High-resolution benthic habitat mapping from machine learning on PlanetScope imagery and ICESat-2 data

Abstract This study proposed a comprehensive approach that utilized PlanetScope imagery for classifying tropical-marine benthic habitats after retrieving bathymetry from ICESat-2 data and water-column correction for areas around Lyson Islands, Vietnam. Exact bathymetry derivation and water column correction were applied to the PlanetScope imagery, making it an effective method for mapping marine benthic habitats. Water column correction was achieved by applying Depth Invariant Index (DII) and Bottom Reflectance Index (BRI). Moreover, two conventional machine learning algorithms, including Random Forest and Support Vector Machine, and a current deep Convolutional Neural Network (CNN) was employed to classify the benthic features. The overall accuracy of these classifiers are 80.74%, 84.19%, and 89.80% with the BRI, 80.17%, 82.75%, and 87.85% with the DII compared to 37.64%, 42.5%, and 47.2% of without corrected water columns respectively. The CNN model demonstrated that the approach significantly maximizes the improvement in benthic classification results in coastal region.


Introduction
Marine coastal habitat classification and mapping is a prerequisite for the assessment of the status of marine ecosystems which provides an effective communication tool for environmental management decisions (Fincham et al. 2020;Wilson et al. 2021;Janowski et al. 2022).One of the technology developed in the past three decades that can provide valuable information on coastal geomorphology and a synoptic view of ecology is satellite remote sensing (Green 2000;Mumby et al. 2004;Le Quilleuc et al. 2021).Numerous remote sensing techniques have been employed to map general benthic habitat types (e.g.sand, seagrass, coral reefs, hard substrate) in coastal environments (Mishra et al. 2005;Quang et al. 2015;Wilson et al. 2021;Janowski et al. 2022).Most of them focus on coral reef investigations such as mapping benthic habitats (Stolt et al. 2011;Eugenio et al. 2015;Hedley et al. 2016;Wicaksono and Lazuardi 2018;Li et al. 2019) and monitoring coral reef changes (Mishra et al. 2006;Strong et al. 2006;Andr efou€ et et al. 2007).Nonetheless, mapping the bottom substrate in complex and heterogeneous environments remains challenging (Zajac 2008;Rigby et al. 2010;Lecours et al. 2015).The influence of seawater environment and water depth leads to different reflectance in remote sensing data, which directly affects the benthic mapping results.In other words, water column correction is the essential key to improving the classification results of coastal benthic habitats.
Several methods and models were developed to reduce the water column effect in order to extract the actual bottom reflectance (Lyzenga 1981;Brando et al. 2009;Sagawa. et al. 2010;Zoffoli et al. 2014, Vinayaraj et al. 2016).The two main approaches currently widely used are empirical and physical-based methods.The physical-based approach tries to model the relationships between the water column's inherent optical properties, water depth, bottom reflectance, and remote sensing reflectance (Maritorena et al. 1994;Lee et al. 1998;Le Quilleuc et al. 2021).Afterward, some inversion techniques such as optimal algorithm (Lee et al. 1999;Lee et al. 2001;Klonowski et al. 2007;Brando et al. 2009) and adaptive look-up table (Mobley et al. 2005;Hedley et al. 2009) were applied to retrieve the properties of the water column, bathymetry, and bottom reflectance simultaneously.However, these techniques require complicated inputs which can only obtained through field investigations.
On the other hand, the empirical approaches attempted to simplify their model under some specific assumptions and make them more reliable, compatible with different conditions of the field sites.Therefore, the empirical techniques almost do not require the field water column data, which are often difficult to be acquired by optical or multispectral remote sensing systems (Schimel et al. 2020).The reliable and straightforward Deep-Invariant Index (DII) proposed by Lyzenga (1981) is the most popular method to correct the effect of the water layer (Andr efou€ et et al. 2003;Louchard et al. 2003;Wicaksono et al. 2019).Mumby et al. (1998) recognized that DII could improve the benthic map's accuracy by 13%.However, the technique of Lyzenga is only appropriate for transparent water, which is unsuitable for most coastal regions.Sagawa et al. (2010) proposed an alternative strategy to overcome the low transparency of turbid water.In Sagawa's approach, the water depth data was required to estimate the attenuation coefficient and retrieve bottom reflectance index.This bathymetry-based method is expected to increase the accuracy of the marine habitat classification results (Zoffoli et al. 2014).Clearly, the challenge of this approach is the availability of exact bathymetric data which is usually unavailable in most remote islands.
Traditional bathymetric mapping techniques such as echo sounders or Light Detection and Ranging (LiDAR) can provide reliable measurements with high resolution (Yeu et al. 2018).However, sonar systems mounted on surveying vessels are difficult to operate in shallow water areas due to economic challenges and safety.Satellite remote sensing can be considered an alternative approach to cope with the cost and labor-consuming problems of the field survey.With altitude in hundreds of kilometers, satellites can observe vast areas with reasonable resolution at a glance.Many researchers (e.g.Lyzenga et al. 2006;Pacheco et al. 2015;Vinayaraj et al. 2016) have successfully illustrated using satellite remote sensing to estimate depth in coastal waters.However, most of previous remote sensing based bathymetry models used the medium resolution satellite data such as Landsat and Sentinel which limit the precision of the estimated depth as well as benthic habitat results.This study attempt to investigated the high resolution Planet Scope imagery to enhance the classification results of marine habitat.
In the satellite-derived bathymetry (SDB) approach, ground truth data is essential for supervised training water depth models.The Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) and its sensor Advanced Topographic Laser Altimeter System (ATLAS) are potential solutions when reliable training data is lacking.With a 6-beam green wavelength (532 nm) laser at 10 kHz pulse repetition rate, ICESat-2 provides a dense ground track with a footprint size of 17 m and an along-track interval at 0.7 m (Neumann et al. 2019).The capability of ICESat-2 can address the challenges in bathymetric mapping in nearshore regions (Forfinski-Sarkozi and Parrish 2016;Jasinski et al. 2016;Parrish et al. 2019).
Mapping benthic habitats using machine learning is now more commonly used than other approaches such as manual delineation by expert interpretation, expert manual class assignment and expert-derived ruleset development, and has shown promising accuracy (Mohamed et al. 2018;Wicaksono et al. 2019;Burns et al. 2022).Improving the accuracy is not only limited to finding the suitable classification algorithm but also finding the most suitable benthic habitat classification scheme for specific remote-sensing data.Several models such as conventional machine learning algorithms, including Random Forest (RF) and Support Vector Machine (SVM), and a current deep Convolutional Neural Network (CNN) are more suitable to accommodate these issues and produce high accuracy, since they do not require such assumptions to work effectively.Several studies used CNN to improve the classification results (Makantasis et al. 2015;L€ angkvist et al. 2016;Kussul et al. 2017).Although various studies used deep-learning methods and CNN for coral classification (Fincham et al. 2020;Raphael et al. 2020;Shields et al. 2020), the processed image of these techniques was captured under ideal conditions.Therefore, the potential of CNN in the classification of benthic habitats from satellite images has still not been fully explored.
The present study proposed an approach that solves the confounding influence of depth and water column attenuation on substrate reflectance.Accordingly, a complete scheme for estimating coastal bathymetry from ICESat-2 data, integrated for interpreting marine benthic habitats based on the high-resolution PlanetScope satellite imagery was developed.The proposed approach is applied in Lyson Islands, located in the middle of Vietnam, and validated based on field measurements.Depth Invariant Index (DII) and Bottom Reflectance Index (BRI) are implemented to reduce the water column's effect.The abovementioned classification process was performed using three machine-learning algorithms (RF, SVM, and CNN) in order to classify benthic habitats and obtain the parameter of the most accurate map.The achieved results for benthic cover mapping based on the three machine learning models were then evaluated and compared using the overall accuracy and the Kappa statistical criteria.All the processing and analyses performed on Google Earth Engine (GEE), an open-source cloud computing platform.

Study area
Lyson is an island district of Quang Ngai Province in Central Vietnam.The district comprises two main volcanic islands situated at a distance of 14.5 miles ($25 km) from the mainland (Central Vietnam) between 15 22 0 and 15 26 0 N, with an area of nearly 10 km 2 , surrounded by fringing reefs (Figure 1).Lyson island region has abundant coastal and marine resources including coral reefs and seagrass meadows which perform important biological, ecological, aesthetic, and economic functions.The local coral reefs are rich in diversity, with over 200 species from 60 genera of hard coral (Astakhov 2015), 60 species from 10 genera of soft coral (Ben and Dautova 2012), 200 species of fish, 137 species of seaweed, 70 species of molluscs and 96 species of crustaceans (Ben et al. 2018).In addition, seagrass is often found in the southwest and southeast of the island region with the dominance of two species: Cymodocea rotundata and Thalassia hemprichii.Both volcanic islands are permanently inhabited with mainly the cultivation of garlic.Unfortunately, they are being destroyed at an alarming rate primarily by human activities (Quang et al. 2015).Agricultural and tourism activities have a significantly harmful impact on the environment around the islands (Long and Vo 2013).Consequently, there is a critical need to monitor the conditions of these coastal marine ecosystems not only by developing baseline maps depicting their spatial patterns and associated habitats but also by documenting the changing conditions associated with Lyson coastal ecosystems over time.

PlanetScope data
Also called Dove imagery with high spatial (3 m) and temporal resolution (daily), PlanetScope offers opportunities to monitor and detect coastal region's movement with higher accuracy than previously available satellite images.As the newest high spatial and temporal resolution satellite imaging, this imagery is capable of recording an area on Earch's surface of 150 million km 2 per day.In the present study, the PlanetScope data was obtained freely under the Planet Education and Research program (PlanetTeam 2020) at level 3B (orthorectified scene) (https://www.planet.com/explorer/).The PlanetScope images have four bands in Blue (470 nm), Green (540 nm), Red (610 nm), and Near-Infrared (NIR; 780 nm).The product is available in three formats: digital number, atsensor radiance, and atmospherically corrected surface reflectance.We selected the atmospherically corrected image with minimum cloud coverage or ocean wave contamination.Regarding the data correction, PlanetScope Lab reduces the impact of atmospheric layer by using a 6S radiative transfer model, which has input parameters (Aerosol Optical Depth-AOD, water vapor, and ozone) retrieved from MODIS data in near real-time.Figure 1 presents the PlanetScope satellite image of our site collected on May 13, 2020.

ICESat-2 data
The Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2), launched on September 15, 2018 at an altitude of approximately 500 km, is a descendant of ICESat, which was operated in the 2003-2010 period.The Advanced Topographic Laser Altimeter System (ATLAS) onboard ICESat-2 employs a 6-beam green wavelength (532 nm, 10 kHz repetition frequency) laser (Figure 2), which is organized in three pairs consisting of a "weak" and "strong" beam (energy ratio between them is approximately 1:4).The interval between two adjacent laser points in the along-track direction is 0.7 m (Neumann et al. 2019).Data can be downloaded freely after registration (https://nsidc.org/data/icesat-2/data).Only the strong beam of ICESat-2 was used in this research because it has higher energy and can penetrate deeper than the weak beam.Even though the original design goals of ICESat-2 are focused on polar regions, some researchers have shown the potential to map bathymetry with ICESat-2 in other lower latitude regions (Jasinski et al. 2016;Parrish et al. 2019).In our work, ICESat-2 elevation data was integrated with the PlanetScope images to generate bathymetry charts to overcome the challenge of lacking reliable ground truth information.Although each flight route of ICESat-2 has six beam tracks, only the strong beam track across shallow water is considered for processing bathymetric measurements.In our study area, a total of 13 transects belonging to two flight routes whose IDs are 103 and 1040 were employed (Table 1).

Ground truth samples
Several field surveys on marine biology have been carried out in Lyson Islands during the last decade (every two to three years) which belong to different projects.The results of these campaigns were documented in Ben et al. (2018), Long and Vo (2013), Quang et al. (2015), Tin et al. (2021).Scuba diving for marine biodiversity assessment was carried out in Lyson coastal area in every single survey using the photo-transect survey method (Roelfsema and Phinn 2010) (Figure 3).The underwater surveying method can be summarized as the surveyors being split into two teams, in which the first team walked along the transects and captured substrate photos every ±2 m in intervals in the exposure area, whereas the second team took underwater samples in the subtidal zone.The divers captured photos and recorded data underwater.The coordinates of each image was recorded simultaneously by Garmin GPSMAP 76CSx.It is noted that the average positional accuracy of Garmin GPSMAP 76CSx is about 5.7 meters (Gikunda and Griffith 2019).Each picture was labeled and determined based on the most dominant class with the support of Coral Point Count with Excel extensions (CPCe) (Kohler and Gill 2006).Finally, a set that included 1589 sample points were produced from the collected pictures.These samples were randomly selected for training and validating the classification results, with the ratio of 70 and 30 percent respectively.The deterministic pseudorandom numbers in the range of [0, 1] were generated and used to separate training and validation sets to accomplish this step.
In this study, data on benthic cover, with corresponding locations (Appendix), marine environmental conditions, weather conditions, etc. were collected for the purpose of map production and accuracy assessment.Based on reviewing existing knowledge of and observing the study area, we separated the benthic cover into six classes: Dense Seagrass (dominated by Thalassia hempricii), Sparse Seagrass (dominated by Cymodocea rodundata), Coral in the intertidal zone, Hard Bottom, Sandy Substratum, and Darkness (the objects lying in the deep water and were hard to distinguish).In order to validate our water depth model's efficiency, the bathymetry chart of Lyson coastal area was also reconstructed.This task was accomplished based on a combined topographic map established by US Naval (chart code VN4752) and depth measured by single beam echo sounder Lawrance VP 1000.As the result, a 10 m resolution smooth surface of the bathymetric chart was generated.

Methodology
In this study, the process of bathymetry retrieval and benthic habitat mapping using PlanetScope Imagery and ICESat-2 data was conducted based on an empirical approach, as shown in Figure 4 below.This process includes four main steps: deep-water correction, bathymetry estimation (water depth model), water column correction, and benthic classification.PlanetScope and ICESat-2 data were used to generate the water depth model.These data are important inputs for the water column correction.The BRI extracted from the water column correction integrated with the field measurement data are the critical background for obtaining benthic habitat classification, which is this study's primary objective.

Deep-water correction
In this stage, we utilized the normalized difference water index (NDWI) to mask out the terrestrial and non-water features from the region of interest (Gao 1996) on the PlanetScope imagery.Next, three visible bands were subtracted by the mean value of the NIR band in the deep water to reduce the sun-glint effect of the sea surface.This process is based on a simple assumption that the NIR band has strong water absorption and does not significantly contribute to bottom reflectance in the deep region (Hochberg et al. 2003).
where X(k) i is the transform radiance in the band (i), L(k) i , L 1 (k) NIR are reflectance of visible and NIR in the deep region, respectively.

Extract ICESat-2 bathymetry
In this study, we developed an algorithm that employs the ATL03 product of ICESat-2 to derive bottom elevation.ATL03 combines the data of POD (Precision Orbit Determination), PPD (Precision Pointing Determination), and ATL02 to produce a Level 2 product containing geolocated ellipsoidal heights (Neumann et al. 2019).Our algorithm was encouraged by the technique applied to Multiple Altimeter Beam Experimental Lidar (MABEL) (Brunt et al. 2014), which identifies the bottom signal based on its potential and probability.The processing of determining the bottom signal includes three steps.Firstly, the mean and standard deviation of the high signal confident class with signal confident of 4 were calculated (signal confidence has four levels: 0-noise, 1-background, 2-low, 3-medium, 4-high).Based on the Gauss-Laplace distribution (Andrews and Mallows 1974), signal photons with height values larger than three times standard deviation were considered water surface signals and not counted in retrieving the bottom depth.Secondly, we eliminated the potential water surface signals found in step 1 and generated 3 m along-track segments of data.Here, it should be noted that the interval between two laser pulses in the along-track dimension is 0.7 m.However, in order to be incorporated with spatial resolution of PlanetScope images, the track segments of data were resampled as 3 m of distance interval.Furthermore, the two statistical indices above were again recalculated to discriminate coarse signal photons.Segments with more photon counts than the standard deviation values were considered to contain bottom signal photons potentially.Finally, we selected the median photon elevation for each remaining segment as seafloor bathymetry.
Because of the elevation of ICESat-2 pulse points calculated as a function of travel time from the satellite to Earth's surface, the major source of bathymetric error arised from the dissimilar velocity in the water column and atmosphere due to the difference of refractive index (Parrish et al. 2019).Thus, refraction correction needs to be included in retrieving ICESat-2 bathymetric measurements.The refraction correction was accomplished following Eq. 2 below: where D is water depth derived from ICESat-2 data, H B is bottom surface elevation, H S is mean surface elevation, n air is index refraction of air, n water is index refraction of seawater.
The default values of index refraction of seawater and atmosphere are 1.33469 and 1.00029, respectively (Parrish et al. 2019).The processing of bathymetry extraction from ICESat-2 data has illustrated in Figure 5 below.After processing, a total of 365 water depth training points (Figure 6) were generated from thirteen beam tracks (Table 1).Once the ICESat-2 bathymetry had been retrieved, the corresponding tide levels at the time of ICESat-2 acquisition were obtained from the FES2014 global tide model (Carr ere et al. 2016).This model is among the best barotropic ocean tide models for coastal regions (Stammer et al. 2014).In the next step, we corrected ICESat-2 bathymetry based on the tide elevation obtained.Furthermore, these bathymetric training points were related to multispectral bands to estimate regression coefficients.

Nearshore bathymetry estimation
Bathymetry can provide valuable information for the understanding of coastal processes as well as for discrimination and classification of coral reef habitats.Water depth information allows for estimating bottom albedo, which can improve marine habitat mapping (Mumby et al. 1998).
Lyzenga in the 1970s first introduced an optical approach based on passive remote sensing (Lyzenga 1978(Lyzenga , 1985) ) and expressed the relationship between the observed radiance and water depth as: where R W is observed radiance, R B is bottom reflectance, R 1 is water column reflectance, z is depth, and g is attenuation coefficient according to diffusion and absorption.
The varieties of the bottom characteristics are among the most critical issues that make the estimated thickness of the pure-water column unstable (Vinayaraj et al. 2016).Several researchers have addressed this concern, such as the utilization of all visible spectra in multiple linear regression algorithms (Lyzenga et al. 2006) and the non-linear bathymetric inversion model (Stumpf et al. 2003).Nonetheless, the heterogeneity of the bottom substrate is still unsolved.In our study, to minimize the variety of bottom composition, the study site is classified into different classes based on their spectral characteristics.This task was accomplished by using the K-mean cluster with the support of Google Earth Engine platform.Afterward, the band ratio inversion model of Stumpf et al. (2003) as in Eq (4) and the multiple linear regression algorithm Lyzenga et al. (2006) as in Eq (5) were applied to retrieve the water depth.In Eq (3) and Eq (4), the regression analysis was carried out between ICESat-2 training depth and multispectral to estimate coefficients.
where m 0 , m 1 , and n are constant coefficients, and R(k 1 ) and R(k 2 ) are remote sensing radiances for spectral bands k 1 and k 2 .where a 0 , a 1 , a 2 , a 3 are coefficients determined from multiple linear regression.X(k 1 ), X(k 2 ), X(k 3 ) are log-transformed radiances defined by Eq (1).

Water column correction
Technically, different benthic habitats can be distinguished based on their spectral behavior at the terrestrial scale.If they are placed underwater, their reflectance will decrease following the exponential function of Beer's law (Zoffoli et al. 2014).This situation would be worsened by the increase in the depth and the appearance of suspended matters, which often happens in coastal regions.Therefore, poor transparency of shallow water surrounding Lyson Islands makes benthic discrimination more complicated.Lyzenga's algorithm (1981) is one of the most widespread approaches owing to its easy access without requiring local depth for processing.The Lyzenga method is based on these simple assumptions that the water properties do not change in the entire scene and the variation of radiance in the uniform substrate is due to differences in water depth.The DII transformation is explained as follows: where R i and R j are remote sensing radiances for spectral bands i, j; k i /k j is ratio of attenuation coefficients; r ii , r jj are variances of the bands i,j; and r ij is covariance between band i and band j.Sagawa et al. (2010) suggested a method to overcome the poor transparency of coastal waters by using regression between the radiance and the depth on various pixels of a homogenous substrate (sand) to estimate the attenuation coefficient.In our study, we utilized water depth and attenuation coefficient estimates from Sagawa's technique to substitute into Eq (3) to derive the bottom reflectance index for each band as in Eq (7).

BRI
Then, supervised classification was applied from the BRI in order to obtain benthic mapping results.Besides, the DII technique of Lyzenga was conducted for comparison with the BRI approach.

Image classification and machine learning models
As mentioned previously, classifying and mapping benthic habitats are challenging because spectra variation depending on bottom substrate, water depth, and water attenuation coefficient.Therefore, the inclusion of bathymetry information in the classification scheme is crucial.The present study used the BRI extracted from four optical bands of PlanetScope imagery for benthic habitat classification based on a pixel-based approach.The BRI after the water column correction integrated with field measurement data were used as input for a supervised classification method.Two traditional machine learning algorithms, including Random Forest (RF), Support Vector Machine (SVM), and a current deep Convolutional Neural Network (CNN) were employed to classify marine benthic features with the support of Google Earth Engine (GEE).
RF is based on the idea of utilizing multiple decision trees when training samples and variables are selected randomly (Breiman 2001).The most crucial advantage of RF is computational efficiency and higher accuracy even in high-dimensional remote sensing data.In order to obtain the optimal parameters of the RF model and avoid overfitting, we tuned the model with different decision trees.Totally, 128 trees were selected which showed the best for distinguishing benthic habitats.
SVM has also proven its effectiveness in dealing with complex remote sensing data (Mountrakis et al. 2011;Mather and Tso 2016).The basic concept of SVM is transferring the dataset into a higher-dimensional space through non-linear transformers to maximize the distance between dataset features.This characteristic is precious especially when training samples are limited.In this study, we selected SVM parameters according to the suggestion of previous studies with Gaussian Radial Basis Function (RBF) kernel and the cost (C) parameter equal to 1 (Eugenio et al. 2015;Wicaksono et al. 2019).Both RF and SVM were working on the GEE platform.
Besides, CNN has also demonstrated its effectiveness in many application domains such as speech recognition, object recognition, and other applications (LeCun et al. 2015).In this study, we used the TensorFlow package (Abadi et al. 2016) in Python to train a convolutional neural network with visible spectral bands of PlanetScope as input.ReLU (Rectified Linear Unit) fuction which is the most used activation fuction in neural network has been constructed on four hidden layers with 512 nodes, based on reviewing of existing studies on CNN method and suggestion of optimal parameters (Fincham et al. 2020, Burns et al. 2022).For the output layer, Softmax activation function which is the most used fuction for multi-classification purpose has been deployed in this CNN model.This study used ADAM optimization algorithm and the categorical cross-entropy loss function to compile the model and fitted the network coefficients to the trained data.The network training lasted 100 epochs with a batch size of 30 samples.
The quality of benthic habitat maps derived from the PlanetScope imagery and ICESat-2 data was determined by the process of accuracy assessment.The validation data set from surveying was used independently for evaluation and not included in the supervised classification training.The performance of classification was evaluated in terms of the Confusion Matrix (Foody 2002).The agreement between the output classification map and ground-truth data was measured by using two statistical indices: overall accuracy (OA) and the Kappa coefficient.

Results and discussion
4.1.Nearshore bathymetry derived from combination of PlanetScope and ICESat-2 The bottom substrate variation is one of the significant factors that lead to the uncertainties of estimated depth (Su et al. 2014).To address this issue, we divided the study site into several classes based on spectral similarity.Hence the heterogeneity of bottom types will decrease within each class.Subsequently, two bathymetry estimation algorithms including band ratio inversion as in Eq (4) and multiple linear regression were applied to generate the optimal parameters for each class.This stage is expected to address the inadequacy of the conventional water depth model by adapting the model parameters according to the classification result.
In order to produce coefficients of the water depth model, a total of 365 ICESat-2 processed bathymetric points were related to the spectral bands of the PlanetScope image through regression analysis.After that, these coefficients were used to estimate bathymetry.Figure 7a and b represent strong agreement (R 2 ¼0.883 and R 2 ¼0.885) between ICESat-2 bathymetric and water depth estimates from two modelsmultiple linear regression algorithm and ratio transform.These results also confirm that employing ICESat-2 bathymetric data for water depth retrieval is reasonable.
The results of bathymetry mapping from PlanetScope and ICESat-2 are shown in Figure 8.In order to evaluate our approach's efficiency, a random sample was used to extract the values from the ground bathymetric and estimated water depth map.This dataset consists of 1000 points and accounts for all bottom types and depths across the islands.The root means square error (RMSE) and the coefficient of determination were used to assess the accuracy of bathymetry results.As Figure 7c and d above illustrate, there is a considerable improvement in the ratio transform model (R 2 ¼0.824,RMSE ¼ 1.22m) compared to the multiple linear regression model (R 2 ¼0.8, RMSE ¼ 1.35m) in the validation stage as well as the training stage.It can be presumed that band ratio inversion is more responsive to the change of depth than the bottom variation.The shift in bottom albedo affects both bands similarly; thus, substrate variability is reduced through the ratio of logarithm-transformed radiances.Furthermore, the band ratio inversion method can also enhance the separation between the seagrass and algae, which have similar spectra in deep water, making the bathymetry result biased (Su et al. 2008).

Classification results
The bottom type retrieval from remote sensing data in Lyson Islands becomes more challenging due to the water quality.When water depth exceeds 10 m, due to the low signal-to-noise phenomenal of PlanetScope imagery, bottom reflectance response to satellite sensors becomes too noisy to be detected effectually.Hence, in this study, we only investigated, and mapped the benthic habitats in shallow water under 10 m.
The field surveys conducted in Lyson coastal area were to discover benthic cover of Lyson.Accordingly, six different benthic habitat classes were identified: Dense Seagrass-SG1, Sparse Seagrass-SG2, Coral in the intertidal zone-CR1, Hard Bottom-HB, Sandy Substratum-SD, and Darkness (optical deep water)-DK.The survey points including 1,589 points for bottom type and water level were divided randomly into two sets for training and validation, with 70% for training and 30% for validation.
The accuracy of benthic habitat mapping is affected by many factors such as habitat distribution, variation of water depth, and water properties (Li et al. 2019).Thus, the correction steps before classification are more critical.Regarding the water column correction, although the DII of Lyzenga can improve the accuracy of the bottom map (Mumby et al. 1998;Andr efou€ et et al. 2003;Louchard et al. 2003), it is sometimes not adequate for the complex shallow environment, particularly in the low transparency water.In shallow water, suspended matters and phytoplankton concentrations reduce the water depth penetration and increase the difficulty of distinguishing substratum (Vahtm€ ae and Kutser 2013).Due to the appearance of more suspended sediments in turbid water, the distribution of the water column proportion to remote sensing reflectance is higher than clear water, hence derived bottom reflectance before classifying is essential to enhance classification efficiency.
The BRI requires water depth to retrieve bottom information and is expected to improve benthic mapping performance.Both DII and BRI were utilized to reduce the water column's effect in supervised classification.Table 2 shows the results of the accuracy assessment correspondence by employing BRI and DII for classification.The table clearly illustrates the benefit of using the BRI to discriminate the benthic features with the highest performance (OA: 89.80%, kappa: 0.87) compared to (OA: 87.85%, kappa: 0.84) of DII and (OA: 47.2%, kappa: 0.38) of without water column correction in the case of using the CNN method.The outcomes again demonstrate the advantage of employing the BRI index to handle the effect of water column over the conventional DII approach, particularly when various benthic objects have similar spectral characteristics.However, the reliability of the BRI index depends directly on the accuracy of bathymetric information that is hard to obtain in remote islands.Therefore, ICESat-2 bathymetric data with its reliability and availability on the global scale can be considered an alternative solution to improve benthic habitat classification results.
The overall accuracies of the RF model were 80.74% for the BRI, 80.17% for DII and 37.64% of without water column correction (Table 2).These results were obtained by setting the number of trees to 128.The Kappa coefficient represents the proportion of reliability obtained after removing the proportion of chance agreement (Foody 1992).In the RF result, the Kappa coefficient values were 0.77 and 0.76, corresponding to BRI and DII respectively.These values represent substantial agreement between the output classified image and validation data.The misclassification mainly occurred between SG1 and SG2 with the lowest user accuracy being 57.14% belonging to SG2 (Table 3).This situation might be caused by comparable physical characteristics and spectra of SG1 and SG2.Moreover, SG2 is generally found in the north of the island with low density.Thus, their spectral reflectance is highly affected by the background environment.See Table 3 for the confusion matrix of the RF-BRI result.SVM shows better performance than RF algorithm with an overall accuracy of 84.19% for the BRI, 82.75% for DII and 42.5% for without water column correction.SVM also represents higher agreement between the classification image and validation dataset regarding the kappa coefficient.The confusion between SG1 and SG2 was improved but is still unsolved.Although the overall accuracy improvement of SVM compared to RF is only more than 3%, SVM made a significant impact on resolving the misclassification between CR1 and HB, with user accuracy of CR1 increasing nearly 10% from 70.69% to 81.03%.The confusion matrix of the SVM-BRI result is provided in Table 4.
In our study, the CNN method was also used in benthic habitat classification.The detailed setup of the model was discussed in the previous section.In order to build the model, we used the TensorFlow package integrated with the GEE cloud computing platform.The incorporation makes the model more adjustable and less time processing.As Figure 9 showed, the developed CNN model began to stabilize in the 80 epochs.After that time, the training's accuracy and loss function values did not have any significant change.The OAs of the training stage were 89% in BRI and 88% for DII composition, respectively.
As shown in Table 2, the CNN model demonstrates the highest accuracy, reaching 89.8% with BRI and 87.85% with DII.The CNN results show the highest accuracy in deep water and Dense Seagrass class.This could be due to the homogeneity and wide distribution of these classes.However, in the southern part of the main island, some uncertainty and mixing appeared in the edge of the coral reef intertidal zone and dense grass (Table 5).The reason is presumably that there are non-existent training points in these areas and similar spectra of habitat types in deeper water.However, the CNN, which has been well demonstrated in many studies, also showed superior performance over conventional machine learning classification approaches in this case study.Moreover, our CNN model with uncomplicated architecture proves to be promising for reapplying in various cases of research and optical imagery.The results of benthic habitat mapping by CNN are provided in Table 5 and Figure 10.Our proposed classification scheme has shown several advantages.First, the ICESat-2 bathymetric measurements can solve the lack of reliable water depth information in remote regions, which is difficult to access by other collection methods.Second, the BRI index incorporates sufficient accuracy bathymetric map to reduce the influence of low transparency water column, hence expanding the application range of the technique.The proposed CNN model integrates with cloud computing GEE is not only simple architecture but also time-saving and cost-effective that can be re-applicable.Finally, the integration with the high-temporal resolution imagery as PlanetSCope is beneficial for monitoring and managing the dynamic changes of coastal areas, particularly rapid evaluation after extreme events such as tropical storms.
In conclusion, the CNN method has proved the out performance respect to other machine learning classification approaches.Comparing between BRI and DII indicies, the BRI showed better representation to DII with respect to the overall accuracy and the Kappa coefficient.Table 6 is the statistics of area for six bottom types in Lyson islands using the CNN method for the BRI and DII techniques that extracted from Figure 10.In general, there are not much differences between the classification results by BRI and DII although BRI demonstrated the best performance.There are differences in area around 10 ha to 18 ha on each benthic class extracted by BRI and DII techniques.In Lyson islands, the distribution of seagrass and coral are considerable with totally 487.01 ha occupying around 23.3% of the coastal area of Lyson.Considering the hard bottom (HB) which includes the dead coral, the coverage of seagrass and coral ecosystem in Lyson coastal island is around 34.3%.These are the important ecosystem that need the sustainable preservation in the Lyson island which is one of the most attractive tourist destinations as well as the marine protected areas in Vietnam.

Limitations and recommendations
Although this study has demonstrated several efficiencies in bathymetric retrieval and benthic habitat mapping, some limitations remain in remote sensing applications in marine studies.Due to flight routes by track of ICESat-2 and atmospheric conditions, some regions lack bathymetric points for the training model.However, this limitation can be overcome by increasing the time operation of the satellite.The low signal-to-noise that happens in the PlanetScope image also restricts its application.According to this research, when the depth is over 10 m, the bottom reflectance becomes too noisy to record effectually.Therefore, benthic types that appear at the depth greater than 10 m could not be exactly classified, leading to a reduced overall accuracy of the method.However, if the proposed approach is applied to other case studies with higher water clarity (transparency), the effective depth may be reached deeper than 10 m.
The major limitation when retrieving bathymetry from ICESat-2 data is that the maximum depth retrieved is less than 20 meters.Thus, the survey region needs to be determined manually, and the improvement to the automation procedure needs further investigation.Moreover, in our study, the effect of horizontal geolocation error on ICESat-2 bathymetry was still not explored.Assessing and monitoring these ecosystems is hindered by a lack of reliable bathymetry data.With an increasing amount of data collected by ICESat-2, we can expand opportunities to investigate benthic changes and their effects by combining data from optical images and ICESat-2.Furthermore, with the CNN model developed in this study, we expect to reapply it to map bottom substrates in another area with similar environmental conditions without conducting field measurements.

Conclusions
This work presents the first assessment of PlanetScope imagery for mapping benthic habitats in complex optical shallow water of Lyson Islands, Vietnam.In this study, we proposed a complete scheme to derive coastal water bathymetry using PlanetScope imagery and ICESat-2 data.We applied deep water and water column correction to retrieve bottom reflectance by applying two different techniques, Depth Invariant Index (DII) and Bottom Reflectance Index (BRI).The highest classification accuracy obtained from integrating BRI index and CNN classifier model was 89.8%, while without using any water column index, the accuracy was only 47.2% with CNN.Our results also illustrated the benefit of using the BRI index over the conventional DII approach as the OA for BRI in most of our experiments were higher.
The proposed method was successfully applied in Lyson Islands, where the difference between bottom features and water properties is problematic.The results clearly illustrated the suitability of using ICESat-2 as training data to estimate water depth in shallow water areas.Moreover, the CNN approach with its efficiencies can improve the benthic classification results in coastal regions.Our study also found that utilizing the bottom reflectance index as an input for discrimination can make identifying the benthic habitat more accurate than conventional methods.Based on the above results, with very high spatial and temporal resolutions, PlanetScope combined with ICESat-2 data is promising for monitoring the dynamics and changes in benthic habitats, as well as quick impact assessment for events such as typhoons, storms, climate changes, etc. in far-flung areas where observed data is limited.

Figure 1 .
Figure 1.Location of the study area in Vietnam (red point).Highlighted coastal areas along Lyson Islands based on false composite of PlanetScope scene on May 13, 2020.

Figure 2 .
Figure 2. ATLAS beam pattern and ground separation.

Figure 3 .
Figure 3. Field photos of bottom features in Lyson Islands.

Figure 4 .
Figure 4.The workflow of estimated bathymetry and benthic map retrieval from satellite data.

Figure 7 .
Figure 7. Accuracy assessment of multiple linear regression algorithm (Fig. a, c) and ratio transform (Fig. b, d).

Figure 8 .
Figure 8. Derived bathymetric map (in meters) of Lyson Island generated from ICESat-2 and PlanetScope utilizing the Ratio transform model.

Figure 9 .
Figure 9. Accuracy and loss function value in the training stage (Epoch refers to the iterations over the entire training dataset that the learning algorithm will perform.Both epoch and categorical accuracy are dimensionless.).

Figure 10 .
Figure 10.Benthic habitat mapping from the best classification algorithms, CNN for BRI (a), and CNN for DII (b).

Table 1 .
Detailed ICESat-2 data used in the present study.

Table 3 .
Error matrix of benthic mapping by RF and BRI.

Table 4 .
Error matrix of benthic mapping by SVM and BRI.

Table 5 .
Error matrix of benthic mapping by CNN and BRI.

Table 6 .
Statistics of area (ha) of six bottom types after classification, based on BRI and DII techniques using CNN model.