Application of Sentinel 2 MSI Images to Retrieve Suspended Particulate Matter Concentrations in Poyang Lake

Suspended particulate matter (SPM) is one of the dominant water constituents in inland and coastal waters, and SPM concnetration (CSPM) is a key parameter describing water quality. This study, using in-situ spectral and CSPM measurements as well as Sentinel 2 Multispectral Imager (MSI) images, aimed to develop CSPM retrieval models and further to estimate the CSPM values of Poyang Lake, China. Sixty-eight in-situ hyperspectral measurements and relative spectral response function were applied to simulate Sentinel 2 MIS spectra. Thirty-four samples were used to calibrate and the left samples were used to validate CSPM retrieval models, respectively. The developed models were then applied to two Sentinel 2 MSI images captured in wet and dry seasons, and the derived CSPM values were compared with those derived from MODIS B1 (λ = 645 nm). Results showed that the Sentinel 2 MSI B4–B8b models achieved acceptable to high fitting accuracies, which explained 81–93% of the variation of CSPM. The validation results also showed the reliability of these six models, and the estimated CSPM explained 77–93% of the variation of measured CSPM with the mean absolute percentage error (MAPE) ranging from 36.87% to 21.54%. Among those, a model based on B7 (λ = 783 nm) appeared to be the most accurate one. The Sentinel 2 MSI-derived CSPM values were generally consistent in spatial distribution and magnitude with those derived from MODIS. The CSPM derived from Sentinel 2 MSI B7 showed the highest consistency with MODIS on 15 August 2016, while the Sentinel 2 MSI B4 (λ = 665 nm) produced the highest consistency with MODIS on 2 April 2017. Overall, this study demonstrated the applicability of Sentinel 2 MSI for CSPM retrieval in Poyang Lake, and the Sentinel 2 MSI B4 and B7 are recommended for low and high loadings of SPM, respectively.


Introduction
In water color remote sensing, case-II waters are defined as waters whose optical properties are significantly influenced by suspended particulate matter (SPM), phytoplankton, and colored dissolved organic matter (CDOM), and SPM and CDOM concentrations do not covary with phytoplankton [1].Most of inland and coastal waters belong to case-II waters.These waters provide various benefits to society through commerce, aesthetics, tourism, recreation, and biodiversity conservation, which are greatly affected by water quality [2][3][4][5].SPM is one of the main constituents and optically active matters in sediment-laden case-II waters.Moreover, SPM concentration (C SPM ) is a key parameter describing water quality because of the impacts of SPM on the light transmission, accumulation, and transportation of nutrients and contaminants in water bodies, and further on an aquatic ecosystem [6][7][8].
Both natural forces and anthropogenic activities may influence the distribution pattern of SPM, making the C SPM values temporally dynamic and spatially heterogeneous [9][10][11].Therefore, obtaining the information of C SPM at high spatiotemporal resolution is necessary for understanding water quality dynamics and identifying their driving forces, and further for managing and protecting aquatic ecosystems [12].In-situ sampling and laboratory analysis can provide accurate measurements; however, field-based analyses are costly and labor-intensive, and may fail to capture detailed spatial variation.On the contrary, satellite remote sensing can provide synoptic observations from visible to near infrared (NIR) spectral regions, which can be used to derive the C SPM values in water.
The spatial, temporal, spectral, and radiometric resolutions of a sensor system determine its capability in water color remote sensing [5].Ocean color sensors, such as Coastal Zone Color Scanner (CZCS), Sea-viewing Wide Field-of-View Sensor (SeaWiFS), Moderate-Resolution Imaging Spectroradiometer (MODIS), and Geostationary Ocean Color Imager (GOCI), have advantages in more frequent revisits, higher signal-to-noise ratio (SNR), and higher radiometric sensitivity.However, those sensors hold lower spatial resolutions ranging from several hundred meters to over 1 km, which are often challenged to monitor fine spatial structure of water parameters in inland and coastal case-II waters [13].Therefore, land-intended satellites, such as Landsat series sensors, SPOT and GF-1 [3,14,15], have been applied to retrieve C SPM values of case-II waters [5,11,12,[16][17][18].
Sentinel 2 Multispectral Imager (MSI), a constellation of two satellites in sun-synchronous polar orbit, targets land and coastal zone monitoring by providing high spatial (10-60 m) and temporal (2-5 days) resolution images [19].Following the launches and completions of in-orbit commissioning phase of Sentinel 2A and Sentinel 2B, they turned operational on 28 November 2015 and 7 July 2017, respectively.MSI has 13 bands from visible to shortwave infrared (SWIR) spectral regions, providing appropriate data source to document C SPM variations in coastal and inland waters [20].
Some studies have been carried out in retrieving water parameters using Sentinel 2 MSI data.Gernez et al. [15] applied a simulated Sentinel 2 MSI dataset to retrieve C SPM values in turbid estuary waters.Manzo et al. [21] investigated the sensitivity of Sentinel 2 MSI to optically active matters in Italian lakes with the simulated reflectance from bio-optical model.Dörnhöfer et al. [22] applied Sentinel 2 MSI to assess water constituents and bottom characteristics over oligotrophic lake waters.Kutser et al. [23] demonstrated the advantage of Sentinel 2 MSI over Landsat 8 OLI in retrieving suspended matters in black lakes.Toming et al. [24] obtained high fitting accuracy between in-situ data and remote-sensed reflectance from Sentinel 2 MSI for chlorophyll-a (R 2 = 0.83) and dissolved organic carbon (R 2 = 0.92).Moreover, the potentials of Sentinel 2 MSI on water quality monitoring in both inland and coastal waters have been anticipated [5,17,[25][26][27][28].
Poyang Lake is the largest freshwater lake in China, which provides important habitats for Siberian cranes [29] and the Yangtze Finless Porpoise [30].As a floodpath lake, SPM in Poyang Lake supplements sediment losses due to sand dredging and scouring.However, high C SPM may deteriorate the water quality and silt channels.Therefore, some studies have been focused on deriving C SPM in this lake.The C SPM monitoring in Poyang Lake has been largely dependent on the images from Landsat and MODIS Terra and Aqua sensors [9,11,12,[31][32][33][34]. Recently, Li et al. [14] investigated the capability of GF 1 WFV in C SPM estimations; however, the atmospheric correction over turbid waters is hindered by lacking SWIR bands, limiting its applications on water quality monitoring.Red and NIR bands have been frequently used to retrieve C SPM due to their sensitivity to C SPM [3,12,35].Sentinel 2 MSI has seven bands in red to NIR spectral regions, which may provide additional potential in C SPM retrieval.Considering the advantages in spatial, temporal, and spectral resolutions of Sentinel 2 MSI, developing Sentinel 2 MSI-based retrieval models should facilitate the C SPM recording in Poyang Lake.This study aimed to: (i) develop Sentinel 2 MSI-based C SPM retrieval models, (ii) compare Sentinel 2 MSI with MODIS in C SPM estimations to validate the developed models, and (iii) to identify appropriate spectral bands for C SPM retrieval in Poyang Lake.

Study Area
Poyang Lake (115 • 47 -116 • 45 E, 28 • 22 -29 • 45 N) is located on the southern bank of the middle Yangtze River (Figure 1).It is a floodpath lake with large fluctuations of seasonal water level variation.Its area ranges from <1000 km 2 in dry season to about 4000 km 2 in wet season.The lake receives water from tributaries of five rivers (Raohe, Xinjiang, Fuhe, Ganjiang, and Xiushui), and drains into the Yangtze River through the lake mouth in the north.Suspended sediment is one of the dominant factors affecting the water quality of Poyang Lake, and its concentration fluctuates largely due to discharges from the five rivers, sand dredging activities, and backflows from Yangtze River [11,36,37].
Remote Sens. 2017, 9, 761 3 of 20 MSI with MODIS in CSPM estimations to validate the developed models, and (iii) to identify appropriate spectral bands for CSPM retrieval in Poyang Lake.

Study Area
Poyang Lake (115°47′-116°45′E, 28°22′-29°45′N) is located on the southern bank of the middle Yangtze River (Figure 1).It is a floodpath lake with large fluctuations of seasonal water level variation.Its area ranges from <1000 km 2 in dry season to about 4000 km 2 in wet season.The lake receives water from tributaries of five rivers (Raohe, Xinjiang, Fuhe, Ganjiang, and Xiushui), and drains into the Yangtze River through the lake mouth in the north.Suspended sediment is one of the dominant factors affecting the water quality of Poyang Lake, and its concentration fluctuates largely due to discharges from the five rivers, sand dredging activities, and backflows from Yangtze River [11,36,37].

In-Situ Data
Fieldworks were carried out on 16-18 October 2010 and 8-10 August 2011.Eighty-five sampling sites (47 in 2010 and 38 in 2011) were distributed along the main channel of Poyang Lake (Figure 1a).At each sampling site, the wind speed and direction were measured using a wind velocity indicator to determine the water-air interface reflectance rate used in remote sensing reflectance calculation.The location was recorded using a global positioning system receiver (Garmin Ltd., Lenexa, USA).Water spectra were measured using a portable ASD Fieldspec Pro Dual VNIR spectrometer (ASD Inc., Longmont, USA).The ASD Fieldspec Pro Dual VNIR spectrometer has a spectral range of 350-1050 nm with a spectral sampling interval of 1.4 nm, and it was used to measure the water spectra following the Ocean Optics Protocols For Satellite Ocean Color Sensor Validation [38] and steps described by Ma et al. [39].

In-Situ Data
Fieldworks were carried out on 16-18 October 2010 and 8-10 August 2011.Eighty-five sampling sites (47 in 2010 and 38 in 2011) were distributed along the main channel of Poyang Lake (Figure 1a).At each sampling site, the wind speed and direction were measured using a wind velocity indicator to determine the water-air interface reflectance rate used in remote sensing reflectance calculation.The location was recorded using a global positioning system receiver (Garmin Ltd., Lenexa, KS, USA).Water spectra were measured using a portable ASD Fieldspec Pro Dual VNIR spectrometer (ASD Inc., Longmont, CO, USA).The ASD Fieldspec Pro Dual VNIR spectrometer has a spectral range of 350-1050 nm with a spectral sampling interval of 1.4 nm, and it was used to measure the water spectra following the Ocean Optics Protocols For Satellite Ocean Color Sensor Validation [38] and steps described by Ma et al. [39].
The C SPM value of each water sample was measured gravimetrically with steps: (1) the water sample was filtered using a pre-dried and weighted Whatman GF/F glass fiber filter with a 0.45 µm pore size; (2) the filter was dried for 2 h at 110 • C and reweighted after cooling to room temperature; and (3) the C SPM was obtained by dividing the difference in weight before and after filtering by the water sample volume [39].

Satellite Images
The Sentinel 2 MSI, with a 290 km field of view and a radiometric quantization of 12-bit, provides a total of 13 spectral bands spanning from visible and near infrared to SWIR spectral region, and it has four bands with a spatial resolution of 10 m, six bands with 20 m, and three bands with 60 m [40].Two Sentinel 2 MSI L1C images captured on 15 August 2016 and 2 April 2017 were downloaded from the website provided by European Space Agency (https://scihub.copernicus.eu/dhus/).The inundation areas of Poyang Lake were appropriately 3200 and 1700 km 2 on 15 August 2016 and 2 April 2017, respectively, which represent the characteristics of Poyang Lake in the wet season and in the dry season.
One MODIS Terra image captured on 15 August 2016 and one MODIS Aqua L1A image captured on 2 April 2017 were downloaded from NASA OceanColor website (http://oceancolor.gsfc.nasa.gov/).The MODIS Terra image, instead of MODIS Aqua, was used for the MODIS Aqua image on 15 August 2016 because it was affected by clouds.The MODIS images were atmospherically corrected by firstly removing the Rayleigh reflectance and then subtracting the reflectance at 1240 nm following the procedures applied by Wu et al. [12].The Sentinel 2 MSI images were resampled to 20 m to facilitate further processing and analyses.

Sentinel 2 MSI Atmospheric Correction
Sentinel 2 MSI L1C product provides the top of atmosphere reflectance (ρ TOA ), which is assumed to be the sum of Rayleigh reflectance (ρ r ), aerosol reflectance (ρ a ) and water-leaving reflectance (ρ w ): where t is the two-way diffuse atmospheric transmittance [41].
The reflectance caused by Rayleigh scattering was estimated with the 6S radiative transfer code [42] for each band, meanwhile the diffuse atmospheric transmittance (t) was also derived.A mid-latitude Summer atmosphere condition was adopted, and the mean values of solar and view zenith and azimuth across the whole granule were calculated for the sun and sensor geometry.Cloud and land masking was performed with a threshold on the reflectance centered at 1610 nm, and pixels were classified as not being water when the Rayleigh-corrected reflectance (ρ c = ρ TOA − ρ r ) of B11 was greater than 0.0215 [28,43].
The SWIR-based aerosol correction method [28,44] was implemented using the following steps: (1) a 15 × 15 moving average filter was applied to improve the SNR of two SWIR bands; (2) the aerosol type (ε) was determined from the ratio of two SWIR bands (ε = ρ c 11 /ρ c 12 ) for each pixel; (3) the ε i,12 of a given band i over band 12 was extrapolated exponentially as follow: and (4) the ρ w at the i th band was derived with the following formula: where t i is the two-way diffuse atmospheric transmittance for band i.

In-Situ Water-Leaving Reflectance Calculation
The abnormal radiance measurements from grey diffuse reflectance standard, water, and sky were first removed considering their spectral characteristics, and the remote sensing reflectance for each sampling site was then calculated using the remaining measurements with the following equation [39]: where the R rs (sr −1 ) is the remote sensing reflectance; L w , L sky and L p (W•m −2 •sr −1 ) are the measured radiances from water, sky, and grey diffuse reflectance standard, respectively; r indicates water-air interface reflectance rate, and its value is 0.022, 0.025, or 0.028 for a wind velocity of 0, 5, or 10 m/s, respectively; ρ p is the reflectance rate of grey diffuse reflectance standard, and its value is 0.25; and π (sr) is solid angle.Only the R rs at the wavelengths of 400-920 nm were employed for further analysis because of low SNR in the edges of the spectral regions and strong absorptions in longer spectral regions.

Sentinel 2 MSI Spectral Simulation
The in-situ R rs spectra were integrated using the relative spectral response (RSR) function of Sentinel 2 MSI (downloaded from the website provided by European Space Agency, https://earth.esa.int/web/sentinel/user-guides/sentinel-2-msi/document-library; accessed on 4 July 2017) to obtain simulated water-leaving reflectance according to following formula: where λ is the wavelength, λ 1 and λ 2 are the lower and upper wavelength bound for each band and ρ w is the simulated water-leaving reflectance.The ρ w values of the first nine bands were obtained for Sentinel 2 MSI.

Model Development
The samples with incorrect spectrum (non-typical case-II water spectrum or negative values), wrong water constituent concentration (negative value), or unstable measurement conditions were removed.The remaining samples were recorded and numbered from 1 to n based on C SPM value from high to low.The odd-numbered samples were used for model calibration, the others for model validation.The C SPM values of both datasets were statistically described.The R rs spectra were visualized and analyzed.The correlation analysis between C SPM and R rs was implemented.The widely used linear, quadratic, exponential, and power models of C SPM against the simulated ρ w spectra were calibrated using the least squares technique, respectively.The calibrated models were applied to estimate the C SPM values of validation dataset.The coefficient of determination (R 2 ), mean absolute percentage error (MAPE), and root mean square error (RMSE) between the measured and estimated values were calculated to assess the fitting and validation accuracy.Six models with better fitting and validation accuracies were selected for further analyses.

C SPM Estimation and Comparison
The selected six C SPM retrieval models were applied to Sentinel 2 MSI images to retrieve C SPM values.To evaluate these models' applicability, the C SPM values deriving from MODIS using an empirical model were used as the quasi-reference values and compared with those deriving from Sentinel 2. The C SPM retrieval model based on MODIS B1 (C SPM = 0.43 × exp(31.46× B1), B1 is centered at 645 nm), which was calibrated and validated with synchronous satellite and in-situ observations [12], was applied to MODIS images to retrieve C SPM values.The C SPM values derived from these sensors were mapped and compared to illustrate their consistence and difference.Moreover, the C SPM values retrieved from each band of Sentinel 2 MSI were downsampled to 250 m by averaging, and plotted against those from MODIS.The correlation coefficient and simple linear regression line were calculated to evaluate the reliability of Sentinel 2 MSI-derived C SPM values.For expression convenience, MODIS was used to represent MODIS Terra and Aqua B1 in the following sections, since only this band was used for C SPM retrieval from MODIS images.
The image processing and mapping were implemented with Python and ArcGIS (ESRI, Inc., Redlands, CA, USA), except that the Rayleigh correction for MODIS images was carried out using SEADAS (http://seadas.gsfc.nasa.gov/).Model calibration and validation were implemented using Matlab (Mathworks, Inc., Natick, MA, USA).

In-Situ Data
Thirteen samples in 2010 were removed from following statistics and analyses, one with a negative C CHL value, one with obviously higher reflectance than others, one with extremely low reflectance, and 10 collected under high wind velocities (>5 m/s) or unstable light illumination condition on 15 October 2010.Four samples in 2011 were identified from their spectral shape and removed.
After removing outliers, a total of 68 in-situ samples were left.The statistical results (Table 1) for C SPM showed that the C SPM values of all the samples have a mean value of 77.75 mg/L, a standard deviation of 60.56 mg/L and a coefficient of variation of 77.89%, with the C SPM ranging from 17.16 to 294.50 mg/L.The calibration dataset (average C SPM = 79.45mg/L, standard deviation = 63.22 mg/L, CV = 79.57%) and validation dataset (average C SPM = 76.05 mg/L, standard deviation = 58.67 mg/L, CV = 77.15%)hold similar statistical characteristics.The R rs spectra show typical spectral characteristics of case-II waters (Figure 2a), in which R rs values increase over 400-550 nm and are much higher than zero in NIR spectral regions.Normally, a higher C SPM value tends to produce a higher reflectance curve.There are two SPM reflectance peaks around 580 and 810 nm. Figure 2b shows the correlations between C SPM and R rs of all samples as well as samples with C SPM greater than and lower than 50 mg/L.When taking all samples into consideration, there exists significantly positive correlations between C SPM and R rs over 400-920 nm (r > 0.6) and strong correlations between 707-900 nm (r ≥ 0.85) at a significance level of 0.05 (Figure 2b).For samples with C SPM greater than 50 mg/L, the correlations were obviously lower in visible spectral region and higher in NIR spectral region than those with C SPM greater than 50 mg/L.However, all of the correlation curves peaked at around 750 nm.

Model Development
The calibration results of CSPM retrieval models are shown in Table 2, and the B4-and B7-based models are illustrated in Figure 3.The models based on Sentinel 2 MSI B6 or longer wavelengths achieved very good fitting performance with R 2 over 0.9, RMSE less than 20 mg/L, and F-value of the model's significance test ranging from 151. 14

Model Development
The calibration results of C SPM retrieval models are shown in Table 2, and the B4-and B7-based models are illustrated in Figure 3.The models based on Sentinel 2 MSI B6 or longer wavelengths achieved very good fitting performance with R 2 over 0.9, RMSE less than 20 mg/L, and F-value of the model's significance test ranging from 151. 14  ).However, B1-, B2-and B3-based models produced poor fitting accuracy with R 2 smaller than 0.60.Therefore, the models calibrated with B4-B8b were selected for further validations.The six selected models were applied to the validation dataset to estimate the C SPM values, and validation results were shown in Figure 4. Similar to the calibration results, the B6 or longer wavelength-based models produced better estimation accuracies, which explained 88-93% of the variation of measured C SPM , and the B7-based model produced the best validation performance (R 2 = 0.93, MAPE = 21.54%,RMSE = 16.06 mg/L).The B4-based model had the lowest estimation accuracy among the six selected, however, it still obtained acceptable performance (R 2 = 0.77, MAPE = 36.87%,RMSE = 32.38 mg/L).Therefore, all the six selected models were applied to retrieve C SPM values of Poyang Lake.The six selected models were applied to the validation dataset to estimate the CSPM values, and validation results were shown in Figure 4. Similar to the calibration results, the B6 or longer wavelength-based models produced better estimation accuracies, which explained 88-93% of the variation of measured CSPM, and the B7-based model produced the best validation performance (R 2 = 0.93, MAPE = 21.54%,RMSE = 16.06 mg/L).The B4-based model had the lowest estimation accuracy among the six selected, however, it still obtained acceptable performance (R 2 = 0.77, MAPE = 36.87%,RMSE = 32.38 mg/L).Therefore, all the six selected models were applied to retrieve CSPM values of Poyang Lake.The six selected models were applied to the validation dataset to estimate the CSPM values, and validation results were shown in Figure 4. Similar to the calibration results, the B6 or longer wavelength-based models produced better estimation accuracies, which explained 88-93% of the variation of measured CSPM, and the B7-based model produced the best validation performance (R 2 = 0.93, MAPE = 21.54%,RMSE = 16.06 mg/L).The B4-based model had the lowest estimation accuracy among the six selected, however, it still obtained acceptable performance (R 2 = 0.77, MAPE = 36.87%,RMSE = 32.38 mg/L).Therefore, all the six selected models were applied to retrieve CSPM values of Poyang Lake.

CSPM Estimation and Comparison
The CSPM values derived from Sentinel 2 MSI B4, B7, and B8b, as well as those from MODIS captured on 15 August 2016 were shown in Figure 5.The distribution patterns of CSPM obtained from the three bands of Sentinel 2 MSI and MODIS Terra were generally consistent across the whole Poyang Lake, while Sentinel 2, with obviously higher spatial resolutions, could resolve more detailed

C SPM Estimation and Comparison
The C SPM values derived from Sentinel 2 MSI B4, B7, and B8b, as well as those from MODIS captured on 15 August 2016 were shown in Figure 5.The distribution patterns of C SPM obtained from the three bands of Sentinel 2 MSI and MODIS Terra were generally consistent across the whole Poyang Lake, while Sentinel 2, with obviously higher spatial resolutions, could resolve more detailed spatial variations than MODIS.The eastern and western parts of the lake had lower C SPM values, while a turbid plume with high loadings of C SPM was observed in the middle part from south to north.However, the Sentinel 2 MSI B4 tended to produce slight lower C SPM values than the other three maps in the very high C SPM regions, while the B8b-derived map showed noise patterns in waters with low C SPM values.Figure 6 shows the scatter plots of the C SPM values derived from Sentinel 2 MSI B4-B8b against those of MODIS on 15 August 2016.The C SPM values derived from Sentinel 2 MSI were highly and significantly correlated with those from the MODIS at a significance level of 0.05 (r > 0.9, p < 0.05), however, the slope of the regression line varied a lot.The slope of the regression line for Sentinel 2 MSI B4 and B5 against MODIS were 0.76 and 0.75, respectively, which means that these two bands tended to produce lower C SPM values than MODIS for high C SPM values.However, the C SPM values retrieved from Sentinel 2 MSI B6-B8b are relatively more consistent with those from MODIS with a higher slope of the regression line between them.Overall, the Sentinel 2 MSI B7 showed the highest consistence with a slope of the regression line of 1.00, though it tended to produce slightly higher C SPM values than MODIS with an intercept of 9.87 mg/L.Figure 7 illustrates the mean and standard deviation values of C SPM across the whole lake derived from Sentinel 2 MSI and MODIS.The mean C SPM values of Sentinel 2 MSI B4-B8b were 22.31, 28.10, 25.51, 29.86, 24.25, and 37.76 mg/L, respectively, which were higher than that of MODIS (19.51 mg/L).The standard deviation values were obviously larger than their corresponding mean values except Sentinel 2 MSI B5, which indicates high spatial variations across the lake.The Sentinel 2 MIS B4 and B5 had lower standard variations than other bands.
The C SPM values derived from Sentinel 2 MSI B4, B7, and B8b as well as those from MODIS on 2 April 2017 were shown in Figure 8.The C SPM values were relatively low in most parts of Poyang Lake on 2 April 2017, except that the central to southern Poyang Lake showed slightly higher SPM loading.Sentinel 2 MSI B4 and B7 produced similar C SPM distribution patterns with MODIS.However, the B8b-derived maps showed noise patterns and obviously held higher C SPM values than the other three maps.Figure 9 illustrates the scatter plots of C SPM values derived from Sentinel 2 MSI B4-B8b against those of MODIS on 2 April 2017.The C SPM values derived from Sentinel 2 MSI were strongly and significantly correlated with those from the MODIS at a significance level of 0.05 (r > 0.8, p < 0.05), however, all of B4-B8b tended to produce higher C SPM values than MODIS.Sentinel 2 MSI B4 showed the highest consistence with MODIS with a correlation coefficient of 0.91, a slope of 0.97, and an intercept of 6.42 mg/L.Sentinel 2 MSI B8b produced the lowest consistence with an intercept of 19.27 mg/L, while B7 produced acceptable results with a correlation coefficient of 0.86 and a regression line of y = 1.11x + 7.63 with MODIS.The mean and standard deviation values of C SPM values from Sentinel 2 MSI and MODIS were illustrated in Figure 10.Similar to results shown in Figure 9, the Sentinel 2 MSI B4 with mean = 13.43 mg/L and standard deviation = 5.21 mg/L was the closest to MODIS (mean = 7.19, standard deviation = 4.87 mg/L), while Sentinel 2 MSI B8b produced an obviously higher mean value (27.18 mg/L) than the other bands.The low standard deviations, ranging from 4.87 to 6.48 mg/L, also indicated low spatial variations of C SPM values on 2 April 2017.The CSPM values derived from Sentinel 2 MSI B4, B7, and B8b as well as those from MODIS on 2 April 2017 were shown in Figure 8.The CSPM values were relatively low in most parts of Poyang Lake on 2 April 2017, except that the central to southern Poyang Lake showed slightly higher SPM loading.Sentinel 2 MSI B4 and B7 produced similar CSPM distribution patterns with MODIS.However, the B8b-derived maps showed noise patterns and obviously held higher CSPM values than the other three maps.Figure 9

Discussion
This study developed CSPM retrieval models with simulated Sentinel 2 MSI spectra and in-situ CSPM values in Poyang Lake.Several models were reported with in-situ optical and CSPM values, and they were successfully applied to satellite images for CSPM retrieval [14,45,46].The models developed in this study should be reliable considering the following two reasons: (i) the Rrs derived from in-situ measurements are less affected by atmospheric interference and are frequently used for spaceborne

Discussion
This study developed C SPM retrieval models with simulated Sentinel 2 MSI spectra and in-situ C SPM values in Poyang Lake.Several models were reported with in-situ optical and C SPM values, and they were successfully applied to satellite images for C SPM retrieval [14,45,46].The models developed in this study should be reliable considering the following two reasons: (i) the R rs derived from in-situ measurements are less affected by atmospheric interference and are frequently used for spaceborne sensor calibration and validation [5,47], and the NASA protocol [38] was strictly followed to ensure high quality optical measurements; and (ii) the synchronous in-situ C SPM values and R rs are more consistent in spatial and temporal scale, while the spatiotemporal scale gap problem always exists between satellite and in-situ observations [3].
This study found that the Sentinel 2 MSI B4-B8b band obtained acceptable to high fitting accuracy and validation performance, and the models derived from Sentinel 2 MSI B1-B3 were less accurate.These calibration results were similar to the models develop for Dongting Lake [3], which is also a large sediment-laden floodpath lake along the middle section of the Yangtze River.According to the bio-optical model [39,48], R rs can be expressed as a function of absorption and backscattering coefficients.The reflectance of short wavelength bands, like Sentinel 2 MSI B1-B3, are affected by the absorption and backscattering characteristics of the water optically active matters (including SPM, chlorophyll, and CDOM) and pure water together, therefore the correlations between C SPM and these three bands are weak (Figure 2).The absorption coefficients of these optically active matters decrease with increasing wavelength, and turn close to zero in NIR regions [49].Therefore, the magnitude and shape of R rs are mainly determined by the particulate backscattering coefficient and pure water absorption in longer spectral regions [50], which explains the better fitting results obtained by Sentinel 2 MSI B6-B8b.The best calibration and validation results obtained by Sentinel 2 MSI B7 (centered at 783 nm) may be explained by the fact that it lies in the left shoulder of the SPM reflectance peak around 810 nm (Figure 2a), which is a result of particulate backscattering and slight decrease in water absorption [51].Such a result is consistent with that of Kutser et al. [23], who demonstrated the capability of Sentinel 2 MSI B7 in retrieving C SPM values over strong absorbing black waters due to its adjacency to 810 reflectance peak.
The absorptions of the optically active matters could still contribute to the reflectance of Sentinel 2 MSI B4 (centered at 665 nm), especially for waters with high loading of SPM or the presence of absorbing substances [3,23].Therefore, the relationships between C SPM and R rs at this band are complicated, and the calibration samples scattered sparsely around the fitting curve in Figure 3, which explains the relatively lower performance obtained by Sentinel 2 MSI B4.Such a statement could also be supported by the findings of Wu et al. [49], in which an obvious phytoplankton absorption peak, a correlation peak between a d and suspended particulate inorganic matter concentration, and a correlation peak between a g and C DOC were observed in this spectral region, while the correlation between a p and C SPM plummeted around 665 nm.However, Sentinel 2 MSI B4 still holds advantage for waters with low C SPM values, which has been widely used for case-II waters [13,16,25].
By applying the models calibrated in this study, the Sentinel 2 MSI B4-B8b produced generally consistent C SPM maps with MODIS.However, compared with MODIS images, the most frequently used data for C SPM monitoring in Poyang Lake, Sentinel 2 MSI can resolve finer variations over small scales.It is understandable that one single MODIS pixel with a 250 m resolution is corresponding to more than 12 × 12 pixels with a 20 m resolution.Therefore, Sentinel 2 MSI can provide more accurate estimations especially in spatially heterogeneous regions.Similar phenomena were also found for Landsat 8 OLI and GF 1 WFV OLI over the MODIS Terra and Aqua [14,28,52].More importantly for the floodpath lake, many parts of Poyang Lake can be narrow in dry seasons, which makes low resolution images unusable, while higher spatial resolution image, such as Landsat 8 OLI and Sentinel 2 MSI, can still work well to resolve small water bodies [53,54].
The Sentinel 2 MSI B8b (centered at 865 nm) might not be suitable for C SPM retrieval because of the noise patterns in the low C SPM waters.This could be explained by the low R rs for waters with low C SPM values due to high water absorption in this band [55].For example, a R rs value of 0.00046 was measured at 865 nm for the water sample with a C SPM value of 23.32 mg/L.Therefore, the water-leaving radiances only contribute to a very small part of the satellite-sensed signal at this band, which could not be separated accurately from the atmospheric interference.Moreover, the relative lower SNR value held by Sentinel 2 MSI B8b may also partially account for the noise pattern.
The C SPM values derived from MODIS were cross-compared with those from Sentinel 2 MSI to determine the reliability and applicability of Sentinel 2 MSI-based models developed in this study, because of lacking synchronous Sentinel image and in-situ C SPM values.The time differences between MODIS and Sentinel 2 MSI overpasses are within 3 h.Therefore, it is reasonable to omit the influence of SPM loadings dynamics on cross-comparisons between Sentinel 2 and MODIS.The MODIS-based C SPM retrieval model was used because it was calibrated and validate with two independent datasets collected in 2007 and 2012, respectively, and found to have acceptable and stable performance [12].Compared with MODIS, Sentinel 2 MSI B4 and B5 tended to underestimate in the regions with higher C SPM values, which might be partially explained by the models' uncertainty and the discrepancy in RSR and SNR of the two sensors [14].The Sentinel 2 MSI B7-derived C SPM values were the most consistent with those from MODIS for turbid waters, which further demonstrates its capability in C SPM retrieval for Poyang Lake.The cross-comparisons between MODIS and other sensors for C SPM retrieval could be found in literatures [14,25].For examples, Wu et al. developed Landsat-5 TM-based C SPM model for Poyang Lake by assuming the MODIS-derived to be the true values [56]; and Tian et al. developed GF-1 WFV-based C SPM model for Deep Bay with assistance of MODIS-derived C SPM values [52].
The high quality SWIR bands of Sentinel 2 MSI also partially contributed to the successful C SPM retrieval in this study, because they facilitate the atmospheric correction over case-II waters like Poyang Lake.The SWIR-based atmospheric correction method has been proven to be reliable and integrated into SEADAS for the operational product generations over turbid waters from MODIS images [57,58].Recently, it is also modified and successfully applied to Landsat 8 OLI, a precursor of Sentinel 2 MSI [28,59,60].Although the Sentinel 2 MSI SWIR bands is about one third of the Landsat 8 OLI SWIR bands in terms of SNR specification, the SNR can be improved through average filtering [44].Therefore, we infer that the SWIR-based atmospheric correction method should produce acceptable performance for Sentinel 2 MSI.
The advantage of high temporal resolution is obvious, since frequent observations would enable C SPM monitoring over short periods.For example, GF 1 WFV with a 4-day revisit period obtained monthly cloud-free observations for Poyang Lake in 2015, while Landsat 8 OLI only had 4 clear scenes over the same period.Sentinel 2 MSI provides a revisit time of 5-day at the equator with the full operations of the two satellites [40].Considering its capability in C SPM retrieval demonstrated in our study, Sentinel 2 MSI should be an ideal data source for the operational C SPM monitoring in Poyang Lake.Novoa et al. [53] developed a switching model for low-to-high turbidity waters, which used red band for low-to-medium turbid waters and NIR band for high turbid waters to retrieve C SPM values.Developing a switching model for Poyang Lake based on Sentinel 2 MSI B4 and B7 using this modeling strategy might work better to capture the complex spatiotemporal C SPM patterns of this floodpath lake.Moreover, further elaborate evaluation of the atmospheric correction and C SPM retrieval accuracy with concurrent in-situ and satellite observations would be a meaningful task.

Conclusions
The Sentinel 2 MSI B4-B8b-based C SPM retrieval models were developed, and they were found acceptable and applicable in estimating C SPM values of Poyang Lake.The Sentinel 2 MSI B4-B8b-derived C SPM maps revealed clear spatial distribution patterns, and a riverine induced turbid plume was observed on 15 August 2016, while the C SPM values were relatively lower across the whole Poyang Lake on 2 April 2017.The consistent results of cross comparisons between MODIS and Sentinel 2 MSI also proved the applicability of the models developed.The Sentinel 2 MSI B7-based power model (C SPM = 2950 × B7 1.357 ) with the highest calibration and validation accuracy, and high consistency with MODIS, is recommended for sediment-laden waters, while the Sentinel 2 MIS B4-based model (4.044 × exp(19.53× B4)) works better for clear waters.This study demonstrated that the Sentinel 2 MSI, with higher spatial resolution than MODIS, more band configurations than GF 1 WFV, and shorter revisiting time than Landsat 8 OLI, should be an appropriate data source for monitoring C SPM over case-II waters such as Poyang Lake.

Figure 2 .
Figure 2. Remote sensing reflectance (Rrs) spectra and their corresponding suspended particulate matter concentrations (mg/L), and the relative spectral response function of the Sentinel 2 MSI (black dash curve) (a); and the correlation coefficient (r) between the Rrs and suspended particulate matter concentration (CSPM) (b).

Figure 2 .
Figure 2. Remote sensing reflectance (R rs ) spectra and their corresponding suspended particulate matter concentrations (mg/L), and the relative spectral response function of the Sentinel 2 MSI (black dash curve) (a); and the correlation coefficient (r) between the R rs and suspended particulate matter concentration (C SPM ) (b).

Figure 4 .
Figure 4. Scatter plots of estimated against measured suspended particulate matter concentration (C SPM ) for the validation dataset: Sentinel 2 MSI B4 (a), B5 (b), B6 (c), B7 (d), B8 (e), and B8b (f).The solid line is the regression line between the estimated and measured values, and the dashed line is 1:1 line.

Figure 5 .
Figure 5. Suspended particulate matter concentrations (CSPM) retrieved from Sentinel 2 MSI B4 (a), B7 (b), B8b (c), and MODIS Terra B1 (d) captured on 15 August 2016.The areas in the red rectangle are zoomed in to show the detailed CSPM variations.

Figure 5 .
Figure 5. Suspended particulate matter concentrations (C SPM ) retrieved from Sentinel 2 MSI B4 (a), B7 (b), B8b (c), and MODIS Terra B1 (d) captured on 15 August 2016.The areas in the red rectangle are zoomed in to show the detailed C SPM variations.

Figure 6 .
Figure 6.Scatter plots of CSPM values derived from Sentinel 2 MSI B4 (a), B5 (b), B6 (c), B7 (d), B8 (e), and B8b (f) against those from MODIS Terra B1 on 15 August 2016.The solid line is the regression line and the dashed line is 1:1 line.The number along the color ramp indicates the pixel number after log transformation (y = log1.05(x)).

Figure 6 . 20 Figure 7 .
Figure 6.Scatter plots of C SPM values derived from Sentinel 2 MSI B4 (a), B5 (b), B6 (c), B7 (d), B8 (e), and B8b (f) against those from MODIS Terra B1 on 15 August 2016.The solid line is the regression line and the dashed line is 1:1 line.The number along the color ramp indicates the pixel number after log transformation (y = log 1.05 (x)).Remote Sens. 2017, 9, 761 12 of 20 illustrates the scatter plots of CSPM values derived from Sentinel 2 MSI B4-B8b against those of MODIS on 2 April 2017.The CSPM values derived from Sentinel 2 MSI were strongly and

Figure 7 .
Figure 7.The mean and standard deviation (Std) of C SPM derived from Sentinel 2 MSI B4-B8b and MODIS Terra B1 on 15 August 2016.

Figure 8 .
Figure 8. Suspended particulate matter concentrations (CSPM) retrieved from Sentinel 2 MSI B4 (a), B7 (b), B8b (c), and MODIS Aqua B1 (d) on 2 April 2017.The areas in the red rectangle are zoomed in to show the detailed CSPM variations.

Figure 8 . 20 Figure 9 .
Figure 8. Suspended particulate matter concentrations (C SPM ) retrieved from Sentinel 2 MSI B4 (a), B7 (b), B8b (c), and MODIS Aqua B1 (d) on 2 April 2017.The areas in the red rectangle are zoomed in to show the detailed C SPM variations.

Figure 10 .
Figure 10.The mean and standard deviation (Std) of C SPM derived from Sentinel 2 MSI B4-B8b and MODIS Aqua B1 sensed on 2 April 2017.

Table 1 .
Statistics describing the concentrations of suspended particulate matter (C SPM , mg/L) of the calibration and validation dataset as well as all the samples.Std.Dev.means standard deviation, and CV is coefficient of variation (%).

Table 2 .
Regression models with the goodness of fitting between suspended particulate matter concentration (CSPM, mg/L) and the simulated water-leaving reflectance at Sentinel-2 MSI B1-B8b.

Table 2 .
Regression models with the goodness of fitting between suspended particulate matter concentration (C SPM , mg/L) and the simulated water-leaving reflectance at Sentinel-2 MSI B1-B8b.
Note: MAPE is mean absolute percentage error (%), RMSE is the root mean square error (mg/L), F is the F-value of the model's significance test.