A Remote-Sensing Method to Estimate Bulk Refractive Index of Suspended Particles from GOCI Satellite Measurements over Bohai Sea and Yellow Sea

Featured Application

The study proposed a multiple-step hybrid remote sensing method to estimate bulk refractive index (n_p) of suspended particles in the Bohai Sea and Yellow Sea from GOCI satellite measurements. This proposed method can be applied to study the spatial and temporal variations of n_p in the Bohai Sea and Yellow Sea, and thereby to understand the particulate biogeochemical properties (e.g., composition and size) and their role in exploring the changes of marine environments.

Abstract

The bulk refractive index (n_p) of suspended particles, an apparent measure of particulate refraction capability and yet an essential element of particulate compositions and optical properties, is a critical indicator that helps understand many biogeochemical processes and ecosystems in marine waters. Remote estimation of n_p remains a very challenging task. Here, a multiple-step hybrid model is developed to estimate the n_p in the Bohai Sea (BS) and Yellow Sea (YS) through obtaining two key intermediate parameters (i.e., particulate backscattering ratio, B_p, and particle size distribution (PSD) slope, j) from remote-sensing reflectance, R_rs(λ). The in situ observed datasets available to us were collected from four cruise surveys during a period from 2014 to 2017 in the BS and YS, covering beam attenuation (c_p), scattering (b_p), and backscattering (b_bp) coefficients, total suspended matter (TSM) concentrations, and R_rs(λ). Based on those in situ observation data, two retrieval algorithms for TSM and b_bp were firstly established from R_rs(λ), and then close empirical relationships between c_p and b_p with TSM could be constructed to determine the B_p and j parameters. The series of steps for the n_p estimation model proposed in this study can be summarized as follows: R_rs (λ) → TSM and b_bp, TSM → b_p → cp → j, b_bp and b_p → B_p, and j and B_p → n_p. This method shows a high degree of fit (R² = 0.85) between the measured and modeled n_p by validation, with low predictive errors (such as a mean relative error, MRE, of 2.55%), while satellite-derived results also reveal good performance (R² = 0.95, MRE = 2.32%). A spatial distribution pattern of n_p in January 2017 derived from GOCI (Geostationary Ocean Color Imager) data agrees well with those in situ observations. This also verifies the satisfactory performance of our developed n_p estimation model. Applying this model to GOCI data for one year (from December 2014 to November 2015), we document the n_p spatial distribution patterns at different time scales (such as monthly, seasonal, and annual scales) for the first time in the study areas. While the applicability of our developed method to other water areas is unknown, our findings in the current study demonstrate that the method presented here can serve as a proof-of-concept template to remotely estimate n_p in other coastal optically complex water bodies.

1. Introduction

The bulk refractive index (n_p) of suspended particles in natural waters is a critical physical parameter that describes particulate intrinsic properties such as material composition, shape, texture, and structure [1,2,3,4,5]. This parameter carries much critical information that significantly supports the research on the underwater light properties of particulate assemblages, and it is closely related to many marine ecological and biogeochemical processes and is, thus, capable of contributing to the knowledge of regional and even global ocean ecosystems [4,6]. Consequently, acquiring n_p information, such as its spatiotemporal distribution, is of great significance to us.

At present, several measurement methods for the bulk refractive index of suspended particles are available to us, such as the laser diffraction method, electrical resistance technique, and flow cytometry analysis [7,8,9]. These methods show relatively high observational accuracy and yet are time-consuming and laborious and need a large quantity of water sample collections. More importantly, they struggle to provide variations in large-scale synoptic and temporal distributions. Actually, the n_p data records remain scarce, especially for large-scale marine waters. The remote sensing technique provides a possibility for filling this gap.

A large quantity of algorithms was developed for remote sensing estimation on water condition parameters such as chlorophyll a [10,11,12,13,14], suspended particulate matter [15,16,17], turbidity [18,19], and even colored dissolved organic matter [20,21,22]. However, it is unfortunate that fewer algorithms were developed to derive the bulk refractive index of suspended particles. Based on the Twardowski et al. (2001) [2] model, i.e., modeling the n_p using the particulate backscattering ratio (B_p) and hyperbolic slope of the particle size distribution (j), Suresh et al. (2006) [23] developed an empirical band ratio method to model B_p and j and then further estimated the n_p in the Arabian Sea from Remote Sensing Satellite (IRS-P4), Ocean Color Monitor (OCM)satellite data. They found that the n_p values were low in the open ocean and relatively high in the coastal waters. Based on Mie theory, Nasiha et al. (2014) [4] developed a retrieval model to estimate the n_p to understand particulate assemblage dynamics in coastal waters. To apply this inversion model to satellite data, Nasiha et al. (2015) [5] further proposed a proof-of-concept method to derive the n_p from Moderate-Resolution Imagine Spectroradiometer (MODIS)-Aqua images. This method firstly uses an empirical relationship to estimate turbidity based on single green band (551 nm) reflectance and then deduces the parameters, including B_p, j, and particulate apparent density (ρ_a), which are finally used to calculate the refraction index.

Clearly, the n_p estimation models using remotely sensed data remain very limited. Additionally, the existing methods are mostly based on empirical relationships such as those between R_rs(λ) band ratios with the parameters B_p and j of Suresh et al., (2006) [23] and fitted empirical relationships between R_rs(λ) with turbidity and B_p [5]. Although case studies showed good performance, the lack of a necessary physical basis still limits the use of these methods, despite their proof of concept. Therefore, more validation should be undertaken to evaluate the previous limited methods for application to other water areas of interest. Importantly, new n_p estimation models should be developed to cope with different water conditions, especially for turbid coastal waters.

This study collects an adequate bio-optical dataset that covers the measurements of inherent optical properties, such as particulate backscattering (b_bp), scattering (b_p), and beam attenuation (c_p) coefficients, remote sensing reflectance (R_rs(λ)) measurement, and total suspended matter (TSM) measurement. The investigated water areas of this study were the Bohai Sea and Yellow Sea. The objective of the current study was to develop a method for remote sensing estimation of n_p by using satellite ocean color data. This study proposes a multiple-step hybrid model that firstly depends on the TSM and b_bp retrievals, then the established close relationships between b_p and c_p with TSM, and ultimately upon the Twardowski et al. (2001) [2] model derived from Mie theory calculations. By using independent datasets including in situ data and satellite retrievals, the proposed method is assessed, and then applied to Geostationary Ocean Color Imager (GOCI) data. The spatiotemporal distribution patterns of n_p are documented. At last, some necessary discussions are appended.

2. Data and Methods

2.1. The Study Water Areas

The investigated water areas of this study were the marginal seas of the northwest Pacific Ocean along with China, i.e., the Bohai Sea (BS) and Yellow Sea (YS). The BS and YS are typically large shallow semi-enclosed seas with a water depth from several meters to about 100 m (Figure 1). These seas are highly turbid and optically complex waters that are significantly influenced by terrigenous discharge. The datasets used in the present study were collected from four cruise surveys in November 2014, August 2015, July 2016, and January 2017 in the BS and YS (Figure 1).

2.2. Bio-Optical Measurements

By using a profiling package, bio-optical measurements were performed to measure various optical parameters. The package included a Seabird ABE911P conductivity–temperature–depth (CTD) profiler that was used to measure hydrological characteristics of the water bodies and some optical devices, including a WET Labs AC-S and a HOBI Labs Hydroscat-6 (HS-6). The particulate absorption coefficient (a_p(λ)) and beam attenuation coefficient (c_p(λ)) in the spectral range of 400–700 nm were observed with the AC-S instrument, and then the particulate scattering coefficient (b_p(λ)) could be derived using the relationship b_p(λ) = c_p(λ) − a_p(λ). The particulate backscattering coefficient, b_bp(λ), was observed by the HS-6 instrument which has six spectral channels, i.e., 442, 488, 550, 620, 700, and 852 nm, while the blue and green bands (i.e., 442, 488, and 550 nm) were used as delegates for the analysis in this study. The detailed measurement methods for the AC-S and HS-6 instruments can be seen in the Sun et al. (2016) [24] study.

The remote sensing reflectance (R_rs(λ)) spectra were collected with a Satlantic Hyper-Profiler II radiometer [25]. Meanwhile, water samples were simultaneously collected along with the above optical measurements and then filtered immediately in the lab onboard. The filtered particulates on GF/F filters were used to analyze total suspended matter (TSM) concentrations [24]. The collected bio-optical parameters during the four cruises are summarized in

Table 1. Description of in situ observed datasets collected during four cruise surveys in the study area. TSM—total suspended matter.

Cruise Date	Measured Parameters	Sample Numbers
7 November 2014–23 November 2014	Rrs(λ) (sr−1)	27
	TSM (mg∙L−1)	108
17 August 2015–5 September 2015	Rrs(λ) (sr−1)	37
	TSM (mg∙L−1)	101
29 June 2016–14 July 2016	Rrs(λ) (sr−1)	58
	TSM (mg∙L−1)	123
29 December 2016–13 January 2017	Rrs(λ) (sr−1)	24
	TSM (mg∙L−1)	103
	bbp (m−1)	102
	bp (m−1)	102
	cp (m−1)	103

.

2.3. n_p Calculation

The bulk refractive index of suspended particles (n_p) was calculated using a Mie theory-based relationship model from the so-called input parameters, namely, the particulate backscattering ratio B_P and the particle size distribution (PSD) slope j (Equation (1) by Twardowski et al., (2001) [2]).

$$n_{\mathrm{p}}=1+B_{{}_{\mathrm{p}}}^{0.5377+0.4867(j-3){}^2}[1.4676+2.2950{(j-3)}^2+2.3113{(j-3)}^4].$$

(1)

Here, the B_p (=b_bp/b_p), b_bp, and b_p were measured by HS-6 and AC-S. The c_p(λ) spectrum can be modeled using a hyperbolic power function [2,26,27].

$$c_{\mathrm{p}}(\lambda )=c_{\mathrm{p}}({\lambda }_{\mathrm{ref}})\times {(\frac{\lambda }{{\lambda }_{\mathrm{ref}}})}^{-\beta },$$

(2)

where λ_ref is a reference wavelength, and β is the spectral slope parameter. By using the c_p at different wavelengths, we were accordingly able to calculate parameter β. According to Boss et al. [28], the parameter β is found to covary with the parameter j, which can be refined as the following equation for more PSD cases:

$$j=\beta +3-0.5e^{-6\beta }.$$

(3)

Therefore, we can obtain the parameter j to calculate the n_p.

2.4. GOCI Data Collection and Processing

In this study, GOCI Level-1B data were used to assess the application of the developed refractive index (n_p) model. As a geostationary ocean color remote sensing satellite, GOCI receives images eight times a day from 12:15 a.m. to 7:45 a.m. Greenwich Mean Time (GMT) (8:15 to 15:15 local time) with 1-h temporal resolution. The GOCI satellite data have a spatial coverage of about 2500 km × 2500 km covering the northwest Pacific Ocean, a spatial resolution of 500 m, and eight spectral bands with the central wavelengths of 412, 443, 490, 555, 660, 680, 745, and 865 nm. In this study, a total of 2825 GOCI satellite images during the one-year period from December 2014 to November 2015 were downloaded from the Korea Ocean Satellite Center (KOSC). By utilizing the GOCI Data Processing System (GDPS, version 1.3), the image data focusing on the BS and YS regions were firstly extracted and then processed by the default atmospheric correction method of Wang & Gordon (1994) [29] to output R_rs(λ) data. These R_rs(λ) data were further quality controlled by using the various flags such as stray light and cloud coverage to avoid interference from invalid satellite data signals.

2.5. Development of n_p Estimation Model

2.5.1. Estimation of TSM from $R_{\mathrm{rs}}\left(\lambda \right)$

Estimation of TSM is a key step in retrieving the final target, n_p, by R_rs(λ). This study firstly analyzed the correlation between single R_rs(λ) and TSM in the spectral range of 400–700 nm, the results of which showed that the red bands were the most sensitive to TSM vis-à-vis the other bands (Figure 2). For instance, the correlation coefficients (R) showed high values (0.838 and 0.839) in the 660-nm and 680-nm bands of the GOCI image data, respectively. These findings also agree with those of previous studies [30,31]. These bands can be subsequently considered for use in establishing a TSM model.

Empirical methods are simple and easy to model and, more importantly, straight and efficient, particularly for local regions [32]. Thus, this study made use of single bands, band ratios, and band combinations to develop the TSM model. Specifically, eight band forms, as shown in

Table 2. Correlations between TSM and band form X derived from GOCI (Geostationary Ocean Color Imager) six bands. X1 to X8 indicate eight band forms, respectively; R is the correlation coefficient.

X	Detailed Band Form	Optimal Band Combination	R
X1	Rrs(λ1)	λ1 = 680 nm	0.839
X2	log10(Rrs(λ1))	λ1 = 680 nm	0.588
X3	Rrs(λ1) − Rrs(λ2)	λ1 = 660 nm, λ2 = 680 nm	0.725
X4	$\frac{R_{\mathrm{rs}}\left({\lambda }_1\right)}{R_{\mathrm{rs}}\left({\lambda }_2\right)}$	λ1 = 660 nm, λ2 = 555 nm	0.606
X5	$\frac{{\log }_{10}{R}_{\mathrm{rs}}\left({\lambda }_1\right)}{{\log }_{10}{R}_{\mathrm{rs}}\left({\lambda }_2\right)}$	λ1 = 490 nm, λ2 = 660 nm	0.581
X6	$\frac{R_{\mathrm{rs}}{\lambda }_1-R_{\mathrm{rs}}\left({\lambda }_2\right)}{R_{\mathrm{rs}}{\lambda }_1/R_{\mathrm{rs}}\left({\lambda }_2\right)}$	λ1 = 660 nm, λ2 = 680 nm	0.739
X7	$\frac{R_{\mathrm{rs}}{\lambda }_1+R_{\mathrm{rs}}\left({\lambda }_2\right)}{R_{\mathrm{rs}}\left({\lambda }_1\right)/R_{\mathrm{rs}}\left({\lambda }_2\right)}$	λ1 = 555 nm, λ2 = 660 nm	0.847
X8	$\frac{R_{\mathrm{rs}}{\lambda }_1-R_{\mathrm{rs}}\left({\lambda }_2\right)}{R_{\mathrm{rs}}\left({\lambda }_1\right)+R_{\mathrm{rs}}\left({\lambda }_2\right)}$	λ1 = 660 nm, λ2 = 555 nm	0.555

, were tested. For each band form, the correlation with TSM was examined for all possible combinations from six GOCI R_rc(λ) bands, and the optimal band combinations with the highest R are given in Table 2. After comparing the correlations among these band forms, we determined X7, i.e., (R_rs(555) + R_rs(660))/(R_rs(555)/R_rs(660)), with the highest R of 0.847 among those band forms for establishing the TSM model. This method to find the optimal band form and optimal band combination was similar to that used in previous studies [33,34]. Several mathematical methods, including linear, power, exponential, and logarithmic functions, were used to model TSM, and their accuracies were then intercompared to obtain a good retrieval model. The obtained results showed that the simple linear model continued to perform best (Figure 3), and it was subsequently used to estimate TSM in our whole model framework.

2.5.2. Estimation of $b_{\mathrm{bp}}$ and $B_{\mathrm{p}}$ from $R_{\mathrm{rs}}\left(\lambda \right)$

Estimation of the particulate backscattering coefficient, b_bp(λ), is another important step. Similar to TSM estimation, this study firstly carried out a correlation analysis and then demonstrated that the X7 band combination form was the best indicator for the b_bp(λ) estimation (Figure 4B). Note that the 488-nm channel was used to denote the parameter due to its better performance when compared to those of the other channels. After testing several different mathematical functions, we selected and established two strong relationships using the X7 form to model the b_bp(488). As shown in Figure 4C, the fitted determination coefficient, R², was 0.820 (p < 0.001), with relatively low predictive errors.

2.5.3. Derivation of the PSD Slope, j, from TSM

A close relationship between TSM and b_p(488) could be established with relatively high fitting accuracy (Figure 5A), which was also consistent with that demonstrated in the He et al. [32] study. Meanwhile, very close relationships between b_p(λ) and c_p(λ) could also be found in our study area, as observed in previous studies [5,35,36,37]. As shown in Figure 5B, their relationships could be modeled well, with an extremely high R² of 0.998 (p < 0.001). Such good relationships provide a stable basis for accurately deriving c_p(λ) at different wavelengths from TSM. According to Equations (2) and (3), we obtained the parameter j, which was then used as an input of the model estimating n_p in this study.

As a summary, Figure 6 shows the main processes of our developed model in the current study to estimate the bulk refractive index n_p of particles by R_rs(λ). A Mie-theory-based relationship model was essentially adopted to derive n_p values from the so-called input parameters, namely, B_P and j (Equation (1) by Twardowski et al. (2001) [2]), considering the fact that there are currently hardly any methods that can be used to directly measure the particle bulk refraction. Although some empirical relationships were used in the framework of our model development, the proposed n_p estimation model is a meaningful attempt to obtain satellite-derived n_p records.

2.6. Performance Metrics

To evaluate the models’ performances, this study utilized the root-mean-square error (RMSE), mean absolute error (MAE), mean relative error (MRE), determination coefficient (R²), and relative bias (Bias). These indicators can be calculated as follows:

$$\mathrm{RMSE}=\sqrt{\frac{1}{N}{\displaystyle {\sum }_{i=1}^N{(y_i^{\prime }-y_i)}^2}},$$

(4)

$$\mathrm{MAE}=\frac{1}{N}{\displaystyle {\sum }_{i=1}^N\left|y_i^{\prime }-y_i\right|},$$

(5)

$$\mathrm{MRE}=\frac{1}{N}{\displaystyle {\sum }_{i=1}^N\left|\frac{y_i^{\prime }-y_i}{y_i}\right|},$$

(6)

$$\mathrm{Bias}=\frac{1}{N}{\displaystyle {\sum }_{i=1}^N\left(y_i-y_i^{\prime }\right)},$$

(7)

where y_i and y_i’ are the measured and predictive values for the i-th sample, respectively, and N represents the total number of samples.

3. Results

3.1. Water Bio-Optical Conditions

The bio-optical parameters in our investigated water regions showed very large dynamic ranges, which were collected from the four cruises (Figure 7). TSM varied from 0.6 to 223.9 mg∙L⁻¹ and had a mean of 11.2 ± 22. 7 mg∙L⁻¹. The large variation coefficient (CV)value (202.51%) indicates its large variation. The particulate backscattering coefficient at 488 nm, b_bp (488), ranged between ~0 and 1.21 m⁻¹, had a mean of 0.19 ± 0.17 m⁻¹, and had a large CV (141.7%).

Correspondingly, the particulate scattering and beam attenuation coefficients, b_p(488), c_p(532), and c_p(555), also showed wide variation ranges, as well as large CV values. The particulate backscattering ratio B_p(488), i.e., b_bp/b_p, changed from ~0 to 0.10 when the PSD slope, j, was observed to be in the range of 3.2–5.0. Note that a large span also appeared for the bulk refractive index of particles n_p. Meanwhile, the collected in situ R_rs(λ) showed large variations in terms of both magnitude and spectral shapes (Figure 8). These R_rs(λ) spectra had similar spectral properties to those seen in previous reports on coastal waters, for example, the studies of He et al. (2013) [32] and Sun et al. (2014) [36], which indicate typical optically complex turbid water conditions in the study area.

3.2. Validation of n_p Estimation Model

To evaluate the performance of the developed n_p model, we firstly used an independent validation dataset with 24 samples collected from the BS and YS in January 2017. As shown in Figure 9A, the modeled and in situ n_p values generally showed good clustering along the 1:1 line, with relatively low predictive errors. Most of the data points (approximately 84% percentage) were distributed within the ±5% range. On the other hand, this study carried out satellite synchronization verification. After matching GOCI satellite-derived R_rs data with in situ measurements using a set of strict constraints, a limited seven synchronous samples (±1 h) were obtained for use in comparisons between the in situ measured and satellite-derived n_p. Figure 9B shows the generally good agreement between them, with very low predictive errors (such as MRE = 2.32%), despite the modeled values being on the high side to some extent. This implies that the developed n_p model has the potential ability to produce accurate and acceptable results.

Additionally, we compared the spatial distributions of particle bulk refraction produced by in situ measurements and satellite derivation. Note that the satellite data used were from the specific monthly coverage concurrent with the cruise in January 2017. Figure 9C,D show the spatial patterns derived by in situ observations and satellite retrievals, respectively. Although there was no absolute real-time synchronization between the in situ measured and satellite-derived monthly n_p, the cruise sampling covered a period of nearly one month and was able to roughly provide an accurate reference. The generally similar spatial distributions of the in situ observations and satellite retrievals imply the accuracy of our developed n_p model.

3.3. Model Application to Satellite Data

We could produce hourly n_p products by applying the developed model into the hourly GOCI R_rs(λ) data. Those hourly products were then synthesized into daily, monthly, seasonally, and annually averaged products. As shown in Figure 10, the general distribution pattern for 2015 is available. High n_p values generally dominated nearshore waters, while low values were mainly distributed in the offshore waters. The whole BS region usually had relatively high refraction values of approximately >1.10, whereas most areas in the YS, except coastal regions, showed low values (roughly <1.05). Note that a large water region that includes Subei shoal and the Yangtze River estuary appeared to have very high refraction, which even exceeded 1.15 for some areas.

The seasonal and monthly variations in n_p in this study water areas were further generated as shown in Figure 11. On the whole, the n_p during the period from December 2014 to April 2015, as well as November 2015, appeared to have the most high-valued distribution throughout the entirety of the BS and YS regions, whereas relatively low values showed up between May and October 2015. In detail, in several particular water regions, including the Yellow River estuary, Yangtze River estuary, Subei shoal, and very nearshore areas, the n_p values remained almost static over time and were always high (>1.15) throughout the entire year. In most regions of the BS, the n_p values showed distinct seasonal variations, namely, high values in winter, early spring, and late autumn, whereas the values in summer were low. Similar temporal trends in n_p variation also appeared in the YS. However, the difference with that in the BS was that there existed generally lower n_p values in the YS for the same month.

4. Discussion

4.1. Advantages and Disadvantages of the n_p Model

The n_p model proposed in this study is a proof-of-concept remote sensing method for retrieving particulate refraction from GOCI satellite measurements in the study area. It is essentially dependent on the accurate retrievals of TSM and b_bp, which are achieved by establishing stable empirical models. The variety of field measurements from the four cruises represents a prominent advantage in this study and provides a steady basis for developing the TSM and b_bp empirical models. As shown in Figure 12, a leave-one-out cross-validation (LOO-CV) [38,39] was conducted to evaluate the performances of the TSM and b_bp models. Briefly, we randomly selected one sample from the full dataset (sample number, n) that served for the validation of the model, while the remaining n − 1 samples were trained for model calibration. Based on the LOO-CV method, all samples were tried for one round, deriving relatively low and acceptable predictive errors. Thus, relatively accurate TSM and b_bp retrievals, as well as their close relationships with b_p and c_p, provided a feasible route for achieving the estimation of n_p in the study areas. Another advantage lies in that there are more adequate remote sensing signals for model inputs when compared with the remote sensing method proposed by Nasiha et al. (2015) [5], which utilizes a single band reflectance (i.e., a green band at 551 nm) to derive the intermediate and target parameters. In addition to green band reflectance (555 nm), this study introduced red band reflectance, i.e., 660 nm, as another input, which assures more adequate valid remote sensing information that could potentially derive better outputs.

The developed n_p model may show limitations when it is applied to other coastal water areas. This is due to different optical properties that may vary with the change of study area because of the diversity and dynamics of in-water constituents [24,40,41,42,43,44]. Such local characteristics on water conditions generally determine region-specific empirical relationships between TSM (and b_bp) and remote sensing reflectance and between b_p (and c_p) and TSM [16,17,45,46]. Nevertheless, these so-called limitations should not hinder the use of the proof-of-concept method proposed in this study for other regions. The same method with necessary local parameterizations can be employed to derive the particulate refraction using remote sensing data only if valid regional data can be obtained. However, in future, detailed investigations are still required to thoroughly assess the model’s applicability in various coastal water conditions. In addition, this study mainly focused on developing a method for estimating n_p from satellite data; thus, we used GOCI data as a case study. The applicability of the method to other satellite sensors is not clear, and further studies are also required to apply the similar method of this study to develop n_p estimation models specifically for MODIS, Visible infrared Imaging Radiometer (VIIRS), Landsat, etc., according to their band specifications, and then compare and cross-validate their performances.

4.2. Comparison with the Method Using a Straight B_p Empirical Model

The particulate backscattering ratio, B_p, is a straight input parameter to calculate n_p. In our developed n_p model, this parameter was obtained through the ratio by the modeled b_bp and b_p. However, we may still have another alternative, i.e., developing a B_p empirical model. By employing steps similar to those in Section 2.5.1 and Section 2.5.2, we developed an empirical relationship between B_p(488) and X8 that performed better than other functional models (see Figure 13A). By using the new method nested with the B_p empirical model, this study then obtained new n_p results, which showed roughly similar performance with that of the previous method (Figure 13B). However, note that the n_p spatial distribution derived by the new method was not very desirable, as there existed obvious overestimations for the offshore areas (Figure 13C). This was probably due to the fitted over-empirical relationship between R_rs and B_p of Yang et al., (2011) [47] because there is no physical linkage between them in theory, considering that B_p (= b_bp/b_p), especially b_p therein, is essentially not retrievable from remote sensing owing to the fact that remote sensing covers no information on the forward scattering of suspended particles by Sun et al., (2016) [24]. By comparison, our proposed n_p retrieval method still showed superior performance in the study water areas.

4.3. Driving Factors of n_p Spatiotemporal Variation

By applying the developed n_p model to time-series GOCI satellite data, the n_p spatiotemporal variations in the study areas could be obtained, with the derived values spanning a range of 1.01–1.22 (Figure 10 and Figure 11). This variation in magnitude is generally consistent with that reported in previous studies [2,5,23,48,49]. In general, the n_p in the nearshore regions was higher than that offshore, noting that the nearshore regions showed relatively stable high values, whereas the offshore areas changed a lot during different seasons (or months). High-content mineral particles in the nearshore waters (e.g., the Yellow River estuary and Yangtze River estuary) induced by terrestrial discharge, sediment resuspension, and even shore erosion can explain the high n_p (1.12–1.18, see Figure 10) distribution there.

Regarding the offshore areas, the n_p values in the BS were higher than those in the YS. This is because the BS is prone to sediment resuspension under the action of wind waves due to its relatively shallow depths [50,51]. Song et al. (2014) [52] demonstrated that inorganic minerals and silts controlled the suspended particles in the BS. Therefore, those hard mineral and sediment particles may explain the high n_p values in the BS (except those nearshore regions), which were mainly in the range of 1.10–1.15 (Figure 10) [2,5,53]. Fortunately, the derived n_p distributions generally agree with the previous report [5]. By contrast, the YS waters (except those nearshore regions) are easily affected by algal particles, where diatoms and dinoflagellates are the main algal species [54,55,56]. Importantly, these algal particles generally show a relatively low bulk refraction (approximately 1.01–1.07) due to their high water content [57,58,59]. Therefore, the low refraction distribution in most regions of the YS (i.e., n_p < 1.07, see Figure 10) can be attributed to the dominance of organic algal particles.

From the perspective of the seasonal variations, the n_p showed the lowest distribution in summer for most areas of the BS and YS, whereas the highest distribution appeared in winter. The seasonal variations are probably related to water column mixing and stratification, as well as phytoplankton growth and extinction during different seasons [60,61]. Actually, the factors influencing the n_p distribution could be very complex, which may be linked with many marine physical and biogeochemical processes such as wind waves, river discharge, hydrodynamic conditions, water column mixing and stratification, sediment suspension, phytoplankton growth, and zooplankton grazing. In short, the precise understanding on the mechanisms influencing the n_p spatiotemporal variations remains challenging owing to the coupling influence of these multiple factors.

4.4. Implications for Marine Environmental Changes

The bulk refractive index of suspended particles (n_p) is a key parameter that reflects the intrinsic characteristics of suspended particles, is closely related to the particulate composition and size structures [2], and significantly affects the optical properties of suspended particles. Actually, different n_p distribution ranges are indicative of different types of suspended particles present in water bodies such as 1.01–1.09 mainly for organic algal particles, >1.15 for water regions dominated by inorganic particles and detritus, and 1.09–1.15 for complex and mixed particulates [2,3,48,49]. Satellite-derived n_p values in the study areas spanned a wide range (approximately 1.01–1.22), thus implying a large spatial heterogeneity. Three types of water regions can be generalized as (1) offshore oceanic water areas dominated by phytoplankton, (2) river plume and sediment-laden nearshore regions dominated by inorganic mineral particles, and (3) coastal regions controlled by organic and inorganic mixed materials. Although no in situ method exists for directly measuring the refractive index of suspended particles, the validation revealed the robustness of the currently proposed method to detect n_p for the study areas and further distinguish waters dominated by different types of suspended particles. In short, the spatiotemporal detection of n_p provides a new data record that enriches suspended particulate properties from space, and importantly improves understanding of the particulate biogeochemical properties and their role in exploring the changes of marine environments.

5. Conclusions

The current study proposed and validated a multiple-step hybrid remote sensing method to estimate n_p in the BS and YS from GOCI satellite measurements. This n_p estimation model depends on three crucial steps, namely, (1) the TSM and b_bp retrievals, (2) establishing close relationships between b_p and c_p with TSM, and (3) utilizing the Twardowski et al. (2001) [2] model derived from Mie theory. The first two steps were calibrated and validated using a large quantity of in situ observed data from four cruises in the study area, whereas the third step was directly utilized, as it is widely recognized by the community. These steps jointly constructed our developed n_p estimation model in this study, which led to reliable performances with reasonable uncertainties (2.55% and 2.32% for MRE, validated using in situ observations with the leave-one-out cross-validation method and satellite-derived results, respectively) and spatial distribution patterns. By means of 2825 GOCI satellite images collected during one year from December 2014 to November 2015, this n_p estimation model documented n_p distribution patterns in the BS and YS at different temporal scales for the first time, such as monthly, seasonal, and annual scales. Although this method is possibly region-specific and may need more validation for its application to other water areas, the developed method here provides a proof-of-concept template for n_p remote sensing estimation in other coastal turbid water bodies.