Characterizing the spatial distribution of coral reefs in the South-Central Coast region of Viet Nam using Planetscope imagery

Study sites

The study areas (i.e., ST (108°13′58.1″E–16°08′54.4″N and 108°20′22.4″E–16°05′16.2″N) and CLC (108°24′42.8″E–15°58′57.7″N and 108°33′04.2″E–15° 53′58.4″N)) are nature reserve areas located in the South-Central coast region of Viet Nam (Nguyen, Luong & Ho, 2015; Nguyen, Huynh & Zhang, 2015; Hoang et al., 2019) (Fig. 1). CLC was recognized as a world biosphere reserve by UNESCO in 2009 (Nguyen, Luong & Ho, 2015; Nguyen, Huynh & Zhang, 2015). The study sites are located within the tropical monsoon region with a distinct dry season (January–July) receiving 600 mm rainfall over these 6 months and a rainy season, where 80% of the annual rainfall (2,000 mm) falls between the month of August to December. The annual mean temperature at the study areas is about 26 °C with little seasonal variation. The region is regularly disturbed by typhoons and river run-off in the wet season, which can lead to destruction of reefs.

The reefs at study area are shallow fringing reefs, with a diversity of 226 and 227 coral species in ST and CLC, respectively (Kimura, Tun & Chou, 2014; Kimura, Tun & Chou, 2018). Deposited materials on the ocean floor in this region is sand and rubble, which deposit between the inner and outer reef regions. Both regions are important to their local economies providing revenue from coral reef-based tourism, and small-scale fisheries. However, these reefs are under threat from an accumulation of stressors such as human disruption, typhoons, new outbreaks of crown of thorns starfish, excess terrigenous sediment output due to land use change, pollution, overfishing, destructive fishing, and the mining of reefs for limestone (Nguyen & Vo, 2013; Quach, 2018). A lack of regular monitoring has impacted on effective coral reef management and protection (Le, 2020).

Data collection and Processing

The Planetscope images (3 m spatial resolution) used in this study were obtained from the Planet Education and Research Program server (https://www.planet.com/markets/education-and-research/) (Table 1). The wavelengths of Planetscope imagery cover the blue (455–515 nm), green (500–590 nm), red (590–670 nm), and near-infrared (780–860 nm) spectral regions (Gabr, Ahmed & Marmoush, 2020). These images were acquired on 28 June 2019 (ST) and 08 August 2019 (CLC) with 0% of cloud cover and coordinate system projection of WGS 84/ UTM zone 49N. The images, acquired by a Dove satellite with additional postprocessing applied, known as the PlanetScope Ortho Scene Product (level 3B), were ortho-rectified, atmospherically corrected, and scaled to surface reflectance (Lemajic, Vajsová & Aastrand, 2018).

We carried out field surveys during August and September 2019 to gather information about the depth and features of the ocean bottom. The bottom features were measured following the 25’s method, proposed by (Roelfsema, Phinn & Joyce, 2006), in a grid of 5 m × 5 m using an underwater camera (GoPro HERO 8, USA). Based on this approach, five photographs including the four corners and one center positions of a 1 m × 1 m quadrat were taken. The underwater photographs were ortho-rectified to control the spatial uncertainty to <3 m in order to match the resolution of the satellite data. The depth and the coordinates of each photograph were recorded using a Hondex PS-7FL (Honda Electronics Ltd., Toyohashi, Japan) and Garmin GPS MAP 62sc (Garmin Ltd., Lenexa, Kansas, USA), respectively. In addition to photo capture in the 5 m × 5 m grid, underwater 360 video transects (CND713, CND Co., Ltd., Korea) were taken to determine the ocean floor classes (Table 2) in deep and homogenous areas before collecting substrate types (i.e., sand, and hard bottom) to reduce the cost of field survey. A total of 1,543 photographs and 167 video transects at ST and 1,560 photographs and 174 video transects at CLC were collected during the field campaign. The photos were filtered to select for good quality photographs demonstrating their ability to clearly identify the bottom feature. Five photographs were overlapped to create a transect and each transect was geo-referenced based on captured coordinates before analyzing the benthic composition by the Coral Point Count Excel (CPCe) software (Kohler & Gill, 2006; Roelfsema, Phinn & Joyce, 2006).

Table 2:

Identification scheme of bottom types and coral reef health.

Class	Detail	Class description	Coral reef health criteria	Label	Depth range (m)	Sample image
Live coral (LC)	hard and soft coral	0 < LC ≤ 25%	Poor	1LC	ST: 0.8–8.5 CLC: 0.6–11.5
		25 < LC ≤ 50%	Fair	2LC	ST: 0.5–8.2 CLC: 0.9–9.9
		50 < LC ≤ 75%	Good	3LC	ST: 1.3–7 CLC: 0.8–10.2
		LC > 75%	Excellent	4LC	CLC: 0.9–10.5
Sand (S)	Sand dominated	LC = 0% and S ≥ 50% S		S	ST: 0.5–13.1 CLC: 0.4–12.5
Hard bottom	Rock, rubble, dead structure	LC = 0% LC and HB > 50%		HB	ST: 1–8.2 CLC: 0.7–10.5
Deep zone		Depth > 14 m		D	ST: 14–24 CLC: 14–29.5

DOI: 10.7717/peerj.12413/table-2

Bottom types, derived from the photographs of 1 m × 1 m quadrats, were identified into four classes (i.e., live coral [hard and soft coral], sand, hard bottom [rock, rubble, dead coral], deep water (>14 m)) (Hill & Wilkinson, 2004), and were used to calculate proportion of substrate composition. The coral reef health in 5 m × 5 m grids was categorized as the ‘poor’, ‘fair’, ‘good’ or ‘excellent’ condition based on live coral cover (Chou et al., 1995; Hill & Wilkinson, 2004) (Table 2). Finally, a part of the ground-truth samples was randomly selected for image classification training (see the sub-section in Supervised classification and coral cover mapping). A total of 323 and 369 randomly selected ground-truth samples in the ST and CLC, respectively, were identified to assess classification accuracy. It included each classed category as follow: 1LC (69), 2LC (64), 3LC (47), HB (43), S (51), and D (49) in Son Tra and 1LC (62), 2LC (56), 3LC (54), 4LC (55), HB (43), S (55), and D (44) in Cu Lao Cham. (Table 2 and the Supplemental Files).

Sun glint correction and water column correction

Sun glint occurs in images under conditions of clear skies, shallow waters (when wind generated waves disturb the surface), more so when images are acquired at a high spatial resolution (Hedley, Harborne & Mumby, 2005). This phenomenon could be defined as the specular reflection of light from the water surface, when the component of sensor-received radiance due to the surface reflection is greater than the water-leaving radiance from sub-surface features (Kay, Hedley & Lavender, 2009). Previous studies proposed sun-glint removal methods for ocean color applications with resolutions on the scale of 100–1,000 m (Wang & Bailey, 2001; Hedley, Harborne & Mumby, 2005; Kay, Hedley & Lavender, 2009). However, these models could only correct moderate glint and large errors remain in the brightest glint areas (Kay, Hedley & Lavender, 2009). Hence, a separate set of revised methods were developed for benthic remotely sensed data at high spatial resolutions (<10 m) (Hochberg, Andréfouët & Tyler, 2003; Kay, Hedley & Lavender, 2009). These approaches use data from the near-infrared (NIR) to determine the amount of glint in the received signal (Hedley, Harborne & Mumby, 2005; Kay, Hedley & Lavender, 2009), thus improving data retrieval for bathymetry or coral habitat classification (Hochberg, Andréfouët & Tyler, 2003; Hedley, Harborne & Mumby, 2005; Kay, Hedley & Lavender, 2009). In this study, the sun light correction of the Planetscope images were undertaken using a revised ‘de-glinting’ method (Hochberg, Andréfouët & Tyler, 2003; Hedley, Harborne & Mumby, 2005). Following this approach, the relationships between NIR and visible bands were explored by linear regression based on a sample of the image pixels. The pixel sample was collected from one or more regions of the image where a range of sun glint is evident, but where the underlying spectral brightness would be expected to be consistent (normally deep-water areas) (Hedley, Harborne & Mumby, 2005). If the slope of the regression line for band i is b_i, then all pixels in the image could be de-glinted in band i, following Eq. (1): (1)${\displaystyle {\mathrm{L}}_{\mathrm{i}}^{{}^{\prime }}={\mathrm{L}}_{\mathrm{i}}+{\mathrm{b}}_{\mathrm{i}}\left({\mathrm{L}}_{\mathrm{NIR}}-{\mathrm{Min}}_{\mathrm{NIR}}\right)}$

where, L’i is the de-glinted pixel of visible band i; L_i is reflectance from the visible band i; b_i is the regression slope; L_NIR is reflectance from the NIR band; and Min_NIR is the minimum value of the NIR band.

In addition to the sun glint phenomenon, we needed to correct for the water column so that bottom classes were not misclassified due to different water column conditions (Zoffoli, Frouin & Kampel, 2014). Among various algorithms, Lyzenga’s equation (Lyzenga, 1978; Lyzenga, 1981) is widely applied to empirically diminish water column attenuation impacts. This method enhances mapping accuracy in digital classification processes of bottom types, and does not require knowing the local depth across every pixel of the scene (Green et al., 2000; Pahlevan, Valadanzouj & Alimohamadi, 2006; Pu et al., 2012; Zoffoli, Frouin & Kampel, 2014). The algorithm uses multispectral image datasets to generate ‘depth –invariant’ coefficients of bottom types from each pair of spectral bands. The result of the correction creates multiple ’depth - invariant bands, replacing original bands to classify bottom features (Green et al., 2000). The depth-independent composition of corrected radiance (DII_ij) in bands i and j (pseudo-color band) was generated by Eq. (2): (2)${\displaystyle {\mathrm{DII}}_{\mathrm{ij}}=ln\left({\mathrm{L}}_{\mathrm{i}}\right)-\left[\left(\frac{{\mathrm{k}}_{\mathrm{i}}}{{\mathrm{k}}_{\mathrm{j}}}\right)ln\left({\mathrm{L}}_{\mathrm{j}}\right)\right]}$

where L_i is de-glinted pixel of band i; L_j is de-glinted pixel of band j; $\left(\frac{{\mathrm{k}}_{\mathrm{i}}}{{\mathrm{k}}_{\mathrm{j}}}\right)$ is the ratio of attenuation coefficients.

In this study, three visible bands of PlanetScope imagery with high water penetration were plotted against transformed band pairs (i.e., band 1 and band 2, band 2 and band 3, band 1 and band 3). The slope values of the ratio of attenuation coefficients for each band were applied to generate three DII bands (i.e., DII₁₂, DII_13, and DII₂₃) to class the bottom types.

Supervised classification and coral cover mapping

Different types of supervised classification methods have been used to identify coral reefs with water column corrected images and bottom classes in similar marine regions. The Maximum Likelihood Classifier was used due to its effectiveness and adaptability, compared to other classifiers (Ali, Qazi & Aslam, 2018; Cabral et al., 2018). In the current study, the Maximum Likelihood Classifier was subsequently applied to each image. This classifier was based on a probability density function and guessed the probability with which a specific pixel belongs to a specific category (Richards, 1999).The standard false colour satellite images were composed before applying the supervised classification approach. The standard implementation of supervised maximum likelihood classification required training samples representing the feature types to be classified (Lillesand, Kiefer & Chipman, 2015). Taking this a-priori knowledge into account, each pixel was labelled as 1LC, 2LC, 3LC, 4LC, HB, S and D (Table 2). Finally, we used a 3 × 3 focal majority filter to minimize the “salt-and-pepper effect” in the coral reef maps (Booth & Oldfield, 1989).

Finally, to validate the classification results, an error matrix was produced to assess the performance. We used independent training sites to assess the accuracy of each classification. The error matrix showed seven classes using four metrics: overall accuracy (OA), producer’s accuracy (PA), user’s accuracy (UA), and the Cohen Kappa coefficient. The Kappa coefficient was computed using Eq. (3): (3)${\displaystyle K=\frac{ObservedAccuracy-ChangeAgreement}{1-ChangeAgreement}}$

The Kappa coefficient measures the agreement between the classification and ground-truth values. A Kappa value of 1 represents perfect agreement, while a value of 0 represents no agreement. We interpreted the Kappa values as per Table 3 following (McHugh, 2012; Nguyen, Luong & Ho, 2015; Nguyen, Huynh & Zhang, 2015). All the image processing and post-classification steps, presented in Fig. 2, were processed using ArcGIS v. 10.8 software packages (ESRI, USA).

Descriptive field data analysis

The correlation coefficients and ratio of attenuation coefficients among visible bands at good water penetration (Fig. 3). To perform the water column correction, homogenous sand substrates of varying depths collected from field surveys were applied to calculate the ratio of attenuation coefficients. The very high correlation coefficients of 0.95 above were explored between bands 1 and 2, while the relationship between bands 1 and 3 showed high R² values of 0.85 in both ST and CLC.

In addition to the water column correction, the study analyzed the spectral signatures among the bottom types (i.e., coral lives and non-coral classes) based on 04 visible bands after corrected sun glint. Many studies measured the spectral characteristics from in situ data or from remotely sensed data (Green et al., 2000; Hochberg, Andréfouët & Tyler, 2003; Yamano & Tamura, 2004; Leiper et al., 2014; Giardino et al., 2019). These studies considered that separating spectral features among coral reef bottom classes is challenging. The classes with good separation are sand dominated vs deep water classes representing light and dark objects, respectively (Green et al., 2000; Hochberg, Andréfouët & Tyler, 2003). Our analysis shows that good discriminating ability is also found between the D and S classes, while HB bottom easily overlapped with 1LC and 2LC classes in the blue and green bands in both study areas. For live corals, the spectral characteristics utilized a separation among bands blue, red, and NIR (Fig. 4). The reflectance of bottom types among these classes decreased in the blue to red bands while increasing at the NIR band. Therefore, the high separation among bands permits the classification of live coral and bottom types in both areas.

The patterns of coral reefs distribution

From the analysis of field survey results, the most common coral reef types in the study areas were fringing reefs, reefs directly attached to the shoreline of the peninsula and the islands. This reef type varied in relation to a combination of coastline geomorphology and seabed characteristics, normally ranging between 30 m and 300 m in width. The vertical pattern of coral distribution in the study area reached to 14 m in depth and was densest between 2 and 6 m (Fig. 5).

The depth distribution of coral reefs and other substrates is presented in Figs. 6 and 7 for ST and CLC, respectively. Coral reefs in ST were located along the shoreline from Tien Sa in the northwest to Hon Sup in the southwest. The northwestern most coral reefs in this region were not included in this study due to the high turbidity of the Han estuary. High coral cover was found in the inner coral reefs of the northern ST, while coral cover was patchier in the southern region as shown by the alternating distribution of three classes (1LC, 2LC and 3LC; poor, fair, and good coral cover respectively) across the reef (Fig. 6). In southwest ST including Bai Nom, Bai But, and Hon Sup (Figs. 6D–6E), the reef width ranged between 140–300 m, while across the rest of ST reefs were narrower (30–90 m wide).

In CLC, coral reefs were located throughout the western and northern shores of Hon Lao (main island) and around small islands (i.e., Hon Kho, Hon La, Hon Mo, Hon Tai, and Hon Dai) (Figs. 7A, 7B, 7G, and 7H). High coral cover was found predominantly in the southern part of the smaller islands, while the northern shores of the islands recorded lower coral cover (Fig. 7). The widest reef (150 m) was the SunGroup reef (near the tourist area managed by SunGroup) while the reef width at the rest of the island locations ranged between 50–100 m.

Following the classification of the bottom cover into different categories (Table 2), we estimated that the live coral areas were 0.47 km² and 0.58 km² in the ST and CLC, respectively. Bottom cover proportion is reported in Table 4. In ST, the health condition of ranged from ‘poor’ to ‘fair condition’ with 60.25% classified as a coral reef in poor condition (1LC), 31.88% in ‘fair condition’ (2LC), and only 7.87% in ‘good condition’ (3LC). Several regions (e.g., Vung Da Ban and Tay Bai Bac - Dong Bai Bac) recorded over 70% of coral cover in poor condition, while none of the reefs were in good condition or above.

Compared to ST, the health condition of coral reefs in CLC was better and more consistent across reefs. Despite the better condition, the percentage of reefs in poor health still accounted for 34.28%, followed by 2LC and 3LC at 28.51% and 25.75%, respectively. However, the proportion of coral reefs in ‘excellent condition’ in the CLC was 11.46%, compared to none in the ST. This is perhaps due to several programs to protect and recover coral reefs in the CLC (Hua et al., 2015; Quach, 2018) such as the reef recovery program at some regions of Cu Lao Cham Island (e.g., Hon Tai, Hon La and Hon Dai, Bai Bac and Bai Huong), which have contributed to coral reef health (Hua et al., 2015).

Performance assessment

The Planetscope images provided three visible spectral bands that enabled us to penetrate to the water bottom and perform the classification of the benthos (Green et al., 2000; Mumby et al., 2004) and a NIR band that we used for sun-glint removal, subsequently improving the classification accuracy (Hedley, Harborne & Mumby, 2005; Anggoro, Siregar & Agus, 2016). Assessments of accuracy were carried out on the coral cover and bottom classes by comparison with geo-rectified ground reference data collected during field surveys. Error matrices were generated for each of the classified images in ST and CLC (Tables 5 and 6, respectively). The overall classification accuracies (Kappa coefficients) for ST and CLC were 78.08% (0.78) and 76.78% (0.76), respectively. Among the various coral and bottom types, deep zone and sands class showed the highest producer’s accuracies (PA) and user’s accuracy (UA) in both regions because of both the exponential decrease of light intensity with increasing depth and the high light absorption through the column before reaching to the deep bottom (Green et al., 2000; Goodman, Purkis & Phinn, 2013). In the coral cover classification, the highest PA (78.57%) and UA (80.85%) were found for 2LC and 3LC, respectively in ST, followed by the 1LC at 76.56% and 71.01%, respectively. Whereas these indicators were quite different in CLC. Major areas of confusion and misinterpretation were found to roughly coincide with the transitional areas between hard bottom (HB) and coral classes.