The Capability of Sentinel-MSI ( 2 A / 2 B ) and Landsat-OLI ( 8 / 9 ) 1 for Seagrass and Algae Species Differentiation using Spectral 2 Reflectance 3

Abstract. This paper assesses the reflectance difference values between the homologous visible and near-infrared (VNIR) spectral bands of Sentinel-MSI-2A/2B and Landsat-OLI-8/9 sensors for seagrass, algae, and mixed species discrimination and monitoring in a shallow marine environment southeastern of Bahrain in the Arabian Gulf. To achieve these, a field survey was conducted to collect samples of seawater, underwater sediments, seagrass (Halodule uninebell.netrvis and Halophila stipulacea) and algae (green and brown). As well, an experimental mode was established in a Goniometric-Laboratory to simulate the marine environment, and spectral measurements were performed using an ASD spectroradiometer over each separate and different case of seagrass and algae mixed species at different coverage rate (0, 10, 30, 75, and 100 %) considering the bottom sediments with clear and dark colors. All measured spectra were analyzed and transformed using continuum-removed reflectance spectral (CRRS) approach to assess spectral separability among separate or mixed species at varying coverage rates. Afterward, the spectra were resampled and convolved in the solar-reflective spectral bands of MSI and OLI sensors and converted into water vegetation indices (WVI) to investigate the potential of red, green, and blue bands for seagrass and algae species discrimination. For comparison and sensor differences quantification, statistical fits (p < 0.05) were conducted between reflectances in homologous bands and also between homologous WVI; as well as the coefficient of determination (R2) and root mean square difference (RMSD) were calculated. The results of spectral and CRRS analyses highlighted the importance of the blue, green, and NIR wavelengths for seagrass and algae detection and probable discrimination based on hyperspectral measurements. However, when resampled and convolved in MSI and OLI bands, spectral information loses the specific and unique absorption features and becomes more generalized and less precise. Therefore, relying on the multispectral bandwidth of MSI and OLI sensors, it is difficult or even impossible to differentiate or to map seagrass and algae individually at the species level. Additionally, instead of the red band, the integration of the blue or the green bands in WVI increases their discriminating power of submerged aquatic vegetation (SAV), particularly Water Adjusted Vegetation Index (WAVI), Water Enhance Vegetation Index (WEVI), and Water Transformed Vegetation Index (WTDVI) indices. These results corroborate the spectral analysis and the CRRS transformations that the blue and green electromagnetic radiation allows better marine vegetation differentiation. However, despite the power of blue wavelength to penetrate deeper into the water, it also leads to a relative overestimation of dense SAV coverage due to the higher scattering in this part of the spectrum. Furthermore, statistical fits between the reflectance in the VNIR homologous bands of SMI and OLI revealed excellent linear relationships (R2 of 0.999) with insignificant RMSD (≤ 0.0015). Important agreements (0.63 ≤ R2 ≤ 0.96) were also obtained between homologous WVI regardless of the integrated spectral bands (i.e., red, green, and blue), yielding insignificant RMSD (≤ 0.01). Accordingly, these results pointed out that MSI and OLI sensors are spectrally similar, and their data can be used jointly to monitor accurately the spatial distribution of SAV and its dynamic in time and space in shallow marine environment, provided that rigorous data pre-processing issues are addressed.


capabilities of the Earth's surface and ecosystems (Drusch et al., 2012). Their spectral resolutions and configurations 108 are designed in such a way that there is a significant match between the homologous spectral bands (Drusch et al.,  117 of these differences depends on the application (spectral characteristics of the observed target) and on the approach 118 adopted to perform time-series analyses, mapping, or change detection exploiting these instruments (Flood, 2017).

119
For instance, it is plausible that the extraction of seagrass and/or algae information in time over shallow water areas 120 using surface reflectances, empirical, semi-empirical, and/or physical approaches, may affect the comparison of the

132
Finally, 4) efficiency and accuracy analysis of the examined WVI to discriminate between species (seagrass, algae 133 and mixed) by integrating the green and blue bands instead of the red band. Further, according to these analyses 134 results, it will be clear whether it possible for these sensors to differentiate between seagrass and algae effectively and 135 precisely at the species level, or simply and generally to discriminate among submerged aquatic vegetation (SAV) 136 cover at different density classes. Traditional seagrass in-situ surveys require time and intensive field sampling, which is generally lack the spatial 139 coverage and precision that are required to detect changes before they become irreversible or very difficult to maintain 140 year after year Fourqurean, 2001, Yang andYang, 2012). Over the recent decades, remote sensing 141 science and sensors technology has played an essential role in seagrass mapping and monitoring (Dean and     Landsat TM and ETM+ data time-series analysis enabled seagrass distribution to be appropriately assessed in the 169 context of its spatial and temporal history and provide the ability to not only quantify change but also describe the 170 type of change. Moreover, a regional-scale mapping of seagrass habitat in the Wider-Caribbean region was achieved 171 with acceptable accuracies using a total of 40 Landsat scenes acquired with TM and ETM+ sensors, and applying 172 different images processing methods, i.e., segmentation, contextual editing, supervised classifications, etc.

212
For Posidonia oceanica mapping in the Mediterranean region, the random forests method gives more accurate results 213 than SVM approaches when compared with in-situ observations (Bakirman and Gumusay, 2020). Whereas, using a 214 high spatial resolution of WorldView-2 imagery acquired over a coastal area in Florida, the neural network classifier 215 performed better than SVM for seagrass mapping (Perez et al., 2020). According to Uhrin and Townsend (2016), 216 linear spectral mixture analysis (LSMA) can be used with photo interpretation to generate spatially resolved maps suitable for seagrass spatial distribution and provide improved estimates of seagrass classes. Nevertheless, Chen et al.

218
(2016) revealed the difficulty and limitation of LSMA for mapping the fraction of scattered and heterogeneous 219 seagrass patches that are smaller than the pixel size. At Ritchie's archipelago within the Andaman and Nicobar group 220 of Islands, Bayyana et al. (2020) showed that Sentinel-MSI data can detect, and map submerged benthic habitat and 221 seagrass beds present at a depth of 21 m using random forest, SVM, and K-nearest-neighbour classification algorithms.

225
Since the emergence of remote sensing as a new scientific discipline in the early 1970s, vegetation indices (VI's)

226
were involved as radiometric measurements of the spatial and temporal distribution of land vegetation photo-227 synthetically active. They use the red and near-infrared (NIR) bands, the normalized difference vegetation index 228 (NDVI) was proposed by Rouse et al. (1974) at the dawn of remote sensing. Since these two spectral bands are 229 generally present on Earth observation and meteorological satellites, and often containing more than 90% of the 230 information relating to vegetation canopy (Bannari et al., 1995), the NDVI had taken a privileged place in the 231 NASA/NOAA Pathfinder project (James and Kalluri, 1994). Thus, it was daily derived from NOAA-AVHRR data at 232 the Earth scale. Subsequently, it was also derived every day from MODIS and SPOT-Vegetation data to produce time-233 series products for global vegetation assessment and monitoring at the regional and global scales. Due to this glorious 234 history and its simplicity, the NDVI has become the most widely used to assess vegetation canopy. Then, this index

239
Likewise, the transformed difference vegetation index (TDVI) was developed by Bannari et al. (2002) to describe the 240 vegetation cover fraction independently to the background artefacts, to reduce the saturation problem, and to enhance 241 the vegetation dynamic range linearly. These indices (NDVI, SAVI, EVI, and TDVI) were used to establish a close 242 relationship between radiometric responses and land vegetative cover densities, and they were implemented in the 243 ENVI image processing system.

244
In marine applications, these indices were tested by several scientists for seagrass and algae discrimination and    However, although VNIR bands are generally available in optical remote sensing satellites, it is well known that only 251 the visible bands can penetrate ocean water deeper than NIR which is largely absorbed by the water surface (Kirk,

299
The area under investigation in this research is the water boundary of the Kingdom of Bahrain (25º 32' and 26º00'N, 300 50º 20' and 50º 50'E) which is a group of islands located in the Arabian Gulf, east of Saudi Arabia and west of Qatar 301 (Fig. 3). The archipelago comprises 33 islands, with a total area of 8269 km 2 , 9% of it is a land area (778.4 km 2 ).

302
Along the southeast coast of Bahrain, the continental plateau extends for kilometers with a depth of less than one or          (Fig. 6). In this step, the resampling procedure considers the nominal width of each spectral band (Table 1). Then, the 417 convolution process was executed using the CAM5S transfer radiative code (Teillet and Santer, 1991). This 418 fundamental step simulates the signal received by the considered sensors at the top of the atmosphere from a surface 419 reflecting solar and sky irradiance at sea level, considering the filter of each band (Fig. 6), and assuming ideal 420 atmospheric conditions without scattering or absorption (Zhang and Roy, 2016). Accordingly, the equivalent   As discussed previously, the MSI and OLI relative spectral response profiles characterizing the filters of each spectral 473 band are relatively different (Fig. 6). To examine the impact of this difference, statistical analyses were computed 474 using "Statistica" software. The relationships between the product values (reflectances and WVI's) derived from MSI 475 against those obtained from OLI were analyzed between homologous bands using a linear regression model (p < 0.05).

476
As well, the R 2 was used to evaluate the strength of this linear relationship. For this process, the resampled and

498
For scattered and low coverage (~ 10 %), the shapes of all spectra are relatively similar, without the possibility to 7a). The highest reflectance values vary between 10% and 15% across NIR wavelengths with a difference reflectance 501 (NIR) around 5%, while in the visible all the reflectance values are below 5% with visible are also ˂ 5%. For this 502 low and sparse cover, it is observed that the reflectance is influenced by spectral properties of the underlying 503 sediments, fragments of vegetation, light shading, etc., thus contributing to the confusion between spectral signatures.

504
Definitely, under such conditions, it is a challenge to distinguish between seagrass and/or algae species based only on 505 their spectral signatures. Whereas, the measurements acquired over somewhat denser coverage rates (~ 30 %) show 506 analogous spectral behaviour and patterns with overlap among spectra in visible wavelengths (400 to 700 nm), but a 507 slight separability between species stands out relatively in NIR (Fig. 7b).

603
However, despite its near-perfect linearity and insignificant RMSD between MSI and OLI values (0.001), the TGI

604
show a very weak and limited spatial variability with a range between 0 for pure seawater and 0.05 for a very dense 605 coverage (100%) of seagrass or algae (Fig. 10e). This range cannot allow the differentiation among the marine 606 environment classes, because this index was not developed for biomass sensing but was designed for crop nitrogen 607 requirements detection. Likewise, although the scatter-plot of VARI shows an excellent coefficient of determination (R 2 of 0.99), this index overestimates the predicted values by MSI sensor compared to those estimated by OLI, 609 resulting in the data not fitting the 1:1 line very well (Fig. 10f). Moreover, the difference 610 values of VARI derived from MSI and OLI data vary between 0 and 0.14 depending on the sample species and its  As discussed previously, when integrating the blue and green bands, the indices WDVI, WAVI, WEVI, and

658
but it underestimates the SAV as shown in Fig. 10 Table 2 shows that the indices WAVI, WEVI, and WTDVI provided relatively identical results 666 when integrating the blue or green bands. Nevertheless, the scatter plots in Fig. 11 (a, b, and c) illustrate that when the 667 green band is considered instead of the blue, the majority of sampled points are located closer to line 1:1, especially 668 when the coverage rate becomes denser. This can be explained by the fact that despite the power of blue wavelengths 669 to penetrate deeper into the water, this band also leads to an overestimation of indices values due to its higher scattering 670 (Fig. 11), mainly in turbid environments.   will be approximately 0.08 to 0.10 (reflectance unit). Therefore, this total RMSE is greater than the achieved difference

779
(sensors calibration, atmospheric corrections, sun-glint corrections, and BRDF normalization) must be addressed 780 before the joint use of acquired data with these sensors. Furthermore, we demonstrated that blue and green bands are 781 better than red for seagrass and algae biomass discrimination, providing the best R 2 and the most insignificant RMSD 782 for the investigated indices. Nevertheless, it is preferable to consider the green band integration due to its sensitivity 783 to pigment content within seagrass and algae tissues, for its ability to penetrate water, and for its low sensibility to 784 atmosphere and water column scattering compared to the blue band.