinguished by LSWI, MMRI, and to a limited extent bands SWIR1 and SWIR2. In contrast, open-canopy
mangrove was best distinguished by CMRI, MNDWI, NIR, SWIR1, and SWIR2. While there are meaningful
and distinct differences between the two canopy-based mangrove classes, there is spectral overlap—this
speaks to the advantage of capturing the variability within mangrove forests while subsequently merging
into a single class post-classification. Field work is required to confidently define the boundaries between
these sub-mangrove types to make them final map classes—following classification and prior to validation,
mangroves were merged into a single class (i.e., mangrove). Taken together, the combined mangrove class
exhibited some confusion with terrestrial forest and non-forest vegetation classes in EVI, the visible bands,
and SWIR1 and SWIR2. The exposed/barren class had the most separability in indices CMRI, MMRI,
and LSWI, and the most confusion with non-forest vegetation and terrestrial forest notably in MNDWI
and residual water in the visible bands. Residual water was easily distinguished with MNDWI, and the
non-visible bands, but was confused with exposed/barren in the visible bands, non-forest vegetation
within CMRI, and all classes within EVI.
For both historic and contemporary classifications, resubstitution accuracies were 100%, indicating
all training data was assigned to the correct land-cover class. Based on accuracy assessments using
independent validation data, overall accuracies for historic and contemporary classifications were
97.0 and 98.5%, respectively (Table 7). For the contemporary classification, there was slight confusion
between terrestrial forest and mangroves. Additionally, there was a small amount of two-way
confusion between non-forest vegetation and terrestrial forest. The greatest source of error for the
historic classification was the non-forest vegetation class, which was at-times confused with mangroves
and the exposed/barren class.
Remote Sens. 2020, 12, 3758 22 of 35

Figure 7. The spectral separability of all target classes as represented by CRAs across Landsat red (B1), green (B2), blue (B3), NIR (B4), SWIR1 (B5), and SWIR2 (B7)
bands. The set of bar and whisker plots shows the min, max, and interquartile range.
Remote Sens. 2020, 12, 3758 23 of 35

Figure 8. Example of index-specific overview of spectral values by land-cover class as represented by CRAs. MMRI is shown for each the historic and contemporary
HOT and LOT datasets. The bar-whisker plots represent the min, max, and interquartile range (IQR) for each class.
Remote Sens. 2020, 12, 3758 24 of 35

Table 7. Final Historic and Contemporary Validation Error Matrices, using validation CRAs pixels.

Historic Classification Validation Error Matrix

Terrestrial Forest Mangrove Exposed/Barren Residual Water Non-Forest Vegetation Total User’s Accuracy
Terrestrial Forest 24 0 0 0 0 24 100.0
Mangrove 0 54 0 0 0 54 100.0
Exposed/Barren 0 0 25 0 0 25 100.0
Residual Water 0 0 0 33 0 33 100.0
Non-Forest
0 4 1 0 26 31 83.9
Vegetation
Total 24 58 26 33 26 167
Producer’s
100.0 93.1 96.2 100.0 100.0
Accuracy
Overall Accuracy 162/167 97.0
Contemporary Classification Validation Error Matrix
Terrestrial Forest Mangrove Exposed/Barren Residual Water Non-Forest Vegetation Total User’s Accuracy
Terrestrial Forest 77 0 0 0 2 79 97.5
Mangrove 1 122 0 0 0 123 99.2
Exposed/Barren 0 0 33 0 0 33 100.0
Residual Water 0 0 0 24 0 24 100.0
Non-Forest
2 0 0 0 71 73 97.3
Vegetation
Total 80 122 33 24 73 327
Producer’s
96.3 100.0 100.0 100.0 97.3
Accuracy
Overall Accuracy 327/332 98.5
Remote Sens. 2020, 12, 3758 25 of 35

3.2.3. Module 3—Dynamics and QAA

Classification results indicate that circa 2004–2008, Myanmar contained 995,412 ha of mangroves.
In contrast, by 2014–2018, Myanmar contained 642,659 ha of mangroves. These results suggest that
from 2004–2008 to 2014–2018 there was 551,570.99 ha of loss and 198,818.42 ha of gain (i.e., net loss
352,752.57 ha or 35.4%) (Figures 4 and 9, Table 5). As compared to Estoque et al. [56] and Giri et al. [70],
estimated rates of loss are within reported trends and ranges; however, other studies reported lower
rates of loss often coinciding with lower total estimates of mangrove cover (i.e., Bunting et al. [74];
Hamilton and Casey [78]; Richards and Friess [47]; Estoque et al. [56]). Figure 9 shows LPG from
GEEMMM results within the loss hotspots identified through existing literature (i.e., Figure 1).

Figure 9. (Left) panels: Known mangrove loss hotspots (Figure 1). (Top left) shows loss, persistence,
and gain (LPG) from 2004–2008 to 2014–2018 in Rakhine State; (middle left) panel shows the Ayeyarwady
Region; (bottom left) shows Tanintharyi Region. (Right) Panel: contemporary high tide (HOT) image
composite, false colour (R: NIR G: SWIR B: Red) with boundaries of left panels highlighted in cyan.

While there was a substantial net loss based on GEEMMM results, the reported gain seems relatively
high. Portions of this likely reflect actual natural processes and increases in mangrove extent; however,
the overall gain estimate is likely an overestimation. Exaggerated gain likely reflects the desk-based process
of deriving CRAs. Clearly any classification is only as good as the examples used to calibrate the algorithm,
and a limitation of this pilot was no direct access to field observations or ground truth, and constrained
access to historical high spatial resolution satellite imagery. Disproportionate mangrove gain therefore likely
reflects an underrepresentation of lower stature, less dense mangroves in the historic classification, which in
turn exaggerates the amount of supposed gain (i.e., many of these areas were likely actually mangroves in
both dates). Extensive field work and ground verification is required to confirm.
The GEEMMM QAA was conducted for the contemporary map, then repeated for the historic
map. As part of the contemporary QAA, spot-checks were conducted over 108 sub-grid cells
across Myanmar (Figure 1). The mangrove class was generally well-represented; however, at-times
under-represented in favor of classes depicting portions of areas in the variable agricultural mosaic,
i.e., non-forest vegetation, and exposed/barren. In both the contemporary, and less so the historic map,
the agricultural mosaic was depicted as a patchwork of these two classes, on a pixel-by-pixel basis,
given the inherent variability within the seasonal window. This resulted in some confusion between
the two classes, and to some extent an under-representation of mangroves. In the contemporary map,
sparser mangroves at the ecosystem periphery were at-times misclassified as non-forest vegetation,
thereby under-representing mangrove and over-representing non-forest vegetation. Terrestrial forest
Remote Sens. 2020, 12, 3758 26 of 35

was also at-times over-represented, occasionally at the expense of actual mangrove areas. Overall,
the contemporary classification appeared to best represent Myanmar’s south (i.e., the Tanintharyi
coastline). The historic QAA, while not quite as comprehensive as the contemporary QAA (mainly due
to the absence of historic imagery in GEE), found the mangrove class to be generally well-represented,
though at-times over-represented at the expense of classes depicting the agricultural mosaic—an inverse
to the contemporary map. Some portions of the agricultural mosaic were also found to be misclassified
as terrestrial forest. As with the contemporary map, Myanmar’s southern Tanintharyi coastline seemed
best represented. Notably, most existing studies did not provide standard quantitative accuracy
assessments, and no existing studies went beyond these and further qualitatively assessed resulting
maps. While quantitative accuracy assessments should be a standard part of reporting, QAAs also
help further assess resulting maps and identify areas for improvement. As such, the GEEMMM goes
beyond standard accuracy—for which GEEMMM results were very high—and allows users to more
closely examine actual distributions and subsequently dynamics.

3.2.4. Dissemination and Improvement

The code is available in a GitHub repository (see Supplementary Materials Section), with a GNU
GPLv3 license permitting free use, modification, and sharing, provided that the source is disclosed and not
used for commercial purposes. The code runs based on provided links, or is copied-and-pasted into GEE,
which remains available for free non-profit and educational use. The tool itself continues to be adjusted and
updated, as the GEE library evolves and as new mangrove remote sensing techniques become available.
While the tool performs well there are always potential improvements. Notably, CRAs are a key
input for the workflow, and highly influence the outcome of the classifications. Future applications
of the GEEMMM would benefit from direct access to field-based ground truth when deriving CRAs,
particularly when it comes to confidently using mangrove sub-types as final map classes. While the
need for and merit of isolating tidal conditions is proven, which tidal conditions are best requires
further exploration—we used combined HOT and LOT in this pilot, whereas HOT or LOT on its
own could also be employed. Furthermore, the choice of tidal condition depends on the intended
application. For example, mangrove carbon projects may favor using HOT composites on their own
for more conservative estimates of mangrove extent and change.
While going beyond standard accuracy metrics, the QAA is a somewhat complex component
requiring significant user interaction; however, it too will evolve as the GEEMMM is further tested
with other settings and applied to other AOIs. GEE itself also has notable limitations: the AOI can
be as large or as small as the user requires but GEE has computational limits. Google shares its
cloud processing among all GEE users, which means that if the task requested to process is too large
(e.g., a long complex coastline, with collections containing hundreds of images) the user’s allocated
capacity may be exceeded and error(s) returned. Additionally, the functioning of this tool requires a
relatively stable and reasonably strong internet connection, especially to view images and products
within the GUI. If internet connectivity is limited, there may be latency issues loading data or even
time-out errors. One of the benefits of working within the GEE environment however is that once a
data product export has begun it will be completed on Google’s server side. This means that internet
access can be interrupted while using the tool, and it will continue to run. It was this feature of GEE,
the server-side image/vector data exporting that drove the current configuration of three modules,
where intermediate data products are exported to the user’s assets, effectively saving their progress
through the tool.
The GEEMMM is currently designed around the use of Landsat data—this was a conscious choice
based around data availability. Sentinel imagery—which is also available through GEE—offers an
increased revisit time (i.e., higher temporal resolution) and finer spatial resolution; however, it remains
limited by a 2015-present temporal window. In contrast, the Landsat archive in GEE offers >35 years
of imagery which facilitates more historically meaningful and robust dynamics assessments while
also providing enough imagery to draw from multiple years to produce composites within preferred
Remote Sens. 2020, 12, 3758 27 of 35

seasonal windows. Given the added benefits—especially once the archive spans 10+ years—future
versions of the GEEMMM should also offer the choice of Sentinel imagery to users as an option.

4. Conclusions
We present a new tool—the GEEMMM—for mapping and monitoring mangrove ecosystems.
By leveraging GEE, this new tool circumvents many traditional barriers to conventional methods. In addition,
it presents an internal, image-based approach for tidal calibration. The GEEMMM—including the well
commented source code—is available online and is ready to be used by practitioners anywhere mangrove
ecosystems exist; please see information in Supplementary Material Section on how to access the GEEMMM.
While operational, the GEEMMM is not without its limitations: the larger the area the more complex the
mapping task, particularly when it comes to creating optimal imagery composites within defined seasonal
windows. In addition, the upper limits of GEE and internet connectivity present a challenge in terms of the
time associated with and reliability of running the GEEMMM; however, when compared to the conventional
processing times associated with standalone workstations it remains much faster, and once a part of the
GEEMMM starts running it will continue to run even if the internet connection is lost. In any application,
the resulting maps and dynamics assessments will only ever be as good as the examples of target map
classes provided. Coastal managers will normally have such information available to them and GEEMMM
provides them with a framework through which to capitalizes on this local knowledge, rather than relying
on external datasets, which allow little to no customization, to map and monitor their mangroves.
The GEEMMM makes a significant and ready-to-go contribution toward accessible mangrove
mapping and monitoring. It also remains a living tool wherein non-profit users are encouraged by
the authors to make useful suggestions for modifications or additions, or modify the tool directly
themselves to meet their own customized needs. While piloting the GEEMMM for Myanmar is an
important first step, additional applications and tests are required, particularly for smaller areas of
interest, wherein the GEEMMM can help fill a critical sub-national mapping gap. The authors welcome
the opportunity to receive feedback from and work with users to more comprehensively assess the
tool and gauge areas for improvement. A series of in-person and online instructional materials will go
a long way toward ensuring the maximum and optimal utility of the GEEMMM. This first iteration of
the GEEMMM further sets the stage for a comparatively more automated and even more accessible
version to be deployable completely on mobile devices.

Supplementary Materials: The GEEMMM tool is freely available within the GitHub repository: https://github.
com/Blue-Ventures-Conservation/GEEMMM.
Author Contributions: The GEEMMM was conceived of by, T.G.J., S.R.G., L.G., C.F., and J.M.M.Y. Contributions
to the methodology were made by T.G.J., S.R.G., and C.F., with J.M.M.Y. developing the key tidal detection
methods. J.M.M.Y. wrote all of the code in GEE for the GEEMMM tool, with the work reviewed by C.F. The results
of this paper were validated by J.M.M.Y., T.G.J., S.R.G., C.F., and A.L. Formal analysis was performed by J.M.M.Y.,
T.G.J., and S.R.G., using the analysis tools developed by J.M.M.Y. Investigation for this work was conducted by
J.M.M.Y., T.G.J., S.R.G., C.F., and A.L., J.M.M.Y. performed all of the data curation for this paper. The original
manuscript writing was conducted by T.G.J. and J.M.M.Y.; with T.G.J. writing the introduction, discussion points,
and conclusion and J.M.M.Y. writing the bulk of the methods and results. All authors, T.G.J., J.M.M.Y., S.R.G., C.F.,
A.L., and, L.G. were involved in writing—review and editing. Visualizations were generated by J.M.M.Y., A.L.,
and S.R.G. The project was administrated and supervised by T.G.J. and L.G. All authors have read and agreed to
the published version of the manuscript.
Funding: This research was funded by Blue Ventures Conservation, with support from the UK Government’s
International Climate Fund, part of the UK commitment to developing countries to help them address the
challenges presented by climate change and benefit from the opportunities.
Acknowledgments: We thank the following authors of studies referenced in this paper: Ake Rosenqvist, of solo
Earth Observation (soloEO), for provision of and support regarding GMW data; J. Ronald Eastman and James
Toledano, of Clark Labs, for guidance regarding the Clark Labs Aquaculture dataset; Chandra Giri, of United States
Environmental Protection Agency, for provision of data and associated guidance; Edward L. Webb, of National
University Singapore, for data provision and guidance; and Ate Poortinga, of Spatial Informatics Group,
for provision of and guidance regarding SERVIR-Mekong data.
Conflicts of Interest: The authors declare no conflict of interest.
Remote Sens. 2020, 12, 3758 28 of 35

Appendix A

Table A1. Correlation matrices for the calculated spectral indices for both HOT and LOT historic and contemporary imagery extracted from the CRAs.

Contemporary HOT Index Band Correlation

SR NDVI NDWI MNDWI CMRI MMRI SAVI OSAVI EVI MRI SMRI LSWI NDTI EBBI
SR 1 0.654 −0.61 −0.42 0.256 −0.46 0.654 0.654 0.13 0.018 −0.14 0.498 0.719 −0.71
NDVI 0.654 1 −0.98 −0.85 0.177 −0.7 0.999 0.999 0.203 −0.11 −0.38 0.163 0.755 −0.76
NDWI −0.61 −0.98 1 0.907 0.005 0.679 −0.98 −0.98 −0.18 0.118 0.392 −0.06 −0.69 0.698
MNDWI −0.42 −0.85 0.907 1 0.222 0.475 −0.85 −0.85 −0.14 0.222 0.342 0.328 −0.43 0.448
CMRI 0.256 0.177 0.005 0.222 1 −0.18 0.177 0.177 0.14 0 8.709 * 0.562 0.407 −0.4
MMRI −0.46 −0.7 0.679 0.475 −0.18 1 −0.7 −0.7 −0.08 −0.04 0.154 −0.31 −0.72 0.617
SAVI 0.654 0.999 −0.98 −0.85 0.177 −0.7 1 0.999 0.203 −0.11 −0.38 0.163 0.755 −0.76
OSAVI 0.654 0.999 −0.98 −0.85 0.177 −0.7 0.999 1 0.203 −0.11 −0.38 0.163 0.755 −0.76
EVI 0.13 0.203 −0.18 −0.14 0.14 −0.08 0.203 0.203 1 −0.01 −0.09 0.032 0.145 −0.15
MRI 0.018 −0.11 0.118 0.222 0 −0.04 −0.11 −0.11 −0.01 1 0.033 0.227 0.067 0.085
SMRI −0.14 −0.38 0.392 0.342 8.709 * 0.154 −0.38 −0.38 −0.09 0.033 1 0.028 −0.27 0.175
LSWI 0.498 0.163 −0.06 0.328 0.562 −0.31 0.163 0.163 0.032 0.227 0.028 1 0.566 −0.6
NDTI 0.719 0.755 −0.69 −0.43 0.407 −0.72 0.755 0.755 0.145 0.067 −0.27 0.566 1 −0.79
EBBI −0.71 −0.76 0.698 0.448 −0.4 0.617 −0.76 −0.76 −0.15 0.085 0.175 −0.6 −0.79 1
*—Denotes an error output from the GEE servers for the index correlations.
Contemporary LOT Index Band Correlation
SR NDVI NDWI MNDWI CMRI MMRI SAVI OSAVI EVI MRI SMRI LSWI NDTI EBBI
SR 1 0.718 −0.67 −0.44 0.286 −0.47 0.718 0.718 0.158 0.054 −0.14 0.509 0.717 −0.76
NDVI 0.718 1 −0.97 −0.79 0.234 −0.78 0.999 0.999 0.223 −0.1 −0.25 0.267 0.75 −0.79
NDWI −0.67 −0.97 1 0.871 −0.02 0.774 −0.97 −0.97 −0.23 0.119 0.236 −0.14 −0.67 0.714
MNDWI −0.44 −0.79 0.871 1 0.27 0.492 −0.79 −0.79 −0.22 0.251 0.221 0.337 −0.3 0.374
CMRI 0.286 0.234 −0.02 0.27 1 −0.13 0.234 0.234 −0.03 0.033 −0.1 0.612 0.459 −0.46
MMRI −0.47 −0.78 0.774 0.492 −0.13 1 −0.78 −0.78 −0.21 −0.04 0.137 −0.37 −0.69 0.65
SAVI 0.718 0.999 −0.97 −0.79 0.234 −0.78 1 0.999 0.223 −0.1 −0.25 0.267 0.75 −0.79
OSAVI 0.718 0.999 −0.97 −0.79 0.234 −0.78 0.999 1 0.223 −0.1 −0.25 0.267 0.75 −0.79
EVI 0.158 0.223 −0.23 −0.22 −0.03 −0.21 0.223 0.223 1 −0.02 −0.04 0.006 0.126 −0.17
MRI 0.054 −0.1 0.119 0.251 0.033 −0.04 −0.1 −0.1 −0.02 1 0.033 0.253 0.081 −0.02
SMRI −0.14 −0.25 0.236 0.221 −0.1 0.137 −0.25 −0.25 −0.04 0.033 1 −0.01 −0.14 0.134
LSWI 0.509 0.267 −0.14 0.337 0.612 −0.37 0.267 0.267 0.006 0.253 −0.01 1 0.716 −0.67
NDTI 0.717 0.75 −0.67 −0.3 0.459 −0.69 0.75 0.75 0.126 0.081 −0.14 0.716 1 −0.89
EBBI 0.76 −0.79 0.714 0.374 −0.46 0.65 −0.79 −0.79 −0.17 −0.02 0.134 −0.67 −0.89 1
Remote Sens. 2020, 12, 3758 29 of 35

Table A1. Cont.

Historic HOT Index Band Correlation

SR NDVI NDWI MNDWI CMRI MMRI SAVI OSAVI EVI MRI SMRI LSWI NDTI EBBI
SR 1 0.844 −0.78 −0.56 0.603 −0.74 0.844 0.844 0.406 −0.17 −0.26 0.123 0.373 −0.83
NDVI 0.844 1 −0.98 −0.83 0.418 −0.82 0.999 0.999 0.38 −0.25 −0.45 −0.26 0.248 −0.69
NDWI −0.78 −0.98 1 0.894 −0.26 0.773 −0.98 −0.98 −0.34 0.232 0.49 0.38 −0.2 0.595
MNDWI −0.56 −0.83 0.894 1 0.027 0.618 −0.83 −0.83 −0.26 0.239 0.435 0.718 −0.06 0.244
CMRI 0.603 0.418 −0.26 0.027 1 −0.54 0.418 0.418 0.341 −0.21 0.03 0.56 0.341 −0.77
MMRI −0.74 −0.82 0.773 0.618 −0.54 1 −0.82 −0.82 −0.45 0.184 0.279 0.031 −0.33 0.719
SAVI 0.844 0.999 −0.98 −0.83 0.418 −0.82 1 0.999 0.38 −0.25 −0.45 −0.26 0.248 −0.69
OSAVI 0.844 0.999 −0.98 −0.83 0.418 −0.82 0.999 1 0.38 −0.25 −0.45 −0.26 0.248 −0.69
EVI 0.406 0.38 −0.34 −0.26 0.341 −0.45 0.38 0.38 1 −0.09 −0.02 0.064 0.151 −0.38
MRI −0.17 −0.25 0.232 0.239 −0.21 0.184 −0.25 −0.25 −0.09 1 −0.16 0.104 −0.02 0.231
SMRI −0.26 −0.45 0.49 0.435 0.03 0.279 −0.45 −0.45 −0.02 −0.16 1 0.267 0.089 0.16
LSWI 0.123 −0.26 0.38 0.718 0.56 0.031 −0.26 −0.26 0.064 0.104 0.267 1 0.23 −0.45
NDTI 0.373 0.248 −0.2 −0.06 0.341 −0.33 0.248 0.248 0.151 −0.02 0.089 0.23 1 −0.38
1
EBBI −0.83 −0.69 0.595 0.244 −0.77 0.719 −0.69 −0.69 −0.38 0.231 0.16 −0.45 −0.38

Historic LOT Index Band Correlation

SR NDVI NDWI MNDWI CMRI MMRI SAVI OSAVI EVI MRI SMRI LSWI NDTI EBBI
SR 1 0.852 −0.77 −0.42 0.529 −0.77 0.852 0.852 0.135 0.158 −0.3 0.412 0.714 −0.82
NDVI 0.852 1 −0.97 −0.74 0.325 −0.86 0.999 0.999 0.183 0.037 −0.39 0.084 0.705 −0.69
NDWI −0.77 −0.97 1 0.842 −0.11 0.824 −0.97 −0.97 −0.18 0.056 0.398 0.072 −0.63 0.574
MNDWI −0.42 −0.74 0.842 1 0.28 0.539 −0.74 −0.74 −0.14 0.27 0.24 0.586 −0.28 0.086
CMRI 0.529 0.325 −0.11 0.28 1 −0.38 0.325 0.325 0.041 0.429 −0.08 0.724 0.485 −0.71
MMRI −0.77 −0.86 0.824 0.539 −0.38 1 −0.86 −0.86 −0.14 −0.15 0.341 −0.25 −0.67 0.728
SAVI 0.852 0.999 −0.97 −0.74 0.325 −0.86 1 0.999 0.183 0.037 −0.39 0.084 0.705 −0.69
OSAVI 0.852 0.999 −0.97 −0.74 0.325 −0.86 0.999 1 0.183 0.037 −0.39 0.084 0.705 −0.69
EVI 0.135 0.183 −0.18 −0.14 0.041 −0.14 0.183 0.183 1 0.004 −0.01 0.008 0 −0.11
MRI 0.158 0.037 0.056 0.27 0.429 −0.15 0.037 0.037 0.004 1 −0.16 0.414 0.167 −0.32
SMRI −0.3 −0.39 0.398 0.24 −0.08 0.341 −0.39 −0.39 −0.01 −0.16 1 −0.11 −0.25 0.322
LSWI 0.412 0.084 0.072 0.586 0.724 −0.25 0.084 0.084 0.008 0.414 −0.11 1 0.405 −0.71
NDTI 0.714 0.705 −0.63 −0.28 0.485 −0.67 0.705 0.705 0 0.167 −0.25 0.405 1 −0.72
EBBI −0.82 −0.69 0.574 0.086 −0.71 0.728 −0.69 −0.69 −0.11 −0.32 0.322 −0.71 −0.72 1
Remote Sens. 2020, 12, 3758 30 of 35

References
1. Saenger, P. Mangrove Ecology, Silviculture and Conservation; Springer Science & Business Media: Berlin,
Germany, 2002; Volume 3, ISBN 9789048160501.
2. Food and Agriculture Organization of the United Nations; Forestry Department (Rome). Global Forest
Resources Assessment 2000: Main Report; Food and Agriculture Organization of the United Nations: Rome,
Italy, 2001.
3. Xia, Q.; Qin, C.Z.; Li, H.; Huang, C.; Su, F.Z.; Jia, M.M. Evaluation of submerged mangrove recognition index
using multi-tidal remote sensing data. Ecol. Indic. 2020, 113. [CrossRef]
4. Lugo, A.E. Mangrove Ecosystems: Successional or Steady State? Biotropica 1980, 12, 65–72. [CrossRef]
5. Spalding, M.; Kainuma, M.; Collins, L. World Atlas of Mangroves; Earthscan: New York, NY, USA, 2010; ISBN
978-1-84407-657-4.
6. Duke, N.; Nagelkerken, I.; Agardy, T.; Wells, S.; van Lavieren, H. The Importance of Mangroves to People: A Call
to Action; van Bochove, J.-W., Sullivan, E., Nakamura, T., Eds.; United Nations Environment Programme
World Conservation Monitoring Centre: Cambridge, UK, 2014; ISBN 9789280733976.
7. Scales, I.R.; Friess, D.A.; Glass, L.; Ravaoarinorotsihoarana, L. Rural livelihoods and mangrove degradation
in south-west Madagascar: Lime production as an emerging threat. Oryx 2018, 52, 641–645. [CrossRef]
8. Blue Ventures Conservation. Value Chain Analysis: The wild capture mud crab fishery of Madagascar’s
Menabe region. Available online: https://blueventures.org/publication/value-chain-analysis-the-wild-
capture-mud-crab-fishery-of-madagascars-menabe-region/ (accessed on 1 October 2020).
9. Aye, W.N.; Wen, Y.; Marin, K.; Thapa, S.; Tun, A.W. Contribution of mangrove forest to the livelihood of local
communities in Ayeyarwaddy Region, Myanmar. Forests 2019, 10, 414. [CrossRef]
10. Nagelkerken, I.; Blaber, S.J.M.; Bouillon, S.; Green, P.; Haywood, M.; Kirton, L.G.; Meynecke, J.O.; Pawlik, J.;
Penrose, H.M.; Sasekumar, A.; et al. The habitat function of mangroves for terrestrial and marine fauna:
A review. Aquat. Bot. 2008, 89, 155–185. [CrossRef]
11. Gopal, B.; Chauhan, M. Biodiversity and its conservation in the Sundarban mangrove ecosystem. Aquat. Sci.
2006, 68, 338–354. [CrossRef]
12. Gardner, C.J. Use of Mangroves by Lemurs. Int. J. Primatol. 2016, 37, 317–332. [CrossRef]
13. Alongi, D.M. Carbon cycling and storage in mangrove forests. Annu. Rev. Mar. Sci. 2014, 6, 195–219.
[CrossRef]
14. 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]
15. Sanderman, J.; Hengl, T.; Fiske, G.; Solvik, K.; Adame, M.F.; Benson, L.; Bukoski, J.J.; Carnell, P.;
Cifuentes-Jara, M.; Donato, D.; et al. A global map of mangrove forest soil carbon at 30 m spatial
resolution. Environ. Res. Lett. 2018, 13. [CrossRef]
16. Rakotomahazo, C.; Ravaoarinorotsihoarana, L.A.; Randrianandrasaziky, D.; Glass, L.; Gough, C.;
Boleslas Todinanahary, G.G.; Gardner, C.J. Participatory planning of a community-based payments for
ecosystem services initiative in Madagascar’s mangroves. Ocean Coast. Manag. 2019, 175, 43–52. [CrossRef]
17. Ahmed, N.; Glaser, M. Coastal aquaculture, mangrove deforestation and blue carbon emissions: Is REDD+ a
solution? Mar. Policy 2016, 66, 58–66. [CrossRef]
18. Valiela, I.; Bowen, J.L.; York, J.K. Mangrove forests: One of the world’s threatened major tropical environments.
BioScience 2001, 51, 807–815. [CrossRef]
19. Gandhi, S.; Jones, T.G. Identifying mangrove deforestation hotspots in South Asia, Southeast Asia and
Asia-Pacific. Remote Sens. 2019, 11, 728. [CrossRef]
20. Richards, D.R.; Thompson, B.S.; Wijedasa, L. Quantifying net loss of global mangrove carbon stocks from
20 years of land cover change. Nat. Commun. 2020, 11, 4260. [CrossRef]
21. Global Mangrove Watch Mangrove Atlas. Available online: https://www.globalmangrovewatch.org/
(accessed on 8 September 2020).
22. Friess, D.A.; Rogers, K.; Lovelock, C.E.; Krauss, K.W.; Hamilton, S.E.; Lee, S.Y.; Lucas, R.; Primavera, J.;
Rajkaran, A.; Shi, S. The State of the World’s Mangrove Forests: Past, Present, and Future. Ann. Rev. Environ.
Resour. 2019, 44, 89–115. [CrossRef]
23. 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]
Remote Sens. 2020, 12, 3758 31 of 35

24. UN Millennium Project. Investing in Development: A Practical Plan to Achieve the Millennium Development Goals.
Overview; United Nations Development Programme: Washington, DC, USA, 2005; ISBN 1844072177.
25. Ramsar Convention Secretariat. The Fourth RAMSAR Strategic Plan 2016–2024. Available online: https:
//www.ramsar.org/the-ramsar-strategic-plan-2016--24 (accessed on 1 October 2020).
26. Wang, L.; Jia, M.; Yin, D.; Tian, J. A review of remote sensing for mangrove forests: 1956–2018.
Remote Sens. Environ. 2019, 231. [CrossRef]
27. 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]
28. 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]
29. 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]
30. Sagar, S.; Roberts, D.; Bala, B.; Lymburner, L. Extracting the intertidal extent and topography of the Australian
coastline from a 28 year time series of Landsat observations. Remote Sens. Environ. 2017, 195, 153–169.
[CrossRef]
31. Bishop-Taylor, R.; Sagar, S.; Lymburner, L.; Beaman, R.J. Between the tides: Modelling the elevation of
Australia’s exposed intertidal zone at continental scale. Estuar. Coast. Shelf Sci. 2019, 223, 115–128. [CrossRef]
32. Murray, N.J.; Phinn, S.R.; DeWitt, M.; Ferrari, R.; Johnston, R.; Lyons, M.B.; Clinton, N.; Thau, D.; Fuller, R.A.
The global distribution and trajectory of tidal flats. Nature 2019, 565, 222–225. [CrossRef]
33. Martin, P.J.; Smith, S.R.; Posey, P.G.; Dawson, G.M.; Riedlinger, S.H. Use of the Oregon State University Tidal
Inversion Software (OTIS) to Generate Improved Tidal Prediction in the East-Asian Seas; Stennis Space Center:
Hancock, MS, USA, 2009.
34. 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]
35. U.S. Geological Survey. Landsat 8 Collection 1 (C1) Land Surface Reflectance Code (LaSRC) Product Guide;
LSDS-1368; U.S. Geological Survey, Dempartment of the Interior: Sioux Falls, SD, USA, 2020.
36. Wulder, M.A.; Coops, N.C.; Roy, D.P.; White, J.C.; Hermosilla, T. Land cover 2.0. Int. J. Remote Sens. 2018, 39,
4254–4284. [CrossRef]
37. Hansen, M.C.; Potapov, P.V.; Moore, R.; Hancher, M.; Turubanova, S.A.; Tyukavina, A.; Thau, D.; Stehman, S.V.;
Goetz, S.J.; Loveland, T.R.; et al. High-Resolution Global Maps of 21st-Century Forest Cover Change. Science
2013, 342, 850–853. [CrossRef]
38. Vos, K.; Splinter, K.D.; Harley, M.D.; Simmons, J.A.; Turner, I.L. CoastSat: A Google Earth Engine-enabled
Python toolkit to extract shorelines from publicly available satellite imagery. Environ. Model. Softw. 2019,
122, 104528. [CrossRef]
39. Chen, B.; Xiao, X.; Li, X.; Pan, L.; Doughty, R.; Ma, J.; Dong, J.; Qin, Y.; Zhao, B.; Wu, Z.; 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]
40. Pimple, U.; Simonetti, D.; Sitthi, A.; Pungkul, S.; Leadprathom, K.; Skupek, H.; Som-ard, J.; Gond, V.;
Towprayoon, S. Google Earth Engine Based Three Decadal Landsat Imagery Analysis for Mapping of
Mangrove Forests and Its Surroundings in the Trat Province of Thailand. J. Comput. Commun. 2018, 6,
247–264. [CrossRef]
41. Tieng, T.; Sharma, S.; Mackenzie, R.A.; Venkattappa, M.; Sasaki, N.K.; Collin, A. Mapping mangrove forest
cover using Landsat-8 imagery, Sentinel-2, Very High Resolution Images and Google Earth Engine algorithm
for entire Cambodia. In Proceedings of the IOP Conference Series: Earth and Environmental Science,
4th International Forum on Sustainable Future in Asia/4th NIES International Forum, Pan Pacific Hanoi,
Vietnam, 23–24 January 2019; IOP Publishing: Bristol, UK, 2019; Volume 266.
42. Mondal, P.; Trzaska, S.; de Sherbinin, A. Landsat-derived estimates of mangrove extents in the Sierra Leone
coastal landscape complex during 1990–2016. Sensors 2018, 18, 12. [CrossRef] [PubMed]
43. Diniz, C.; Cortinhas, L.; Nerino, G.; Rodrigues, J.; Sadeck, L.; Adami, M.; Souza-Filho, P.W.M. Brazilian
mangrove status: Three decades of satellite data analysis. Remote Sens. 2019, 11, 808. [CrossRef]
Remote Sens. 2020, 12, 3758 32 of 35

44. 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]
45. Ellison, A.M.; Farnsworthf, E.J.; Merkt, R.E. Origins of Mangrove Ecosystems and the Mangrove Biodiversity
Anomaly. Glob. Ecol. Biogeogr. 1999, 8, 95–115.
46. Food and Agricultural Organization (FAO). Loss of Mangroves Alarming. Available online: http://www.fao.
org/newsroom/en/news/2008/1000776/index.html (accessed on 30 August 2020).
47. Richards, D.R.; Friess, D.A. Rates and drivers of mangrove deforestation in Southeast Asia, 2000–2012.
Proc. Natl. Acad. Sci. USA 2016, 113, 344–349. [CrossRef]
48. Farnsworth, E.J.; Ellison, A.M. The global conservation status of mangroves. AMBIO 1997, 26, 328–334.
[CrossRef]
49. Primavera, J.H. Development and conservation of Philippine mangroves: Institutional issues. Ecol. Econ.
2000, 35, 91–106. [CrossRef]
50. Dahdouh-Guebas, F. The use of remote sensing and GIS in the sustainable management of tropical coastal
ecosystems. Environ. Dev. Sustain. 2002, 4, 93–112. [CrossRef]
51. Primavera, J.H. Mangroves, Fishponds, and the Quest for Sustainability. Science 2005, 310, 57–59. [CrossRef]
52. Primavera, J.H. Overcoming the impacts of aquaculture on the coastal zone. Ocean Coast. Manag. 2006, 49,
531–545. [CrossRef]
53. Gilman, E.L.; Ellison, J.; Duke, N.C.; Field, C. Threats to mangroves from climate change and adaptation
options: A review. Aquat. Bot. 2008, 89, 237–250. [CrossRef]
54. Walters, B.B.; Rönnbäck, P.; Kovacs, J.M.; Crona, B.; Hussain, S.A.; Badola, R.; Primavera, J.H.; Barbier, E.;
Dahdouh-Guebas, F. Ethnobiology, socio-economics and management of mangrove forests: A review.
Aquat. Bot. 2008, 89, 220–236. [CrossRef]
55. Webb, E.L.; Jachowski, N.R.A.; Phelps, J.; Friess, D.A.; Than, M.M.; Ziegler, A.D. Deforestation in
the Ayeyarwady Delta and the conservation implications of an internationally-engaged Myanmar.
Glob. Environ. Chang. 2014, 24, 321–333. [CrossRef]
56. Estoque, R.C.; Myint, S.W.; Wang, C.; Ishtiaque, A.; Aung, T.T.; Emerton, L.; Ooba, M.; Hijioka, Y.; Mon, M.S.;
Wang, Z.; et al. Assessing environmental impacts and change in Myanmar’s mangrove ecosystem service
value due to deforestation (2000–2014). Glob. Chang. Biol. 2018, 24, 5391–5410. [CrossRef]
57. De Alban, J.D.T.; Jamaludin, J.; Wong De Wen, D.; Than, M.M.; Webb, E.L. Improved estimates of mangrove
cover and change reveal catastrophic deforestation in Myanmar. Environ. Res. Lett. 2020, 15, 034034.
[CrossRef]
58. Alongi, D.M. Present state and future of the world’s mangrove forests. Environ. Conserv. 2002, 29, 331–349.
[CrossRef]
59. Alongi, D.M. Mangrove forests: Resilience, protection from tsunamis, and responses to global climate change.
Estuar. Coast. Shelf Sci. 2008, 76, 1–13. [CrossRef]
60. Alongi, D.M. The Impact of Climate Change on Mangrove Forests. Curr. Clim. Chang. Rep. 2015, 1, 30–39.
[CrossRef]
61. Sitoe, A.A.; Mandlate, L.J.C.; Guedes, B.S. Biomass and carbon stocks of Sofala Bay mangrove forests. Forests
2014, 5, 1967–1981. [CrossRef]
62. Field, C.D. Impact of expected climate change on mangroves. Hydrobiologia 1995, 295, 75–81. [CrossRef]
63. Krauss, K.W.; Lovelock, C.E.; McKee, K.L.; López-Hoffman, L.; Ewe, S.M.L.; Sousa, W.P. Environmental
drivers in mangrove establishment and early development: A review. Aquat. Bot. 2008, 89, 105–127.
[CrossRef]
64. Chan, H.T.; Baba, S. Manual on Guidlines for Rehabilitation of Coastal Forests Damaged by Natural Hazards in
the Asia-Pacific Region; International Society for Mangrove Ecosystems and International Tropical Timber
Organization: Okinawa, Japan, 2009; ISBN 9784906584130.
65. Suzuki, T.; Zijlema, M.; Burger, B.; Meijer, M.C.; Narayan, S. Wave dissipation by vegetation with layer
schematization in SWAN. Coast. Eng. 2012, 59, 64–71. [CrossRef]
66. Di Nitto, D.; Neukermans, G.; Koedam, N.; Defever, H.; Pattyn, F.; Kairo, J.G.; Dahdouh-Guebas, F. Mangroves
facing climate change: Landward migration potential in response to projected scenarios of sea level rise.
Biogeosciences 2014, 11, 857–871. [CrossRef]
Remote Sens. 2020, 12, 3758 33 of 35

67. Giesen, W.; Wulffraat, S.; Zieren, M.; Scholten, L. Mangrove guidebook for Southeast Asia; Food and Agriculture
Organization of the United Nations, Regional Office for Asia and the Pacific: Bangkok, Thailand, 2006.
68. 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]
69. Romañach, S.S.; DeAngelis, D.L.; Koh, H.L.; Li, Y.; Teh, S.Y.; Raja Barizan, R.S.; Zhai, L. Conservation and
restoration of mangroves: Global status, perspectives, and prognosis. Ocean Coast. Manag. 2018, 154, 72–82.
[CrossRef]
70. Giri, C.; Zhu, Z.; Tieszen, L.L.; Singh, A.; Gillette, S.; Kelmelis, J.A. Mangrove Forest Distributions and
Dynamics (1975-2005) of the Tsunami-Affected Region of Asia. J. Biogeogr. 2008, 35, 519–528. [CrossRef]
71. Giri, C.; Long, J.; Abbas, S.; Murali, R.M.; Qamer, F.M.; Pengra, B.; Thau, D. Distribution and dynamics of
mangrove forests of South Asia. J. Environ. Manag. 2015, 148, 101–111. [CrossRef]
72. Bhattarai, B. Assessment of mangrove forests in the Pacific region using Landsat imagery. J. Appl. Remote Sens.
2011, 5, 053509. [CrossRef]
73. Saah, D.; Tenneson, K.; Poortinga, A.; Nguyen, Q.; Chishtie, F.; Aung, K.S.; Markert, K.N.; Clinton, N.;
Anderson, E.R.; Cutter, P.; et al. Primitives as building blocks for constructing land cover maps. Int. J. Appl.
Earth Obs. Geoinf. 2020, 85, 101979. [CrossRef]
74. Bunting, P.; Rosenqvist, A.; Lucas, R.M.; Rebelo, L.M.; Hilarides, L.; Thomas, N.; Hardy, A.; Itoh, T.;
Shimada, M.; Finlayson, C.M. The global mangrove watch—A new 2010 global baseline of mangrove extent.
Remote Sens. 2018, 10, 1669. [CrossRef]
75. Clark Labs. Coastal Habitat Mapping: Mangrove and Pond Aquaculture Conversion. Available online:
https://clarklabs.org/aquaculture/ (accessed on 30 August 2020).
76. Stibig, H.-J.; Belward, A.S.; Roy, P.S.; Rosalina-Wasrin, U.; Agrawal, S.; Joshi, P.K.; Hildanus; Beuchle, R.;
Fritz, S.; Mubareka, S.; et al. A Land-Cover Map for South and Southeast Asia Derived from
SPOT-VEGETATION Data. J. Biogeogr. 2007, 34, 625–637. [CrossRef]
77. Blasco, F.; Aizpuru, M.; Gers, C. Depletion of the mangroves of Continental Asia. Wetl. Ecol. Manag. 2001, 9,
245–256. [CrossRef]
78. Hamilton, S.E.; Casey, D. Creation of a high spatio-temporal resolution global database of continuous
mangrove forest cover for the 21st century (CGMFC-21). Glob. Ecol. Biogeogr. 2016, 25, 729–738. [CrossRef]
79. Google Earth Pro; Google LLC: Mountain View, CA, USA, 2020.
80. Global LSIB Polygons Detailed 2017. Available online: https://catalog.data.gov/dataset/global-lsib-polygons-
detailed-2017dec29 (accessed on 22 September 2020).
81. Tadono, T.; Ishida, H.; Oda, F.; Naito, S.; Minakawa, K.; Iwamoto, H. Precise Global DEM Generation by
ALOS PRISM. ISPRS Ann. Photogramm. Remote Sens. Spat. Inf. Sci. 2014, II-4, 71–76. [CrossRef]
82. GADM Maps and Data. Available online: https://www.gadm.org/ (accessed on 22 September 2020).
83. Li, P.; Feng, Z.; Jiang, L.; Liao, C.; Zhang, J. A review of swidden agriculture in Southeast Asia. Remote Sens.
2014, 6, 1654–1683. [CrossRef]
84. Streets, D.G.; Yarber, K.F.; Woo, J.-H.; Carmichael, G.R. Biomass burning in Asia: Annual and seasonal
estimates and atmospheric emissions. Glob. Biogeochem. Cycles 2003, 17. [CrossRef]
85. Davis, J.H. The Forests of Burma; New York Botanical Garden: New York, NY, USA, 1964.
86. U.S. Geological Survey. Landsat 8 (L8) Data Users Handbook; LSDS-1574; U.S. Geological Survey, Department
of the Interior: Sioux Falls, SD, USA, 2019.
87. 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]
88. Hogarth, P.J. Mangrove Ecosystems. In Encyclopedia of Biodiversity, 2nd ed.; Elsevier: Amsterdam,
The Netherlands, 2013; pp. 10–22. ISBN 9780123847195.
89. Eastman, J.R.; Crema, S.C.; Sangermano, F.; Cunningham, S.; Xiao, X.; Zhou, Z.; Hu, P.; Johnson, C.;
Arakwiye, B.; Crone, N. A Baseline Mapping of Aquaculture and Coastal Habitats in Thailand, Cambodia and
Vietnam; Aquaculture and Coastal Habitats Report No. 1; Clark Labs: Worcester, MA, USA, 2015.
90. Htway, O.; Matsumoto, J. Climatological onset dates of summer monsoon over Myanmar. Int. J. Climatol.
2011, 31, 382–393. [CrossRef]
91. Jones, T.G.; Glass, L.; Gandhi, S.; Ravaoarinorotsihoarana, L.; Carro, A.; Benson, L.; Ratsimba, H.R.; Giri, C.;
Randriamanatena, D.; Cripps, G. Madagascar’s mangroves: Quantifying nation-wide and ecosystem specific
dynamics, and detailed contemporary mapping of distinct ecosystems. Remote Sens. 2016, 8, 106. [CrossRef]
Remote Sens. 2020, 12, 3758 34 of 35

92. Younes Cárdenas, N.; Joyce, K.E.; Maier, S.W. Monitoring mangrove forests: Are we taking full advantage of
technology? Int. J. Appl. Earth Obs. Geoinf. 2017, 63, 1–14. [CrossRef]
93. Zhang, X.; Tian, Q. A mangrove recognition index for remote sensing of mangrove forest from space.
Curr. Sci. 2013, 105, 1149–1154.
94. U.S. Geological Survey. Landsat 4-7 Collection 1 (C1) Surface Reflectance (LEDAPS) Product Guide; LSDS-1370;
U.S. Geological Survey, Department of the Interior: Sioux Falls, SD, USA, 2020.
95. Jordan, C.F. Derivation of Leaf-Area Index from Quality of Light on the Forest Floor. Ecology 1969, 50,
663–666. [CrossRef]
96. Tarpley, J.D.; Schneider, S.R.; Money, R.L. Global Vegetation Indices from the NOAA-7 Meteorological
Satellite. J. Appl. Meteorol. Climatol. 1984, 23, 491–494. [CrossRef]
97. Gao, B.-C. NDWI A Normalized Difference Water Index for Remote Sensing of Vegetation Liquid Water from
Space. Remote Sens. Envrion. 1996, 58, 257–266. [CrossRef]
98. 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]
99. Huete, A.R. A Soil-Adjusted Vegetation Index (SAVI). Remote Sens. Environ. 1988, 25, 295–309. [CrossRef]
100. Rondeaux, G.; Steven, M.; Baret, F. Optimization of soil-adjusted vegetation indices. Remote Sens. Environ.
1996, 55, 95–107. [CrossRef]
101. Huete, A.R.; Didan, K.; van Leeuwen, W. Modis Vegetation Index (MOD 13) Algorithm Theoretical Basis Document;
University of Arizona: Tucson, AZ, USA, 1999; Version 3.
102. Chandrasekar, K.; Sesha Sai, M.V.R.; Roy, P.S.; Dwevedi, R.S. Land Surface Water Index (LSWI) response
to rainfall and NDVI using the MODIS vegetation index product. Int. J. Remote Sens. 2010, 31, 3987–4005.
[CrossRef]
103. Van Deventer, A.P.; Ward, A.D.; Gowda, P.M.; Lyon, J.G. Using thematic mapper data to identify contrasting
soil plains and tillage practices. Photogramm. Eng. Remote Sens. 1997, 63, 87–93.
104. As-syakur, A.R.; Adnyana, I.W.S.; Arthana, I.W.; Nuarsa, I.W. Enhanced built-UP and bareness index (EBBI)
for mapping built-UP and bare land in an urban area. Remote Sens. 2012, 4, 2957–2970. [CrossRef]
105. Weber, S.J.; Keddell, L.; Kemal, M. Myanmar Ecological Forecasting: Utilizing NASA Earth Observations to
Monitor, Map, and Analyze Mangrove Forests in Myanmar for Enhanced Conservation. Available online:
https://ntrs.nasa.gov/citations/20140006912 (accessed on 1 October 2020).
106. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
107. Akoglu, H. User’s guide to correlation coefficients. Turk. J. Emerg. Med. 2018, 18, 91–93. [CrossRef] [PubMed]
108. Goodwin, L.D.; Leech, N.L. Understanding correlation: Factors that affect the size of r. J. Exp. Educ. 2006, 74,
249–266. [CrossRef]
109. Xie, Y.; Sha, Z.; Yu, M. Remote sensing imagery in vegetation mapping: A review. J. Plant Ecol. 2008, 1, 9–23.
[CrossRef]
110. Giri, C.P. Remote Sensing of Land Use and Land Cover: Principles and Applications; CRC Press: Boca Raton, FL,
USA, 2016; ISBN 9781420070750.
111. Li, W.; El-Askary, H.; Qurban, M.A.; Li, J.; ManiKandan, K.P.; Piechota, T. Using multi-indices approach to
quantify mangrove changes over the Western Arabian Gulf along Saudi Arabia coast. Ecol. Indic. 2019, 102,
734–745. [CrossRef]
112. Jones, T.G.; Ratsimba, H.R.; Ravaoarinorotsihoarana, L.; Glass, L.; Benson, L.; Teoh, M.; Carro, A.; Cripps, G.;
Giri, C.; Gandhi, S.; et al. The dynamics, ecological variability and estimated carbon stocks of mangroves in
Mahajamba Bay, Madagascar. J. Mar. Sci. Eng. 2015, 3, 793–820. [CrossRef]
113. Chen, N. Mapping mangrove in Dongzhaigang, China using Sentinel-2 imagery. J. Appl. Remote Sens. 2020,
14, 1. [CrossRef]
114. Stehman, S.V.; Foody, G.M. Accuracy Assessment. In The SAGE Handbook of Remote Sensing; Warner, T.A.,
Nellis, M.D., Foody, G.M., Eds.; SAGE Publications Ltd.: London, UK, 2009; pp. 297–314. ISBN 9780857021052.
115. Heumann, B.W. Satellite remote sensing of mangrove forests: Recent advances and future opportunities.
Prog. Phys. Geogr. 2011, 35, 87–108. [CrossRef]
116. Rosenfield, G.H. Analysis of thematic map classification error matrices. Photogramm. Eng. Remote Sens. 1986,
52, 681–686.
Remote Sens. 2020, 12, 3758 35 of 35

117. Lonneville, B.; Schepers, L.; Fernández Bejarano, S.; Vanhoorne, B.; Tyberghein, L. Marine Regions. Available
online: https://www.marineregions.org/ (accessed on 22 September 2020).
118. Jones, T.G.; Ratsimba, H.R.; Carro, A.; Ravaoarinorotsihoarana, L.; Glass, L.; Teoh, M.; Benson, L.; Cripps, G.;
Giri, C.; Zafindrasilivonona, B.; et al. The Mangroves of Ambanja and Ambaro Bays, Northwest Madagascar:
Historical Dynamics, Current Status and Deforestation Mitigation Strategy. In Estuaries of the World; Diop, S.,
Scheren, P., Ferdinand Machiwa, J., Eds.; Springer: Cham, Switzerland, 2016; pp. 67–85.
119. Jones, T.G.; Ratsimba, H.R.; Ravaoarinorotsihoarana, L.; Cripps, G.; Bey, A. Ecological Variability and Carbon
Stock Estimates of Mangrove Ecosystems in Northwestern Madagascar. Forests 2014, 5, 177–205. [CrossRef]

Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional
affiliations.

© 2020 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/).