International Journal of Applied Earth Observations and Geoinformation 99 (2021) 102320

Contents lists available at ScienceDirect

International Journal of Applied Earth

Observations and Geoinformation
journal homepage: www.elsevier.com/locate/jag

Optimal and robust vegetation mapping in complex environments using

multiple satellite imagery: Application to mangroves in Southeast Asia
Han Xiao a, b, c, Fenzhen Su a, b, c, d, *, Dongjie Fu a, b, c, *, Vincent Lyne e, Gaohuan Liu a,
Tingting Pan a, b, c, Jiakun Teng a, b
a
State Key Laboratory of Resources and Environmental Information System, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of
Sciences, Beijing 100101, China
b
College of Resources and Environment, University of Chinese Academy of Sciences, Beijing 100049, China
c
Collaborative Innovation Center of South China Sea Studies, Nanjing University, Nanjing 210093, China
d
Faculty of Geomatics, Lanzhou Jiaotong University, Lanzhou 730070, China
e
IMAS-Hobart, University of Tasmania, Tasmania 7004, Australia

A R T I C L E I N F O A B S T R A C T

Keywords: A band selection model was described for efficient and accurate remotely-sensed vegetation mapping in cloudy
Remote sensing band selection mixed-vegetation areas, demonstrated with an application on mapping mangroves in Southeast Asia (SE Asia).
Large-scale mapping We show how to use multi-source satellite imagery and Cloud Computing Platforms to improve mapping and
Mangrove mapping
computational efficiency in complex environments. A key element of the method relies upon field surveys to
Multi-source data
establish a detailed sample database that includes easily-confused land cover. The Maximal Separability and
Information (MSI) model was developed to select key bands for target land cover classification from multiple
satellite imagery based on two principles: 1. maximize separability of the target cover from other land cover; and
2. maximize and prioritize information from band combinations. Application of the MSI model to map man­
groves in SE Asia using three optical and SAR data systems (Landsat OLI, Sentinel-2 and Sentinel-1) showed: 1.
Sentinel-2 is better at classifying mangrove than Landsat and Sentinel-1; and 2. SWIR, NIR and Red bands (with
SWIR in particular) are effective in separating mangrove from other vegetation. The MSI-mapped mangroves
showed lower computation cost compared to using all bands from individual satellites, and higher accuracy
(above 90%) when applied to SE Asia. It was robust in tolerating smaller sample sizes, thereby demonstrating
computational feasibility and substantial improvements with the MSI model for large-scale land cover mapping
in complex environments.

1. Introduction information redundancy from multi-data systems, maximizing the

relevance of information used, and better targeting of field operations
Regional and global land cover mapping is being facilitated by for classification training.
increasing access to remotely-sensed big data (Ma et al. 2015). Realizing In practice, for non-hyperspectral data, most researchers still use of
these opportunities for large scale mapping also demands judicious use all spectral bands from the datasets, and few studies pay attention to
of high-performance computational resources from cloud platforms, select bands first in the application of data with a limited number of
such as Google Earth Engine (GEE) –which provides access to global bands or lack for references on band selection for specific areas and
high-resolution multi-temporal and multi-satellite data, high- target land cover. With the increase of data sources, the need for large-
performance processing, and mapping facilities (Dong et al. 2016, area mapping, and the need for real-time updates, band selection for
Midekisa et al. 2017). While timely processing operations in GEE are comprehensive applications of multi-source data is gradually becoming
expensive, data-intensive, and user-limited (Gorelick et al. 2017). necessary (Mandanici and Bitelli 2016, Astola et al. 2019).
Hence, efficient large-scale land-cover mapping requires minimizing Existing band selection methods include: the optimum index factor

* Corresponding authors at: State Key Laboratory of Resources and Environmental Information System, Institute of Geographic Sciences and Natural Resources
Research, Chinese Academy of Sciences, Beijing 100101, China.
E-mail addresses: [email protected] (H. Xiao), [email protected] (F. Su), [email protected] (D. Fu), [email protected] (V. Lyne), [email protected] (G. Liu),
[email protected] (T. Pan), [email protected] (J. Teng).

https://doi.org/10.1016/j.jag.2021.102320
Received 7 September 2020; Received in revised form 26 December 2020; Accepted 28 February 2021
Available online 12 March 2021
1569-8432/© 2021 Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences. Published by Elsevier B.V. This is an open
access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
H. Xiao et al. International Journal of Applied Earth Observations and Geoinformation 99 (2021) 102320

Fig. 1. Extent of the study area and field sample points distribution showing recorded parameters (see Legend).

model (Patel and Kaushal 2011), the maximum entropy model (Lin et al. specific areas. These two principles were motivated to meet the
2014) and the regression model (Verrelst et al. 2016, Long et al. 2019). increasing global need for practical, efficient and rapid high-precision
Previous studies on band selection used methods based, in part, on the land cover mapping from medium-to-high spatial resolution optical
principle of reducing information redundancy. Verrelst et al. (2016) and Synthetic-aperture radar (SAR) data (Benz et al. 2004, Wang et al.,
introduced a band analysis tool (BAT) based on Gaussian process 2019), and aided by cloud platforms such as GEE (Gorelick et al. 2017) if
regression (GPR), and Sotoca et al. (2007) selected bands by minimizing computational loads can be minimized.
the dependent information between spectral bands while maximizing We demonstrate the application of our band selection model in
the conditional entropies of the selected bands. These band selection mapping mangroves from one of the most diverse, complex and rapidly
methods consider the information amount and lack consideration for changing regions – Southeast Asia. Mangroves of Southeast Asia are rich
improving the classification accuracy. Also machine learning algo­ in species diversity and density (Giesen et al. 2007), and provide a va­
rithms, such as Random Forest, rate the importance of variables when riety of high-value ecosystem services for millions of people (Brander
applied to band selection (Demarchi et al. 2020) but are not effective in et al., 2012). However, mangroves are variously located next to con­
dealing with information redundancies and customized applications. founding vegetation, water environments, marshes, human land use,
Across these references, the principles used in band selection methods deforestation and regions of population growth and commodity demand
can be summarized as: 1. Maximizing the information content of the (Richards and Friess 2016). At the same time mangroves are sensitive to
band combination; 2. Minimizing information redundancy; and, 3. changing climate, storms, and sea-level rise (Woodroffe 2018). These
Optimizing the classification accuracy (through sample training). factors have intensified the urgency for identifying, monitoring, and
However, we suggest that existing band selection methods do not mapping to better manage the remaining narrow dynamic coastal
address the full practical needs as two main principles are lacking: 1. mangrove resources scattered across extensive regions.
With fragmented and complicated spatial land cover, calibration data is Giri et al. (2011) drew the first global mangrove map using Landsat
required that specifically focuses the separability between the target images. Since then, for mangrove monitoring and management pur­
land cover and the surrounding easily-confused land cover; 2. Effective poses, efficient accurate mapping and quickly updating are key re­
band selection should comprehensively consider minimizing computa­ quirements (Oliphant et al. 2019). In addition, mangroves are easily
tional loads and optimizing classifications based on the characteristic of confused with other vegetation and confounded by tides, the rich sur­
specific land cover and/or areas. rounding vegetation diversity, and wet environments in Southeast Asia
We therefore developed a comprehensive and effective band selec­ (Giri et al. 2015, Xia et al. 2018). Consequently, mangroves mapping in
tion method by implementing two modified principles: 1. maximize Southeast Asia is difficult and prone to high error (Grainger 2008), and
separability of target land cover from surrounding easily-confused other requires combined multi-data to resolve (M. Brander et al., 2012, Clerici
land cover; and 2. prioritize and select as few bands as possible that et al. 2017).
maximize the information content of the combined bands, especially for In previous studies, researchers tried to classify mangroves from
areas lacking data. Implementation of the first principle requires a sci­ Landsat (Son et al. 2015), Sentinel (Argamosa et al. 2018), WorldView
entific sample database that is appropriate for target land cover, the (Tian et al. 2017), ALOS-1 (Hamdan et al. 2014), and other satellite
surrounding easily-confused cover, and specific areas – as demonstrated images (Xia et al. 2018). Freely available medium-to-high resolution
by supervised band selection experience (Yang et al. 2011, Cao et al. satellite data, such as Landsat and Sentinel, are the main data sources for
2016). Resolution of the second principle requires redundancy assess­ large-scale mangrove mapping (Wang et al., 2018) as they provide
ments of bands from multi-source data for the target land cover, and for suitable spatial resolution for large-scale mapping (Giri et al. 2011,

2
H. Xiao et al. International Journal of Applied Earth Observations and Geoinformation 99 (2021) 102320

Fig. 2. High quality image coverage in Southeast Asia from 2018 to 12–01 to 2019–03–01. High quality refers to the criteria of cloud coverage less than 20% of LS
\S2. a)-c) are the image boundaries of LS\S2\S1 respectively; d) is the overlay of the three; ①-③ are the composite images of LS\S2\S1 respectively in Lalang river of
Thailand, with white areas denoting lack of data.

Hansen et al. 2013). Several studies compared three satellite data sys­ We began the 2019 field program by mapping mangroves in small
tems used for mangrove mapping — Landsat OLI-8 (LS), Sentinel-2 (S2) areas (including the Andaman Sea in Myanmar, and Thailand), and then
and Sentinel-1 (S1) — and showed that radar data combined with op­ throughout Southeast Asia (see Appendix Fig. A1 for field photos from
tical data improves land cover mappings (Clerici et al. 2017, Zhao and the study area and see Fig. 1. for the study area). Details of the field
Qin 2020). Also, Forkuor et al. (2017) showed that compared to LS, S2′ s program are presented following the description of the remotely sensed
red-edge bands improve discrimination of Land Use and Land Cover data used in the study.
(LULC) classes, and that complementary use of S2 and LS is beneficial for
LULC mapping.
2.1. Low quality of data due to cloud
Optical images from LS and S2, commonly used over the cloudy re­
gion of Southeast Asia, suffer from high correlation between bands
Cloud and rainy conditions in Southeast Asia, result in poor quality
resulting in information redundancy and insufficient emphasis on
optical images, especially for Landsat—as seen in the image-quality
informative bands for land cover mapping (Mandanici and Bitelli 2016,
coverage of (a) and (b), and mapped areas in ①-② of Fig. 2. Sentinel-
Korhonen et al. 2017). Computational and algorithmic benefits are
1 images were not affected by cloud, which could potentially make up
therefore potentially possible by avoiding or minimizing these two
for this lack of coverage (see (c) and ③ in Fig. 2). Images of LS\S2\S1
problems. A key objective therefore of this study was to identify the
that meet the criteria of cloud coverage less than 20% of LS\S2 in
informative bands from multiple satellites such as LS, S2, and S1 using
Southeast Asia from 01.12.2018 to 01.03.2019 (Fig. 2 a-d) show that S1
the two proposed principles previously discussed. Our application of
and S2 are rich in good quality images, while there are obvious missing
band selection from three satellites (LS/S2/S1) to mangrove mapping of
areas in LS, especially in Indonesia and Malaysia.
Southeast Asia begins with a description of the study area, the mapping
issues, and the ground sampling across several countries and a variety of
environments where mangroves are found. We then detail the band 2.2. Mapping confusion in confounding environments
selection algorithm and the results of its application.
Although the implementation of our study is to mangroves, the field The complexity of vegetation types in Southeast Asia necessitated a
and analytical methodology we developed are generic and can be field survey planned according to a-priori knowledge and mappings of
applied to select informative bands from multi-source images to achieve mangroves and confounding vegetation. To facilitate band separation
efficient high-accuracy mapping within specific regions of easily- analyses, we sampled both the easily-confused land cover and the target
confused vegetation such as mangroves. land cover. This sample database allowed us to analyze band selection
for spectral characteristics for separating target land cover in complex
2. Study area, data and field sampling local environments. A detailed sample database for classification was
built from field investigations in mangrove areas conducted in Thailand
The diversity of mangroves of Southeast Asia (SEA), accounts for from 2018 to 09–07 to 2018–09–16, Malaysia from 2019 to 11–29 to
35% of global mangrove diversity (Giesen et al. 2007). However, the 2019–12–07, Indonesia from 2019 to 09–10 to 2019–09–20 and
region is undergoing rapid changes from land conversion and other Myanmar from 2019 to 09–20 to 2019–09–27.
human activities (Richards and Friess 2016), and Southeast Asia is Field investigation routes were planned before the trip using Google
therefore a global focus for mangrove management and monitoring (see “My Maps” (https://www.google.com/maps/about/mymaps/) and
Fig. 1 for the location and extent of SEA). driven according to the route. Accompanying personnel recorded the
sample point category and coordinates of mangrove surrounding

3
H. Xiao et al. International Journal of Applied Earth Observations and Geoinformation 99 (2021) 102320

Fig. 3. Flow chart showing the processing sequence for the 2019 field and satellite information (left hand column), the list of processed variables used, database
components, analyses, and final mapped product (right hand column).

features with GPS and OSM tracker software (www.openstreetmap.org).

The main work during the field investigation was recording sample Table 1
points of mangrove and surrounding features. After driving to a sample Band information for the three satellite sensors.
location, we collected detailed sample points and boundary coordinates Sensor Description Band Wavelength Resolution
of the mangrove and focused on manually mapping the easily-confused Name range
features around the mangrove.
Landsat Band2: Blue blue_LS 0.45–0.51 µm 30 m
Based on the sample points collected during the field survey and the OLI 8 Band3: Green green_LS 0.53–0.59 µm 30 m
high-resolution images in Google Earth software, the most confusing Band4: Red nir_LS 0.64–0.67 µm 30 m
features around mangroves in the field investigation are summarized as Band5: Near red_LS 0.85–0.88 µm 30 m
infrared
follows (see the field photos in Appendix Fig. A1):
Band6: Shortwave swir1_LS 1.57–1.65 µm 30 m
infrared 1
1. Vegetation confusion: mangroves extending inland are confused with Band7: Shortwave swir2_LS 2.11–2.29 µm 30 m
other vegetation, and the inner boundary of mangroves are not easily infrared 2
separable from palm vegetation, rice, and natural forest; Sentinel-2 Band2: Blue blue_S2 458–523 nm 10 m
Band3: Green green_S2 543–578 nm 10 m
2. Salt marsh confusion: Salt marshes are greatly affected by tides
Band4: Red red_S2 650 – 680 nm 10 m
leading to significant confusion with mangroves, especially in salt Band5: Red Edge 1 re1_S2 698–713 nm 20 m
marshes with sparse mangrove vegetation. Band6: Red Edge 2 re2_S2 733–748 nm 20 m
Band7: Red Edge 3 re3_S2 773–793 nm 20 m
Band8: NIR nir_S2 785–900 nm 10 m
3. Material and methods
Band8A: Red Edge nir2_S2 855–875 nm 20 m
4
Specific steps for mapping are summarized as follows, and shown Band9: Water WV_S2 935–955 nm 60 m
schematically in the flow chart of Fig. 3: vapor
Band11: SWIR 1 swir1_S2 1565–1655 nm 20 m
Step 1. The data was managed and processed by establishing a
Band12: SWIR 2 swir2_S2 2100–2280 nm 20 m
sample points database used to develop a spatial mask of mangrove Sensor Description Band Centre Resolution
presence (from extreme value analysis of temperature, elevation, and Name frequency
distance-to-water) that restricted the area for mapping, and band se­ Sentinel-1 VH-ASCENDING VH_asc 5.405 GHz X: 10 m;
lection analysis; Y:10 m
VH-DESCENDING VH_desc 5.405 GHz X: 10 m;
Step 2. Images from Landsat OLI, Sentinel-2, and Sentinel-1, were
Y:10 m
filtered, processed, standardized, and band values were extracted for the VV-ASCENDING VV_asc 5.405 GHz X: 10 m;
sample points; Y:10 m
Step 3. The band selection model (a key technology of this study) was VV-DESCENDING VV_desc 5.405 GHz X: 10 m;
developed to analyze band reflectance of Landsat OLI, Sentinel-2, and Y:10 m

Sentinel-1 for each land cover from the sample database;

Step 4. The selected key/priority bands were used to map the database for the easily-confused vegetation and water categories as
regional-scale mangrove distribution of Southeast Asia. follows:

3.1. Mapping mask and sample points database 1. The classes of the sample database were determined according to
both the field survey and high-resolution images from Google Earth.
Published mangrove distribution show that its growth is affected by The sample database included mangrove, four “Other vegetation”
temperature, elevation, and the water environment (Cavanaugh et al., classes (Grassland, Farmland, Natural forests, Artificial forests) and
2014). Thresholds of these factors were obtained through extreme value five detailed water classes (Flowing water, Pond Seawater, River,
statistical analyses of published mangrove products. Mapping masks of Salt marsh with vegetation, Salt marsh without vegetation). The
these thresholds were applied to the satellite images and intersected study area and sample point distribution are shown in Fig. 1.
with the river buffer zone in the coastal zone to determine the potential 2. In addition to 259 field sample points, we randomly generated 2000
(meta) environment of mangrove—see Appendix Table A1 for the spe­ points within the mapping mask, with a minimum distance of 1 km,
cific threshold parameters used for determining the mapping mask. and adjusted the position and set the attributes according to the high-
Based on the characteristics of the above-mentioned field in­ resolution image. 204 points were added to the mangrove class, and
vestigations and observations, we established a detailed sample the total resulting sample database contained 2463 points. Appendix

4
H. Xiao et al. International Journal of Applied Earth Observations and Geoinformation 99 (2021) 102320

Table 2 After checking whether the band value at the sample points conforms
Detailed description of bands of Sentinel-1. to the normal distribution, three indicators were computed for the
Band Transmitter Receiver Polarization Orbit signal model, these being: the difference of band value between mangrove and
Name Properties other land cover (DIFF); the standard deviation of bands (STDEV); and
VH_asc Dual-band cross-polarization, vertical ASCENDING the correlation coefficient between bands (CC). These components and
transmit/ horizontal receive direction the formula used are described as follows:
VH_desc Dual-band cross-polarization, vertical DESCENDING 1. The difference between band values for mangrove and other land
transmit/ horizontal receive direction cover (DIFF): The fundamental purpose of band selection is to improve
VV_asc Single co-polarization, vertical transmit/ ASCENDING
vertical receive direction
the separability of mangroves from confounding ground features in the
VV_desc Single co-polarization, vertical transmit/ DESCENDING sample database. We used the band-value difference between mangrove
vertical receive direction and classes of Other Vegetation/Water/Bare Land according to the for­
mula (with larger differences indicating greater separability):
∑N
Table A2 shows the detailed classes and number of each class in the diff n = abs(xni − xnmangrove ) (1)
sample database.
i=1

3. A 60 m radius buffer was created at each point to contain 4 × 4 pixels Where,

of Landsat, 2 × 2, 6 × 6, 12 × 12 pixels of Sentinel-2, and 12 × 12 i is the land cover class number from the sample database;
pixels of Sentinel-1—see Table 1 for band resolution. The purpose of n is the band serial number;
this processing was to ensure good representation from the satellite N is the total number of land cover classes;
images around each sample point. xni is the average band value of Other land cover class;
xnmangrove is the average band value of mangrove.
3.2. Image preparation 2. Standard deviation of bands (STDEV): We used standard deviation
as a measure of the “information potential” of a band, with higher
Data from three online datasets, Landsat OLI (LS), Sentinel-2 (S2), standard deviations implying greater accommodation, through greater
and Sentinel-1 (S1) were obtained using Google Earth Engine (GEE); see variations, of the band in response to changes in underlying environ­
Appendix Table A3, Table 1, and Table 2 for information on the satellites ments and vegetation. So, the standard deviation of each band value of
and bands used. Image standardization and image sampling methods total sample points is used as an index of its information potential:
were conducted as follows: √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
((xn1 − x)2 + (xn2 − x)2 + ......(xnr − x)2 ))
stdevn = , (2)
1. Image selection: We chose images from 2018 to 12–01 to r− 1
2019–03–01 as this is the dry season in Southeast Asia, and we Where,
anticipated less problems with cloud and rain both for the field r is total sample point number;
survey and for image clarity from the satellites. We used surface n is the band serial number;
reflectance products from Landsat and Sentinel-2, and the back­ x is the sample average for that band.
scatter coefficient of Ground Range Detected (GRD) scenes from 3. Correlation coefficient between bands (CC): Correlations exist
Sentinel-1. The surface reflectance products were further screened to between the medium-resolution satellite bands; for example, Landsat
select images with cloud coverage less than 20%. Table A3 lists the and Sentinel-2 have SWIR, Red, Blue, and other bands, so when selecting
collection time and condition, and Table 1. lists the band names of bands, the size of the correlation between the bands should be consid­
the two optical images. Data from ascending and descending, and ered. Smaller correlations between a band and all other bands imply that
two polarizations were combined in the 4 bands from S1—see this band contains information that is distinct from other bands, and
Table 1. for the band names, and Table 2. for the detailed description hence potentially has additional information relevant for band separa­
of polarizations and orbit directions. tion or cover characterization. The pairwise correlation for the band is
2. Image standardization: The median() function of GEE was used to defined as:
obtain the median pixel value images from the acquisitions collected
∑N
from LS, S2, and S1. We then combined the three median value im­ abs(CCmn )
CCn = m (3)
ages using the merge() function of GEE. This resulted in one Image­ N− 1
Collection including median value images of LS, S2 and S1, for 21 Where,
bands (6 from LS, 11 from S2, and 4 from S1) in total—see Table 1 for m is band serial number except n;
band names; N is total number of bands;
3. Get band value: We used the sampleRegions() function of GEE to CCmn is the Correlation Coefficient value between bandm and bandn.
extract the mean value from bands from each sample’s buffer zone 4. The band selection index MSI, standing for Maximal Separability
(of 60 m radius centered on the sample point). Sampled values were and Information index, was created by combining the above three fac­
entered into a csv table with bands as the columns. We first removed tors; with factors in the numerator increasing with greater separability
5% of extreme values, from 2.5% of extremes from either end of the or information content potential, and vice versa for the denominator
distribution, and each value was divided by the absolute maximum factor:
band value, so that the value of each column ranged from − 1 to 1.
MSI n = diff n × (stdevn Ã⋅CCn ) (4)
3.3. Band selection Other formulations are possible but MSI serves as a simple and easily
calculable formula for band selection which captures both separability
The purpose of band selection was to ensure mangrove separability and information content.
and to reduce information redundancy. We developed a combined The above four steps were carried out in RStudio (https://rstudio.
formulation (described below) for the maximum separability of com/).
mangrove and maximum information content of bands, based on In later sections we describe tests on the effectiveness of the number
commonly used indicators: reflectance difference; band standard devi­ and order of MSI bands by examining the variation of overall mapping
ation; correlation between bands. The same principle can be applied to accuracy as the number of the input bands was increased one-by-
other target vegetation as the methodology is generic.

5
H. Xiao et al. International Journal of Applied Earth Observations and Geoinformation 99 (2021) 102320

Fig. 4. Variation by band of the reflectance from land cover classes characterized by a box plot for each of the 21 bands and for each land cover class (shown in the
legend as “Class”). For each box plot, the lower bar is the minimum value, the lower quartile is the lower bound of the box, the median is the bar in the box, the upper
quartile (third quarter) is the upper bound of the box, and the top bar is the statistical maximum value, and dots represent outliers.

one—from high to low MSI rank order. Furthermore, as described later, described later in the Results section) based on the MSI model rankings
we checked the fault tolerance of the selected bands by reducing the considering the requirement for reasonable operation speed and cloud
percentage of sample points in the supervised classification. computing capacity. Classifications were carried out on the GEE plat­
form with sample points for the supervised classification chosen
3.4. Classification processing randomly from the study area. We first randomly split the sample points
into two parts, 80% for training and 20% for validation. Then the su­
For classification purposes, Southeast Asia was divided into latitude pervised classification based on Random Forest (Pal 2007) with 50 times
× longitude grids of 10◦ × 10◦ , representing the basic classification unit. iteration was carried out using the training sample points. Four classes,
The sample points were subsequently increased to 11,315 within the Mangrove\Other vegetation\Water\Bare Land, were generated as the
mapping mask to enhance the sample needs for classification and vali­ classification results through 50 iterations to reduce sample
dation. We selected 10 bands for classification (optimal band selection is dependency.
The accuracy parameters for classifying were obtained by using the
validation samples in the Confusion Matrix (Stehman 1997) to generate
the Overall Accuracy (OA), Producer’s Accuracy (PA), and Customer’s
Accuracy (CA). Then an overall F1 score (Deus 2018) (generated by the
formula shown below) was applied to combine PA and CA in a unique
score that combined mapping accuracies among each land cover class.
The accuracy parameters and classification results were aggregated as
follows:
F1 Score = 2 × (CA × PA)/(CA + PA) (5)

4. Results

4.1. Key bands for mangrove classification

For each band, differences between mangrove and other land cover
were analyzed with the MSI model. The MSI results for all bands were
ranked to determine the key bands for mangrove classification, as
detailed below.
Box plots in Fig. 4 and the DIFF values shown in Table A4 comparing
Fig. 5. Band MSI ranking diagram. The top-down ordered sequence of MSI the mean Class I band reflectance value of land and vegetation cover
rankings (rank value displayed in each horizontal bar and corresponding to the classes show:
horizonal length of each bar) shows the importance of bands from high to low
in terms of its overall value for mangrove classification taking into consider­ i. The general order of band difference between Mangrove and
ation separability and information content of the bands. Other Vegetation/Water/Bare Land classes is: Other Vegetation

6
H. Xiao et al. International Journal of Applied Earth Observations and Geoinformation 99 (2021) 102320

Fig. 6. Mangrove mapping contrasts in different key areas— mapped as polygons along the land side of water edges. The zones at the top of the image are the
example areas. The LS/S2/MSI on the left are respectively results classified using only LS bands, only S2 bands, and MSI-selected bands. The (X,Y) on the bottom are
the center lat/lon coordinates of the areas. The indices on the right are representation of the average calculation load and accuracy of the three areas (average by
row). CV is the computation amount used in GEE in units of Megabytes (Mb); OA is Overall Accuracy (Accuracy terms are defined in the Google Earth Engine).

< Water < Bare Land; that is, the band difference between this study is specifically for mangrove mapping for Southeast Asia. The
Mangrove and Other Vegetation is the smallest implying that higher the MSI index, the more informative the band is for mangrove
Mangrove and Other Vegetation are most easily confused; classification (as discussed previously). It is well known that Red and
ii. In differentiating between Other vegetation and Mangrove, NIR bands are important in vegetation classification and the MSI rank­
Sentinel-2 (S2) discriminates best in the Shortwave Infrared ings showed that for Southeast Asia, the key bands for distinguishing
(SWIR) band, followed by the Near Infrared (NIR) band. The mangrove are Red and NIR. In addition, the following results can be
order of band difference between Mangrove and Other Vegeta­ drawn from the MSI rankings in Fig. 5: bands from S2 are generally
tion is: swir1_S2 > nir2_LS > nir_S2 > swir1_LS > re3_S2 > re2_S2 ranked higher than LS—an indication of technological improvement in
> nir_LS; the remote sensing capabilities of more modern satellite systems; the
iii. The four SAR bands show greatest band differences between red, nir2, nir, swir2 and swir1 bands of S2 are ranked highest by MSI,
Mangrove and Water, with ranking: VV_asc > VV_desc > VH_asc followed by the red and swir2 bands of Landsat (LS), and then the VV_asc
> VH_desc, which are much higher than the optical bands—­ band of Sentinel 1 (S1); not all the red-edge bands of S2 are ranked high
where the NIR band and Red Edge band of S2 show differences (such as the red-edge 1 band (named re1_S2) and the red-edge 2 band
between Mangrove and Water: nir2_S2 > nir_S2 > re3_S2 > (named re2_S2)); not all the nir bands ranked high, such as the NIR band
re2_S2; of LS.
iv. The difference between Bare Land and Mangrove is greatest in the
optical SWIR band, with ranking: swir2_S2 > swir1_S2 >
4.2. Comparison of mangrove classification results
swir1_LS > swir2_LS > red_S2 > re1_S2.
Turning now to mapping of mangroves in Southeast Asia, we used
The MSI ranking of bands is shown in Fig. 5 and listed in Appendix
the top ten bands of the MSI index, and the 10◦ × 10◦ grids as the
Table A4. It should be noted that the MSI ranking of bands determined in
classification unit. Classifications were carried out using the Random

7
H. Xiao et al. International Journal of Applied Earth Observations and Geoinformation 99 (2021) 102320

Fig. 7. Spatial comparison of man­

groves in different areas between maps
from the Global Mangrove Watch
(GMW-2016, with the base map of LS
median image from 20151201 to
20160301) and the MSI classifications
(with the base map of LS median
image from 20181201 to 20190301):
a) - c) are in the Lalang river of
Thailand corresponding to the yellow
bounded regions in Fig. 2 ③; d) - e) are
in Ayeyarwady area of Irrawaddy in
Myanmar; e) - f) correspond to the
yellow bounded regions in d). (For
interpretation of the references to
colour in this figure legend, the reader
is referred to the web version of this
article.) (For interpretation of the ref­
erences to colour in this figure legend,
the reader is referred to the web
version of this article.)

8
H. Xiao et al. International Journal of Applied Earth Observations and Geoinformation 99 (2021) 102320

Fig. 8. Graphs of Overall Accuracy (OA) change. a) The change in OA by including an increasing number of bands ranked by MSI from high to low; b) The change in
OA as the number of sample points are reduced (the first 10 bands ranked by MSI represent the MSI results, while all bands are used for LS/S2/S1). The inflection
point in the number of bands in a) is shown by the arrow.

Forest classifier as described in Section 3.4 to derive the 2019 mangrove the computational load for the MSI bands is significantly reduced
maps for Southeast Asia. In this section, we will describe the mapping (compared to the sum of LS and S2) for regional-scale mapping.
results and the accuracy from key areas and the whole area.
4.2.2. Visual comparison of MSI results versus GMW product
4.2.1. MSI bands vs. LS and S2 In order to further verify that the MSI-selected bands are superior, we
First, as shown in Fig. 6, example results for a grid of the Andaman compared our classification results with the 2016 classification from the
Sea in Myanmar and Thailand show that natural forest, rice, oil palm Global Mangrove Watch (Bunting et al., 2018) (mapping results for
and other vegetation in this area are mixed—and hence caused confu­ different areas shown in Fig. 7, the yellow boundary is the MSI result,
sion for mangrove mapping. Mangroves growing in the Intertidal zone and the green boundary is the GMW result). The GWM classification was
and River zone are mixed with plantations along the Andaman Coast, based on Landsat images, while MSI-selected bands are composed of LS/
and mixed with rice in the Irrawaddy Delta (Fig. 6). S2/S1. Compared with GMW, the classification results appear similar in
In comparing mangrove maps classified by the same method and large coastal areas but as mangrove patches extending to the river
same sample points only by LS, S2, and bands selected by MSI, we can decrease, and as the surrounding environment (especially vegetation)
summarize that: The MSI results (Fig. 6) show better mangrove sepa­ becomes more complex, the classification results of GMW are inferior to
rability in the intertidal, river, and mixed-vegetation zones than results MSI: this difference can be seen in the comparison between the small
from only using LS and S2, especially in identifying small patch man­ patch mangroves in c-1) and c-2) of Fig. 7 in river channels, and the
groves. Mangrove classifications are missing if only S2 bands are used in mixed areas of the rice and small patch mangroves in e-1) and e-2) in
the intertidal zone (S2-Intertidal Zone in Fig. 6). There is classification Irrawaddy. In area f), GMW did not extract any mangroves in contrast to
confusion between river water and mangrove in the river zone if only many small patches identified by MSI.
using LS or S2 (LS\S2-River Zone in Fig. 6). And there are missing
mangrove classifications in small patches in the mixed-vegetation area 4.2.3. Band number determination and fault tolerance test
(LS\S2-Mixed-vegetation area in Fig. 6). Advantages displayed by the To further test the effectiveness of the MSI model in determining the
MSI-selected bands include: complete patch mapping, clear boundaries order of importance of the bands, we examined the variation of overall
between mangrove, water, and other vegetation, especially in the clas­ accuracy as the number of the input bands were increased one-by-
sification of small patches (see for example the comparison of Mixed- one—from high to low MSI rank order. The change of overall accuracy
vegetation areas in Fig. 6). Furthermore, accuracies from the MSI- with increasing MSI bands (Fig. 8-a) shows an inflection point of OA of
selected bands are the highest. As for computational load, the running six bands (five S2 and one LS) as a minimum (and optimal) number of
speed of the MSI-selected bands is less than the sum of the three kinds of bands necessary for accurate mangrove classification—representing
images running separately. Overall, in addition to improved accuracy, significant savings in computational load.

Fig. 9. Mangrove mapping results and accuracy table of Southeast Asia using MSI-selected bands. N315-N424 are the IDs of the 10◦ × 10◦ classification grid shown
in the map. Columns are: OA %: Mean and standard deviation value of Overall accuracy from the 50-times iteration; F1 score % (Man/ OV/ Wat/ BL) are the
respective F1 scores for Mangrove, Other Vegetation, Water, and Bare Land.

9
H. Xiao et al. International Journal of Applied Earth Observations and Geoinformation 99 (2021) 102320

Furthermore, we randomly diluted the sample points for training by Table A1

percentage (e.g. 80% in Fig. 8-b means retaining 80% of the sample Data used for the mangrove mapping mask.
points to participate in the supervised classification), and compared the Class Data name
OA-accuracy changes with the same percent of sample points for veri­
a) Climate data TerraClimate: Monthly Climate (Abatzoglou et al., 2018) 1
fication (always 20%). The experimental results in Fig. 8-b) shows that b) Mangrove Global mangrove forests distribution, 2000 (Giri et al.,
the OA for MSI bands are higher than that of LS, S2 and S1, indicating Products 2011) 2
higher fault tolerance, or robustness, of the MSI-selected bands, partic­ c) Other Products Nature Resources (River_lake_10m; Coastline_10m) 3
ularly in the low sample range from 50% to 20%. SRTM Digital Elevation Data 30 m 4
Global Intertidal Change Classification (Murray et al., 2019)
5

4.3. Application to Southeast Asia 1 and 2

Temperature zone for mangrove growth - the maximum and minimum
temperatures of 20170101–20190101 extracted by TerraClimate: monthly
After practicing in a small area, the selected bands from MSI were climate were sampled to get temperature thresholds and to delimit the tem­
used for regional mangrove mapping in Southeast Asia. We trained the perature range for mangrove extraction.
classification on a grid of 10◦ × 10◦ using the Random Forest algorithm 3and 5
Within 50 km of the coastal zone, the horizontal range of the river is 30 km,
(see Section 3.4). For 2019, the average of the mean value of overall and the distance between the river and the estuary is 100 km.
2 and 4
accuracy (OA) over the total 16 grids of Southeast Asia was 92.84%, the Mangrove growth elevation zone: SRTM 30 m DEM is sampled to obtain
average of F1 score of Mangrove, Other Vegetation, Water, and Bare the elevation range of mangrove growth in the world.
Land were: 93.34%, 92.41%, 95.36%, and 91.80% respectively. Accu­
racy results for individual grids are shown in the table of Fig. 9. Man­ number of variable values and variable similarity affect the RF variable-
groves of Southeast Asia mapped via MSI-selected bands show an overall importance measures (Strobl et al., 2007). The weight will be biased
accuracy higher than 90%—which represents an accurate regional towards similar bands, such as the Red Edge1,2,3 bands used in our
mapping. research; 2. RF variable weight processing is a black-box process and
cannot be customized for specific separation of features, and specific
5. Discussion areas. In contrast, in the MSI model we added the difference as an in­
dicator to improve mangrove separability, especially for separating the
5.1. First selecting then mapping confusing vegetation when mapping mangroves in Southeast Asia. In
comparing the RF and MSI ranking for the whole of Southeast Asia, we
Compared to the MSI method, machine learning methods such as the find that for the top ten bands, six bands are common (red_S2, nir2_S2,
typically-used Random Forest algorithm (RF) provide variable- nir_S2, swir2_S2, swir1_LS, re3_S2), but RF ranked four similar bands
importance weight, based on the principle of best training accuracy into its top ten; such as Red Edge1,2,3 and nir_S2 (see the RF rank and
through large-scale sample training (Abdel-Rahman et al., 2013). Users weights in Table A5). Computationally, if we directly use RF for classi­
can directly use RF for classification (weights are embedded into the fication without band selection, the calculation at the 10◦ × 10◦ scale on
classification) or use the weights as the importance rank for band se­ the Google Earth Engine exceeded the user memory. In contrast, the MSI
lection. But, RF has disadvantages in practical applications: 1. The calculation for the same 10◦ × 10◦ region passed using 10 selected

Fig. A1. Photos of confounding land cover collected during the field investigations: a) Ngwesaung, Myanmar; b) Ayeyarwady, Myanmar; c) Samarinda, Indonesia; d)
Samarinda, Indonesia.

10
H. Xiao et al. International Journal of Applied Earth Observations and Geoinformation 99 (2021) 102320

Table A2 priority bands.

Land cover classes and identification categories used for the detailed sample As mentioned in the introduction, there are two major problems
point database. when mapping mangroves in Southeast Asia: vegetation confusion and
Class I Class II Class III ID: lack of quality optical imagery. Hence, we formulated the two principles
Amount used in the MSI method to generate a customized combination of pri­
1 Mangrove – – 1: 877 ority bands for mangroves in Southeast Asia: one for maximizing sepa­
2 Other 21 Forest 211 Natural forests 211: 254 rability between mangrove and other easily-confused features, and
vegetation 22 Grassland 212 Artificial forests 212: 187 another for maximizing the information content of selected bands. In
23 Farmland 22: 131
brief, to maximize separability we used supervised band selection (Cao
23: 124
3 Water 31 Flowing 311 Seawater 311: 111 et al. 2016) to analyze differences of the bands between classes (Zhang
water 312 River 312: 103 et al. 2017). To overcome high correlations between bands (Claverie
32 Salt marsh 321 Salt marshes with 321: 112 et al. 2018) we chose the correlation and standard deviation as two
33 Pond vegetation 322: 108 indices in the MSI model.
322 Salt marsh without 33: 134
Overall, we find that the MSI model successfully improves mangrove
vegetation
4 Bare Land – – 4: 322 separability, efficiently fuses multiple data, and improves tolerance of
Total 1: 877 points; 2: 696 points; 3: 568 points; 4: 322 All: 2463 low sample numbers, whilst maintaining acceptable classification ac­
points curacy. It is suited for regional-scale target land cover mapping, espe­
Description of detail categories: cially in complex regions such as Southeast Asia. However, what needs
212 Artificial Forest includes oil palm, coconut, rubber, fruit tree, etc.; to be made clear is that implementation of MSI must be guided by the
311 salt marshes with vegetation means salt marshes with glassworts or cord­ following requirement and restriction: 1. a detailed and sufficient sam­
grass; ple database must be established that includes target features and easily
33 Pond includes salt pan, fish pond, and shrimp pond. confused features; 2. MSI works best for target land cover mapping
within specific areas where the sampling requirements are met, but it is
not suitable for full land cover classification where the aim is for overall
Table A3 accuracy averaged across large disparate regions.
Satellite data information.
Class Sensor ImageCollection ID of Timing and screening
GEE conditions 5.2. Mapping experience, band selection results, and future directions
Optical Landsat OLI LANDSAT/LC08/C01/ 2018–12–01 to 2019–03–01
8 T1_SR Cloud < 20% For the regional supervised classification, we also faced the difficulty
Sentinel-2 COPERNICUS/S2_SR of field sampling of mangroves due to the confounding influence of the
SAR Sentinel-1 COPERNICUS/S1_GRD
surrounding environment, especially in the mixed vegetation areas, such
as Southeast Asia. As examples: in the field investigations we observed

Table A4
Band analysis and MSI results – see main text for description of column variables.
Band DIFF STDEV CC MSI

MAN_OV MAN_WAT MAN_BL

blue_LS 0.0679 0.1028 0.1379 0.1799 0.2867 0.1937

blue_S2 0.0015 0.1609 0.1146 0.1071 0.3819 0.0777
green_LS 0.0718 0.0644 0.1548 0.1776 0.2977 0.1736
green_S2 0.0237 0.1107 0.1512 0.1090 0.2791 0.1116
nir_LS 0.1201 0.3512 0.0509 0.2277 0.6032 0.1971
nir_S2 0.1504 0.4076 0.0253 0.2151 0.5738 0.2186
nir2_S2 0.1533 0.4272 0.0417 0.2223 0.5750 0.2406
re1_S2 0.0806 0.0372 0.3378 0.1438 0.4454 0.1471
re2_S2 0.1284 0.3655 0.0279 0.1942 0.5722 0.1771
re3_S2 0.1442 0.4045 0.0195 0.2122 0.5728 0.2105
red_LS 0.0735 0.0604 0.1944 0.1848 0.2858 0.2123
red_S2 0.0119 0.1744 0.3638 0.1582 0.3295 0.2641
swir1_LS 0.1452 0.0913 0.4042 0.1675 0.4954 0.2167
swir1_S2 0.1899 0.1038 0.4265 0.1473 0.4887 0.2172
swir2_LS 0.0947 0.0172 0.3724 0.1605 0.3725 0.2087
swir2_S2 0.1114 0.0075 0.4866 0.1261 0.3508 0.2176
VH_asc 0.0057 0.4483 0.1132 0.1879 0.5296 0.2013
VH_desc 0.0077 0.4452 0.0891 0.1898 0.5269 0.1953
VV_asc 0.0069 0.4611 0.0917 0.1996 0.5261 0.2123
VV_desc 0.0193 0.4590 0.0762 0.2007 0.5316 0.2094
waterVapor_S2 0.1026 0.1603 0.0223 0.1642 0.4366 0.1073

Table A5
The top ten bands’ weight value ranked by the Random forest algorithm.
Band swir2_S2 swir1_LS nir_S2 swir2_LS nir_LS nir2_S2 re3_S2 re2_S2 re1_S2 red_S2

Weight 27.05 19.03 15.29 15.18 14.64 14.48 13.51 12.90 12.11 11.44

11
H. Xiao et al. International Journal of Applied Earth Observations and Geoinformation 99 (2021) 102320

large area of rice planted around mangroves in Myanmar; coconuts and efficiency, classification accuracy and robustness to sample
palms planted around mangroves in Indonesia; and, natural forests, size—making it suitable for regional scale high-precision mapping;
fishponds and salt fields distributed around mangroves in Thailand. 2. Effective bands for mangrove classification in Southeast Asia: For
Our overall findings from the field work and band selection can be mangrove, the MSI model shows that Sentinel-2 bands are better at
summarized as the following generic recommendations: First, to char­ classifying mangrove than Landsat and Sentinel-1; the more effective
acterize the local environment, make detailed mappings around sample bands for mangrove mapping are SWIR, NIR and Red bands, with
points, especially where they are easily-confused features; Second, sta­ SWIR particularly effective in distinguishing mangrove forest from
tistics of vegetation, topography and other characteristics in each other vegetation; SAR bands are effective in discriminating regional
mapping unit are needed in order to use different suitable band com­ forests from water bodies, thereby rectifying leakage of classifica­
binations for different units and to mask out the area of interest for tions into the water environment;
mapping. For example, the unit in this study was a 10◦ world grid, which 3. The MSI band selection method relies upon a sample database con­
can be changed into ecological division or watershed division in other taining target and easily-confused land cover, and it is more suitable
studies. The characteristics of the different regions will require different for target land cover extraction applications.
band combinations for the classification—as MSI statistics will vary by
region; Thirdly, for extremely confusing land cover, after the selection of CRediT authorship contribution statement
key bands, customized indices may be needed to enhance the separa­
bility between land cover types; Fourth, to distinguish temporal changes Han Xiao: Conceptualization, Methodology, Validation, Formal
and plan field sampling, for example between mangroves and seasonal analysis, Investigation, Resources, Data curation, Visualization, Writing
vegetation (such as rice, as in the Irrawaddy region shown in Fig. 7), - original draft, Writing - review & editing. Fenzhen Su: Conceptuali­
time-series information is needed—which has been well reflected in our zation, Methodology, Validation, Formal analysis, Investigation, Re­
subsequent research. sources, Supervision, Project administration, Funding acquisition.
The results from many previous studies on band selection are varied. Dongjie Fu: Conceptualization, Methodology, Validation, Formal
Some have used combinations of optical and SAR to improve the map­ analysis, Investigation, Resources, Data curation, Visualization, Writing
ping accuracy, especially in tropical, cloudy, and coastal areas: Vaglio - original draft, Writing - review & editing, Supervision. Vincent Lyne:
Laurin et al. (2013) mapped forest and land cover in a tropical site and Writing - review & editing, Supervision. Gaohuan Liu: Conceptualiza­
Yang et al. (2019) made vulnerability assessment of marine disaster in tion. Tingting Pan: Supervision. Jiakun Teng: Supervision.
cloudy coastal areas by combining optical and SAR images; at the same
time, other studies compared two commonly-used optical satellites, LS
Declaration of Competing Interest
and S2, for vegetation mapping using the Red/Red-edges/NIR/SWIR
bands (Mandanici and Bitelli 2016, Clark 2017, Forkuor et al. 2017,
The authors declare that they have no known competing financial
Shang and Zhu 2019). In contrast to these studies, the MSI model ad­
interests or personal relationships that could have appeared to influence
vances a generic systematic approach for taking advantage of different
the work reported in this paper.
satellites in improving the accuracy of mapping. For mangrove mapping,
the MSI ranking of bands shows that S2 is more important than LS and
Acknowledgements
S1. S2 bands perform best in vegetation classification, especially in areas
with confounding vegetation. The advantage of Landsat is that the
This work was supported by the Strategic Priority Research Program
computation cost is minimal. For large mangrove patches with simple
of Chinese Academy of Sciences (Grant No. XDA19060304), Science and
surrounding environment, Landsat satellite is therefore the most
Technology Basic Resources Investigation Program of China (Grant No.
economical choice. However, for mangroves along rivers the narrow
2017FY201401), and the President’s International Fellowship Initiative
river water is often classified as mangrove. Likewise, for tidal mangrove,
of Chinese Academy of Sciences (Grant Number: 2020VEA0009).
the classifications show missing patches. Sentinel-1 can rectify these
errors because of its strong ability for water identification. Therefore, for
mangrove mapping in tropical, cloudy, coastal areas in Southeast Asia, Appendix A
combining these elements in the MSI-selected bands gives efficient, su­
perior, and more robust results than those from individual satellite See Fig. A1 and Tables A1-A5
systems.
The results also suggest future research directions for improving References
mangrove mapping accuracy. This finding is also reflected in previous
Abatzoglou, J.T., Dobrowski, S.Z., Parks, S.A., Hegewisch, K.C., 2018. TerraClimate, a
studies that show the effectiveness of SWIR and NIR bands in estimating high-resolution global dataset of monthly climate and climatic water balance from
vegetation water content (Chen et al. 2005). Jia et al. (2019) utilized the 1958–2015. Sci. Data 5, 170191.
red-edge bands in S2 to establish a new vegetation index that is sensitive Abdel-Rahman, E.M., Ahmed, F.B., Ismail, R., 2013. Random forest regression and
spectral band selection for estimating sugarcane leaf nitrogen concentration using
to submerged mangrove forests. Similarly, we found that the NIR band EO-1 Hyperion hyperspectral data. Int. J. Remote Sens. 34 (2), 712–728.
and the red-edge band of S2 have the best ability for distinguishing Argamosa, R., Blanco, A., Baloloy, A., Candido, C., Dumalag, J.B.L., Dimapilis, L.,
mangroves and water. This suggests that in future, the discriminating of Paringit, E., 2018. “MODELLING ABOVE GROUND BIOMASS OF MANGROVE
FOREST USING SENTINEL-1 IMAGERY.” ISPRS Annals of Photogrammetry, Remote
multi-tidal mangroves can be improved by combining these bands with
Sensing and Spatial. Inf. Sci. IV-3, 13–20.
the water-discriminating VV and VH bands of S1. Astola, H., Häme, T., Sirro, L., Molinier, M., Kilpi, J., 2019. Comparison of Sentinel-2 and
Landsat 8 imagery for forest variable prediction in boreal region. Remote Sens.
Environ. 223, 257–273.
6. Conclusion Benz, U.C., Hofmann, P., Willhauck, G., Lingenfelder, I., Heynen, M., 2004. Multi-
resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready
Our overall conclusions are as follows: information. ISPRS J. Photogramm. Remote Sens. 58 (3), 239–258.
Bunting, P., Rosenqvist, A., Lucas, R., Rebelo, L.-M., Hilarides, L., Thomas, N., Hardy, A.,
Itoh, T., Shimada, M., Finlayson, C., 2018. The Global Mangrove Watch—A New
1. Principles for band selection and classification: before classification, 2010 Global Baseline of Mangrove Extent. Remote Sens. 10 (10).
establish a sample database including confounding features, and use Cao, X., Xiong, T., Jiao, L., 2016. Supervised Band Selection Using Local Spatial
the MSI model to select the key bands for improving the separability Information for Hyperspectral Image. IEEE Geosci. Remote Sens. Lett. 1–5.
Cavanaugh, K.C., Kellner, J.R., Forde, A.J., Gruner, D.S., Parker, J.D., Rodriguez, W.,
of confounding features and reducing redundancy from multi-source Feller, I.C., 2014. Poleward expansion of mangroves is a threshold response to
remote sensing data. This will significantly improve the operation decreased frequency of extreme cold events. Proc. Natl. Acad. Sci. 111 (2), 723–727.

12
H. Xiao et al. International Journal of Applied Earth Observations and Geoinformation 99 (2021) 102320

Chen, D., Huang, J., Jackson, T., 2005. Vegetation Water Content Estimation for Corn Mandanici, E., Bitelli, G., 2016. Preliminary Comparison of Sentinel-2 and Landsat 8
and Soybeans Using Spectral Indices From MODIS Near- and Short-Wave Infrared Imagery for a Combined Use. Remote Sens. 8 (12).
Bands. Remote Sens. Environ. 98, 225–236. Midekisa, A., Holl, F., Savory, D.J., Andrade-Pacheco, R., Gething, P.W., Bennett, A.,
Clark, M.L., 2017. Comparison of simulated hyperspectral HyspIRI and multispectral Sturrock, H.J.W., 2017. Mapping land cover change over continental Africa using
Landsat 8 and Sentinel-2 imagery for multi-seasonal, regional land-cover mapping. Landsat and Google Earth Engine cloud computing. PLoS ONE 12 (9), e0184926.
Remote Sens. Environ. 200, 311–325. Murray, N.J., Phinn, S.R., DeWitt, M., Ferrari, R., Johnston, R., Lyons, M.B., Clinton, N.,
Claverie, M., Ju, J., Masek, J.G., Dungan, J.L., Vermote, E.F., Roger, J.-C., Skakun, S.V., Thau, D., Fuller, R.A., 2019. The global distribution and trajectory of tidal flats.
Justice, C., 2018. The Harmonized Landsat and Sentinel-2 surface reflectance data Nature 565 (7738), 222–225.
set. Remote Sens. Environ. 219, 145–161. Oliphant, A.J., Thenkabail, P.S., Teluguntla, P., Xiong, J., Gumma, M.K., Congalton, R.G.,
Clerici, N., Valbuena Calderón, C.A., Posada, J.M., 2017. Fusion of Sentinel-1A and Yadav, K., 2019. Mapping cropland extent of Southeast and Northeast Asia using
Sentinel-2A data for land cover mapping: a case study in the lower Magdalena multi-year time-series Landsat 30-m data using a random forest classifier on the
region, Colombia. J. Maps 13 (2), 718–726. Google Earth Engine Cloud. Int. J. Appl. Earth Obs. Geoinf. 81, 110–124.
Demarchi, L., Kania, A., Ciężkowski, W., Piórkowski, H., Oświecimska-Piasko, Z., Pal, M., 2007. Random forest classifier for remote sensing classification. Int. J. Remote
Chormański, J., 2020. Recursive Feature Elimination and Random Forest Sens. 26 (1), 217–222.
Classification of Natura 2000 Grasslands in Lowland River Valleys of Poland Based Patel, N., Kaushal, B., 2011. Classification of features selected through Optimum Index
on Airborne Hyperspectral and LiDAR Data Fusion. Remote Sens. 12 (11). Factor (OIF) for improving classification accuracy. J. For. Res. 22 (1), 99–105.
Deus, D., 2018. Assessment of Supervised Classifiers for Land Cover Categorization Based Richards, D.R., Friess, D.A., 2016. Rates and drivers of mangrove deforestation in
on Integration of ALOS PALSAR and Landsat Data. Adv. Remote Sens. 7 (2), 47–60. Southeast Asia, 2000–2012. Proc. Natl. Acad. Sci. 113 (2), 344–349.
Dong, J., Xiao, X., Menarguez, M.A., Zhang, G., Qin, Y., Thau, D., Biradar, C., Moore, B., Shang, R., Zhu, Z., 2019. Harmonizing Landsat 8 and Sentinel-2: A time-series-based
2016. Mapping paddy rice planting area in northeastern Asia with Landsat 8 images, reflectance adjustment approach. Remote Sens. Environ., 235.
phenology-based algorithm and Google Earth Engine. Remote Sens. Environ. 185, Son, N., Chen, C., Chang, N., Chen, C., Chang, L., Thanh, B., 2015. Mangrove Mapping
142–154. and Change Detection in Ca Mau Peninsula, Vietnam, Using Landsat Data and
Forkuor, G., Dimobe, K., Serme, I., Tondoh, J.E., 2017. Landsat-8 vs. Sentinel-2: Object-Based Image Analysis. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 8 (2),
examining the added value of sentinel-2’s red-edge bands to land-use and land-cover 503–510.
mapping in Burkina Faso. GIScience & Remote Sens. 55 (3), 331–354. Sotoca, J.M., Pla, F., Snchez, J.S., 2007. Band Selection in Multispectral Images by
Giesen, W., Wulffraat, S., Zieren, M., Scholten L., 2007. Mangrove guidebook for Minimization of Dependent Information. IEEE Trans. Syst., Man Cybernet., Part C
Southeast Asia. Bangkok, FAO Regional Office for Asia and the Pacific. (Appl. Rev.) 37 (2), 258–267.
Giri, C., Long, J., Abbas, S., Murali, R.M., Qamer, F.M., Pengra, B., Thau, D., 2015. Stehman, S.V., 1997. Selecting and interpreting measures of thematic classification
Distribution and dynamics of mangrove forests of South Asia. J. Environ. Manage. accuracy. Remote Sens. Environ. 62 (1), 77–89.
148, 101–111. Strobl, C., Boulesteix, A.-L., Zeileis, A., Hothorn, T., 2007. Bias in random forest variable
Giri, C., Ochieng, E., Tieszen, L.L., Zhu, Z., Singh, A., Loveland, T., Masek, J., Duke, N., importance measures: Illustrations, sources and a solution. BMC Bioinf. 8 (1), 25.
2011. Status and distribution of mangrove forests of the world using earth Tian, J., Wang, L., Li, X., Gong, H., Shi, C., Zhong, R., Liu, X., 2017. Comparison of UAV
observation satellite data. Glob. Ecol. Biogeogr. 20 (1), 154–159. and WorldView-2 imagery for mapping leaf area index of mangrove forest. Int. J.
Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., Moore, R., 2017. Appl. Earth Obs. Geoinf. 61, 22–31.
Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sens. Vaglio Laurin, G., Liesenberg, V., Chen, Q., Guerriero, L., Del Frate, F., Bartolini, A.,
Environ. 202, 18–27. Coomes, D., Wilebore, B., Lindsell, J., Valentini, R., 2013. Optical and SAR sensor
Grainger, A., 2008. Difficulties in tracking the long-term global trend in tropical forest synergies for forest and land cover mapping in a tropical site in West Africa. Int. J.
area. Proc. Natl. Acad. Sci. 105 (2), 818–823. Appl. Earth Obs. Geoinf. 21, 7–16.
Hamdan, O., Khali Aziz, H., Mohd Hasmadi, I., 2014. L-band ALOS PALSAR for biomass Verrelst, J., Rivera, J.P., Gitelson, A., Delegido, J., Moreno, J., Camps-Valls, G., 2016.
estimation of Matang Mangroves, Malaysia. Remote Sens. Environ. 155, 69–78. Spectral band selection for vegetation properties retrieval using Gaussian processes
Hansen, M.C., Potapov, P.V., Moore, R., Hancher, M., Turubanova, S.A., Tyukavina, A., regression. Int. J. Appl. Earth Obs. Geoinf. 52, 554–567.
Thau, D., Stehman, S.V., Goetz, S.J., Loveland, T.R., Kommareddy, A., Egorov, A., Wang, D., Wan, B., Qiu, P., Su, Y., Guo, Q., Wang, R., Sun, F., Wu, X., 2018. Evaluating
Chini, L., Justice, C.O., Townshend, J.R.G., 2013. High-Resolution Global Maps of the Performance of Sentinel-2, Landsat 8 and Pléiades-1 in Mapping Mangrove
21st-Century Forest Cover Change. Science 342 (6160), 850–853. Extent and Species. Remote Sensing 10 (9).
Jia, Wang, Wang, Mao and Zhang, 2019. A New Vegetation Index to Detect Periodically Wang, L., Jia, M., Yin, D., Tian, J., 2019. A review of remote sensing for mangrove
Submerged Mangrove Forest Using Single-Tide Sentinel-2 Imagery. Remote Sens., forests: 1956–2018. Remote Sens. Environ. 231, 111223.
11, 17. Woodroffe, C.D., 2018. Mangrove response to sea level rise: palaeoecological insights
Korhonen, L., Hadi, P Packalen, Rautiainen, M., 2017. Comparison of Sentinel-2 and from macrotidal systems in northern Australia. Mar. Freshw. Res. 69 (6), 917–932.
Landsat 8 in the estimation of boreal forest canopy cover and leaf area index. Remote Xia, Q., Qin, C.-Z., Li, H., Huang, C., Su, F.-Z., 2018. Mapping Mangrove Forests Based on
Sens. Environ. 195, 259–274. Multi-Tidal High-Resolution Satellite Imagery. Remote Sens. 10 (9).
Lin, J., Liu, X., Li, K., Li, X., 2014. A maximum entropy method to extract urban land by Yang, F., Yang, X., Wang, Z., Lu, C., Li, Z., Liu, Y., 2019. Object-based classification of
combining MODIS reflectance, MODIS NDVI, and DMSP-OLS data. Int. J. Remote cloudy coastal areas using medium-resolution optical and SAR images for
Sens. 35 (18), 6708–6727. vulnerability assessment of marine disaster. J. Oceanol. Limnol. 37 (6), 1955–1970.
Long, Y., Rivard, B., Rogge, D., Tian, M., 2019. Hyperspectral band selection using the N- Yang, H., Du, Q., Su, H., Sheng, Y., 2011. An Efficient Method for Supervised
dimensional Spectral Solid Angle method for the improved discrimination of Hyperspectral Band Selection. IEEE Geosci. Remote Sens. Lett. 8 (1), 138–142.
spectrally similar targets. Int. J. Appl. Earth Obs. Geoinf. 79, 35–47. Zhang, X., Treitz, P.M., Chen, D., Quan, C., Shi, L., Li, X., 2017. Mapping mangrove
Brander, M.L., Wagtendonk, A.J., Hussain, S.S., McVittie, A., Verburg, P.H., de Groot, R. forests using multi-tidal remotely-sensed data and a decision-tree-based procedure.
S., van der Ploeg, S., 2012. Ecosystem service values for mangroves in Southeast Int. J. Appl. Earth Obs. Geoinf. 62, 201–214.
Asia: A meta-analysis and value transfer application. Ecosyst. Serv. 1 (1), 62–69. Zhao, C., Qin, C.-Z., 2020. 10-m-resolution mangrove maps of China derived from multi-
Ma, Y., Wu, H., Wang, L., Huang, B., Ranjan, R., Zomaya, A., Jie, W., 2015. Remote source and multi-temporal satellite observations. ISPRS J. Photogramm. Remote
sensing big data computing: Challenges and opportunities. Future Gener. Comp. Sens. 169, 389–405.
Syst. 51, 47–60.

13