Multi-temporal change detection of seagrass beds ...

Ecological Informatics 35 (2016) 43–54

Contents lists available at ScienceDirect

Ecological Informatics journal homepage: www.elsevier.com/locate/ecolinf

Multi-temporal change detection of seagrass beds using integrated Landsat TM/ETM +/OLI imageries in Cam Ranh Bay, Vietnam Chi-Farn Chen a,b, Va-Khin Lau a,c,⁎, Ni-Bin Chang d, Nguyen-Thanh Son b, Phuoc-Hoang-Son Tong c, Shou-Hao Chiang b a

Department of Civil Engineering, National Central University, Taoyuan City 32001, Taiwan Center for Space and Remote Sensing Research, National Central University, Taoyuan City 32001, Taiwan Institute of Oceanography, Vietnam Academy of Science and Technology, 01 Cau Da, Nha Trang City, Khanh Hoa Province, Vietnam d Department of Civil, Environmental, and Construction Engineering, University of Central Florida, Orlando, FL 32816, USA b c

a r t i c l e

i n f o

Article history: Received 24 December 2015 Received in revised form 23 July 2016 Accepted 24 July 2016 Available online 26 July 2016 Keywords: Seagrass beds Landsat data Linear mixture model Change analysis Cam Ranh Bay

a b s t r a c t Seagrass beds comprise a unique marine ecosystem that acts as a biofilter in marine environments and serves as a spawning ground and nursery for various species of fish. Long-term monitoring of seagrass beds is critical to understanding the dynamic relationships between the ecosystems and the stresses from natural systems and society. This study investigated temporal changes of seagrass beds in Cam Ranh Bay (CRB), Vietnam using multitemporal Landsat data from 1996 to 2015. The data were processed through 5 main steps including: (1) image preprocessing to convert Landsat data to the top of atmosphere reflectance (TOA) and to correct atmospheric effects, (2) water column correction to eliminate effects on remotely sensed data of aquatic environments, (3) image classification using a linear mixed model, (4) accuracy assessment using the ground reference data, and (5) change detection of seagrass beds. The classification results compared with the ground reference data indicated that the overall accuracies and Kappa coefficients were higher than 91.7% and 0.8, respectively, in all cases. From 1996 to 2015, the total area of seagrass beds had declined by approximately 25% (66 ha), mainly attributed to coastal development and infrastructure construction. © 2016 Elsevier B.V. All rights reserved.

1. Introduction Seagrasses are flowering plants commonly distributed in shallow water along coastlines, estuaries, bays, and lagoons. They serve as direct and indirect food for marine animals, including fish, dugong, and green turtles (Dai, 2011; Duarte, 2002; Fortes, 1990; Heck et al., 2003; Short et al., 2001). Seagrass beds provide the habitat for marine animals, stabilize sediments, and prevent soil erosion (Cabaço et al., 2008; Dai, 2011; Dai et al., 1999; Fortes, 1990), and their leaves can act as a biofilter by absorbing nutrients from coastal run-off (Spalding et al., 2014; Stapel et al., 1996; Vonk et al., 2008). They can also store organic carbon at levels two times higher than typical terrestrial forests per each km2 (Fourqurean et al., 2012). In recent years, the seagrass loss rate has been increasing due to impacts of economic development, infrastructure, aquaculture farm constructions, urbanization, dredging, and turbidity and eutrophication in many locations such as, in the Gulf of Mexico, Indonesia, Philippines, Singapore, Thailand and Vietnam (Duarte, 2002; Erftemeijer and Robin Lewis Iii, 2006; Green and Short, 2003; Vo et al., 2013; Walker, 1996; Walker and McComb, 1992). It is thus necessary to monitor seagrass beds for environmental management and conservation.

http://dx.doi.org/10.1016/j.ecoinf.2016.07.005 1574-9541/© 2016 Elsevier B.V. All rights reserved.

Indicators of seagrass distribution are critically important for monitoring coastal ecosystem health. The presence/absence and spatial distribution of seagrasses are commonly used ecological indicators representing the status of seagrass ecosystems and the response to surrounding environments at the landscape scale. These response patterns or changes may be intimately tied to anthropogenic impacts, such as eutrophication, land-use changes, coastal development, boating, dredging, and agriculture. Cam Ranh Bay (CRB) is a region in Vietnam that has a large meadow of seagrasses with a high diversity of species. Because of aquaculture activities and socioeconomic development, however, seagrasses in the region have been seriously degraded; approximately 20–30% of the total area of seagrass was lost between 1998 and 2002 (Dai et al., 2002). Studies indicated that Enhalus acoroides, a dominant seagrass species usually flowering and fruiting during July to August, almost disappeared in this region because of anthropogenic activities (Dai et al., 2002). Despite these observations, no formal ecological study of seagrass monitoring has been conducted in the region. Thus, understanding the spatiotemporal changes of seagrasses is critical to provide ecologists and economists in the region with information to improve sustainable management strategies for marine ecosystems.

44

C.-F. Chen et al. / Ecological Informatics 35 (2016) 43–54

Mapping seagrass beds in the study region is traditionally implemented through costly and time-consuming field surveys limited to small areas in either shallow (b10 m) or deeper (N 10 m) waters by snorkeling or diving using 50 × 50 cm quadrat frames, cover sheets, and waterproof cameras (Green et al., 1996, 2000; Komatsu et al., 2003a; Komatsu et al., 2003b; Mumby et al., 1999; Sagawa et al., 2010). Remote sensing technologies such as aerial photography and satellites have been an indispensable tool for marine ecosystem monitoring, including change detection of seagrasses, because they can acquire data over larger regions. Mapping seagrass distributions by remote sensors could be influenced by various contributions from the atmosphere, water column, and sea bottom. Because the bottom signal is not always distinct, the signal received by remote sensors may be compromised by bottom features that appear as variations in the radiance directed toward the sensor. For example, seagrasses present in shallow waters with a light sandy bottom are distinguished easily by remote sensing images (Andréfouët et al., 2001; Andréfouët et al., 2003; Hochberg et al., 2003), yet a wide dynamic range of colors are required to distinguish a dark silty bottom with mixed seagrasses, mussel beds, and other covers types in deeper waters (Botha et al., 2013; Lyzenga, 1981). Images of turbid and deep water with other dark features such as mussel beds, stones, or macroalgae are much more difficult to interpret than clear and shallow water environments where seagrasses grow in dense beds and constitute the only dark features on a sandy bottom (Andréfouët et al., 2001; Andréfouët et al., 2003; Hochberg et al., 2003). Low spatial resolution satellite images can be used only for macroscale mapping to catalogue the presence and absence of seagrass beds or coarsely assess the area distribution of seagrass beds (Ferguson and Korfmacher, 1997; Pu et al., 2014; Wabnitz et al., 2008). Many mapping methods use high spatial resolution satellite images to map macroalgae and seagrasses in the intertidal regions at a scale of 2 to 20 m pixel size to investigate the distribution of seagrass beds for change detection or to estimate the biomass (Mumby and Edwards, 2002; Phinn et al., 2008; Sagawa et al., 2010; Valle et al., 2015). The use of high spatial and spectral resolutions of hyperspectral satellite images for seagrass mapping usually produces results with wide coverage and are easily georectified, enabling a photo-interpreter to differentiate between objects with colors that appear identical (Table 1). Even high-resolution satellite images have several limitations: (1) narrow coverage of spectral bands in hyperspectral remote sensing; (2) limited temporal resolution; (3) high photographic distortion; (4) low radiometric resolution; (5) cloud contamination (i.e., in optical remote sensing); (6) mapping inaccuracies of seagrass meadows caused by the growth of epiphytes, seagrasses cover density, varying water depth, and changing optical properties of overlying water driven by seagrass die-back in anoxia and high temperature environment; (7) interpretation difficulty in deep and turbid waters, especially in low light or when water transparency is disturbed by high nutrient concentrations; (8) highly variable sun-glint reflection from all directions in image (i.e., especially in air-borne remote sensing); (9) errors due to converting analogue air-borne photos to digital images, and (10) high costs when high spatial and spectral resolutions are required (Mumby et al., 1999). Nevertheless, multispectral Landsat data with 30 m spatial resolution are a good candidate for this monitoring purpose over a multi-decadal scale because they are free of charge, and historic

archives from Landsat TM to Landsat ETM + and to Landsat OLI have existed since the 1970s. Integration of these three satellite sensors without the need for extra bias corrections in the cross-sensor data merging allows homogeneous multisensor image processing to support a longterm seagrass monitoring mission. Sophisticated feature extraction and content-based mapping are essential to retrieving useful information in many circumstances, especially when interpretation becomes difficult in deep and turbid waters, triggering formulation of additional semiempirical or empirical feature extraction models. A number of methods have been developed for feature extraction and for seagrass mapping, including principal component analysis (PCA) (Ferguson and Korfmacher, 1997; Pasqualini et al., 2005), normalized difference vegetation index (NDVI) (Barillé et al., 2010), and leaf area index (LAI) combined with additional in-situ optical data of water leaving radiance and attenuation coefficient (Yang et al., 2011). These studies did not consider effects of the water column, however, and ignoring this processing step could reduce the mapping accuracy by approximately 22% and 17% when using an airborne hyperspectral imaging device such as Compact Airborne Spectrographic Imager (CASI) and satellite data, respectively (Mumby et al., 1998). The depth invariant index (DII) (Lyzenga, 1981) and bottom reflectance index (BRI) (Sagawa et al., 2010) are two commonly used water column correction methods based on the bottom reflectance equation that considers the reflectance through water decreasing exponentially with an increase in water depth. The DII used to indicate sea bottom types without using bathymetry data is based on the unchanged characteristic of the y-intercept value (i.e., invariant index) of relations between two visual bands associated with the water depth on the same substrate (Lyzenga, 1981). The BRI uses bathymetry data of a substrate to obtain attenuation coefficients from a reflectance function using exponential regression analysis; the coefficients are then used to obtain the BRI, which is then used to indicate the bottom type (Sagawa et al., 2010). The classification of seagrass beds using BRI could improve the overall accuracy by 21–36% compared to DII (Sagawa et al., 2010; Sagawa et al., 2012). The main objective of this study was to investigate the potential use of multi-temporal Landsat data with appropriate water column corrections for seagrass mapping and conduct multi-temporal change detection of five key seagrass beds in CRB of Vietnam (i.e., five key study regions) during the periods 1996–2001, 2001–2005, 2005–2010, 2010–2013, and 2013–2015. The case study allows us to determine if the integrated Landsat data could support a holistic seagrass mapping over the anticipated spatial and/or temporal time scales. To overcome water column impact on reflectance, we used the BRI to remove the water column effects followed by the linear mixed model (LMM) to quantify the abundance fraction of seagrass beds in each pixel (Gokul et al., 2014; Philpot, 1989; Schroeder et al., 2006). This type of multitemporal change detection may help answer the following questions: 1) which key study region suffered from the biggest net loss of the seagrass bed over the five study periods; and 2) can the new seagrass bed outweigh the loss of seagrass bed over the five study periods? The quantitative information achieved from this study might be useful for resources managers to formulate long-term management strategies of seagrass ecosystems in the region.

Table 1 Accuracy consideration for mapping seagrass beds with satellite and air-borne sensors (Blakey et al., 2015; Lyons et al., 2012; Mumby et al., 1997; Pasqualini et al., 2005; Sagawa et al., 2010; Wabnitz et al., 2008). Type of images

Accuracy of the map (%) Coverage per scene (km)

Landsat TM/ETM

SPOT XS/5

CASI

Aerial photography

Space-borne high resolution multispectral images

Space-borne high resolution multispectral images

Air-borne high resolution hyperspectral images

Air-borne high resolution multispectral images

≤88 185 × 185

≤96 60 × 60

b90 Variable

≤90 Variable


2. Study area Cam Ranh Bay (CRB) is located in the south of Khanh Hoa province, Vietnam, between 11°48′–11°58′N and 109°06′–109°13′E, and is approximately 19 km long and 1.5–6 km wide (Fig. 1). The bay is connected to the sea and belongs to the tropical monsoon zone with two distinct seasons. The rainy season (September–December), contributing approximately 70–80% of annual precipitation, has an average sea surface temperature of 26.5 °C and salinity concentration of 26–33.3‰, whereas the dry season (January–August) has an average sea surface temperature and salinity concentration of 28.2 °C and 30.6–34.2‰, respectively (Long et al., 2011). The region is also affected by the tidal current, which plays a significant role in water exchange between the bay and sea. The sea surface in the bay is steady, with N80.6% area of current b 10 cm/s (Long et al., 2011). The total seagrass area was approximately 300 ha, classified into 7 species: Enhalus acoroides, Halophila ovalis, Halophila minor, Thalassia hemprichii, Halodule pinifolia, Halodule uninervis, and Ruppia maritima (Dai, 2011; Hoa, 2009; Vo et al., 2013) (Fig. 2). Enhalus acoroides is the dominant species, listed as endangered by the International Union for Conservation of Nature and Natural Resources (IUCN) (Dai et al., 2002; Hoa, 2009), and commonly appears in waters shallower than 5 m (Short and Waycott, 2010). The total area of seagrass beds in the region has been declining due to coastal development, dredging, marina developments, and impacts of climate change over the last two decades (Dai et al., 2002; Hoa, 2009; Pham et al., 2006; Vo et al., 2013). 3. Data collection Landsat data (path/row: 123/52) used in this study were acquired from the United States Geological Survey (USGS), including Landsat

Fig. 1. Map of the study region showing the reference sites where the ground reference seagrass data were collected and used for accuracy assessment of the mapping results.

45

TM (30 May 1996), Landsat ETM + (17 April 2001), Landsat TM (6 May 2005), Landsat TM (7 July 2010), Landsat OLI (29 June 2013), and Landsat OLI (16 April 2015). The Landsat TM data have 7 spectral bands, with a spatial resolution of 30 m for bands 1–5 and 7; TM band 6 (thermal infrared) is acquired at 120 m resolution. The Landsat ETM + data consist of 6 spectral bands with a spatial resolution of 30 m for bands 1–5 and 7; ETM+ band 6 (thermal infrared) is acquired at 60 m resolution, and band 8 (panchromatic band) has a resolution of 15 m. The Landsat OLI data have 8 spectral bands with a spatial resolution of 30 m for bands 1–7 and 9, 1 panchromatic band with spatial resolution of 15 m for band 8, and 2 thermal infrared bands with a spatial resolution of 100 m for bands 10 and 11. In this study, 6 spectral bands for each set of data were used for the classification: bands 1–5 and 7 for Landsat TM and ETM+, and bands 2–7 for Landsat OLI. The rationale for selecting these periodic images was driven by the developmental stages of shrimp ponds, gross output of breeding shrimps and fishes, and the data quality (i.e., cloud-free Landsat images) in the dry season when the seagrass is in a good condition. The number of shrimp ponds in the study region had increased up to 5560 ponds in 2001 (Dai et al., 2002), reaching 12,000 ponds in 2005 with an estimated production of 1865 tons gross output of breeding shrimps. The production slightly decreased to 1393 tons in 2010. The total gross output of breeding fishes had increased from approximately 755 tons in 2005 to 1071 tons in 2010 (Khanh Hoa Statistical Yearbook, 2011). Consequently, multi-temporal changes of seagrass distributions may be intimately tied to landuse/cover changes over the last 2 decades in this coastal region. The seagrass distribution could be an ecological indicator in a preliminary coastal sustainability assessment. The distribution of seagrass beds may generally reflect or imply seagrass cover density rather than a “shoot” density, which is defined as seagrass abundance in a coastal region. Yet, seagrass distribution cannot address colonization depth data, one of the best-known seagrass indicators of water quality that could forecast seagrass status in a coastal region. We also collected other ancillary data. (1) Seagrass distribution data were collected in 2009 (Fig. 3b) (Hoa, 2009) and were used to construct the ground reference data of seagrass beds for 1996, 2001, and 2005. The ground reference data used for accuracy assessment of the 2010 mapping results were generated using a field survey data collected in 2013. Two more reference datasets were collected in 2013 and 2015 using small fishing boat in low tide period with the aid of Global Positioning System (GPS) for position, cameras, and a sheet note at each site chosen (Fig. 1). (2) The sedimentation map was collected from the Nha Trang Institute of Oceanography, Vietnam. (3) N 4000 data points were collected in April 2013 using a single beam echo sounder Lawrance VP 1000 to acquire the bathymetric information needed for water column correction. The bathymetry map was generated from these depth data points using the natural neighbor interpolation method (Fig. 3a). In tropical regions, seagrass sampling is often carried out during both rainy and dry seasons. The rainy season in the study area is short (September–December), with precipitation (198–350 mm/month) concentrated in October–November; the mean sea surface temperature and salinity during this period are 26.5 °C and 31.9‰, respectively (Long et al., 2011). Although Enhalus acoroides is dominant during this period, investigating changes of seagrass beds using the optical satellite data is challenging because of cloud cover contamination. This seagrass species also flowers and fruits during July–August (Dai et al., 2002), with a growth rate of 0.93 cm/day (Estacion and Fortes, 1988). Thus, the field survey can also be conducted once a year during the dry season to trace the recolonization of seagrass beds to verify mapping results.

46


Fig. 2. Major seagrass species in the study region (Hoa, 2009): (a) Enhalus acoroides, (b) Halophila ovalis, (c) Halophila minor, (d) Thalassia hemprichii, (e) Halodule pinifolia, and (f) Halodule uninervis.

4. Methods The methodology used 5 main steps of data processing to investigate temporal changes of seagrass beds in the study region (Fig. 4). First, the data were converted to top of atmosphere (TOA) reflectance and corrected for atmospheric effects. Second, the water column correction was performed to remove water depth effects. Third, seagrass beds were classified using LMM, in part because patches of seagrass beds were small and fragmented in some locations across the study region, and a Landsat pixel (0.09 ha) often covers more than one land-use/ land cover class (e.g., seagrass, sand, muddy sand, mud, and coral reef). Although machine learning algorithms such as support vector machines (Vapnik, 1995) and artificial neural networks (Benediktsson et al., 1990) can perform complex classification tasks, including

mixed-pixel problems (Carpenter et al., 1996; Thornton et al., 2006), a primary limiting factor for benthic mapping (Fallah-Adl et al., 1995), these soft computing classifiers are more suitable for seagrass species discrimination when using moderate resolution satellite data for seagrass mapping. Fourth, the mapping results based on LMM in our study were verified using the ground reference data. As mentioned earlier, because of the unavailability of ground reference data for 2010, 2005, 2001, and 1996, we used the 2009 seagrass bed map to create the ground reference data to validate those years. The reference data for 2013 and 2015 were collected in two field surveys and used to validate the seagrass mapping in 2013 and 2015. To improve the quality of the work, the Google Earth images were used to cross-check the status of seagrass bed degradation in each study period. Finally, change analysis was conducted to detect multi-temporal changes of seagrass beds.

Fig. 3. The reference data: (a) bathymetry map, and (b) seagrass beds reference map in 2009.


47

Table 2 The results of accuracy assessment obtained from the classification of Landsat data for multi-temporal change detection.

Overall accuracy (%) Producer accuracy (%) User accuracy (%) Kappa

1996

2001

2005

2010

2013

2015

94.0 90.7 95.3 0.87

96.4 94.8 97.4 0.92

97.7 96.1 98.4 0.95

94.6 88.1 92.6 0.81

97.9 95.1 98.7 0.94

91.7 82.9 95.1 0.80

exoatmospheric irradiances, θs is the solar zenith angle, and Lλ is radiance calculated using the following equation: Lλ ¼

LMAX λ −LMIN λ ðQCAL−QCALMINÞ þ LMINλ ; QCALMAX−QCALMIN

ð2Þ

where QCAL = digital number, LMINλ is the spectral radiance scales to QCALMIN, LMAXλ is the spectral radiance scales to QCALMAX, QCALMIN is the minimum quantized calibrated pixel value (typically = 1), and QCALMAX is the maximum quantized calibrated pixel value (typically = 255). The Landsat OLI data were converted to TOA reflectance using the following equation:

Fig. 4. An overview of the methodology used in this study for seagrass mapping and change analysis.

4.1. Data pre-processing The land area was first masked out to limit our analysis to the water area. The Landsat TM/ETM+ data were converted to TOA reflectance using Eq. (1). 2

Rλ ¼

πLλ d ; ESUNλ cosðθs Þ

ð1Þ

where Rλ is the TOA planetary reflectance, λ is the band wavelength, d is the earth–sun distance in astronomical units, ESUNλ is the mean solar

Rλ ¼

ρ0λ ; sinðθSE Þ

ð3Þ

where Rλ is the TOA planetary reflectance; λ is the band wavelength; θSE is the local sun elevation angle; and ρλ′ is the TOA planetary reflectance calculated as ρλ′ = MρQcal +Aρ, where Mρ is the band-specific multiplicative rescaling factor, Aρ is the band-specific additive rescaling factor from the metadata, and Qcal is the quantized and calibrated standard product pixel values. 4.2. Water column correction The light intensity through water decreases exponentially with increasing water depth. Thus, the signal received by remote sensor for the same substrate at 2 m depth is relatively different from that at

Fig. 5. Mean spectral profiles of endmembers used for LMM classification: (a) 1996, (b) 2001, (c) 2005, (d) 2010, (e) 2013, and (f) 2015.

48


Fig. 6. Seagrass beds distributions in: (a) 1996, (b) 2001, (c) 2005, (d) 2010, (e) 2013, and (f) 2015.


49

Fig. 7. Changes of seagrass beds in the study region between: (a) 1996–2001, (b) 2001–2005, (c) 2005–2010, (d) 2010–2013, and (e) 2013–2015.

(b)

(a)

Seagrass beds

Aquaculture farms

Fig. 8. Google image of region A1: (a) seagrass beds distribution on 13 August 2003, and (b) seagrass beds on 03 March 2014.

50


(a)

Seagrass beds

(b)

Seagrass beds

Fig. 9. Google image of region A2: (a) seagrass beds on 02 March 2002, and (b) seagrass beds on 03 March 2014. (For interpretation of the references to color in this figure, the reader is referred to the web version of this article.)

20 m. The signal of sand at 20 m depth may be similar to seagrass at 3 m, potentially influencing the results of underwater habitat mapping. To remove these effects, the following bottom reflectance equation was used: Li ¼ Lsi þ ai r i expð−K i gZ Þ;

ð4Þ

where Li is the radiance at band i, Lsi is the radiance over deep water, ai is the solar irradiance at band i, ri is the bottom surface reflectance at band i, Ki is the attenuation coefficient of water of band i, g is the geometric factor accounting for the path length through the water, and Z is the water depth. The BRI was calculated as: BRI i ¼

ðLi −Lsi Þ ; expð−K i gZ Þ

ð5Þ

where Li is the radiance at band i, Lsi is the radiance over deep water of band i, Ki is the attenuation coefficient of water of band i, g is geometric factor, and Z is the water depth. By replacing (Li − Lsi) in Eq. (5) with Eq. (4), we can obtain BRIi = airi, where ri is the bottom surface reflectance. The unknown term Kig in Eq. (5) can be determined by exponential regression analysis from Eq. (4) using bathymetry data and radiance of band i of a substrate. Sand substrate is often considered for the calculation because it is more homogeneous than other bottom types.

endmember for each class used for LMM classification was pure or nearpure, we digitized polygons from homogeneous areas using the ground reference data. The minimum noise fraction transform (MNFT) tool in ENVI 5.1® was used to determine the inherent dimensionality of image data and to segregate noise in the data. The pixel purity index (PPI) was used to determine the purity of pixels from the MNFT data. These pixels were eventually averaged and used as an endmember for the classification. Assuming the spectral reflectance of each pixel can be approximated by a linear mixture of reflectance endmembers, r = (r1, r2, … , rn)T is the spectral reflectance vector to be classified, where n is the number of spectral bands and mi is the signature of the ith endmember, i = 1, …, c. The abundance fraction vector of endmembers α = (α1, α2, … , αc)T for every pixel can be obtained using the following equation: r ¼ Mα þ e;

ð6Þ

where r represents spectral values for a column pixel vector of n spectral band and M = [m1, m2, …, mc] is an n × c signature matrix of endmembers; and e is the error term. Given a set of spectral reflectance r, the abundance fraction of endmembers can then be estimated by this LMM algorithm. The model given in Eq. (6) represents an unconstrained linear mixing problem. The estimation must be completed with two physically reasonable constraints: (1) all abundances must be positive, c

αi ≥ 0, i = [1, …, c], and (2) abundance sum-to-one, ∑ α i ¼ 1. The

4.3. Image classification

i¼1

Based on the sedimentation map, the substrate types in the study region were categorized into 5 classes: sand, sandy mud, muddy sand, mud, and seagrass. Sand was generally distributed along the shoreline in the north part of the bay. Sandy mud and muddy sand were also distributed along the shoreline and the outer part of sandy areas. Mud that had blue-gray and dark blue tones was mainly found in the south part of the bay (Dai et al., 2002; Hoa, 2009; Long et al., 2011). To ensure that the

constrained least-squares method proposed by Heinz and Chein (2001) was applied to estimate the proportion of each component in a pixel by minimizing the sum of the square of errors. The classification result is a composite image scene where the values of each pixel are continuous numbers ranging from 0 to 1, determined from the estimated abundance fractions of the materials in each pixel. The winner-take-all (WTA) method was applied to convert the abundance estimation of a mixed pixel to a pure pixel with respect to a

Seagrass beds

(a)

Aquaculture farms

(b)

Fig. 10. Google image of region A3: (a) aquaculture farms, harbor and transportation density on 02 March 2002, and (b) aquaculture farms, harbor and transportation density on 04 July 2014.


51

(b)

(a)

New construction in 2014

Fig. 11. Google image of region A4: (a) aquaculture farms on 26 March 2013, and (b) aquaculture farms on 04 July 2014.

desired class as:

j ¼ arg max1b jbc α j ðr Þ ;

ð7Þ

where j∗ is the desired class to be assigned and αj(r) is the mixed abundance vector in an observed image pixel vector r. 4.4. Accuracy assessment The mapping results were categorized into two classes: seagrass and non-seagrass (i.e., sand, sandy mud, muddy sand, and mud). The median filter was used to eliminate ‘salt-and-pepper’ noise from the classification maps. We verified the mapping results for each year data using the ground reference data. The overall accuracy, producer accuracy, user accuracy, and Kappa coefficient derived from the confusion matrix (Congalton, 1991) were collectively used to measure the accuracy of the mapping results.

(Chen et al., 2011; Ren, 2000). The classification maps were assessed using the ground reference data, indicating that the overall accuracies and Kappa coefficients were generally higher than 91% and 0.80, respectively, in all cases (Table 2). These results confirmed the validity of our LMM-based approach for seagrass mapping results. In general, several error sources were found to exaggerate the overall accuracies of the mapping results, including characteristics of Landsat data and mixed pixel problems. For instance, studies indicated that spectral responses from Landsat data were better used to map seagrass beds shallower than 5 m (Lyzenga, 1981), whereas seagrasses in the study region may grow in areas deeper than 5 m (Short and Waycott, 2010), creating potential mapping errors. Likewise, seagrasses were spatially scattered across the study region, and some of their patches were smaller than the size of a Landsat pixel. The mixed pixel issue could cause spectral confusion between the seagrass class and other classes, making it difficult to determine an appropriate proportion of seagrass in each pixel.

5. Results and discussion

5.3. Spatial distributions of seagrass beds and decadal changes of seagrasses

5.1. Spectral endmembers

Based on characteristics of the bay and spatial distributions of seagrass beds, we divided the study region into two parts for temporal comparisons of changes of seagrass beds (regions A and B highlighted as red dashed line, Fig. 6). Region A, on the west side of the bay, comprises residential areas where aquaculture farms have been developed. Sites of seagrass meadows, A1–A4, were closely investigated because of their historical presence (Dai et al., 1999; Dai et al., 2002; Hoa, 2009; Vy, 2010). The seagrass area in region A1 was destroyed in 2010 by aquaculture farms constructed directly on seagrass area (Vy, 2010), whereas in region A2, most aquaculture farms were allocated to the shoreline areas, and thus, the spatial distributions of seagrass beds in A2 were likely unchanged during the study period. In region A3, aside from aquaculture development, we observed that the large and busy harbor situated near Cam Ranh City (population of 125,311) could influence seagrass beds. Region A4 is located in the south part of the bay, where substrates of mud, muddy clay, and muddy sand with high turbidity may affect the seagrass health. Region B, situated in the east part of

The results of endmember selection based on MNFT and PPI showed spectral characteristics of endmembers used for LMM classification (Fig. 5). The spectral profile of seagrass was distinguished from those of other classes in the study region. The reflectance signal of seagrass class was lower than those of other classes, facilitating differentiation. Similarly, we could easily differentiate the sand class from other classes because it occupied a large area along the coastline and generally had higher reflectance signal than other classes. The endmembers of other classes including sandy mud, muddy sand, and mud were sequentially derived to improve the LMM classification of seagrass beds. 5.2. Seagrass beds mapping accuracies The LMM-based mapping results showing the abundance of endmembers in each pixel was rationalized by using the WTA method

Fig. 12. Photos taken during the field survey in 2013: (a) farmers collecting clams, and crabs when low tide at region A4, and (b) intact seagrass beds at region B.

52


Fig. 13. Summary of seagrass beds change between 1996 and 2015 including unchanged seagrass area is 99.2 ha, loss seagrass area is 162.1 ha and new seagrass area is 96.1 ha. (For interpretation of the references to color in this figure, the reader is referred to the web version of this article.)

the bay, was strictly controlled by the local government so that no aquaculture activities were observed. The spatiotemporal distributions of seagrasses from 1996 to 2015 (Fig. 6) show that, in general, seagrass beds were spatially distributed throughout the study region but mainly concentrated in the north and south parts of the region. The total area of seagrass was reduced by approximately 1.4% in 2001 compared to 1996, and the loss of seagrass beds was generally found in the southern (A4) and northwestern (A1 and A2) regions (Fig. 6a, b). The seagrass beds remained either unchanged or slightly increased in northeastern parts (region B). In 2005, the area of seagrass had significantly declined in the southern part at regions A4 and A3 (Fig. 6c, b), mainly due to aquaculture development. The number of aquaculture farms in CRB increased from approximately 5560 to 12,000 during 2001–2005. We also observed in the northwest region A1 that seagrass beds had shifted close to the coastline (Fig. 6c, 7b). By 2010, most of the seagrass areas in region A1 were destroyed (Fig. 6d, 7c), possibly because the number of Table 3 Seagrass bed areas and change rates from 1996 to 2015.

Total seagrass areas (ha) Changing rate (%)a a

1996

2001

2005

2010

2013

2015

261.3

257.6 –1.4

222.3 –13.7

189.7 –14.7

187.9 –0.9

195.3 3.9

Change rate (%) = (Later year area (ha) / earlier year area (ha) − 1) ∗ 100%.

aquaculture farms present was similar in 2010 and 2005 (Khanh Hoa Statistical Yearbook, 2011). These results were comparable with those reported in 2010 that seagrass beds in A1 had disappeared (Vy, 2010). From 2010 to 2013, a reduction of seagrass beds between A1 and A2 was observed. The seagrass areas in A3 did not exist based on our measurements, whereas seagrass areas in A4 had shifted to the northwestern part of the region compared to areas observed in 1996, 2001, and 2005 (Fig. 6e, 7d). By 2015, the area of seagrass in A4 had been reduced; however, some seagrass beds were detected in A1 and A3 where no seagrass was found in 2013 (Fig. 6e, f, e). The seagrass beds in region B were largely unchanged in 1996 and 2010 (Fig. 6a–d). The seagrass beds in region A1 in the northwestern part of the bay slowly decreased from 1996 to 2005 but were then seriously degraded by 2010 (Figs. 6, 7), and from 2010 to 2015 they nearly disappeared. Almost no aquaculture activity occurred in this area in 2003 (Fig. 8a), but by 2014, many aquaculture farms had been developed on the seagrass beds, the main cause of seagrass bed degradation on this site (Fig. 8b). The results in 2001 (Fig. 6b) were consistent with those from Google imagery in 2003 (dark blue seagrass area in Fig. 8a). The mapping results in 2015 (Fig. 6f) not only had a high overall accuracy of 91.68%, but also had close agreement with Google imagery in 2014 (Fig. 8b), showing that all lost seagrass areas were occupied by aquaculture farms. During high tide in region A2 in the northwestern area of the bay below A1, the current flows from the bay mouth to the north, stabilizing the water in this area because of a permanent sand dune that protects this region. During low tide, the currents come from the north to A2, changing direction and recessing to the open sea, creating a high exchange rate that results in a clean water environment. In addition, water in this area has been free from aquaculture activities since 1996 (Fig. 9a) to present (Fig. 9b), perhaps explaining why the seagrass beds in this area have maintained the same status since 1996 (Fig. 6). This explanation is corroborated by Google maps of March 2002 and March 2014 (dark blue area in Fig. 9) and confirms the mapping accuracies in 2001 and 2015 (Fig. 6b, f). Region A3 located in the mid-western part of the bay, close to a harbor where transportation and aquaculture activities exist the area of seagrass beds, were remarkably reduced during 2001–2005 (Fig. 6b, c, 7b), and no seagrass was detected in this area in 2010 and 2013. In 2015 a small fraction of seagrass beds was detected, perhaps indicating some recovery. Seagrasses covered the upper part of the harbor in 2002 (shaded area in Fig. 10a), and aquaculture farming was low. By 2014, however, many ships, boats, aquaculture farms, and construction activities had begun in this area (Fig. 10b), an observation supported the mapping accuracy (Fig. 6b, f). The seagrass area of region A4 located in the southern part of the bay was critically reduced between 2001 and 2005 (Fig. 6b, c) and then slowed. The sharp reduction was caused by high rates of sedimentation from muddy sand, mud, and clay substrates coupled with human farming activities, degrading the environmental conditions. From 2005 to 2013 aquaculture development remained stable, and little reduction in the area of seagrass was observed for this region after 2005. From 2013 to 2015, a construction boom in this area (Fig. 11) likely caused the reduced seagrass area in 2015 (Fig. 6f, 7e). Moreover, farmers also collected aquaculture products (e.g., clams and crabs) during low tide (Fig. 12a), activities that markedly destroyed seagrass beds, especially in shallow areas near the shoreline. Region B in the northeastern part of the bay belongs to the government, and therefore no aquaculture activities have occurred, and the seagrass beds remain largely intact (Figs. 6, 7, 12b). Mapping results indicated that A1 and A4 have lost 162.1 ha of seagrass area (red areas in Fig. 13). The growth of 96.1 ha of new seagrass beds appeared mostly in region B in the northeast part of the CRB (blue area in Fig. 13). The unchanged seagrass bed area totals 99.2 ha (green areas in Fig. 13), distributed mainly in region B in the northeastern part of the bay and in region A2. In total, an area of 66 ha


of seagrass beds had been lost by 2015, accounting for approximately 25.25% of the total seagrass area in the study region. The total area of seagrass beds reduced from 1996 to 2015 at different rates (Table 3). From 1996 to 2001, the seagrass area had slightly decreased by approximately 1.4% but was seriously reduced by 13.7% from 2001 to 2005, increasing to a 14.7% reduction during 2005–2010. However, seagrass area slightly increased from 2013 to 2015. On average, approximately 4.5 ha of seagrasses were lost annually. Socioeconomic activities from recent economic development in this study region generate many waste streams discharged into the CRB, where a low water exchange rate (18.9 days in dry season and 16 days in rainy season) is the norm. The water quality in the bay has been degraded by dissolved inorganic nutrients, nitrate, phosphate, and ammonia/ammonium (Thu et al., 2013), and therefore the selfpurification capacity provided by the existing seagrass bed is critical, motivating further recovery of seagrass beds. Our seagrass bed mapping results show that regions A1, A2, A3, and A4 were seriously degraded and should be monitored at least once every two years. 6. Conclusions This study used integrated multi-temporal Landsat images to monitor changes of seagrass in the CRB of Vietnam. The BRI-based water column correction combined with the LMM mapping technique demonstrates its application potential for seagrass beds monitoring in the study region. Although several error sources, such as the quality of satellite data and mixed pixel problems, could exaggerate the classification results, close agreement between the classification maps and the ground reference data were confirmed by the overall accuracies and Kappa coefficients generally higher than 91.7% and 0.8, respectively, in all cases. The total area of seagrasses in the study region was approximately 261.3 ha in 1996 but from 1996 to 2015 had declined by approximately 25% (66 ha), mainly attributed to the high level of coastal development and infrastructure construction. Regions A1 and A2 experienced the biggest dynamics of gain and loss, whereas region A4 had the biggest net loss of seagrass bed over the five study periods. The net gain of seagrass bed in region B almost outweighs the net loss of seagrass bed in the other four regions over the five study periods. The results obtained from this study could provide quantitative information on changes in seagrass beds over 19 years (1996–2015), which natural resources managers could use to derive successful management strategies for marine ecosystems in the region. Acknowledgments This research is mainly supported by National Central University, Taiwan (NSC 102-2923-E-008-001-MY3) and Vietnam Government (VEST 500). The financial support is gratefully acknowledged. We are grateful to the staff of Nha Trang Institute of Oceanography, Vietnam for help in the collection of the ground reference data. We also thank the two anonymous reviewers for their insightful and constructive comments and suggestions. References Andréfouët, S., Muller-Karger, F.E., Hochberg, E.J., Hu, C., Carder, K.L., 2001. Change detection in shallow coral reef environments using Landsat 7 ETM+ data. Remote Sens. Environ. 78, 150–162. Andréfouët, S., Kramer, P., Torres-Pulliza, D., Joyce, K.E., Hochberg, E.J., Garza-Pérez, R., Mumby, P.J., Riegl, B., Yamano, H., White, W.H., Zubia, M., Brock, J.C., Phinn, S.R., Naseer, A., Hatcher, B.G., Muller-Karger, F.E., 2003. Multi-site evaluation of IKONOS data for classification of tropical coral reef environments. Remote Sens. Environ. 88, 128–143. Barillé, L., Robin, M., Harin, N., Bargain, A., Launeau, P., 2010. Increase in seagrass distribution at Bourgneuf Bay (France) detected by spatial remote sensing. Aquat. Bot. 92, 185–194. Benediktsson, J.A., Swain, P.H., Ersoy, O.K., 1990. Neural network approaches versus statistical methods in classification of multisource remote sensing data. IEEE Trans. Geosci. Remote Sens. 28, 540–552.

53

Blakey, T., Melesse, A., Hall, M., 2015. Supervised classification of benthic reflectance in shallow subtropical waters using a generalized pixel-based classifier across a time series. Remote Sens. 7, 5098. Botha, E.J., Brando, V.E., Anstee, J.M., Dekker, A.G., Sagar, S., 2013. Increased spectral resolution enhances coral detection under varying water conditions. Remote Sens. Environ. 131, 247–261. Cabaço, S., Santos, R., Duarte, C.M., 2008. The impact of sediment burial and erosion on seagrasses: a review. Estuar. Coast. Shelf Sci. 79, 354–366. Carpenter, G.A., Gjaja, M.N., Gopal, S., Woodcock, C.E., 1996. ART neural networks for remote sensing: vegetation classification from Landsat TM and terrain data. Geoscience and Remote Sensing Symposium, 1996. IGARSS '96. ‘Remote Sensing for a Sustainable Future’, International vol. 521, pp. 529–531. Chen, C.F., Son, N.T., Chang, L.Y., Chen, C.R., 2011. Classification of rice cropping systems by empirical mode decomposition and linear mixture model for time-series MODIS 250 m NDVI data in the Mekong Delta, Vietnam. Int. J. Remote Sens. 32, 5115–5134. Congalton, R.G., 1991. A review of assessing the accuracy of classifications of remotely sensed data. Remote Sens. Environ. 37, 35–46. Dai, N.H., 2011. Seagrass in Vietnam. Seagrasses: Resource Status and Trends in Indonesia, Japan, Malaysia, Thailand and Vietnam. Seizando-Shoten, Tokyo. Dai, N.H., Hoa, N.X., Tri, P.H., Linh, N.T., Phi, H., 1999. The role of seagrass beds in Khanh Hoa province. The nursery grounds. The Scientific Workshop on Management and Sustainable Use of Natural Resources and the Environment in Coastal Wetland, pp. 26–31. Dai, N.H., Tri, P.H., Linh, N.T., Vy, N.X., 2002. The degradation of seagrass beds in Khanh Hoa province and the ability of restoration of these meadows (in Vietnamese). Scientific Conference of “BIEN DONG – 2002”. Nha Trang, Vietnam. Duarte, C.M., 2002. The future of seagrass meadows. Environ. Conserv. 29, 192–206. Erftemeijer, P.L.A., Robin Lewis Iii, R.R., 2006. Environmental impacts of dredging on seagrasses: a review. Mar. Pollut. Bull. 52, 1553–1572. Estacion, J.S., Fortes, M.D., 1988. Growth rates and primary production of Enhalus acoroides (L.f.) role from Lag-it, North Bais Bay, the Philippines. Aquat. Bot. 29, 347–356. Fallah-Adl, H., JaJa, J., Shunlin, L., Kaufman, Y.J., Townshend, J., 1995. Efficient algorithms for atmospheric correction of remotely sensed data. Supercomputing, 1995. Proceedings of the IEEE/ACM SC95 Conference, p. 12. Ferguson, R.L., Korfmacher, K., 1997. Remote sensing and GIS analysis of seagrass meadows in North Carolina, USA. Aquat. Bot. 58, 241–258. Fortes, M.D., 1990. Seagrasses: A Resource Unknown in the ASEAN Region. WorldFish. Fourqurean, J.W., Duarte, C.M., Kennedy, H., Marba, N., Holmer, M., Mateo, M.A., Apostolaki, E.T., Kendrick, G.A., Krause-Jensen, D., McGlathery, K.J., Serrano, O., 2012. Seagrass ecosystems as a globally significant carbon stock. Nat. Geosci. 5, 505–509. Gokul, E.A., Shanmugam, P., Sundarabalan, B., Sahay, A., Chauhan, P., 2014. Modeling the inherent optical properties and estimating the constituents' concentrations in turbid and eutrophic waters. Cont. Shelf Res. 84, 120–138. Green, E., Short, F., 2003. World Atlas of Seagrasses. Prepared by the UNEP World Conservation Monitoring Centre. University of California Press, Berkeley, CA, USA. Green, E.P., Mumby, P.J., Edwards, A.J., Clark, C.D., 1996. A review of remote sensing for the assessment and management of tropical coastal resources. Coast. Manag. 24, 1–40. Green, E.P., Mumby, P.J., Edwards, A.J., Clark, C.D., 2000. Remote Sensing Handbook for Tropical Coastal Management. UNESCO. Heck Jr., K.L., Hays, G., Orth, R.J., 2003. Critical evaluation of the nursery role hypothesis for seagrass meadows. Mar. Ecol. Prog. Ser. 253, 123–136. Heinz, D.C., Chein, I.C., 2001. Fully constrained least squares linear spectral mixture analysis method for material quantification in hyperspectral imagery. Geoscience and Remote Sensing, IEEE Transactions on 39, 529–545. Hoa, N.X., 2009. Status Assessment of the Area, Species and Distribution on Mangrove and Seagrass Ecosystem and their Role for the Socio-economic in Khanh Hoa Province and Propose Management Solutions Sustainable Development. Hochberg, E.J., Atkinson, M.J., Andréfouët, S., 2003. Spectral reflectance of coral reef bottom-types worldwide and implications for coral reef remote sensing. Remote Sens. Environ. 85, 159–173. Khanh Hoa Statistical Yearbook, 2011. Vietnam General statistics office, Khanh Hoa statistics office. Komatsu, T., Igarashi, C., Tatsukawa, K., Sultana, S., Matsuoka, Y., Harada, S., 2003a. Use of multi-beam sonar to map seagrass beds in Otsuchi Bay on the Sanriku Coast of Japan. Aquat. Living Resour. 16, 223–230. Komatsu, T., Mikami, A., Sultana, S., Ishida, K., Hiraishi, T., Tatsukawa, K.-i., 2003b. Hydroacoustic methods as a practical tool for cartography of seagrass beds. Otsuchi Marine Science 28, 72–79. Long, B.H., Hoan, P.S., Mau, L.D., Vinh, N.K., Tuan, N.V., Chung, T.V., Cong, N.C., 2011. Handbook on Natural Conditions, The Environment - Ecological, Economic, Social and Integrated Coastal Zone Management in South Central of Vietnam. Vietnam Academy of Science and Technology. Lyons, M.B., Phinn, S.R., Roelfsema, C.M., 2012. Long term land cover and seagrass mapping using Landsat and object-based image analysis from 1972 to 2010 in the coastal environment of South East Queensland, Australia. ISPRS J. Photogramm. Remote Sens. 71, 34–46. Lyzenga, D.R., 1981. Remote sensing of bottom reflectance and water attenuation parameters in shallow water using aircraft and Landsat data. Int. J. Remote Sens. 2, 71–82. Mumby, P.J., Edwards, A.J., 2002. Mapping marine environments with IKONOS imagery: enhanced spatial resolution can deliver greater thematic accuracy. Remote Sens. Environ. 82, 248–257. Mumby, P.J., Green, E.P., Edwards, A.J., Clark, C.D., 1997. Measurement of seagrass standing crop using satellite and digital airborne remote sensing. Mar. Ecol. Prog. Ser. 159, 51–60.

54


Mumby, P.J., Clark, C.D., Green, E.P., Edwards, A.J., 1998. Benefits of water column correction and contextual editing for mapping coral reefs. Int. J. Remote Sens. 19, 203–210. Mumby, P.J., Green, E.P., Edwards, A.J., Clark, C.D., 1999. The cost-effectiveness of remote sensing for tropical coastal resources assessment and management. J. Environ. Manag. 55, 157–166. Pasqualini, V., Pergent-Martini, C., Pergent, G., Agreil, M., Skoufas, G., Sourbes, L., Tsirika, A., 2005. Use of SPOT 5 for mapping seagrasses: an application to Posidonia oceanica. Remote Sens. Environ. 94, 39–45. Pham, H.T., Nguyen, H.D., Nguyen, X.H., Nguyen, T.L., 2006. Study on the Variation of Seagrass Population in Coastal Waters of Khanh Hoa Province, Vietnam. Philpot, W.D., 1989. Bathymetric mapping with passive multispectral imagery. Appl. Opt. 28, 1569–1578. Phinn, S., Roelfsema, C., Dekker, A., Brando, V., Anstee, J., 2008. Mapping seagrass species, cover and biomass in shallow waters: an assessment of satellite multi-spectral and airborne hyper-spectral imaging systems in Moreton Bay (Australia). Remote Sens. Environ. 112, 3413–3425. Pu, R., Bell, S., Meyer, C., 2014. Mapping and assessing seagrass bed changes in Central Florida's west coast using multitemporal Landsat TM imagery. Estuar. Coast. Shelf Sci. 149, 68–79. Ren, H., 2000. Unsupervised and Generalized Orthogonal Subspace Projection and Constrained Energy Minimization for Target Detection and Classification in Remotely Sensed Imagery. Department of Electrical Engineering: University of Maryland Baltimore County. Sagawa, T., Boisnier, E., Komatsu, T., Mustapha, K.B., Hattour, A., Kosaka, N., Miyazaki, S., 2010. Using bottom surface reflectance to map coastal marine areas: a new application method for Lyzenga's model. Int. J. Remote Sens. 31, 3051–3064. Sagawa, T., Mikami, A., Aoki, M.N., Komatsu, T., 2012. Mapping seaweed forests with IKONOS image based on bottom surface reflectance (pp. 85250Q-85250Q-85257). Schroeder, T.A., Cohen, W.B., Song, C., Canty, M.J., Yang, Z., 2006. Radiometric correction of multi-temporal Landsat data for characterization of early successional forest patterns in western Oregon. Remote Sens. Environ. 103, 16–26. Short, F.T., Waycott, M., 2010. Enhalus acoroides, The IUCN Red List of Threatened Species 2010: e.T173331A6992567. Short, F.T., Coles, R.G., Pergent-Martini, C., 2001. Chapter 1 - global seagrass distribution. In: Short, F.T., Short, C.A., Coles, R.G. (Eds.), Global Seagrass Research Methods. Elsevier Science, Amsterdam, p. 5-XX.

Spalding, M.D., Ruffo, S., Lacambra, C., Meliane, I., Hale, L.Z., Shepard, C.C., Beck, M.W., 2014. The role of ecosystems in coastal protection: adapting to climate change and coastal hazards. Ocean Coast. Manag. 90, 50–57. Stapel, J., Aarts, T.L., van Duynhoven, B.H.M., de Groot, J.D., van den Hoogen, P.H.W., Hemminga, M.A., 1996. Nutrient uptake by leaves and roots of the seagrass Thalassia hemprichii in the Spermonde Archipelago, Indonesia. Mar. Ecol. Prog. Ser. 134, 195–206. Thornton, M.W., Atkinson, P.M., Holland, D.A., 2006. Sub-pixel mapping of rural land cover objects from fine spatial resolution satellite sensor imagery using superresolution pixel-swapping. Int. J. Remote Sens. 27, 473–491. Thu, P.M., Huan, N.H., Long, B.H., 2013. Assessment of environmental capacity of Thuy Trieu-Cam Ranh waters. J. Mar. Sci. Technol. 13, 371–381. Valle, M., Palà, V., Lafon, V., Dehouck, A., Garmendia, J.M., Borja, Á., Chust, G., 2015. Mapping estuarine habitats using airborne hyperspectral imagery, with special focus on seagrass meadows. Estuar. Coast. Shelf Sci. 164, 433–442. Vapnik, V.N., 1995. The Nature of Statistical Learning Theory. Springer-Verlag, New York, USA. Vo, S.T., Pernetta, J.C., Paterson, C.J., 2013. Status and trends in coastal habitats of the South China Sea. Ocean Coast. Manag. 85, 153–163. Vonk, J.A., Middelburg, J.J., Stapel, J., Bouma, T.J., 2008. Dissolved organic nitrogen uptake by seagrasses. Limnol. Oceanogr. 53, 542–548. Vy, N.X., 2010. Monitoring Biodiversity of Seagrass and Mangrove Ecosystems Along Coastal of Khanh Hoa Province. Institute of Oceanography Nha Trang, Nha Trang Vietnam, p. 64. Wabnitz, C.C., Andréfouët, S., Torres-Pulliza, D., Müller-Karger, F.E., Kramer, P.A., 2008. Regional-scale seagrass habitat mapping in the Wider Caribbean region using Landsat sensors: applications to conservation and ecology. Remote Sens. Environ. 112, 3455–3467. Walker, N.D., 1996. Satellite assessment of Mississippi River plume variability: causes and predictability. Remote Sens. Environ. 58, 21–35. Walker, D.I., McComb, A.J., 1992. Seagrass degradation in Australian coastal waters. Mar. Pollut. Bull. 25, 191–195. Yang, D., Yang, Y., Yang, C., Zhao, J., Sun, Z., 2011. Detection of seagrass in optical shallow water with Quickbird in the Xincun Bay, Hainan province, China. Image Processing, IET 5, 363–368.