Satellite-Based Coral Reef Habitat Mapping in Weno Island Using Water Column Corrected High Spatial Resolution Image

Bara Samudra Syuhada; Deukjae Hwang; Taihun Kim; Jongkuk Choi

doi:10.7780/kjrs.2024.40.5.1.14

Quick links

Most View Most Download Archives Go to e-Submission

Research Article

Split Viewer

Korean J. Remote Sens. 2024; 40(5): 589-600

Published online: October 31, 2024

https://doi.org/10.7780/kjrs.2024.40.5.1.14

Satellite-Based Coral Reef Habitat Mapping in Weno Island Using Water Column Corrected High Spatial Resolution Image

Bara Samudra Syuhada^1,2, Deukjae Hwang³, Taihun Kim⁴, Jongkuk Choi^5,6*

Author Affiliations

¹UST Student, Korea Ocean Satellite Center, Korea Institute of Ocean Science and Technology, Busan, Republic of Korea
²Master Student, Major in Ocean Science, University of Science and Technology, Daejeon, Republic of Korea
³Postdoctoral Scientist, Korea Ocean Satellite Center, Korea Institute of Ocean Science and Technology, Busan, Republic of Korea
⁴Senior Research Scientist, Tropical & Subtropical Research Center, Korea Institute of Ocean Science and Technology, Jeju, Republic of Korea
⁵Principal Research Scientist, Korea Ocean Satellite Center, Korea Institute of Ocean Science and Technology, Busan, Republic of Korea
⁶Professor, Major in Ocean Science, University of Science and Technology, Daejeon, Republic of Korea

Correspondence to : Jongkuk Choi
E-mail: jkchoi@kiost.ac.kr

Received: October 7, 2024; Revised: October 24, 2024; Accepted: October 25, 2024

This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (https://creativecommons.org/licenses/by-nc/4.0/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract

Go to

Abstract
1. Introduction
2. Materials and Methods
3. Results
4. Discussion
5. Conclusions
Acknowledgments
Conflict of Interest
REFERENCES

Coral reefs play significant roles in marine ecosystems, and recently, they have been experiencing degradation primarily due to global warming. Monitoring the coral reef ecosystem is crucial to rehabilitation and preventing further degradation. Here, we used high spatial resolution multispectral image data from the QuickBird sensor and in-situ measurements acquired around 2011 to derive a benthic habitat map around the coral reef ecosystem in Weno Island, Micronesia. Water column correction was performed to eliminate the effect of the light attenuation within the water column from the satellite image, employing a band combination approach known as the depth invariant index (DII) transformation. The combination of the new images generated by the DII transformation was used for image segmentation. This approach was conducted to apply object-based image classification. To determine the accuracy of the classification, we separate the in-situ data into 174 training data (70% of the data) and 75 testing data (30% of the data). This study produced classification results with an overall accuracy of 84% and a kappa value of 0.77, using a scale parameter of 5 for the object-based classification, which supported the reliability of the resultant coral reef habitat map. The findings of this study demonstrate that applying the depth invariant algorithm for water column correction on Weno Island is an appropriate step before conducting benthic habitat classification.

Keywords Coral reef habitat mapping, QuickBird, Water column correction, Object-based image classification, Weno Island

1. Introduction

Go to

Abstract
1. Introduction
2. Materials and Methods
3. Results
4. Discussion
5. Conclusions
Acknowledgments
Conflict of Interest
REFERENCES

Coral reefs play a crucial role in marine ecosystems, providing shelter and habitats for diverse fish species (Coker et al., 2014). Unfortunately, these reefs are facing global degradation due to global warming and human activities (El-Naggar, 2020; Feary et al., 2007; Pandolfi et al., 2011). Rising sea surface temperatures and ocean acidification have led to coral bleaching and subsequent habitat loss (Goreau et al., 2012; Pandolfi et al., 2011). As a result, monitoring coral reef ecosystems has become essential for both rehabilitation efforts and the prevention of further degradation (Choi et al., 2021).

Remote sensing technology has been increasingly utilized in studies focused on monitoring and mapping bottom types in coastal waters. For example, Widya et al. (2023) employed several types of satellite imagery, i.e., GeoEye-1, Sentinel-2, and Landsat-8 to map seagrass distribution along the Eastern Coast of South Korea. Similarly, Choi et al. (2021) used high-spatial-resolution satellite images, such as those from the Kompsat-2 satellite, to map changes in coral reef habitats on Weno Island. In another region, da Silveira et al. (2021) combined WorldView-3 and Landsat-8 imagery to map coral reefs around Tamandaré, Brazil. They demonstrated how remote sensing can assist decision-making for coral reef management. Araujo et al. (2023) applied Sentinel-2 imagery to geomorphological mapping in the Costa dos Corais marine protected area, in Brazil. Their results indicated that the study was applicable to support the management and monitoring of the marine protected area.

However, when using remote sensing to gather information about underwater features, a significant challenge arises due to the effects of the water column (Zoffoli et al., 2014). To address this issue, an algorithm was developed to discriminate between bottom features in underwater environments that had similar reflectance spectra (Lyzenga, 1978). This algorithm has since been refined and the resulting output is widely known as the depth invariant index (DII) (Aljahdali and Elhag, 2020; Widya et al., 2023). The method can be applied to the classification of marine habitats using multispectral remote sensing data (Mumby and Edwards, 2000). To identify seagrass distribution along the Eastern Coast of South Korea, Widya et al. (2023) applied the DII calculation to remote sensing data. The results of the study produced a high overall accuracy. Another study by Ahmed et al. (2020), utilized Landsat 7 and Landsat 8 to generate benthic habitat maps in marine protected areas in Kenya. They also implemented the DII model and obtained acceptable accuracy.

In this study, coral reef mapping is conducted on Weno Island using satellite imagery and available in-situ data. The DII method is applied specifically to habitat mapping in the coral reef ecosystem. A set of train data and test data are selected from the in-situ observations on the bottom types in the study area. New images derived from the DII algorithm are applied for the bottom type using the object-based classification method, a method proven effective for classification based on high spatial resolution remote sensing images (Choi et al., 2010). The accuracy of the resulting classification is assessed to evaluate its effectiveness. This study will be able to support the policy decisions for managing and preserving coral reef systems.

2. Materials and Methods

Go to

Abstract
1. Introduction
2. Materials and Methods
3. Results
4. Discussion
5. Conclusions
Acknowledgments
Conflict of Interest
REFERENCES

2.1. Study Area

The Federated States of Micronesia (FSM) is situated in the western Pacific Ocean, near the equator, and has continuously experienced a tropical climate. FSM is composed of four states, i.e., Kosrae, Yap, Pohnpei, and Chuuk. The nation relies heavily on coral reefs for its tourism, fisheries, and recreational activities (George et al., 2008). The average air temperature in the FSM is approximately 28°C, with two distinct seasons. The dry season spans from November to April, while the wet season extends from May to October. Since 1951, the annual and seasonal mean air temperatures in the FSM have shown an upward trend. The monthly sea surface temperature averages around 29°C. However, the warming ocean has led to a rise in sea levels of over 10 mm per year since 1993 (Federated States of Micronesia National Weather Service Office, 2011).

Chuuk State is composed of five island regions, i.e., Chuuk Lagoon, Mortlocks, Pattiw, Halls, and Nomunweito. Chuuk Lagoon (Fig. 1a), the region’s hub of human activity, consists of several islands, including Weno Island (Fig. 1b), which covers an area of about 20 square kilometers (George et al., 2008). The specific study area is located on the eastern side of Weno Island (Fig. 1c).

Fig. 1. Location of the study area: (a) Chuuk Lagoon, (b) Weno Island, and (c) area of benthic habitat mapping.

Weno Island is one of the 607 islands in the Federated States of Micronesia (George et al., 2008). It is located in a tropical region, with an annual average atmospheric temperature of 27°C. The annual sea surface temperature ranges from 28–29°C and the area’s annual average precipitation is between 3,000 mm and 10,000 mm (Choi et al., 2021). The coastal area of Weno Island features fringing reefs that extend outward from the reef flat, across the reef crest, and down to the reef slope (Kim et al., 2022).

2.2. Data

The multispectral imagery used in this study comes from QuickBird, which provides high-spatial resolution data. This dataset includes three visible bands as well as one near-infrared band (Table 1). The imagery was acquired on April 19, 2011. After applying water column correction, three combinations of visible bands were produced.

Table 1 QuickBird multispectral product specification

Image bands	Blue: 485 nm
	Green: 560 nm
	Red: 660 nm
	Near-infrared: 830 nm
Resolution	2.44 m (at nadir)
Digitization	11 bits
Metric accuracy	23-meter horizontal
Metric accuracy	Circular error at the 90th percentile (CE90%)

In-situ observations of bottom types were collected from previous studies. A total of 109 sampling points (Choi et al., 2021) were gathered between September 14 and 21, 2011, through snorkeling and underwater photography at each location. These observation locations were recorded using a Garmin Oregon 500 Global Positioning System (GPS). An additional 140 observations (Kim et al., 2022) were collected between February 2011 and October 2012 through walking, snorkeling, or scuba diving, with the coordinates of the positions recorded using a Garmin Oregon 600 GPS tracking device.

For this study, a total of 249 observations from field surveys (Fig. 2a) were divided into training and test datasets. The training dataset comprises 70% of the total, resulting in 174 points (Fig. 2b), while the remaining 75 points were designated for testing (Fig. 2c). The bottom types collected were categorized into six classes: coral, short seagrass, large seagrass, sand + seagrass, sand, and rubble (Table 2).

Fig. 2. Location of sampling points. (a) Location of the points for each class. (b) Sampling points for the train data. (c) Sampling points for the test data.

Table 2 Number of data for each class

Class	Train data	Test data	Overall
Coral	23	7	30
Large seagrass	35	16	51
Short seagrass	5	3	8
Sand + Seagrass	10	6	16
Sand	73	34	107
Rubble	28	9	37
Total	174	75	249

2.3. Pre-Processing of the Remote Sensing Image

Passive remote sensing relies on solar energy to collect data from the Earth’s surface. Each surface feature uniquely interacts with sunlight, producing distinct spectral responses. However, before this signal reaches the satellite, it undergoes several processes that can alter its characteristics. In the atmosphere, the signal is disturbed due to absorption and scattering by atmospheric particles. When mapping underwater features, the signal is further influenced by water constituents. Therefore, appropriate correction methods are essential before classifying benthic habitats in coral reef areas (Zoffoli et al., 2014). The overall methodology for this study is illustrated in Fig. 3.

Fig. 3. Flowchart for benthic habitat mapping.

The raw image data, initially recorded as digital numbers (DN), were first converted into radiance using Eq. (1). The values for Gain_λ and Offset_λ were obtained from the DigitalGlobe documentation related to the absolute radiometric calibration of image products. The abscal factor and effectivebandwith were provided in the metadata.

(1)Lλ=Gainλ×abscal factoreffectivebandwith⋅DNλ+Offsetλ

The next step involved eliminating the atmospheric effect from the signal through atmospheric correction. The Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) algorithm is particularly suited for atmospheric correction in coral reef environments (Kondraju et al., 2022). Given that the study area is located in a tropical region, the tropical atmospheric model was selected from the MODTRAN atmospheric model within the ENVI software. Additionally, the maritime aerosol model was chosen in the preliminary settings for atmospheric correction.

2.4. Water Column Correction

The pixel information in the imagery still reflects the influence of the water column. This effect occurs due to the absorption and scattering processes by the interaction of the electromagnetic radiation and the optically active constituents in the water column (Zoffoli et al., 2014). To address this issue, an appropriate water column correction is necessary.

In this study, a band combination algorithm, specifically the DII algorithm, was employed to correct for water column effects. Lyzenga (1978) developed the algorithm and suggested linearizing the approximate relationship between the radiance and the water depth. The slope of the relation between the transformed model of one band to another indicates the ratio of the attenuation coefficient of those two bands (Fig. 4). The DII algorithm is well-suited for multispectral data (Widya et al., 2023) and was applied using the Sen2Coral toolbox in the SNAP software. A subset of the image was extracted for the water column correction process.

Fig. 4. Transformation model of Lyzenga (1978) method.

The first step in the correction process involved linearizing the radiance of the selected pixels representing the same substrate using a logarithmic transformation (Eq. 2):

(2)Xi=InLi−Lsi

where L_i is the atmospherically corrected radiance which represents the low bottom albedo at band i, L_si is the atmospherically corrected radiance for deep water around the selected mapping area which represent the high bottom albedo at band i, and X_i is the logarithmic transformation of the difference between L_i and L_si. Next, the DII was computed using Eq. (3):

(3)DIIij=Xi−ki kj ⋅Xj

where DII_ij is the Depth Invariant Index for the composition of band i and band j, k_i/k_j is the ratio of attenuation coefficient between band i and band j. This ratio can be calculated using Eq. (4):

(4)kikj=a+a2+1

where,

(5)a=σii−σij2σij

and σ_ii and σ_jj in Eq. (5) are the variance of the X_i and X_j of the selected same substrate pixels, while σ_ij is the covariance of X_i and X_j.

2.5. Object-Based Classification

The classification process utilized object-based classification rather than pixel-based classification. While pixel-based classification assigns classes based on individual pixel values, object-based classification first groups pixels with similar textural and contextual properties through an image segmentation process. This approach is particularly well-suited for classifying coral reef benthic habitats (Choi et al., 2021) as it mitigates the salt-pepper effect commonly associated with pixel-based methods and reduces spectral variation within each class (Liu and Xia, 2010).

The object-based classification was performed using eCognition Developer software for its ability to perform the particular task. The procedure began with the multiresolution segmentation of the water column-corrected images. Following segmentation, classes were assigned to objects that contained training data. The assignment was initially automated and subsequently verified manually. The objects that were assigned classes served as training samples for the classification. The classification was carried out using the nearest neighbor classification method.

3. Results

Go to

Abstract
1. Introduction
2. Materials and Methods
3. Results
4. Discussion
5. Conclusions
Acknowledgments
Conflict of Interest
REFERENCES

3.1. Benthic Habitat Map

Three new images were generated following the water column correction procedure using combinations of visible bands from QuickBird imagery. These three new images were then combined to create a composite image for the classification. Each of the images is based on combinations of the blue and the green bands (DIIb1b2), the green and the red bands (DIIb2b3), and the blue and the red band (DIIb1b3), respectively.

Object-based classification was performed for each of the three DII images by adjusting the scale parameter during the segmentation process, which is a critical setting that controls the size of the image objects. Four maps were generated with different scale parameters for the segmentation process. Different scale parameter settings influence the classification results, as higher scale parameters generate larger objects (Trimble, 2019). The maps were assessed to see which scale parameter was best for benthic habitat mapping in the study area using QuickBird imagery. First, a benthic habitat map was produced using a scale parameter of 3, referred to as SP3 Map (Fig. 5a). SP5 Map (Fig. 5b) was generated using a scale parameter of 5. SP7 Map (Fig. 5c) and SP10 Map (Fig. 5d) were the results of benthic habitat classification with scale parameters set to 7 and 10, respectively.

Fig. 5. Benthic habitat maps: (a) SP3 Map, (b) SP5 Map, (c) SP7 Map, and (d) SP10 Map.

3.2. Classification Accuracy

Error matrices were produced to evaluate the mapping accuracy of the benthic habitat classification. The producer’s accuracy is the accuracy for each class of the map based on the test data, while the user’s accuracy measures the precision of each class of the test data from the perspective of the resulting map (Nicolau et al., 2024).

The overall accuracy of the benthic habitat classification was 77.33% for SP3 Map (Table 3), SP7 Map (Table 5), and SP10 Map (Table 6). The highest accuracy was 84% for the SP5 Map (Table 4). Kappa coefficient (Cohen, 1960) was also calculated to assess the level of agreement between the classification results and the test data. The kappa coefficient for SP3 Map, SP7 Map, and SP10 Map was 0.68, indicating a moderate level of agreement (McHugh, 2012). SP5 Map also has a moderate level of agreement with higher kappa coefficient value of 0.77.

Table 3 Error matrix of the benthic habitat classification in SP3 Map

		Image classification							Producer’s accuracy
		Coral	Large Seagrass	Short Seagrass	Sand + Seagrass	Rubble	Sand	Total	Producer’s accuracy
Test data	Coral	4	1	0	0	1	1	7	0.57
	Large Seagrass	0	12	1	1	2	0	16	0.75
	Short Seagrass	0	1	1	1	0	0	3	0.33
	Sand + Seagrass	1	2	0	2	0	1	6	0.33
	Rubble	0	0	0	0	7	2	9	0.78
	Sand	2	0	0	0	0	32	34	0.94
	Total	7	16	2	4	10	36	75
User’s accuracy		0.57	0.75	0.50	0.50	0.70	0.89
Overall accuracy									0.773333333
Kappa coefficient									0.68

Table 4 Error matrix of the benthic habitat classification in SP5 Map

		Image classification							Producer’s accuracy
		Coral	Large Seagrass	Short Seagrass	Sand + Seagrass	Rubble	Sand	Total	Producer’s accuracy
Test data	Coral	6	0	0	0	1	0	7	0.86
	Large Seagrass	0	15	0	1	0	0	16	0.94
	Short Seagrass	0	1	2	0	0	0	3	0.67
	Sand + Seagrass	1	2	0	1	0	2	6	0.17
	Rubble	1	0	0	1	6	1	9	0.67
	Sand	1	0	0	0	0	33	34	0.97
	Total	9	18	2	3	7	36	75
User’s accuracy		0.67	0.83	1.00	0.33	0.86	0.92
Overall accuracy									0.84
Kappa coefficient									0.77

Table 5 Error matrix of the benthic habitat classification in SP7 Map

		Image classification							Producer’s accuracy
		Coral	Large Seagrass	Short Seagrass	Sand + Seagrass	Rubble	Sand	Total	Producer’s accuracy
Test data	Coral	5	0	0	0	1	1	7	0.71
	Large Seagrass	0	13	0	2	0	1	16	0.81
	Short Seagrass	0	1	1	1	0	0	3	0.33
	Sand + Seagrass	1	1	0	2	0	2	6	0.33
	Rubble	1	0	0	0	6	2	9	0.67
	Sand	2	0	0	0	1	31	34	0.91
	Total	9	15	1	5	8	37	75
User’s accuracy		0.56	0.87	1.00	0.40	0.75	0.84
Overall accuracy									0.773333333
Kappa coefficient									0.68

Table 6 Error matrix of the benthic habitat classification in SP10 Map

		Image classification							Producer’s accuracy
		Coral	Large Seagrass	Short Seagrass	Sand + Seagrass	Rubble	Sand	Total	Producer’s accuracy
Test data	Coral	7	0	0	0	0	0	7	1.00
	Large Seagrass	1	13	0	1	0	1	16	0.81
	Short Seagrass	0	1	1	0	0	1	3	0.33
	Sand + Seagrass	1	2	0	2	0	1	6	0.33
	Rubble	3	0	0	0	4	2	9	0.44
	Sand	3	0	0	0	0	31	34	0.91
	Total	15	16	1	3	4	36	75
User’s accuracy		0.47	0.81	1.00	0.67	1.00	0.86
Overall accuracy									0.773333333
Kappa coefficient									0.68

There are 17 misclassified points out of 75 test points for SP3 Map, SP7 Map, and SP10 Map. SP5 Map has 12 misclassified points. Each result has a different misclassified case as describe in Table 7. The location of the points shown in Fig. 2(c).

Table 7 Misclassified data cases for each result

Case of misclassified data		Misclassified point
Class	Misclassified as	SP3 Map	SP5 Map	SP7 Map	SP10 Map
Coral	Large Seagrass	1 point (TL2-85)	-	-	-
	Rubble	1 point (ch10_234)	1 point (ch10_234)	1 point (ch10_234)	-
	Sand	1 point (ch10_269)	-	1 point (TL2- 64)	-
Large Seagrass	Coral	-	-	-	1 point (TL3-14)
	Short Seagrass	1 point (ch10_249)	-	-	-
	Sand + Seagrass	1 point (TL2-26)	1 point (TL2-26)	2 points (TL3-05 and ch08_211)	1 point (ch08_211)
	Rubble	2 points (TL3-05 and ch10_215)	-	-	-
	Sand	-	-	1 point (TL2-26)	1 point (TL2-26)
Short Seagrass	Large Seagrass	1 point (sp192)	1 point (sp192)	1 point (sp192)	1 point (CH0A)
	Sand + Seagrass	1 point (ch08_172)	-	1 point (ch08_172)	-
	Sand	-	-	-	1 point (sp192)
Sand + Seagrass	Coral	1 point (ch10_255)	1 point (ch10_255)	1 point (ch10_255)	1 point (ch10_255)
	Large Seagrass	2 points (TL2-08 and ch10_245)	2 points (TL2-08 and ch10_245)	1 point (TL2-08)	2 points (Tl2-08 and CH10_220)
	Sand	1 point (TL2-25)	2 points (TL2-25 and ch10_220)	2 points (TL3-05 and ch08_211)	1 point (TL2-25)
Rubble	Coral	-	1 point (ch10_266)	1 point (ch10_266)	3 points (TL2-82, ch10_266, and TL3-47)
	Sand + Seagrass	-	1 point (ch10_274)	-	-
	Sand	2 points (TL3-44 and ch08_074)	1 point (ch08_074)	2 points (TL2-77 and ch08_074)	2 points (TL2-77 and ch08_074)
Sand	Coral	2 points (TL2-72 and ch08_193)	1 point (ch08_193)	2 points (TL2-72 and ch08_193)	3 points (ch10_226, ch10_271, and ch08_193
Sand	Rubble	-	-	1 point (ch10_271)	-
Total		17 points	12 points	17 points	17 points

4. Discussion

Go to

Abstract
1. Introduction
2. Materials and Methods
3. Results
4. Discussion
5. Conclusions
Acknowledgments
Conflict of Interest
REFERENCES

4.1. Benthic Habitat Classification

The water column correction was conducted before classification to eliminate the water column effect and enhance the visibility of benthic features for classification. Various scale parameters were tested for the image segmentation process to find the most suitable parameter for the study area. The four resulting maps, generated using different scale parameters, are presented in Fig. 5.

The benthic habitat map was classified into six classes, same as the in-situ sampling classes. The optimal scale parameter for classification was 5 (SP5 Map), yielding an overall accuracy of 84% and a kappa coefficient of 0.77 (Table 4). Out of 75 test samples, 12 were misclassified. SP5 Map produced the highest producer’s accuracy for the “Short Seagrass”, “Large Seagrass”, and “Sand” classes with only one misclassified sample for each class. The “Coral” class also had one misclassified sample, although SP10 Map achieved 100% producer’s accuracy for the “Coral” class.

These results demonstrate the significant effect of the scale parameter on classification outcomes. Higher scale parameters produced larger segmentation objects. Thus, the selection of the appropriate scale parameter of the segmentation is crucial for accurate object-based classification. In this study area, the best scale parameter with the classification scheme is 5.

While the overall accuracy and kappa coefficient indicate good agreement between the segmented images and the test data, some misclassifications occurred due to various factors. One reason for the misclassification is that in-situ data sampling was based on personal interpretation. In the field, there might be subjective distinctions between short and large seagrass, with short seagrass being perceived as denser in the study area. Another potential factor is that some samples may have been located at the object boundaries within the image segmentation results like TL2-85 in segmentation scheme with scale parameter set to 3 (Fig. 6), which might have led the misclassification in the process of assigning class to the object.

Fig. 6. Example of data located on the boundary of object.

Although visually similar, classes such as “Short Seagrass”, “Large Seagrass”, and “Sand + Seagrass” differ in spectral characteristics. Short seagrass typically has a denser distribution compared to large seagrass, leading to differences in their spectral signatures. In other cases, the “Rubble” and “Sand” classes can be confused due to similar reflectance in the green and blue bands (Choi et al., 2021). Rubble also can be interpreted as dead coral, leading to confusion between “Coral”, “Rubble”, and “Sand” classes. The subjective interpretation of the in-situ sampling observation could have also contributed to misclassifications.

4.2. Water Column Correction Significance

The water column plays a critical role in the classification of benthic habitats by affecting the light that reaches the bottom. As light travels through water, it is absorbed and scattered by water constituents, which can distort the reflectance of benthic features. This scattering and absorption lead to challenges when classifying benthic habitats with varying depths (Zoffoli et al., 2014). To mitigate these effects and improve classification accuracy, a water column correction was applied in this study.

In a previous study conducted by Choi et al. (2021) in the same study area, habitat maps were generated using Kompsat-2 high-spatial resolution imagery from 2008 and 2010. They achieved overall accuracies of 78.6% and 72.4% for the respective years using an object-based classification approach with Red-Green-Blue (RGB) reflectance bands. However, their classification did not incorporate a water column correction.

In contrast, this study applied water column correction using the method introduced by Lyzenga (1978) to produce new corrected images. To see the significance of the water column correction procedure, a classification using the image prior to water column correction was conducted. Using the RGB bands of the corrected image and the same training and test samples, object-based classification was performed by setting the scale parameter to 5 (Fig. 7). This classification produces a map with an overall accuracy of 56% and a Kappa coefficient of 0.4 (Table 8).

Fig. 7. Benthic habitat map result without water column correction procedure.

Table 8 Error matrix of the benthic habitat classification without water column correction procedure

		Image classification							Producer’s accuracy
		Coral	Large Seagrass	Short Seagrass	Sand + Seagrass	Rubble	Sand	Total	Producer’s accuracy
Test data	Coral	2	2	0	0	1	1	6	0.33
	Large Seagrass	2	7	1	2	2	2	16	0.44
	Short Seagrass	0	1	1	1	0	0	3	0.33
	Sand + Seagrass	1	1	1	0	2	1	6	0.00
	Rubble	1	2	0	1	4	0	8	0.50
	Sand	3	0	0	2	1	28	34	0.82
	Total	9	13	3	6	10	32	75
User’s accuracy		0.22	0.54	0.33	0.00	0.40	0.88
Overall accuracy									0.56
Kappa coefficient									0.40

The water column correction improved the classification accuracy to 84% (Table 4). By using 70% of the in-situ data for training (Fig. 2b) and 30% for testing (Fig. 2c) we ensure a more reliable assessment of classification performance. The improvement in accuracy highlights the significance of the water column correction procedure for benthic habitat classification. Additionally, fine-tuning the scale parameter was essential to optimize segmentation and further enhance classification accuracy.

5. Conclusions

Go to

Abstract
1. Introduction
2. Materials and Methods
3. Results
4. Discussion
5. Conclusions
Acknowledgments
Conflict of Interest
REFERENCES

The water column correction procedure has proven to enhance the accuracy of benthic habitat mapping in the coral reef area of Weno Island, using high-spatial resolution imagery and in-situ data. This study could achieve an overall classification accuracy of 84% and a kappa coefficient of 0.77 with a scale parameter set to 5 for the segmentation process. This result indicates a good agreement between the classification output and the test data. Fine-tuning the scale parameter for image segmentation was crucial for optimizing object-based classification results. Additionally, the selection of training and test data also likely contribute to the overall improvement in accuracy.

For future study, it is essential to periodically monitor coral reef areas to assess changes in coral reef extent. We have recently carried out in-situ observations in the study area and comparisons of changes in the areal extent of coral reef between two periods should be estimated in the next study. Analyzing temporal changes in coral reefs in relation to sea surface temperature and other parameters associated with global warming could also provide insights into how these factors affect coral bleaching in Weno Island. Such studies will be valuable for understanding the impacts of global warming on coral reef ecosystems and for developing strategies to mitigate these effects.

Acknowledgments

Go to

Abstract
1. Introduction
2. Materials and Methods
3. Results
4. Discussion
5. Conclusions
Acknowledgments
Conflict of Interest
REFERENCES

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (NRF-2021R1A2C1011416).

Conflict of Interest

Go to

Abstract
1. Introduction
2. Materials and Methods
3. Results
4. Discussion
5. Conclusions
Acknowledgments
Conflict of Interest
REFERENCES

The authors declare no potential conflict of interest related to this article.

REFERENCES

Go to

Abstract
1. Introduction
2. Materials and Methods
3. Results
4. Discussion
5. Conclusions
Acknowledgments
Conflict of Interest
REFERENCES

Ahmed, A. F., Mutua, F. N., and Kenduiywo, B. K., 2020. Monitoring benthic habitats using Lyzenga model features from Landsat multi-temporal images in Google Earth Engine. Modeling Earth Systems and Environment, 7(3), 2137-2143. https://doi.org/10.1007/s40808-020-00960-1

Crossref
Aljahdali, M. H., and Elhag, M., 2020. Calibration of the depth invariant algorithm to monitor the tidal action of Rabigh City at the Red Sea Coast, Saudi Arabia. Open Geosciences, 12(1), 1666-1678. https://doi.org/10.1515/geo-2020-0217

Crossref
Araujo, J. C., Seoane, J. C. S., Lima, G. V., da Silva, E. G., França, L. G., and de Souza Santos, E. E., et al, 2023. High-resolution optical remote sensing geomorphological mapping of coral reef: Supporting conservation and management of marine protected áreas. Journal of Sea Research, 196, 102453. https://doi.org/10.1016/j.seares.2023.102453

Crossref
Choi, J. K., Ryu, J. H., Lee, Y. K., Yoo, H. R., Jun, H., and Hwan, C., 2010. Quantitative estimation of intertidal sediment characteristics using remote sensing and GIS. Estuarine, Coastal and Shelf Science, 88(1), 125-134. https://doi.org/10.1016/j.ecss.2010.03.019

Crossref
Choi, J. K., Ryu, J. H., and Min, J. E., 2021. Application of high-spatial-resolution satellite images to monitoring coral reef habitat changes at Weno Island Chuuk, Micronesia. Korean Journal of Remote Sensing, 37(4), 687-698. https://doi.org/10.7780/kjrs.2021.37.4.2

Crossref
Cohen, J., 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1), 37-46. https://doi.org/10.1177/001316446002000104

Crossref
Coker, D. J., Wilson, S. K., and Pratchett, M. S., 2014. Importance of live coral habitat for reef fishes. Reviews in Fish Biology and Fisheries, 24(1), 89-126. https://doi.org/10.1007/s11160-013-9319-5

Crossref
da Silveira, C. B. L., Strenzel, G. M. R., Maida, M., Gaspar, A. L. B., and Ferreira, B. P., 2021. Coral reef mapping with remote sensing and machine learning: A nurture and nature analysis in marine protected areas. Remote Sensing, 13(15), 2907. https://doi.org/10.3390/rs13152907

Crossref
El-Naggar, H. A., 2020. Human impacts on coral reef ecosystem. In: Rhodes, E. R., Naser, H., (eds.), Natural resources management and biological sciences, IntechOpen, pp. 91-107. https://doi.org/10.5772/intechopen.88841

Crossref
Feary, D. A., Almany, G. R., McCormick, M. I., and Jones, G. P., 2007. Habitat choice, recruitment and the response of coral reef fishes to coral degradation. Oecologia, 153(3), 727-737. https://doi.org/10.1007/s00442-007-0773-4

Pubmed Crossref
Federated States of Micronesia National Weather Service Office, 2011. Pacific-Australia climate change science and adaptation planning program: Current and future climate of the Federated States of Micronesia, Australian Bureau of Meteorology, Commonwealth Scientific and Industrial Research Organisation.
George, A., Luckymis, M., Palik, S., Adams, K., Joseph, E., Mathias, D., and Slingsby, S., 2008. The state of coral reef ecosystems of the Federated States of Micronesia. Federated States of Micronesia, pp. 419-436.
Goreau, T. J., Hayes, R. L., and McAllister, D., 2012. Regional patterns of sea surface temperature rise: Implications for global ocean circulation change and the future of the coral reef and fisheries. World Resource Review, 17(3), 350-374 .
Kim, T., Lee, D. W., Kim, H. J., Jung, Y. H., Choi, Y. U., and Oh, J. H., et al, 2022. Estimation of the benthic habitat zonation by photo-quadrat image analysis along the fringing reef of Weno Island, Chuuk, Micronesia. Journal of Marine Science and Engineering, 10(11), 1643. https://doi.org/10.3390/jmse10111643

Crossref
Kondraju, T. T., Mandla, V. R., Chokkavarapu, N., and Peddinti, V. S. S., 2022. A comparative study of atmospheric and water column correction using various algorithms on Landsat imagery to identify coral reefs. Regional Studies in Marine Science, 49, 102082. https://doi.org/10.1016/j.rsma.2021.102082

Crossref
Liu, D., and Xia, F., 2010. Assessing object-based classification: advantages and limitations. Remote Sensing Letters, 1(4), 187-194. https://doi.org/10.1080/01431161003743173

Crossref
Lyzenga, D. R., 1978. Passive remote sensing techniques for mapping water depth and bottom features. Applied Optics, 17(3), 379. https://doi.org/10.1364/ao.17.000379

Pubmed Crossref
McHugh, M. L., 2012. Interrater reliability: The kappa statistic. Biochemia Medica, 22(3), 276-282 .

Pubmed PMC Crossref
Mumby, P. J., and Edwards, A. J., 2000. Water column correction techniques. In: Green, E. P., Mumby, P. J., Edwards, A. J., Clark, C. D., (eds.), Remote Sensing Handbook for Tropical Coastal Management, UNESCO, pp. 121-128.
Nicolau, A. P., Dyson, K., Saah, D., and Clinton, N., 2024. Accuracy assessment: Quantifying classification quality. In: Cardille, A. J., Crowley, M. A., Saah, D., Clinton, N., (eds.), Cloud-Based Remote Sensing with Google Earth Engine, Springer, pp. 135-145. https://doi.org/10.1007/978-3-031-26588-4

Pubmed Crossref
Pandolfi, J. M., Connolly, S. R., Marshall, D. J., and Cohen, A. L., 2011. Projecting coral reef futures under global warming and ocean acidification. Science, 333(6041), 418-422. https://doi.org/10.1126/science.1204794

Pubmed Crossref
Trimble, 2019. Documentation eCognition suite. Available online: https://docs.ecognition.com/v9.5.0/eCognition_documentation/User%20Guide%20Developer/4%20Basic%20Rule%20Set%20Editing.htm (accessed on Sept. 29, 2024)
Widya, L. K., Kim, C. H., Do, J. D., Park, S. J., Kim, B. C., and Lee, C. W., 2023. Comparison of satellite imagery for identifying seagrass distribution using a machine learning algorithm on the eastern coast of South Korea. Journal of Marine Science and Engineering, 11(4), 701. https://doi.org/10.3390/jmse11040701

Crossref
Zoffoli, M. L., Frouin, R., and Kampel, M., 2014. Water column correction for coral reef studies by remote sensing. Sensors, 14(9), 16881-16931. https://doi.org/10.3390/s140916881

Pubmed PMC Crossref

Research Article

Korean J. Remote Sens. 2024; 40(5): 589-600

Published online October 31, 2024 https://doi.org/10.7780/kjrs.2024.40.5.1.14