Estimation of Chlorophyll and Turbidity Using Sentinel 2A and EO1 Data in Kneiss Archipelago Gulf of Gabes, Tunisia

Multispectral and hyperspectral sensor data of the bio-optical parameters with a high spatial resolution are important for monitoring and mapping of the coastal ecosystems and estuarine areas, such as the Kneiss Islands in the Gulf of Gabes. Sentinel 2 S2A and Hyperion Earth observing-1 (EO1) imaging sensors reflectance data have been used for water quality determination and mapping of turbidity TU and chlorophyll Chl-a in shallow waters. First, we applied a tidal swing area mask based on uncorrelated pixel via 2D scatter plot between 665 nm and 865 nm to eliminate the overestimation of the concentration of water quality parameters due to the effect of the bottom reflection. The processing for mapping and validating Chl-a, Turbidity S2A, and EO1 were performed using a relation between reflectance bands and in situ measurements. Therefore, we were able to validate the performance of the case 2 regional coast colour processor (C2RCC) as well as our region-adapted empirical optical remote sensing algorithms. Turbidity was mapped based on the reflectance of 550 nm band for EO1 (R 2 = 0.63) and 665 nm band for S2A (R 2 = 0.70). Chlorophyll was mapped based on (457/528 nm) reflectance ratio (R 2 = 0.57) for EO1 and (705/665 nm) reflectance ratio (R 2 = 0.72) for the S2A.

plex ecosystems, such as the estuarine system of the Kneiss Islands in the Gulf of Gabes.
Multispectral and hyperspectral remote sensing of the bio-optical parameters can be important data for mapping, classification, and monitoring of the coastal ecosystems and estuarine areas [1]. Hyperspectral imaging and empirical relationships between reflectances obtained by optical remote sensing algorithms have been used to assess water quality [2] [3] [4] [5] [6], allowing chlorophyll a (Chl-a) and turbidity TU mapping in shallow waters.
Several authors have analysed surface water by applying algorithms or empirical relationships between water constituent and reflectance [1] [7] [8] using ocean color sensors like SeaWiFS, MERIS, and MODIS. The semi-analytical, empirical bio-optical models, band ratio or chlorophyll quantification algorithms have been developed mainly according to the concentration range of in situ Chl-a and turbidity (case I or case II water) and also in terms of the specific wavelengths available in the multi or hyperspectral sensors (MODIS, SeaWiFS, EO1, MERIS, SENTINEL, etc.) [9]- [17]. The high spatial resolution and the excellent Chl-a spectral quality bands of the Sentinel 2A (S2A) data provide clear and precise information on the optical properties of water. Various turbidity mapping algorithms have been developed for different regions using multi or hyperspectral satellite data. The assessment of suspended matter relies on field spectrometry data merged with the satellite data and is obtained from a combination of channels at several wavelengths [18] or a single wavelength algorithm [19]. These applications are suitable for the coastal or estuarine areas at a coarse spatial resolution (250 m -1000 m). Besides, Landsat or SPOT data have been used to retrieve case II water's optical properties [16] [20].
Inland, coastal and estuarian shallow water area is dependent on the dynamics of sediments originated by wave erosion, tidal currents, hydrological parameters, biological and chemical constituent of water. For that this region presents a complex optical proprieties of case 2 waters, they are influenced by turbidity TU, Total suspended sediments TSS, dissolved matter, organic matter, microorganism, inorganic particulates matter from minerals [21].
Monitoring of water quality from remote sensing data depends on the spectral characteristic of water. So, to have the real optical properties of the water in shallow water, it is important to apply the atmospheric correction and to remove the effects of the bottom reflection.
Several atmospheric correction algorithms have been created and used for case 2 waters. ACOLITE developed to Landsat 8 OLI and S2A (MSI) in two steps: 1) A Rayleigh correction for scattering by air molecules, using a LookUp Table generated using 6SV. 2) An aerosol correction based on the assumption of black SWIR bands over water caused by the extremely high pure-water absorption, and an exponential spectrum for multiple scattering aerosol reflectance [22]. Case 2 Regional Coast Colour processor C2RCC is developed for optically complex Case 2 waters, which use a large database of simulated water leaving ra-diance and TOA radiances. It is based on neural network technology and has been trained in extreme ranges of scattering and absorption properties [23].
Sen2cor available in the Sentinel 2 toolbox more adapted to extract L2 products and classify image often used continental area [24], but also to give good results in inland and coastal waters [25].
Polynomial based algorithm applied to MERIS POLYMER was developed for MERIS data to remove the influence of sun glint and retrieve ocean color parameters and spectrum of water leaving radiance, [26]. As regards the effect of reflection of the seabed, more correction has been applied [27] [28] and this is to extract the reflectance of the water column and avoid the overestimation of the concentrations of the water quality parameters, especially the turbidity [29].
The objectives of this study are to assess the performances of both hyperspectral Hyperion data EO1 and Sentinel 2A data to retrieve Chl-a and TU validated by in situ observations of these two parameters in estuarine Kneiss archipelago.

Study Area
The Kneiss Islands lie in Skhira bay in the Gulf of Gabes in the southeastern coast of Tunisia at Khaouala village. This archipelago is located about 3.5 Km off the continent with an area of about 5850 ha. Figure 1 shows the geographic location of the study area with the digital seabed model established from the digitization of the bathymetric data. The entire archipelago is a part of the shallow water that has emerged above the waterline. It belongs to a shallow extended platform cut into by tidal channels that become visible at low tide like Oued Eddam [30] [31]. The Kneiss archipelago is made up of four islets: Dziret el Bassila, Dziret El Hjar, Dziret El Laboua, and Dziret El Gharbia. The highest point is only 7 m above the sea-level, and the altitude is usually less than 2 m. The hydrodynamics of this area are influenced by semi-diurnal tides with a maximum amplitude of 2 m [31]. The flow direction is south-west with a speed of 0.1 ms −1 and the main direction of the ebb is northeast with a speed of 0.08 ms −1 [32] [33] [34]. This area is characterized by the presence of submerged tidal channels. Coastal currents have a principal direction of NNE-SSO with a speed of 2 ms −1 . The swelling is reduced by shallow water and bottom vegetation [35]. All the islets have a backbone of calcareous sandstone, and Bassila is occupied by sebkhas chotts, especially maritime marshland [36]. The collective name of Kneiss Islands (plural of Knissa, meaning church in Arabic) originates from the presence of a Christian religious building whose foundation dates back to late antiquity. Indeed, the geo-archaeological studies [37] locate the monastery of Saint Fulgence of Ruspe at the beginning of the VIth century on the island of El Laboua, which is central to this group of islets. The small islets continued to be united until the historical period [33] [38]. The Kneiss archipelago is the outlet part of a huge watershed (Leben). The vast indentation bordered by large Salicornia swamps corresponds to an old estuary [30]. It is an alluvial cone of a paleo Oued International Journal of Geosciences that threw itself into the sea and is remodeled by the littoral drift (according to the configuration of the coast and the direction of the prevailing winds). Sediments accumulate in this area due to the existence of the high depths of the Gulf of Gabes [30] [31] [39]. The archipelago region of Kneiss has been listed as RAMSAR site in 2007, Important Bird Area (IBA) in 2003, and Specially Protected Area of Mediterranean Importance (SPAMI) in 2001.
The Earth Observing 1 (EO1) satellite has three imaging sensors: the multispectral Advanced Land imager (ALI), the hyperspectral Hyperion sensor, and the Atmospheric Corrector (AC). ALI is a new, more compact, and slightly enhanced version of the legacy TM. The short-lived AC was a moderate resolution hyperspectral sensor. Hyperion is a hyperspectral imager capable of resolving 220 spectral bands (from 0.4 to 2.5 micron) with a 30 m ground resolution. The instrument images a 7.5 km by 100 km surface area. Since the launch of EO1 in late 2000, Hyperion is the only source of space-borne hyperspectral imaging data at a fine resolution and is used for mapping water quality in estuarine and coastal regions [43].

Hyperion Data
EO1/Hyperion satellite imagery of January 22, 2009 (only image available for the study area) was downloaded from NASA Website (https://earthexplorer.usgs.gov). Among the 242 bands of EO1, only 39 bands (457 µm b11 to 874 µm b52) were retained. All other bands were dropped because they either do not contain a data record or they are of poor spectral quality due to the presence of streaks. Figure  2 shows all the radiance values along a transect. It can be noticed that, while more than half of Hyperion spectral bands provide vital information on the ground, only a limited set of these bands is usable for the sea. This feature is related to the high absorption in the near-infrared domain. The Hyperion sensor lacks accuracy, and the signal-to-noise ratio is high.

Sentinel 2A Data
In this study, we processed and analyzed a Sentinel 2A (S2 MSI level 1C) image acquired on May 21, 2017. The data were downloaded from the Sentinel Scientific Data Hub (https://scihub.copernicus.eu/). Sentinel application platform

In-Situ Datasets
Estuarine Kneiss Islands areas falling under sub-arid climate are challenging environments for remote sensing assessments due to various land Sebkhas, soil, tidal area, shallow waters, seagrasses, slikke, schore to name a few. Water samples were collected on October 8, 2009, May 21, 2017 between 10 a.m. and 12 p.m. from the sites; turbidity TU (NTU), chlorophyll Chl-a (µg/L), and total suspended matter (TSM) (µg/L). A summary is given in Table 1. For each station, a 10-L water sample was collected and transported to the laboratory for measurement. Turbidity was obtained with a nephelometric method, and the Chl a was measured by the spectrophotometric method. All measurements were realized in the GREEN LAB laboratory (Tunisia). Full details of the method can be found in REVAMP protocols [44].

Methodology
Preprocessing, processing, and validation of the in-situ data were carried out on the EO1 and S2A data. Figure 3 shows a detailed flowchart of the adopted methodology. Pretreatment was carried out to correct the effects of the atmosphere. In fact, the spectral resolution of Hyperion bands makes it necessary to eliminate the disturbances caused by atmospheric effects (i.e., water vapor, aerosols). The corrections were based on the 6s software [45] [46] which has a radiative transfer model in the visible and near-infrared spectrum with a resolution of 5 nm. In order to speed up the calculations, the second-order spherical albedo term was discarded. Thus, the atmospheric correction was reduced to a linear transformation based on absorption and a backscatter term. These coefficients were considered as constants on a few square kilometre's scenes. The corrections depend on three atmospheric parameters: the ozone content, the water vapor, and the optical thickness at 550 nm. Only the first parameter was extracted from the climatology provided by 6s. The water vapor content was provided by the NOAA/NCEP-DOE (https://psl.noaa.gov/) reanalysis. The optical thickness was inferred from the Hyperion measurements at the sea of Gabes and by reversal of the 6s model. The 6SV 2.1 version was downloaded from http://6s.ltdri.org site.

Data Processing of EO1 and S2A Images
Based on the concept that the wavelengths are the vehicles of the information we have try to visualize all the bands and eliminates the one which presents a bad radiometric recording especially for the EO1 data or we have to keep 39 bands of the 242 available. In the second step, we applied the pretreatments and the masks; continent and shallow water to minimize the effect of the bottom reflection in certain areas. Finally, the adopted approach to TU and CHL-a product validation is by "match-up" validation, comparing the data value for all wavelengths pixel with an in situ measurement from a location within that pixel and acquired almost simultaneously. In order to make all scatter plot possible and extract a better regression between them, processing and interpretation approach are illustrated in Figure 3.

Hyperion EO1 Data
A false RGB (711 nm, 620 nm, and 528 nm) Color composition and continental mask of the Hyperion data cube is shown in Figure 4. Four samples of different water color and depth were chosen on the image for extracting the spectral profile. Samples were taken and materialized by decreasing order of depth. Sampling 1 was located in the deepest part of the study area at tidal channels (depth of 50 cm), and sample 4 was located in shallow water (depth of 30 cm). The spectral responses of these different depths clearly show that a shallow waters mask coinciding with areas of swaying tide must be applied at the time of the satellite overpass.

Estimation of Chl-a Using Hyperion EO1 Data
In order to estimate Chl-a from EO1, we adopted the ratio between blue to green bands. This ratio is based on the in situ measurements that were made in the (Table 1) and from the spectral signatures extracted via the satellite EO1. Figure 5 shows the best scatter plot between the generating blue to green ratio and the measured Chl-a contents (R 2 = 0.578). The resulting distribution of Chl-a as calculated from this regression model using EO1 reflectance ( Figure 5(b)) is different from the distribution of the water color as shown in the RGB composition except for the shallow water region (less than 50 cm of depth) which are red-flagged in the map. Therefore, we accepted the blue to green ratio to extract Chl-a concentration. Rrs is the remote sensing reflectance defined by:

Estimation of TU Using Hyperion EO1 Data
The turbidity estimation algorithms via satellite data have often been carried out according to the types of coastal or oceanic water and to the range of turbidity [50] [51] [52]. The best linear regression is R 2 = 0.6382 ( Figure 6) was established between the reflectance at 550 nm wavelength of the available EO1 bands and the in situ turbidity data. The spatial distribution of the TU (Figure 6) data, deduced via the 550 nm wavelength, maps the sediment transport via the tidal channels and the direction of the currents in the study area. A slight overestimation of NTU concentration in the shallow water region (less than 50 cm of water depth, visible on Figure 6(a) in red) can be noticed. Thereby an empirical relation to map turbidity in the Kneiss Islands via EO1 data is given as:

Sentinel 2 Data
For a detailed mapping of Kneiss Archipelago, we used S2A sentinel data at 10 m and 20 m spatial resolution. These data were registered on the same day as that of the in situ samplings (May 21, 2017). The preprocessing was carried out by the SNAP software. We've tested The L2 C2RCC SNAP toolbox product, particularly, Rhow 400 -1200 nm; (atmospherically corrected angular dependent water leaving reflectance Rhow = Rrsπ diffuse attenuation coefficients m −1 ), conc_tsm (total suspended matter dry concentration gm −3 ) and conc_chl (chlorophyll-a concentration µg/l). International Journal of Geosciences In addition, we have the possibility to add additional background information such as salinity, elevation, ozone, temperature, and air pressure [53].

Shallow and Tidal Area Mask
Kneiss islands are a tidal area and several regions have very shallow water for that we apply a relationship between 665 nm and 865 nm that is limited to turbidity with reflectance increasing in TU up to ~20 NTU for Rrs (645) and up to ~1000 NTU for Rrs (859) [19]. To better understand the relationship between these two wavelengths for our area, we established a 2D scatter plot between 665 nm and 865 nm (Figure 8(b)). These two wavelengths were highly correlated except for certain pixels. To delineate the area with high turbidity and to extract all uncorrelated pixel water (Figure 8(a)), identification was carried out to locate those pixels. Their positions coincided with the shallow water region, which was a tidal swing area (deepness less than 50 cm). The same regions showed an overestimation for the concentration of Turbidity due to the effect of bottom reflection. Therefore, we have created a mask based on these non-correlated pixels in order to eliminate the areas influenced by the reflection of the seabed (Figure 8).

Estimation of Chl-a Using Sentinel 2A Data
A visualization of all the bands and S2A C2RCC product was carried out, also we tried to test and seek the best linear relationships and regressions between the wavelengths and the in situ measurements of Chl-a. The correspondence between Chl-a and water leaving reflectance of Rrs 490, 560, 665 and 705 is plotted in Figure 9.   The strongest correlation was detected between Chl-a/665 nm and Chl-a/705 nm with respective R 2 = 0.67 and 0.68. A ratio of S2A 490 nm to 560 nm shows a moderate correlation with in situ Chl-a and presents a coefficient of determination (R 2 = 0.32). However the best regression R 2 = 0.72 correspond to ratio S2A 705 nm/665 nm. Therefore, NIR/Red ratio product C2RCC was selected to map Chl-a in the submerged zone of the estuarine system. by following equation (Figure 10  A relationship between in situ data Chl-a and Conc_chl C2RCC product as a function of the measurement station ( Figure 10(c)) shows us the over-estimation of chlorophyll concentrations throughout the region of the Kneiss islands and the unreliability of the Chl-a algorithm C2RCC for these types of coastal area.

Estimation of TU Using Sentinel 2A Data
Linear regression coefficients were computed to search for a better channel for estimating in situ TU data Figures 11(a)-(f). The water-leaving radiance reflectance for the 665 nm band showed the best regression coefficient with the in situ TU data with an R 2 value of 0.70. Similar results were reported [7] [8] [19] showing a good agreement between field and modeled TU for the analyzed sites.
The product conc_TSM C2RCC also shows a good relationship with the Turbidity TU data measured from water sample with an R 2 value of 0.62 (Figure 11(g) and Figure 12(a)).

( )
despite the overestimation of values of TU. The following empirical relation (Equation (4)) was deduced for the quantification of turbidity in the Kneiss Islands from the Sentinel 2A data (Figure 12(b)) shows us close to that measured values at sea.

Discussion and Conclusions
Kneiss Islands is a very complex study environment due to morphological, hydrodynamic, and biological particularities. Mainly because of the influence of semi-diurnal type tides with amplitudes reaching up to 2 m, we used two types of high-resolution sensors (30 m, 20 m, 10 m) for studying optical properties (Chl-a) and water quality (TU). In the absence of reflectance data at sea in Tunisia, we have developed an indirect method of extracting and validating Chl-a and TU satellite data. We used the reflectance or reflectance ratios at different wavelengths and compared a reduced set of in situ data for Chl-a and TU in µg/L, and NTU is obtained concurrently to the images (18 in 2009 and 10 in 2017). The best linear regressions for the calculation and mapping of Chl-a and TU from satellite data are summarized in Table 2 and Table 3. For a measured chlorophyll with a range of 0.5 -4.7 µg/L, Chl-a EO1 offset is 0 while it is 1 for Sentinel 2. For a measured turbidity range of 0 -38 NTU, offset is 11.48 for TU EO1 and 1 for TU S2A. Large offsets in these linear regressions indicate the influence of bottom reflectance. Several oceanographic studies, biogeochemical, fisheries features [54], amphipods [55], faunal analysis [56] were carried out in the Gulf of Gabes and Kneiss islands. The results are temporary and not continuous over time, which is why the extraction and generation of an algorithm for the continuous quantification of water quality data is a necessity in these types of environment. Data and relationships extracted from Sentinel S2A can effectively help us quantify seawater quality. According to our work, the Sentinel 2 radiometric data with a    (4)).  spatial resolution of 10 m and 20 m gave us the best results for the study area for a slice of water greater than 50 cm with coefficients of determination greater than 0.68 for both TU and Chl-a. Indeed, the reflection of the seabed was identified by very high values of chlorophyll concentration and turbidity. The Chl-a and TU predicted in these areas using EO1 and S2A data clearly delimit the localities where we might have doubtful results. Moreover, we developed a region of interest via 2D scatter plot between 665 nm and 865 nm bands to identify and mask the position of pixels of tidal swing area before processing to eliminate the effect of background reflections at the moment of the satellite passage.
Despite the fineness of the spectral bands of EO1 (we could use only 39 bands among the available 220 bands), the reflectance has a very low radiometric qual-ity, especially in aquatic environments. This can explain the lower coefficients of determination for the prediction of TU and Chl-a using the EO1 (R 2 = 0.57 -0.63). Table 2 and Table 3 show that wavelengths of the best reflectance or reflectance ratios differ between Hyperion EO1 and Sentinel 2A. Nevertheless, the ones used for Chl-a (457 nm/528 nm for EO1 and 705 nm/665 nm for Sentinel 2) and for TU (550 nm for EO1 and 665 nm for Sentinel 2) are remarkably similar to those used in other reported studies.