Contents lists available at ScienceDirect

Int J Appl Earth Obs Geoinformation

journal homepage: www.elsevier.com/locate/jag

Mapping mangrove forests using multi-tidal remotely-sensed data and adecision-tree-based procedure

Xuehong Zhanga,b,⁎, Paul M. Treitzc, Dongmei Chena,c,⁎, Chang Quanb, Lixin Shib, Xinhui Lia

a School of Geography and Remote Sensing, Nanjing University of Information Science & Technology, Nanjing 210044, Chinab Key Laboratory of Meteorology and Ecological Environment of Hebei Province, Meteorological Institute of Hebei Province, Shijiazhuang 050021, Chinac Department of Geography and Planning, Queen’s University, Kingston, ON K7L 3N6, Canada

A R T I C L E I N F O

Keywords:Mangrove forestDecision treesMaximum likelihoodTide levelNormalized Difference Vegetation Index(NDVI)Normalized Difference Moisture Index (NDMI)

A B S T R A C T

Mangrove forests grow in intertidal zones in tropical and subtropical regions and have suffered a dramaticdecline globally over the past few decades. Remote sensing data, collected at various spatial resolutions, providean effective way to map the spatial distribution of mangrove forests over time. However, the spectral signaturesof mangrove forests are significantly affected by tide levels. Therefore, mangrove forests may not be accuratelymapped with remote sensing data collected during a single-tidal event, especially if not acquired at low tide. Thisresearch reports how a decision-tree −based procedure was developed to map mangrove forests using multi-tidal Landsat 5 Thematic Mapper (TM) data and a Digital Elevation Model (DEM). Three indices, including theNormalized Difference Moisture Index (NDMI), the Normalized Difference Vegetation Index (NDVI) andNDVIL·NDMIH (the multiplication of NDVIL by NDMIH, L: low tide level, H: high tide level) were used in thisalgorithm to differentiate mangrove forests from other land-cover and land-use types in Fangchenggang City,China. Additionally, the recent Landsat 8 OLI (Operational Land Imager) data were selected to validate theresults and compare if the methodology is reliable. The results demonstrate that short-term multi-tidal remotely-sensed data better represent the unique nearshore coastal wetland habitats of mangrove forests than single-tidaldata. Furthermore, multi-tidal remotely-sensed data has led to improved accuracies using two classificationapproaches: i.e. decision trees and the maximum likelihood classification (MLC). Since mangrove forests aretypically found at low elevations, the inclusion of elevation data in the two classification procedures was tested.Given the decision-tree method does not assume strict data distribution parameters, it was able to optimize theapplication of multi-tidal and elevation data, resulting in higher classification accuracies of mangrove forests.When using multi-source data of differing types and distributions to map mangrove forests, a decision-treemethod appears to be superior to traditional statistical classifiers.

1. Introduction

Mangrove forests are widely distributed in tropical and subtropicalregions of the world, forming important intertidal ecosystems that linkterrestrial and marine systems (Giri et al., 2011a; Zhang and Tian,2013). They are typically distributed from mean sea level to the highestspring tide (Alongi, 2009). Mangrove ecosystems can provide a widevariety of important ecological and economical ecosystem services tocoastal communities, e.g., water filtration, storm protection, shorelinestabilization (Alongi, 2008; Blasco et al., 1996; Kuenzer et al., 2011).However, their health and existence are seriously threatened by relativesea-level rise as well as coastal development and various forms of an-thropogenic activities, such as conversion to agriculture, aquaculture,tourism, urban development and overexploitation (Farnsworth and

Ellison, 1997; Giri and Muhlhausenet, 2008; Giri et al., 2008; Lovelocket al., 2015). In the past several decades, the world's mangrove eco-systems have been destroyed at a rate of 1%–2% per annum (Joneset al., 2016).

Rapid and accurate mapping techniques are required to effectivelymonitor and manage mangrove resources. Conventional field surveyingis time-consuming and labor-intensive. It is also difficult to determinemangrove distribution and abundance with field surveying due to theinaccessibility of mangrove communities. Rapid progress in sensortechnologies and remote sensing methods have brought land-coverdetection into a new era (Zhang et al., 2013) and have proven to beeffective for monitoring mangrove forests (Everitt et al., 2010; Giriet al., 2011a,b, 2015). Numerous studies have employed remotely-sensed data to analyze the relationship between changes in coastal land

http://dx.doi.org/10.1016/j.jag.2017.06.010Received 15 January 2017; Received in revised form 22 June 2017; Accepted 23 June 2017

⁎ Corresponding authors at: School of Geography and Remote Sensing, Nanjing University of Information Science & Technology, Nanjing 210044, China.E-mail addresses: zxhbnu@126.com (X. Zhang), chendm@queensu.ca (D. Chen).

Int J Appl  Earth Obs Geoinformation 62 (2017) 201–214

0303-2434/ © 2017 Elsevier B.V. All rights reserved.

MARK

http://www.sciencedirect.com/science/journal/03032434http://www.elsevier.com/locate/jaghttp://dx.doi.org/10.1016/j.jag.2017.06.010http://dx.doi.org/10.1016/j.jag.2017.06.010mailto:zxhbnu@126.commailto:chendm@queensu.cahttp://dx.doi.org/10.1016/j.jag.2017.06.010http://crossmark.crossref.org/dialog/?doi=10.1016/j.jag.2017.06.010&domain=pdf

use and mangrove distribution (Bird et al., 2004; Chen et al., 2013;Cornforth et al., 2013; Tran et al., 2015) and to characterize change(Murray et al., 2003; Ramírez-García et al., 1998). In particular,medium-resolution (e.g. 10–30 m) multispectral data provides surfaceinformation at regional scales (Kuenzer et al., 2011) and has been ap-plied to monitor and map mangrove forests over large areas. These datainclude: (i) Landsat Multispectral Scanner (MSS) (Giri and Muhlhausen,2008; Giri et al., 2008; Howari et al., 2009; Prasad et al., 2009; Seto andFragkias, 2007; Tran et al., 2015); (ii) Landsat Thematic Mapper (TM)or Enhanced Thematic Mapper Plus (ETM+) (Andriamparany andFrancois, 2010; Ferreira et al., 2009; Howari et al., 2009; Tran et al.,2015; Zhang et al., 2013); (iii) SPOT (Gao, 1999; Prasad et al., 2009;Thu and Populus, 2007; Pu et al., 2012; Tran et al., 2015; Vo et al.,2013); (iv) Advanced Spaceborne Thermal Emission and ReflectionRadiometer (ASTER) (Giri and Muhlhausenet, 2008; Seto and Fragkias,2007; Vaiphasa et al., 2006); and (v) radar imagery (Cornforth et al.,2013; Rao et al., 1999).

Over the past few decades, various methods have been employed tomap mangrove forests or other aquatic vegetation from remotely-sensed images, ranging from visual to semi-automated and un-supervised approaches, and pixel to object-based approaches(Andriamparany and Francois, 2010; Everitt et al., 2010; Ferreira et al.,2009; Giri and Muhlhausenet, 2008; Giri et al., 2007; Heenkenda et al.,2014; Heumann, 2011; Howari et al., 2009; Kanniah et al., 2015; Palinget al., 2008; Prasad et al., 2009; Seto and Fragkias, 2007; Shapiro et al.,2015; Vo et al., 2013; Wang et al., 2004; Zhang et al., 2013). Morerecently, machine learning algorithms such as neural network classifi-cation (Ferreira et al., 2009; Mas, 2004; Seto and Fragkias, 2007) ordecision trees (Heumann, 2011; Liu et al., 2008; Luo et al., 2014, 2016,2017) have also been applied.

The majority of previous studies have used single-tidal (i.e., single-date) remotely-sensed data to map mangrove forests. However, man-groves are located in nearshore coastal wetlands and are periodicallysubmerged by incoming tides. The moderate/lower intertidal mangroveforests are likely to be submerged at high tide, as demonstrated in Fig. 1(Zhang and Tian, 2013). Consequently, changes in tide levels oftenresult in contrasting spectral signatures for mangrove forests and leadto different mapping results, especially in areas with large tidal ranges.As demonstrated in previous studies the tidal effect frequently results inunderestimating the mapped area of mangrove forests from single-tidalremotely-sensed data, particularly at high tide (Lin and Fu, 1995; Zhangand Sui, 2001). Meanwhile, spectral signatures for mangrove forestscontain a mixture of information related to vegetation and wetlandconditions due to the unique characteristics of nearshore coastal wet-land habitats. As Luo et al. (2017) reported, aquatic vegetation can beeffectively mapped by considering life history information on aquaticvegetation with multi-seasonal satellite images. If we can make gooduse of the unique habitat characteristics of mangrove forests, i.e. thewetland background and periodic variation of the tide level, mappingaccuracy of mangrove forests using remotely-sensed data may be im-proved (Zhang and Tian, 2013; Zhang et al., 2013).

Vegetation Indices (VIs) have been extensively used to describevegetation growing state and cover condition (Yu et al., 2005). Inparticular, the Normalized Difference Vegetation Index (NDVI) is awidely used VI for mapping and monitoring a vast array of ecosystems(Pettorelli et al., 2005; Rouse et al., 1973) including mangrove forests(Green et al., 1998; Jensen et al., 1991; Panigrahy et al., 2012).Moreover, numerous studies have shown that mangrove NDVI valuesare highly correlated with mangrove biomass, canopy cover, and leafarea index (LAI) (Green et al., 1997, 1998; Jensen et al., 1991). Inaddition, the Normalized Difference Moisture Index (NDMI), calculatedusing the near infrared (NIR) and shortwave infrared (SWIR) bands, caneffectively estimate vegetation water content (Chen et al., 2005; Huanget al., 2009; Wilson and Sader, 2002). However, to our knowledge,NDMI has seldom been incorporated into the classification of mangroveforests.

In this paper, we investigated the unique characteristics of near-shore coastal wetland habitats for mapping mangrove forests inFangchenggang City, China and developed and evaluated a decision-tree algorithm for mapping mangrove forests using multi-tidal remotesensing data. Specifically, two objectives were addressed: (1) to eval-uate the ability to discriminate mangrove forests from other land covertypes using two spectral indices, i.e., NDVI and NDMI; and (2) tocompare the decision-tree algorithm to a traditional classification al-gorithm (i.e., maximum likelihood classification − MLC) using multi-tidal remotely-sensed data to map mangrove forests.

2. Study area

The study area is located in Fangchenggang City, Guangxi ZhuangAutonomous Region of China, which is in a tropical rainforest andmonsoon forest zone (Fig. 1). The area covers a wide range of vegeta-tion types, including: coniferous forests (China fir [Cunninghamia lan-ceolata], masson pine [Pinus massoniana]); broad-leaved forests (eu-calyptus urophylla, Schima superba, Camphor [Cinnamomum camphora]);mangrove forests; orchards (Leechee [Litchi chinemis], Longan [Dimo-carpus longgana Lour.], aniseed [Illicium verum]); bushwood (Myrtle[Rhodomyrtus tomentosa], Rhus chinensis, Phyllanthi Fructus, Melastomacandidum); and grassland (Deyeuxia Clarion, Heteropogon contortus, Di-cranopteriis dichotoma (Thunb.) Bernh and Arundinella hirta). Mangroveforests are distributed at Pearl Bay (21°28′–21°37′N, 108°2′–108°16′E)within the Beilunhekou National Nature Reserve Area. The speciescommunities include Avicennia marina, Aegiceras corniculatum, Kandeliacandel, Bruguiera gymnorrhiza and Acanthus ilicifolius. The hierarchy ofmangrove plant communities was simple and most of them were singlelayer. In addition, the height of mangrove plant communities is usuallyless than 5 m in the study area (Liang et al., 2004). The tidal type isregularly diurnal, and the mean and maximum tidal ranges are 2.24 mand 5.64 m above the Fangchenggang datum plane, respectively (Zhangand Tian, 2013).

3. Data and methodology

3.1. Datasets

The data used for this study include satellite remote sensing data(Table 1), high resolution images from Google Earth (acquired on No-vember 16, 2007 and December 25, 2013), in situ measurements, andancillary spatial data (e.g., topographic maps and a DEM). Two-dates ofLandsat 5 TM data were collected on October 30, 2006 (Fig. 1A) andNovember 15, 2006 (Fig. 1B) for tide levels of 417 cm and 282 cm(National Marine Data and information Service, 2005), respectively.The tidal datum was 230 cm below the mean sea level. Moreover, tovalidate the mapping results of mangrove forests, two Landsat 8 OLIdata (https://lta.cr.usgs.gov/L8) were also employed. They were col-lected on December 20, 2013 (Fig. 1C) and December 4, 2013 (Fig. 1D)for tide levels of 362 cm and 265 cm (National Marine Data and in-formation Service, 2012), respectively. Additionally, though there wasa time lag of 16 days between image acquisitions, the land cover/landuse for the study can be considered unchanged. To geometrically cor-rect the Landsat data, a 1:50,000 topographic map was used for pla-nimetric reference. In our study area, mangrove forests are generallylimited to elevations below 10 m above sea level (a.s.l.) and foundwithin 1 km from the coastline in the study area (Liu et al., 2008). Toestablish elevation limits, the SRTM (Shuttle Radar Topographic Mis-sion) DEM data with 90 m spatial resolution, resampled to 30 m spatialresolution, was used. The vertical accuracy of SRTM is approximately2.39 m in the low altitude areas where the elevation value is less than20 m (Du et al., 2013).

The field survey was carried out during the period of November 1–5,2006. The surveying sites are shown in Fig. 1 and include 15 mangroveforest sites, 20 terrestrial vegetation sites (including 2 shaded forest

X. Zhang et al. Int J Appl  Earth Obs Geoinformation 62 (2017) 201–214

202

https://lta.cr.usgs.gov/L8

Fig. 1. The locations of field samples available for calibration and validation purposes and Landsat 5 TM and Landsat 8 OLI (Operational Land Imager) images of the study area at high (Aand C) and low (B and D) tide levels. TM images A and B are false-color composites of the study area displaying channel 4 (0.76–0.90 μm) in red, channel 3 (0.63–0.69 μm) in green, andchannel 2 (0.52–0.60 μm) in blue. OLI images C and D are false-color composites of the study area displaying channel 5 (0.845–0.885 μm) in red, channel 4 (0.630–0.680 μm) in green,and channel 3 (0.525–0.600 μm) in blue. The white polygon outlines the boundaries of the study area whereas the mangrove forests are highlighted by the yellow polygons at differenttide levels. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

X. Zhang et al. Int J Appl  Earth Obs Geoinformation 62 (2017) 201–214

203

sites and 3 sites of mixed areas of water bodies and terrestrial vegeta-tion), 12 tidal flat sites; 10 bare soil sites, and 7 built-up land sites. Thearea of each field site was approximately 100 m × 100 m. The locationand land cover types for each field site were established using a por-table Hi-Target Qcool i7 Global Positioning System (GPS) with an ac-curacy of approximately± 5 m, which helped recognize calibrationareas and assess the mapping accuracy of mangrove forests in the studyarea.

3.2. Image pre-processing

The methodology for this study is outlined in Fig. 2. Image pre-processing, including radiometric and geometric correction, were per-formed in the ENVI/IDL environment (ITT Industries, Inc., USA). First,the image digital number (DN) data were converted to radiance usingthe calibration coefficients provided by the Landsat 5 TM and Landsat 8OLI metadata files.

Second, radiance data were converted to top-of-atmosphere (TOA)reflectance by normalizing for solar elevation and solar spectral

irradiance (Chander et al., 2009) and atmospheric correction was im-plemented by using the improved dark object subtraction technique(Chavez, 1986, 1996). Finally, geometric corrections were performedwith reference to the 1:50,000 topographic map data. For image re-gistration, ground control points (GCPs) were well distributedthroughout the scene and a first-order polynomial transformation witha nearest neighbor resampling algorithm was applied. The average rootmean square error (RMSE) was approximately 0.5 pixels (i.e., 15 m).

3.3. Spectral indices calculation

The NDVI (Rouse et al., 1973) and the NDMI (Wilson and Sader,2002) were derived from the preprocessed Landsat 5 TM data:

=−

+

ρ ρρ ρ

NDVI NIR RNIR R (1)

=−

+

ρ ρρ ρ

NDMI NIR SWIRNIR SWIR (2)

Table 1Characteristics of TM and OLI images.

Acquisition Date Scene Center Scan Time(Local Time) Platform/Sensor Path/Row Pixel Size Tidal stage

2006−10-30 11:05 Landsat5/TM 125/45 30 m 417 cm2006−11-15 11:01 Landsat5/TM 125/45 30 m 282 cm2013−12-04 11:12 Landsat8/OLI 125/45 30 m 265 cm2013−12-20 11:12 Landsat8/OLI 125/45 30 m 362 cm

Fig. 2. Methodology workflow.

X. Zhang et al. Int J Appl  Earth Obs Geoinformation 62 (2017) 201–214

204

where ρR, ρNIR and ρSWIR are the reflectance values for the red, NIR andSWIR channels, respectively. For instance, at a basic level, NDVI is usedto separate vegetation from non-vegetation (Rouse et al., 1973).Moreover, the NIR band (e.g., TM band 4 centering at 830 nm) is sui-table for normalization because it is relatively insensitive to vegetationwater content changes in comparison with the longer NIR and SWIRwavelengths (Chen et al., 2005). Therefore, the SWIR channel (TM band5)) is a water absorption band as compared to the NIR (Chen et al.,2005; Hunt et al., 1987). NDMI, formulated using a contrast betweenthe NIR and SWIR is widely used to estimate vegetation water content(Chen et al., 2005; Huang et al., 2009; Wilson and Sader, 2002). As atypical wetland habitat, the reflectance of mangrove forests at infraredwavelengths, especially for the SWIR, is typically lower than for ter-restrial vegetation. Furthermore, given the fact that mangrove forestsare characterized by wetland background, forest vegetation and peri-odic variation of the tide level, NDVIL·NDMIH (i.e. the multiplication ofNDVIL by NDMIH, L = low tide level, H = high tide level) was devel-oped for describing the distinctive characteristics of mangrove forests.Consequently, this paper attempted to employ the above three indicesto compare their capacity for mapping mangrove forests in the studyarea.

3.4. Calibration and validation data

To analyze the spectral characteristics of typical land cover types inthe study area and assess the accuracy for mangrove forest classifica-tion, calibration (or training) and validation (or reference) pixels for allland cover categories were collected from the Landsat 5 TM images bycombining field survey data with visual interpretation of Google Earthimages and the Landsat data in a stratified random sampling pattern.Seven major classes were delineated: mangrove forests, terrestrial ve-getation, bare soil, built-up land, water bodies, tidal flats and water-veg, which denotes mixed areas of water and terrestrial vegetation(Table 2) (Note: the cropland was combined with bare soil because thecrops had been harvested at the time of both image acquisitions). Ca-libration pixels were selected using visual interpretation, prior knowl-edge of the study area and from the field survey data. For example, theproperties of nearshore coastal wetland habitats, including the periodicvariation of tide level, result in very distinctive signatures for mangroveforests and exhibit high contrast to other land cover types in the studyarea. In addition, since mangrove forests grow at different groundelevations, they are submerged differentially by the tides at differenttimes. As a result, the mangrove forests are further divided into low/moderate intertidal zone and the upper intertidal zone, denoted as low-stand mangrove and high-stand mangrove, respectively.

In total, 2464 pixels (including 98 sample blocks) were selectedfrom Landsat 5 TM images. Furthermore, 75% of these pixels wererandomly selected and assigned to the calibration dataset (to calibrateand construct the decision trees) and the remaining used to validate the

accuracy of the decision-tree algorithm. For the calibration dataset, thenumbers of low-stand mangrove, high-stand mangrove, terrestrial ve-getation, bare soil, built-up land, water bodies, tidal flats and water-vegwere 238, 387, 292, 230, 275, 216, 139 and 71 pixels, respectively.Since identifying mangrove forests was the focus of the decision-treealgorithm developed, the validation data were only classified as man-grove forest and non-mangrove classes with 196 and 420 pixels, re-spectively.

Similarly, 1860 pixels (including 85 sample blocks) were selectedfrom Landsat 8 OLI images by means of visual interpretation of GoogleEarth images acquired on December 25, 2013. 75% of these pixels werealso randomly selected and assigned to the calibration dataset and theremaining used for validation. For the calibration dataset, the numbersof low-stand mangrove, high-stand mangrove, terrestrial vegetation,bare soil, built-up land, water bodies, tidal flats and water-veg were140, 215, 302, 172, 217, 165, 104 and 80 pixels, respectively. Thevalidation data consisted of mangrove forest and non-mangrove classeswith 150and 315 pixels, respectively.

3.5. Spectral analysis of land cover at high and low tide

Based on calibration data, the spectral reflectance and spectral in-dices of all land-cover types at high and low tide were analyzed. Thespectral analysis of land cover is helpful for understanding spectral si-milarities (and differences) among these land-cover types at high andlow tides. It is understood that spectral characteristics of nearshorecover types may change during high and low tides (Cibula et al., 1992;Zhang and Tian, 2013; Zhao et al., 2008), thereby these changes mustbe characterized when deriving the decision-tree algorithm.

3.6. Decision-tree algorithm

The decision-tree method has been employed in previous studies toextract mangrove forests (Liu eta al., 2008; Heumann, 2011). A decisiontree is a classification procedure that repeatedly divides a set of trainingdata into smaller subsets based on tests to one or more of the featurevalues (Liu eta al., 2008; Tooke et al., 2009). Unlike many other sta-tistical approaches, e.g. MLC, the decision tree does not depend onassumptions specific to variable distribution or the independence of thevariables from one another (Liu eta al., 2008; Quinlan, 1993). The ‘nopresumption condition of the decision tree (Quinlan, 1993; Friedl andBrodley, 1997; Pal and Mather, 2003) is an advantage for incorporatingancillary GIS (Geographic Information System) data which often exhibitdifferent forms/distributions as well as being highly correlated (Jensen,2005; Liu eta al., 2008). As implemented in ENVI/IDL, remote sensingdata are divided sequentially by the decision tree, with each pixeleventually being assigned to a specific class (Liu eta al., 2008; Quinlan,1986; Xu et al., 2005). Considering the unique nature of mangroves innearshore coastal wetland habitats, the partitions are based on re-peatedly applying threshold values of NDVI, NDWI, elevation andspectral reflectance, visually interpreted from calibration data and ex-pert knowledge of the study area.

3.7. Comparison between classification approaches

To evaluate the performance of the decision-tree (DT) algorithm formapping mangrove forests, the results from the decision-tree methodwere compared to those from a MLC using different feature combina-tions (Table 3). Furthermore, DT_SIM_OLI, a decision-tree classificationusing multi-tidal Landsat OLI and DEM data were designed to validatethe mapping results. The classification features consisted of DEM,NDVIL, NDVIH, NDMIL, NDMIH, NDVIL·NDMIH and the spectral re-flectance at high and low tide (denoted as ρH and ρL, respectively).Overall accuracy, producer’s accuracy, user’s accuracy, and the overallkappa coefficient from each classification are reported (Richards andJia, 2006).

Table 2Class definitions.

Classes Classification class definitions

Mangroves Areas covered by both closed and open mangrove forests

Non-mangrove Terrestrialvegetation

Areas covered by forests, shrublands,grasslands and croplands

Bare soil Exposed soilBuilt-up land Areas covered by all kinds of artificial

facilitiesWater bodies Areas of open water with no emergent

vegetationTidal flat Areas of nearly flat alternately covered

and exposed by the tidesWater-veg Mixed areas of water and terrestrial

vegetation

X. Zhang et al. Int J Appl  Earth Obs Geoinformation 62 (2017) 201–214

205

DT: decision-tree classification, MLC: maximum likelihood classifi-cation, SI: spectral index, L: low tide level, H: high tide level, M: mul-tiple tide levels, ρL: TM reflectance data at low tide level, ρH: TM re-flectance data at high tide level, NDVIL·NDMIH: the multiplication ofNDVIL by NDMIH, DT_SIM_TM: the decision-tree algorithm from multi-tidal remotely-sensed TM data, DT_SIM_OLI: the decision-tree algorithmfrom multi-tidal remotely-sensed OLI data (only DT_SIM_OLI used OLIimagery and others used TM imagery).

4. Results and discussion

4.1. Spectral characteristics of typical land covers in the study area

Spectral reflectance curves derived from TM data for typical land-cover types during high and low tides are presented in Fig. 3. It is clearthat there are significant similarities between spectral signatures ofmangrove forests and other vegetated classes such as terrestrial vege-tation and water-veg as reported in previous studies (FGDC, 1992;Ozesmi and Bauer, 2002). However, the canopy reflectance values ofmangrove forests are typically lower than terrestrial vegetation in the

NIR and SWIR bands, likely attributed to the differences in backgroundreflectance (Kuenzer et al., 2011; Xiao et al., 2007). Different mangrovespecies often have similar spectral reflectance patterns (Kamal andPhinn, 2011). However, spectral reflectance may slightly differ due tovariable environmental conditions affecting their biophysical and che-mical properties (e.g., canopy density, chlorophyll content) (Gao, 1998;Kuenzer et al., 2011). There are a number of examples of spectralconfusion between mangrove forests and other land cover types; e.g.,mixed river water and terrestrial vegetation (Fig. 4A); mixed aquaticwater and vegetation (Fig. 4B); and mangrove forests and shaded for-ests located in mountainous areas. In addition, mangrove forests lo-cated at the fringes of patches (Fig. 4C), especially for the pioneermangrove species (e.g., Aricennia marina), or the very sparse mangroveforests (Fig. 4D), were easily confused with the background of man-grove forests such as tidal flat or water bodies.

The spectral characteristics of some land cover types change sig-nificantly during high and low tides (Fig. 3). For example, the NIR re-flectance for low-stand mangrove and tidal flats at the high tide wereless than 65% of corresponding values at the low tide (Table 4).However, the changes of other terrestrial land cover types were lessobvious (i.e., less than 22%). Due to its prevalence in the upper inter-tidal zone and as a function of its canopy height, the difference in re-flectance for high-stand mangrove during high and low tides werelower than those of low-stand mangrove. This indicates that the ter-restrial land cover types, including terrestrial vegetation, built-up land,and bare soil, had relatively stable reflectance values during low andhigh tides (from October 30, 2006 to November 15, 2006). Fluctuatingtidal conditions (i.e., water levels) may produce different elevations ofwaterlines (Zhao et al., 2008) and frequently alter the spectral re-flectance of vegetation (Cibula et al., 1992). In addition, larger tidalranges (i.e., the difference between the low tide and the high tide), willresult in larger spectral changes for mangrove forests. Therefore, water-submergence can be considered the primary reason for changes in re-flectance values of nearshore land cover types, thereby giving rise touncertainty in pixel classification. These changes account for the dis-crepancy (i.e., reduced area) between the measured area of mangroveforests using single-tidal remote sensing data and field surveys inGuangxi Province, China (Lin and Fu, 1995; Zhang and Sui, 2001).

4.2. Spectral indices for typical land cover types

NDVIL/H and NDMIL/H were extracted from Landsat 5 TM data andLandsat 8 OLI data acquired during low and high tides. Similar to re-flectance values, the NDVI values for low-stand mangrove and tidal flatswere much lower during high tide (Table 5). However, NDVI values forother land cover types had no obvious changes during the high and lowtides. Moreover, the differences in absolute mean values of NDVI forlow-stand mangrove and tidal flats during low and high tides were allgreater than 0.2; however, those of other land covers were all less than0.1 (Table 6). This can be attributed to the greater proportion of seawater in the nearshore area during high tide, resulting in lower NIRreflectance, especially for low-stand mangrove and tidal flats.

The contrast between the NIR (760–900 nm) and SWIR(1550–1750 nm; 2080–2350 nm) bands is due to the high absorption ofSWIR by water contained within vegetation (Chen et al., 2005; Huntet al., 1987; Wilson and Sader, 2002; Zhao et al., 2008). Hence thecontrast between the NIR and the SWIR bands for NDMI was used toestimate vegetation water content (Chen et al., 2005; Huang et al.,2009; Wilson and Sader, 2002) and characterize the moisture in-formation for a pixel. As a result, NDMIs of coastal wetlands such asmangrove forests, water bodies, and tidal flats, were higher than otherterrestrial classes (Table 5) due to the presence of higher levels ofmoisture in these environments. The NDMIs of nearshore land covertypes significantly changed during the low and high tides due to theirmoisture differences between tidal conditions, especially for low-standmangrove and tidal flats. It was found that the differences in absolute

Table 3Classification approaches and feature combinations compared in this study.

Approach Classifier Feature combinations

DT_SIL DT DEM+NDVIL+NDMILDT_SIM_TM DT DEM+NDVIL+NDMIH+NDVIL·NDMIHDT_SIM_OLI DT DEM+NDVIL+NDMIH+NDVIL·NDMIHML_DEM_L MLC DEM+ρLML_DEM_M MLC DEM+ρL+ρHML_L MLC ρLML_M MLC ρL+ρH

Fig. 3. Spectral reflectance curves from Landsat TM data for typical land covers at: (A)high tide; and (B) low tide. The curves indicate mean reflectance. Error bar shows± onestandard deviation. Water-veg denotes mixed areas of water and terrestrial vegetation.

X. Zhang et al. Int J Appl  Earth Obs Geoinformation 62 (2017) 201–214

206

mean values of NDMI for low-stand mangrove and tidal flats duringhigh and low tides were all greater than 0.13, and those of other landcovers were all less than 0.09 (Table 6).

NDVIL·NDMIH, can effectively integrate vegetation information

(especially for wetland vegetation) with moisture condition. It is helpfulfor discriminating vegetation from non-vegetation, e.g. built-up land,water, tidal flat and bare soil (Table 5). In particular, NDVIL·NDMIHtakes into account the tide level information, ensuring that the value of

Fig. 4. Examples of misclassification between mangrove forests andother cover types (left − TM data; right − Google Earth©): (A) mixedriver water and terrestrial vegetation; (B) mixed aquatic water andvegetation; (C) mangrove forests located at the fringes of patches; and(D) sparse mangrove forests.

X. Zhang et al. Int J Appl  Earth Obs Geoinformation 62 (2017) 201–214

207

NDVIL·NDMIH for mangrove forests will not change significantly withthe change in tide level.

4.3. Decision-tree algorithm

Combining the analyses of Sections 4.1 and 4.2, three decision-treealgorithms, namely DT_SIM_TM, DT_SIL and DT_SIM_OLI (Table 3), weredeveloped. Three rules were applied to the decision-tree classificationfor DT_SIM_TM (Fig. 5). First, given the fact that the vegetation ischaracterized by high NDVI values as well as high NDMI values,NDVI·NDMI, rather than NDVI/NDMI or other permutations, can beused to distinguish vegetation from non-vegetation. Meanwhile, NDVIvalues and NDMI values for mangrove forests at low tide are generallyhigher than those at high tide due to the tidal effect. Therefore,NDVIL·NDMIH was adopted to separate terrestrial vegetation, mangroveforest and water-veg from built-up land, water bodies, tidal flats andbare soil in Rule 1. If the NDVIL·NDMIH of a pixel was equal to, orgreater than 0.05, it was terrestrial vegetation, mangrove forest orwater-veg; otherwise, it was considered built-up land, water bodies,tidal flats or bare soil. Second, Rule 2 was identified based on NDMIH incombination with elevation (i.e., DEM). As Liu et al. (2008) observed,field investigations found that mangrove forests were not present above10 m elevation in this study area. Consequently, a 10 m elevation value,derived from the DEM, was utilized to exclude non-mangrove pixelsthat had similar spectral attributes to mangrove pixels, e.g. shadedterrestrial forests and mixed pixels of inland water and terrestrial ve-getation. Therefore, if the NDMIH of a pixel was equal to, or greaterthan 0.43, and its DEM was less than 10, the pixel was classified asmangrove forest or water-veg; otherwise it was classified as terrestrialvegetation. Finally, Rule 3 was generated to further separate mangroveforest from water-veg using the NDVIL threshold value of 0.25.

To compare with DT_SIM_TM, the decision-tree classification ofDT_SIL was also performed by adopting the remotely sensed data at lowtide. Two rules were established (Fig. 6). First, Rule 1 was produced toseparate built-up land, water bodies, tidal flat, bare soil and water-vegfrom terrestrial vegetation and mangrove forest using the NDVILthreshold value of 0.31. Second, Rule 2 was derived to further distin-guish mangrove forest from water-veg based on NDMIL in combinationwith DEM data. Similarly, if the NDMIL of a pixel was equal to, orgreater than 0.42 and its DEM was less than 10, the pixel was classified

as mangrove forest; otherwise, it was assigned to terrestrial vegetation.Meanwhile, the reliability of the decision-tree classification was

examined by applying DT procedure to the recent multi-tidal LandsatOLI data (DT_SIM_OLI). Similarly, three rules were applied (Fig. 7).First, Rule 1 was created to separate terrestrial vegetation, mangroveforest and water-veg from built-up land, water bodies, tidal flats andbare soil by employing a NDVIL·NDMIH threshold value of 0.05. Second,Rule 2 was established based on NDMIH (threshold value = 0.44) incombination with elevation. Finally, Rule 3 was built to further dis-tinguish mangrove forests from water-veg employing the NDVILthreshold value of 0.24. Based on these analyses, ground cover wascategorized into mangrove forests and non-mangrove classes. Theabove–mentioned thresholds were derived by comparing the value ofμ ± 2σ (μ = mean value and σ= standard deviation) of spectral fea-tures for each cover class (Table 5).

4.4. Comparison of classifications

4.4.1. Classification resultsGiven the goal of this research was to identify mangrove forests, the

final classification maps simply display the mangrove forests super-imposed on image data collected at low tide. The thematic maps pre-sented in Fig. 8 are generated with the following combinations: a) adecision-tree classification using Landsat TM spectral indices (low tide)and the DEM (DT_SIL) (Fig. 8A); b) a decision-tree classification usingLandsat TM spectral indices (multi-tidal) and the DEM (DT_SIM_TM)(Fig. 8B); c) a MLC using Landsat TM reflectance data (low tide) and theDEM (ML_DEM_L) (Fig. 8C); d) a MLC using Landsat TM reflectancedata (multi-tidal) and the DEM (ML_ DEM_M) (Fig. 8D); e) a MLC usingLandsat TM reflectance data (low tide) (ML_SIL) (Fig. 8E); f) a MLCusing Landsat TM reflectance data (multi-tidal) (ML_SIM) (Fig. 8F); andg) a decision-tree classification using Landsat OLI spectral indices(multi-tidal) and the DEM (DT_SIM_OLI).

Because the TM and OLI data were acquired during the same seasonin 2006 and 2013 and the mangrove forests are situated within naturereserves, there were no obvious changes observed in the land cover/land use in the study area, particularly in mangrove forests, i.e., exceptfor a small increase in built-up land during that period. Visual inspec-tion of the mangrove forests maps indicates the maps produced fromdecision-tree classifications using multi-tidal spectral indices and the

Table 4Percent difference in reflectance for mangrove forests in the NIR from TM data collected at high and low tide (16 days between acquisition dates).

Percent difference Terrestrial vegetation High-stand mangrove Low-stand mangrove Build-up land Water bodies Tidal flats Bare soil Water-veg

−NIR NNIR

( H IRL)L

4.9% −13.2% −70.0% −3.1% −21.5% −66.1% −13.1% −10.9%

Note: NIRL and NIRH are the reflectance values of the NIR band at the low tide and high tide, respectively.

Table 5Mean and two standard deviation values of spectral indices for typical land cover classes.

AcquisitionYear/Sensor

Spectral indices Terrestrialvegetation

High-standmangroveb

Low-standmangroveb

Built-up land Water Tidal flats Bare soil Water-veg

2006/TM NDVILa 0.58 ± 0.06 0.46 ± 0.04 0.39 ± 0.04 0.17 ± 0.08 −0.12 ± 0.03 0.06 ± 0.02 0.18 ± 0.03 0.15 ± 0.05NDVIHa 0.68 ± 0.05 0.53 ± 0.08 0.00 ± 0.07 0.21 ± 0.11 −0.19 ± 0.05 −0.20 ± 0.02 0.22 ± 0.04 0.17 ± 0.09NDMIL 0.33 ± 0.05 0.46 ± 0.04 0.47 ± 0.05 −0.10 ± 0.06 0.50 ± 0.05 −0.13 ± 0.04 −0.06 ± 0.05 0.54 ± 0.08NDMIH 0.34 ± 0.05 0.51 ± 0.04 0.59 ± 0.06 −0.08 ± 0.06 0.48 ± 0.06 0.48 ± 0.06 −0.13 ± 0.06 0.63 ± 0.08NDVIL·NDMIH 0.20 ± 0.04 0.24 ± 0.03 0.23 ± 0.04 −0.02 ± 0.02 −0.06 ± 0.02 0.03 ± 0.01 −0.03 ± 0.01 0.10 ± 0.03

2013/OLI NDVIL 0.73 ± 0.07 0.66 ± 0.03 0.60 ± 0.03 0.29 ± 0.13 −0.35 ± 0.02 0.07 ± 0.01 0.20 ± 0.04 0.16 ± 0.04NDVIH 0.76 ± 0.09 0.74 ± 0.04 0.25 ± 0.03 0.28 ± 0.15 −0.36 ± 0.02 −0.10 ± 0.03 0.25 ± 0.02 0.19 ± 0.10NDMIL 0.35 ± 0.05 0.53 ± 0.02 0.51 ± 0.03 −0.00 ± 0.06 0.40 ± 0.03 −0.18 ± 0.02 −0.05 ± 0.07 0.53 ± 0.07NDMIH 0.34 ± 0.05 0.60 ± 0.03 0.66 ± 0.02 0.01 ± 0.08 0.43 ± 0.02 0.28 ± 0.06 −0.11 ± 0.03 0.61 ± 0.08NDVIL·NDMIH 0.25 ± 0.07 0.40 ± 0.04 0.40 ± 0.02 0.00 ± 0.02 −0.15 ± 0.02 0.02 ± 0.01 −0.02 ± 0.02 0.10 ± 0.03

a The subscript L and H denote low tide level and high tide level, respectively.b The ‘low-stand mangrove’ and ‘high-stand mangrove’ are the mangrove forests in the moderate/lower intertidal and upper intertidal zones, respectively.

X. Zhang et al. Int J Appl  Earth Obs Geoinformation 62 (2017) 201–214

208

DEM (Figs. 8B and G) are more closely related to the mangrove dis-tribution of the study area than those produced from the other fiveclassifications (Fig. 8A, C-F). Moreover, the multi-tidal informationclearly helps to discriminate mangrove forests from non-mangroveclasses. For instance, the classification results of Fig. 8D were markedlybetter than those of Fig. 8C, and F had less terrestrial forest mis-classified as mangrove forests in comparison to Fig. 8E. Also as shownin Fig. 4A and B, some of the water-veg class has been misclassified asmangrove forests, especially in the coastal aquaculture areas and nearthe estuaries (i.e., in Fig. 8C and D.

Table 7 displays the mangrove forest area statistics derived from the

seven classification schemes. ML_DEM_L and ML_DEM_M (mangroveforests = 3717.1 ha and 1880.2 ha, respectively and percentagearea = 4.17% and 2.11%, respectively) resulted in a significantlygreater area being classified as mangrove forests than other classifica-tions. This difference can, in part, be attributed to MLC requiring theassumption that the distribution of a class sample is normal and ele-vation data do not satisfy that assumption. Further, it is anticipated thatthere is considerable spectral overall with similar classes i.e., water-vegand other terrestrial vegetation.

Table 6The difference in mean values of spectral indices with the change of tides and the time lag of 16 days.

Acquisition Year/Sensor

Spectral indices Terrestrialvegetation

High-standmangrove

Low-standmangrove

Build-upland

Waterbodies

Tidal flats Bare soil Water-veg

2006/TM NDVILeNDVIH −0.1 −0.07 0.39 −0.04 0.07 0.26 −0.04 −0.02NDMILeNDMIH −0.01 −0.05 −0.13 −0.02 0.02 −0.60 0.07 −0.09

2013/OLI NDVILeNDVIH −0.03 −0.08 0.35 0.01 0.01 0.17 −0.05 −0.03NDMILeNDMIH 0.01 −0.07 −0.15 −0.01 −0.03 −0.46 0.06 −0.08

Fig. 5. Decision tree for mangrove forest identification using multi-tidal TM data (i.e. DT_SIM_TM).

Fig. 6. Decision tree for mangrove forest identification using low-tidalTM data (i.e. DT_SIL).

X. Zhang et al. Int J Appl  Earth Obs Geoinformation 62 (2017) 201–214

209

4.4.2. Accuracy assessmentIn addition to the visual inspection, a stratified random sampling

method was applied to perform an accuracy assessment for each clas-sified thematic map. Since the accuracy of mangrove forest was ofprimary concern, only two classes (mangrove forest and non-mangrove)were considered in the accuracy assessment. Confusion matrices wereconstructed by comparing the classified data with validation data. Thesummary statistics are reported in Table 8.

The accuracy assessment results demonstrate that the multi-tidalinformation can significantly improve the accuracy of classifyingmangrove forests. For instance, all the accuracy values of DT_SIM_TM,DT_SIM_OLI, ML_DEM_M and ML_M were higher than those of DT_SIL,ML_DEM_L and ML_L, respectively. Moreover, the decision-tree map-ping approach exhibited higher accuracy values in comparison with theMLC. Although the producer's accuracies for mangrove forests forML_DEM_L, ML_DEM_M and ML_M (93.3%, 96.9% and 92.4%, respec-tively) were slightly higher than that for DT_SIM_TM and DT_SIM_OLI(91.8% and 90.7%, respectively), other accuracy values (e.g., overallaccuracy, Kappa and user's accuracy of mangroves) for DT_SIM_TM andDT_SIM_OLI were higher (Table 8). This means that more non-mangrovepixels were misclassified as mangrove class for all the MLC approaches(i.e. ML_DEM_L, ML_DEM_M, ML_M and ML_M) than those ofDT_SIM_TM and DT_SIM_OLI., a result supported by the visual appear-ance of these maps (Fig. 8).

Additionally, though mangrove forests are characterized by lowelevations, the accuracy values for MLC significantly decreased whenDEM data were included, with the exception of the producer's ac-curacies. For example, the Kappa value for ML_L, was 0.215 higher thanthat for ML_DEM_L and the Kappa value for ML_M was also 0.06 higherthan that for ML_DEM_M.

With respect to the classification features, the mapping approachesemploying the spectral indices (including NDVIL/H, NDMIL/H andNDVIL·NDMIH) exhibited higher accuracies, compared to reflectancedata alone. In particular, DT_SIM (including DT_SIM_TM andDT_SIM_OLI) had the highest overall accuracy, Kappa and user's accu-racy for mangroves for all the seven classifications.

4.4.3. Comparison of classification approachesAmong the six classifications of TM data, DT_SIM_TM performed

very well (Table 8). To validate the classification results for mangrove

forests and compare classification approaches, recent Landsat 8 OLIdata were used to compare with DT_SIM_TM. The similar accuracies,area and percentage area of mangrove forests for DT_SIM_TM andDT_SIM_OLI indicate that the decisiontree-based procedure is stable andreliable (Table 9).

Mangroves are located in low coastal wetlands and are periodicallysubmerged by incoming tides. Hence, the fluctuating tidal conditionsproduce different mapping results for mangrove forests (Lin and Fu,1995; Zhang and Sui, 2001) and may restrict the mapping accuracy ofmangrove forests with large tidal ranges when using single-tidal re-motely-sensed data. As demonstrated in Table 8 and Fig. 8, the multi-tidal remotely-sensed data have led to improved classification ac-curacies over single-tidal acquisitions.

Meanwhile, the DEM data can also represent the low coastal wet-land environment of mangrove forests. As Liu et al. (2008) observed,mangrove forests were not distributed above a certain elevation.However, the classification accuracy from MLC decreased when DEMdata were incorporated with the reflectance data (Table 8). Multi-source data, e.g. the reflectance data, spectral derivatives and ancillaryspatial data, often exhibit different forms/distribution (Jensen, 2005;Liu et al., 2008) and do not satisfy the parametric assumptions (e.g.,normal distribution) of the MLC (Liu et al., 2008; Quinlan, 1993; Tookeet al., 2009). This probably contributes to the decreased accuracies forMLC using multi-source data in the study.

Compared to the MLC and other parametric statistical approaches,the decision-tree method does not assume any specific data form ordistribution (Liu et al., 2008; Quinlan, 1993; Tooke et al., 2009) andtherefore can easily incorporate multi-source data of various types intothe analysis. This is the possible reason why DT_SIM_TM and DT_SIM_OLIexhibited the higher accuracy values in comparison with those fromMLC. In addition, given the consistent nature of mangrove forest ha-bitats, the decision-tree method using multi-source data can be applieduniversally for mapping mangrove forests. However, the threshold va-lues used in the decision tree analysis in this study were selected ac-cording to the calibration data. It should be noted that the thresholdvalues derived for the DEM, NDVIL, NDMIH and NDVIL·NDMIH, maychange with different study areas and different sensors (i.e., possessingdifferent spectral bands. For instance, feature thresholds may vary dueto the tidal type, difference in tide level during image acquisitions,mangrove species, and so on.

Fig. 7. Decision tree for mangrove forest identification using multi-tidal OLI data (i.e. DT_SIM_OLI).

X. Zhang et al. Int J Appl  Earth Obs Geoinformation 62 (2017) 201–214

210

The mangrove forest habitats are characterized by wetland back-ground, forest vegetation and periodic variation of the tide level. In thisstudy, the DT_SIM_TM and DT_SIM_OLI combinations achieved thehighest classification accuracies, indicating that the characteristics of

mangrove forest habitats can be more effectively depicted by NDVI,NDMI and NDVIL·NDMIH.

In addition, the classification approaches using multi-tidal re-motely-sensed data have higher classification accuracies of mangrove

Fig. 8. Mangrove forest distribution maps in thestudy area produced from: (A) DT_SIL; (B)DT_SIM_TM; (C) ML_DEM_L; (D) ML_DEM_M; (E)ML_L; (F) ML_M; and (G) DT_SIM_OLI. Each man-grove classification is superimposed on Landsat TM(A-F) or OLI (G) data acquired during low tide.

X. Zhang et al. Int J Appl  Earth Obs Geoinformation 62 (2017) 201–214

211

forests, regardless of classification method, e.g., DT_SIM_TM,DT_SIM_OLI and ML_M (Table 8). The decision-tree procedure requiresfewer features and exhibits higher classification accuracies compared tothe MLC. However, the threshold values of DT_SIM_TM and DT_SIM_OLIneed to be selected based on the calibration data. Therefore, it can bestated that the decision tree-based procedure using multi-source data isa more appropriate approach for classifying mangrove forest con-sidering the number of input features required and classification ac-curacy.

5. Conclusions

In this paper, a decision-tree-based procedure has been developed tomap mangrove forests using multi-tidal remotely-sensed data. Giventhat this procedure is able to incorporate various forms and distribu-tions of spatial data that do not necessarily satisfy the strict require-ments of parametric statistical classifiers, it can take incorporate arange of data characterizing the distinctive environmental and re-flectance characteristics of mangrove forests. The decision-tree proce-dure is based on three decision rules employing input features ofNDMIH, NDVIL, NDVIL·NDMIH and elevation (i.e., DEM). The classifi-cation results from the proposed procedure were compared to thosefrom the traditional MLC approach using different feature combina-tions.

It was found that the spectral signatures of mangrove forests wouldsignificantly change on images acquired at different tide levels.Therefore, mangrove forests cannot be accurately mapped using single-tidal remotely-sensed data, especially when collected at high tide. Thetidal effect is one of the primary features of mangrove forest habitats.This study demonstrates that the application of multi-tidal remotely-

sensed data results in higher classification accuracies over single-tidaldata for the decision-tree method as well as the MLC approach.Mangrove forests are also characterized by low elevations. However,the incorporation of elevation data with reflectance data did not im-prove the classification accuracy when MLC was applied, a result at-tributed to the elevation data not exhibiting a normal distribution.Meanwhile, the decision-tree method achieved the highest classificationaccuracy when incorporating multi-tidal images of NDMI, NDVI andDEM data. This approach (i.e., incorporating multi-source data and thedecision-tree method) was repeated (and successfully validated) withLandsat 8 OLI data collected in 2013. Therefore, when using remotely-sensed data and ancillary spatial data to map mangrove forests, thedecision-tree method appears to be superior to MLC. Meanwhile, thisalso indicates that vegetation indices such as NDVI, NDMI andNDVIL·NDMIH have a strong capacity for describing the unique char-acteristics of mangrove forests.

Acknowledgements

The research work was funded by the Open Fund of Key Laboratoryof Meteorology and Ecological Environment of Hebei Province (GrantNo. Z201607Y), the National Key Research and Development Programof China (Project Ref. No. 2016YFB0501501), China PostdoctoralScience Foundation (Grant No. 2017M610338), the National NaturalScience Foundation of China (Grant No. 41201461) and the JiangsuGovernment Scholarship. We are grateful to the editor and the anon-ymous reviewers for their constructive comments that helped to im-prove this manuscript. The Landsat 5 TM data, Landsat 8 OLI data andSRTM DEM data were provided by International Scientific & TechnicalData Mirror Site, Computer Network Information Center, Chinese

Table 7Comparisons of mangrove forest cover by different classification methods.

Type of classification

DT_SIL DT_SIM_TM ML_DEM_L ML_DEM_M ML_L ML_M DT_SIM_OLI

Area (ha) 1008.5 1198.3 3717.1 1880.2 1089.4 1183.5 1179.5Area percentage 1.13% 1.34% 4.17% 2.11% 1.22% 1.33% 1.32%

Table 8Accuracy measures for seven classifications of TM and OLI data.

Classification Approach Actual Category Producer's Accuracy User's Accuracy Overall Accuracy Kappa

Non-mangrove Mangroves

DT_SIL Classified Category Non-mangrove 389 28 92.6% 93.2% 90.4% 0.78Mangroves 31 168 85.7% 84.4%

DT_SIM_TM Non-mangrove 390 16 92.9% 96.1% 92.5% 0.831Mangroves 30 180 91.8% 85.7%

DT_SIM_OLI Non-mangrove 300 14 95.2% 95.5% 93.8% 0.858Mangroves 15 136 90.7% 90.1%

ML_DEM _L Non-mangrove 299 13 71.2% 95.8% 78.2% 0.563Mangroves 121 183 93.3% 60.2%

ML_DEM _M Non-mangrove 353 6 84.1% 98.3% 88.1% 0.748Mangroves 67 190 96.9% 73.9%

ML_L Non-mangrove 380 21 90.5% 94.8% 90.1% 0.778Mangroves 40 175 89.3% 81.4%

ML_M Non-mangrove 382 15 91.0% 96.2% 91.4% 0.808Mangroves 38 181 92.4% 82.6%

Table 9Comparison of DT_SIM_TM and DT_SIM_OLI for mapping mangrove forest.

Classification Approach Producer's Accuracy of mangroves User's Accuracy of mangroves Overall Accuracy Kappa Area of mangroves Area percentage

DT_SIM_TM 91.80% 85.70% 92.50% 0.831 1198.3 1.34%DT_SIM_OLI 90.70% 90.10% 93.80% 0.858 1179.5 1.32%

X. Zhang et al. Int J Appl  Earth Obs Geoinformation 62 (2017) 201–214

212

Academy of Sciences (http://www.gscloud.cn). This research wasconducted while the principal author was a Visiting Professor in theDepartment of Geography and Planning, Queen’s University, Kingston,Ontario, Canada.

References

Alongi, D.M., 2008. Mangrove forests Resilience, protection from tsunamis, and responsesto global climate change. Estuarine Coast. Shelf Sci. 76, 1–13.

Alongi, D.M., 2009. Introduction in the Energetics of Mangrove Forests. Springer Scienceand Business Media BV, New York, USA.

Andriamparany, R., Francois, F., 2010. Dynamics of mangrove forests in the MangokyRiver Delta, Madagascar, under the influence of natural and human factors. For. Ecol.Manag. 259, 1161–1169.

Bird, M., Chua, S., Fifield, L.K., Teh, T.S., Lai, J., 2004. Evolution of the sungei Buloh-Kranji mangrove coast, Singapore. Appl. Geogr. 24, 181–198.

Blasco, F., Saemger, P., Janodet, E., 1996. Mangroves as indicators of coastal change.Catena 27, 167–178.

Chander, G., Markham, B.L., Helder, D.L., 2009. Summary of current radiometric cali-bration coefficients for landsat MSS, TM, ETM+, and EO-1 ALI sensors. Remote Sens.Environ. 113, 893–903.

Chavez, P.S., 1986. An improved dark-object subtraction technique for atmosphericscattering correction of multispectral data. Remote Sens. Environ. 24, 459–479.

Chavez, P.S., 1996. Image-based atmospheric corrections −revisited and improved.Photogramm. Eng. Remote Sens. 62, 1025–1036.

Chen, D.Y., Huang, J.F., Jackson, T.J., 2005. Vegetation water content estimation for cornand soybeans using spectral indices derived from MODIS near- and short-wave in-frared bands. Remote Sens. Environ. 98, 225–236.

Chen, C.F., Son, N.T., Chang, N.B., Chen, C.R., Chang, L.Y., Valdez, M., Centeno, G.,Thompson, C.A., Aceituno, J.L., 2013. Multi-decadal mangrove forest change detec-tion and prediction in Honduras central america, with landsat imagery and a markovchain model. Remote Sens. 5, 6408–6426.

Cibula, W.G., Zetka, E.F., Rickman, D.L., 1992. Response of The matic Mapper bands toplant water stress. Int. J. Remote Sens. 13, 1869–1880.

Cornforth, W.A., Fatoyinbo, T.E., Freemantle, T.P., Pettorelli, N., 2013. Advanced landobserving satellite phased array type L-band SAR (ALOS PALSAR) to inform theconservation of mangroves: sundarbans as a case study. Remote Sens. 5, 224–237.

Du, X., Guo, H., Fan, X., Zhu, J., Yan, Z., Zhan, Q., 2013. Vertical accuracy assessment ofSRTM and ASTER GDEM over typical regions of China using ICESat/GLAS. EarthSci.—J. China Univ. Geosci. 38, 887–897 (in Chinese with English abstract).

Everitt, J.H., Yang, C., Judd, F.W., Summy, K.R., 2010. Use of archive aerial photographyfor monitoring black mangrove populations. J. Coast. Res. 26, 649–653.

FGDC, 1992. Application of satellite data for mapping and monitoring wetlands. In:Technical Report 1, Wetlands Subcommittee. Federal Geographic Data Committee(FGDC). Washington, DC, USA..

Farnsworth, E.J., Ellison, A.M., 1997. The global conservation status of mangroves.Ambio 26, 328–334.

Ferreira, M.A., Andrade, F., Banderiba, S.O., Cardoso, P., Nogueira Mendes, R., Paula, J.,2009. Analysis of cover change (1995–2005) of Tanzania/Mozambique trans-boundary mangroves using Landsat imagery. Aquat. Conserv. 19, 38–45.

Friedl, M.A., Brodley, C.E., 1997. Decision tree classification of land cover from remotelysensed data. Remote Sens. Environ. 61, 399–409.

Gao, J., 1998. Hybrid method toward accurate mapping of mangroves in a marginalhabitat from SPOT Multispectral data. Int. J. Remote Sens. 19, 1887–1899.

Gao, J., 1999. A comparative study on spatial and spectral resolutions of satellite data inmapping mangrove forests. Int. J. Remote Sens. 20, 2823–2833.

Giri, C., Muhlhausen, J., 2008. Mangrove forest distribution and dynamics in Madagascar(1975–2005). Sensors 8, 2104–2117.

Giri, C., Pengra, B., Zhu, Z., Singh, A., Tieszen, L., 2007. Monitoring mangrove forestdynamics of the Sundarbans in Bangladesh and India using multi-temporal satellitedata from 1973 to 2000. Estuar. Coast. Shelf Sci. 73, 91–100.

Giri, C., Zhu, Z., Tieszen, L.L., Singh, A., Gillette, S., Kelmelis, J.A., 2008. Mangrove forestdistributions and dynamics (1975–2005) of the tsunami-affected region of Asia. J.Biogeogr. 35, 519–528.

Giri, C., Ochieng, E., Tieszen, L.L., Zhu, Z., Singh, A., Loveland, T., Masek, J., Duke, N.,2011a. Status and distribution of mangrove forests of the world using earth ob-servation satellite data. Global Ecol. Biogeogr. 20, 154–159.

Giri, C., Long, J., Tieszen, L., 2011b. Mapping and monitoring Louisiana’s mangroves inthe aftermath of the 2010 Gulf of Mexico oil spill. J. Coast. Res. 27, 1059–1064.

Green, E.P., Mumby, P.J., Edwards, A.J., Clark, C.D., Ellis, A.C., 1997. Estimating leafarea index of mangroves from satellite data. Aquat. Bot. 58, 11–19.

Green, E.P., Mumby, P., Clark, C.D., Mumby, P.J., : Edwards, A.J., Ellis, A.C., 1998.Remote sensing techniques for mangrove mapping. Int. J. Remote Sens. 19, 935–956.

Heenkenda, M.K., Joyce, K.E., Maier, S.W., Bartolo, R., 2014. Mangrove species identi-fication: comparing world view-2 with aerial photographs. Remote Sens. 6,6064–6088.

Heumann, B.W., 2011. Object-based classification of mangroves using a hybrid decisiontree—support vector machine approach. Remote Sens. 3, 2440–2460.

Howari, F.M., Jordan, B.R., Bouhouche, N., Wyllie-Echeverria, S., 2009. Field and remote-sensing assessment of mangrove forests and seagrass beds in the northwestern part ofthe United Arab Emirates. J. Coast. Res. 25, 48–56.

Huang, J.F., Chen, D.Y., Cosh, M.H., 2009. Sub-pixel reflectance unmixing in estimatingvegetation water content and dry biomass of corn and soybeans cropland usingnormalized difference water index (NDWI) from satellites. Int. J. Remote Sens. 30,

2075–2104.Hunt, E.R., Rock, B.N., Nobel, P.S., 1987. Measurement of leaf relative water content by

infrared reflectance. Remote Sens. Environ. 22, 249–435.Jensen, J.R., Lin, H., Yang, X., Ramset, E., Davis, B.A., Thoemke, C.W., 1991. The mea-

surement of mangrove characteristics in south-west Florida using SPOT multispectraldata. Geocartogr. Int. 2, 13–21.

Jensen, J.R., 2005. Introductory Digital Image Processing: A Remote Sensing Perspective,third edition. Prentice Hall, Inc, NJ, USA.

Jones, T.G., Glass, L., Gandhi, S., Ravaoarinorotsihoarana, L., Carro, A., Benson, L.,Ratsimba, H.R., Giri, C., Randriamanatena, D., Cripps, G., 2016. Madagascar’smangroves: quantifying nation-wide and ecosystem specific dynamics, and detailedcontemporary mapping of distinct ecosystems. Remote Sens. 8 (2), 106. http://dx.doi.org/10.3390/rs8020106.

Kamal, M., Phinn, S., 2011. Hyperspectral data for mangrove species mapping: a com-parison of pixel-based and object-based approach. Remote Sens. 3, 2222–2242.

Kanniah, K.D., Sheikhi, A., Cracknell, A.P., Goh, H.C., Tan, K.P., Ho, C.S., Rasli, F.N.,2015. Satellite images for monitoring mangrove cover changes in a fast growingeconomic region in southern peninsular Malaysia. Remote Sens. 7, 14360–14385.

Kuenzer, C., Bluemel, A., Gebhardt, S., Quoc, T.V., Dech, S., 2011. Remote sensing ofmangrove ecosystems: A review. Remote Sens. 3, 878–928.

Liang, S., Liu, J., Liang, M., 2004. Ecological study on the mangrove communities inBeilun Hekou national nature reserve. J. Guangxi Normal Univ. 22 (2), 70–76 (inChinese with English abstract).

Lin, P., Fu, Qin., 1995. Environment, Ecology and Economic Utilization of Mangrove inChina. Higher Education Press, Beijing, China. (in Chinese with English abstract).

Liu, K., Li, X., Shi, X., Wang, S., 2008. Monitoring mangrove forest changes using remotesensing and GIS data with decision-tree learning. Wetlands 28, 336–346.

Lovelock, C.E., Cahoon, D.R., Friess, D.A., Guntenspergen, G.R., Krauss, K.W., Reef, R.,Rogers, K., Saunders, M.L., Sidik, F., Swales, A., Saintilan, N., Thuyen, L.X., Triet, T.,2015. The vulnerability of Indo-Pacific mangrove forests to sea-level rise. Nature 526,559–563.

Luo, J., Ma, R., Duan, H., Hu, W., Zhu, J., Huang, W., Lin, C., 2014. A new method formodifying thresholds in the classification of tree models for mapping aquatic vege-tation in taihu lake with satellite images. Remote Sens. 6 (8), 7442–7462.

Luo, J., Li, X., Ma, R., Li, F., Duan, H., Hu, W., Qin, B., Zhu, J., Huang, W., 2016. Applyingremote sensing techniques to monitoring seasonal and interannual changes of aquaticvegetation in Taihu Lake, China. Ecol. Indic. 60, 503–513.

Luo, J., Duan, H., Ma, R., Jin, X., Li, F., Hu, W., Shi, K., Huang, W., 2017. Mapping speciesof submerged aquatic vegetation with multi-seasonal satellite images and consideringlife history information. Int. J. Appl. Earth Obs. Geoinf. 57, 154–165.

Mas, J.F., 2004. Mapping land use/cover in a tropical coastal area using satellite sensordata, GIS and artificial neural networks. Estuar. Coast. Shelf Sci. 59, 219–230.

Murray, M.R., Zisman, S.A., Furley, P.A., Munro, D.M., Gibson, J., Ratter, J., Bridgewater,S., Minty, C.D., Place, C.J., 2003. The mangroves of Belize: part 1 distribution,composition and classification. For. Ecol. Manag. 174, 265–279.

National Marine Data and information Service,, 2005. Tide Tables in 2006: Vol. 3 FromTaiwan Straits to the Beibu Gulf. Cartographic Publishing House of Shandong, Jinan,China (in Chinese with English abstract).

National Marine Data and information Service,, 2012. Tide Tables in 2013: Vol. 3 FromTaiwan Straits to the Beibu Gulf. China Ocean Press, Beijing, China (in Chinese withEnglish abstract).

Ozesmi, S.L., Bauer, M.E., 2002. Satellite remote sensing of wetlands. Wetl. Ecol. Manag.10, 381–402.

Pal, M., Mather, P.M., 2003. An assessment of the effectiveness of decision tree methodsfor land cover classification. Remote Sens. Environ. 86, 554–565.

Paling, E.I., Kobryn, H.T., Humphreys, G., 2008. Assessing the extent of mangrove changecaused by Cyclone Vance in the eastern Exmouth Gulf, northwestern Australia.Estuar. Coast. Shelf Sci. 77, 603–613.

Panigrahy, Sushma, Murthy, T.V.R., Patel, J.G., Singh, T.S., 2012. Wetlands of India:inventory and assessment at 1: 50,000 scale using geospatial techniques. Curr. Sci.102, 852–856.

Pettorelli, N., Vik, J.O., Mysterud, A., Gaillard, J.-M., Tucker, C.J., Stenseth, N.C., 2005.Using the satellite-derived NDVI to assess ecological responses to environmentalchange. Trends Ecol. Evol. 20, 503–510.

Prasad, P.R.C., Reddy, C.S., Rajan, K.S., Raza, S.H., Dutt, C.B.S., 2009. Assessment oftsunami and anthropogenic impacts on the forest of the North Andaman Islands,India. Int. J. Remote Sens. 30, 1235–1249.

Pu, R., Bell, S., Meyer, C., Baggett, L., Zhao, Y., 2012. Mapping and assessing seagrassalong the western coast of Florida using Landsat TM and EO-1 ALI/Hyperion imagery.Estuar. Coast. Shelf Sci. 115, 234–245.

Quinlan, J.R., 1986. Induction of decision trees. Mach. Learn. 1, 81–106.Quinlan, J.R., 1993. C4.5 Programs for Machine Learning. Morgan Kaufmann Publishers,

Inc., San Mateo, USA.Ramírez-García, P., López-Blanco, J., Ocaña, D., 1998. Mangrove vegetation assessment

in the Santiago River Mouth, Mexico, by means of supervised classification usingLandsat TM imagery. For. Ecol. Manag. 105, 217–229.

Rao, B.R., Dwivedi, R.S., Kushwaha, S.P., Bhattacharya, S.N., Anand, J.B., Dasgupta, S.,1999. Monitoring the spatial extent of coastal wetland using ERS-1 SAR data. Int. J.Remote Sens. 20, 2509–2517.

Richards, J.A., Jia, X.P., 2006. Remote Sensing Digital Image Analysis: an Introduction,4th ed. Springer, Berlin, German.

Rouse, J.W., Haas R.H., Schell J.A., Deering D.W., 1973. Monitoring vegetation systems inthe Great Plains with ERTS, Third ERTS Symposium, NASA SP-351 I, 309–317.

Seto, K.C., Fragkias, M., 2007. Mangrove conversion and aquaculture development inVietnam: a remote sensing-based approach for evaluating the Ramsar Convention onWetlands. Global Environ. Change 17, 486–500.

X. Zhang et al. Int J Appl  Earth Obs Geoinformation 62 (2017) 201–214

213

http://www.gscloud.cnhttp://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0005http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0005http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0010http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0010http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0015http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0015http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0015http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0020http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0020http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0025http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0025http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0030http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0030http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0030http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0035http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0035http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0040http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0040http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0045http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0045http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0045http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0050http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0050http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0050http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0050http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0055http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0055http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0060http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0060http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0060http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0065http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0065http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0065http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0070http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0070http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0075http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0075http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0075http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0080http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0080http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0085http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0085http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0085http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0090http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0090http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0095http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0095http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0100http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0100http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0105http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0105http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0110http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0110http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0110http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0115http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0115http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0115http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0120http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0120http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0120http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0125http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0125http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0130http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0130http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0135http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0135http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0140http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0140http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0140http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0145http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0145http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0150http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0150http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0150http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0155http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0155http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0155http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0155http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0160http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0160http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0165http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0165http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0165http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0170http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0170http://dx.doi.org/10.3390/rs8020106http://dx.doi.org/10.3390/rs8020106http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0180http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0180http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0185http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0185http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0185http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0190http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0190http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0195http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0195http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0195http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0200http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0200http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0205http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0205http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0210http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0210http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0210http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0210http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0215http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0215http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0215http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0220http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0220http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0220http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0225http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0225http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0225http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0230http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0230http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0235http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0235http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0235http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0240http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0240http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0240http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0245http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0245http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0245http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0250http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0250http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0255http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0255http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0260http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0260http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0260http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0265http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0265http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0265http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0270http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0270http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0270http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0275http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0275http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0275http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0280http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0280http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0280http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0285http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0290http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0290http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0295http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0295http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0295http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0300http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0300http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0300http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0305http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0305http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0315http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0315http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0315

Shapiro, A.C., Trettin, C.C., Küchly, H., Alavinapanah, S., Bandeira, S., 2015. The man-groves of the zambezi delta: increase in extent observed via satellite from 1994 to2013. Remote Sens. 7, 16504–16518.

Thu, P.M., Populus, J., 2007. Status and changes of mangrove forest in Mekong Delta:case study in Tra Vinh, Vietnam. Estuar. Coast. Shelf Sci. 71, 98–109.

Tooke, T.R., Coops, N.C., Goodwin, N.R., Voogt, J.A., 2009. Extracting urban vegetationcharacteristics using spectral mixture analysis and decision tree classifications.Remote Sens. Environ. 113, 398–407.

Tran, H., Tran, T., Kervyn, M., 2015. Dynamics of land cover/land use changes in themekong delta, 1973–2011: a remote sensing analysis of the tran van thoi district, Camau province, Vietnam. Remote Sens. 7, 2899–2925.

Vaiphasa, C., Andrew, S.K., Willem, F.B., 2006. A post-classifier for mangrove mappingusing ecological data. ISPRS J. Photogr. Remote Sens. 61, 1–10.

Vo, Q.T., Oppelt, N., Leinenkugel, P., Kuenzer, C., 2013. Remote sensing in mappingmangrove ecosystems — an object-based approach. Remote Sens. 5, 183–201.

Wang, L., Sousa, W., Gong, P., 2004. Integration of object-based and pixel-based classi-fication for mangrove mapping with IKONOS imagery. Int. J. Remote Sens. 25,5655–5668.

Wilson, E.H., Sader, S.A., 2002. Detection of forest harvest type using multiple dates ofLandsat TM imagery. Remote Sens. Environ. 80, 385–396.

Xiao, H., Zeng, H., Zan, Q., Bai, Y., Cheng, H., 2007. Decision tree model in extraction ofmangrove community information using hyperspectral image data. J. Remote Sens.11, 531–537 (in Chinese with English abstract).

Xu, M., Watanachaturaporn, P., Varshney, P.K., Arora, M.K., 2005. Decision tree re-gression for soft classification of remote sensing data. Remote Sens. Environ. 97,322–336.

Yu, T., Li, X., Zhang, Y., Zhao, F., Gu, X., Zhu, L., Wang, P., Min, X., 2005. Comparison ofthe influence factors on NDVI for CCD camera and WFI imager on CBERS-02. Sci.China Ser. E 100–115.

Zhang, Q., Sui, S., 2001. The mangrove wetland resources and their conservation inChina. J. Nat. Resourc. 16, 28–36 (in Chinese with English abstract).

Zhang, X., Tian, Q., 2013. A mangrove recognition index for remote sensing of mangroveforest from space. Curr. Sci. 105, 1149–1155.

Zhang, X., Zhou, J., Wei, Y., Zhu, Y., 2013. Remote sensing identification of mangroveforest combined tidal level information. J. Nanjing Univ. Inf. Scie. Technol.: Nat. Sci.Ed. 5, 501–507 (in Chinese with English abstract).

Zhao, B., Guo, H., Yan, Y., Wang, Q., Li, B., 2008. A simple waterline approach for ti-delands using multi-temporal satellite images: a case study in the Yangtze Delta.Estua. Coast. Shelf Sci. 77, 134–142..

X. Zhang et al. Int J Appl  Earth Obs Geoinformation 62 (2017) 201–214

214

http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0320http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0320http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0320http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0325http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0325http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0330http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0330http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0330http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0335http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0335http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0335http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0340http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0340http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0345http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0345http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0350http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0350http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0350http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0355http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0355http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0360http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0360http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0360http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0365http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0365http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0365http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0370http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0370http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0370http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0375http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0375http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0380http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0380http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0385http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0385http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0385http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0390http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0390http://refhub.elsevier.com/S0303-2434(17)30136-8/sbref0390

Mapping mangrove forests using multi-tidal remotely-sensed data and a decision-tree-based procedureIntroductionStudy areaData and methodologyDatasetsImage pre-processingSpectral indices calculationCalibration and validation dataSpectral analysis of land cover at high and low tideDecision-tree algorithmComparison between classification approaches

Results and discussionSpectral characteristics of typical land covers in the study areaSpectral indices for typical land cover typesDecision-tree algorithmComparison of classificationsClassification resultsAccuracy assessmentComparison of classification approaches

ConclusionsAcknowledgementsReferences