HY-1C/D CZI Image Atmospheric Correction and Quantifying Suspended Particulate Matter

Abstract

HY-1C/D both carry a coastal zone imager (CZI) with a spatial resolution of 50 m and a swath width of 950 km, two observations can be achieved in three days when two satellites operating in a network. Accurate atmospheric correction is the basis for quantitative inversion of ocean color parameters using CZI However, atmospheric correction in estuarine and coastal waters with complex optical properties is a challenge due to the band setting of CZI. This paper proposed a novel atmospheric correction algorithm for CZI images applicable to turbid waters in estuarine and coastal zone. The Rayleigh scattering reflectance of CZI was calculated based on a vector radiative transfer model. Next, a semi-empirical radiative transfer model with suspended particle concentration as the parameter is used to model the water-atmosphere coupling. Finally, the parameters of the coupling model are solved by combining a global optimization method based on a genetic algorithm. The results indicate that the CZI-derived remote-sensing reflectance (R_rs) are in good agreement with the quasi-synchronous Landsat-8/9 operational land imager (OLI) derived R_rs in the green and red bands (R² > 0.96). Validation using in situ data revealed that the RMSE of the CZI-derived R_rs in the green and red bands was 0.0036 sr⁻¹ and 0.0035 sr⁻¹. More importantly, the values and spatial distributions of suspended particulate matter (SPM) estimated by CZI and those estimated by OLI in the Subei Shoal and the Yangtze River Estuary are basically consistent, and the validation using in situ data revealed that the inversion of SPM concentration by CZI was effective (R² = 0.86, RMSE = 0.0362 g/L), indicating that CZI has great potential and broad application prospects for monitoring the spatial and temporal dynamics of SPM in estuarine and coastal waters. The study results will lay the foundation for further estimating suspended sediment fluxes and carbon fluxes, thus providing data support and scientific basis for promoting resource development, utilization and conservation strategies in estuarine and coastal areas.

1. Introduction

The estuary is the hub of the basin and the ocean, and the coastal zone is the link between the land and the ocean. Estuarine and coastal areas are the concentrated areas of land-sea interaction, with complex evolutionary mechanisms of various physical, chemical, biological and geological processes, and the ecological environment is sensitive and fragile [1,2]. At the same time, the estuarine and coastal areas are also economically developed and densely populated. Strong human activities continue to exert influence on estuarine and coastal areas [3]. Therefore, the research, development, and protection of estuarine and coastal areas is a hot issue of great concern to coastal states and scientists around the world, and research objectives, plans, and governance measures have been proposed.

The estuarine and coastal waters gather various terrestrial (basin) materials: freshwater runoff, sediment and chemical substances, which change its lateral and bottom boundaries and its regional ecological environment under the action of powerful oceanic waves, currents and tides [4]. Estuarine and coastal waters are often turbid due to the input of terrestrial sediments and resuspension of bottom sediments [5]. Moreover, the distribution of suspended particulate matter (SPM) is highly uneven [6,7], which not only has a great impact on the regulation of channels [8], but also will affect the shoreline variation and tidal flats evolution in the long term. Remote sensing technology, with the advantages of large-scale synchronous observation, rapidity, and economical efficiency, enables dynamic monitoring of SPM [9]. High-resolution remote sensing images are important for SPM monitoring in estuarine and coastal waters [10] However, as a trade-off in terms of emission cost, signal-to-noise ratio, band settings, revisit period, etc., conventional ocean color sensors typically have coarse spatial resolution [11], which is insufficient to evaluate the details of SPM variability in estuarine and coastal waters.

Moreover, the revisit period of widely-used high-resolution sensors such as the operational land imager (OLI) on-board Landsat-8/9 is 16 days, and that of the multispectral instrument (MSI) on-board Sentinel-2A/B is 10 days, which is not conducive to observe the SPM changes of highly dynamic estuarine and coastal waters [12]. Meanwhile, the width of OLI and MSI (but at present, only partial images of the subdivided width can be downloaded) are 180 km and 290 km, respectively, which is inadequate to SPM estimation on a large scale in the coastal zone. Fortunately, the coastal zone imager (CZI) on-board HY-1C and HY-1D satellites launched in 2018 and 2020 has a width of about 950 km, a spatial resolution better than 50 m [13], and a revisit period of two observations in three days after the HY-1C/D network operation. Therefore, the CZI sensor is very suitable for the study of estuarine and coastal waters [14,15,16,17,18].

However, the lack of a suitable atmospheric correction method limits the application of CZI images in the assessment of ocean color elements over turbid waters. The atmospheric contribution mainly includes Rayleigh scattering of atmospheric molecules and aerosol scattering of aerosol molecules [12]. Currently, the Rayleigh scattering reflectance can be accurately and efficiently calculated by solving the radiative transfer equation to construct a look-up table (LUTs). However, the accurate calculation of aerosol scattering, which is the basis of reliable atmospheric correction, has been challenging [19]. The standard atmospheric correction algorithm [20] uses two or more near-infrared (NIR) bands in the Sea-viewing Wide Field-of-View Sensor Data Analysis System (SeaDAS) to estimate the contribution of aerosols in the visible band. The contribution of this approach to visible band aerosols from open ocean waters is sufficient because the NIR ocean signal in the open ocean is assumed to be zero and can be regarded as fully atmospheric [21], but it is challenging to conduct atmospheric correction operations in turbid estuarine and coastal waters [22,23] where the signal in the NIR spectrum is not negligible due to high SPM concentrations [13,24]. To solve this problem, a common approach is to use the short-wave infrared (SWIR) band which strongly absorbs water signal even in turbid waters [25,26,27], and the water-leaving radiances in the visible bands can be obtained by extrapolation [28]. However, these algorithms typically require two infrared bands [29,30], while CZI has only one NIR band at 825 nm. Therefore, accurate atmospheric correction of turbid water with CZI images is a challenge.

In this study, a novel atmospheric correction method for CZI image applicable to turbid estuarine and coastal waters is proposed, and it is verified using OLI data and in situ data, to further evaluate its potential for SPM inversion. The key points are as follows: (1) The Rayleigh scattering reflectance of CZI image is calculated based on a vector radiative transfer model. (2) a semi-empirical radiative transfer model with suspended particle concentration as the parameter is used to model the water-atmosphere coupling, the parameters of the coupling model are solved by combining a global optimization method based on a genetic algorithm. (3) the potential of CZI image to estimate SPM is explored. These results will provide unique insights to obtain SPM distributions and further estimate suspended sediment fluxes and carbon fluxes in estuarine and coastal waters.

The main contents of this study were managed as follows: The introduction of the study area, in situ data and satellite data are given in Section 2. Section 3 presents the methods of CZI atmospheric correction and SPM inversion used in this study. The results are shown in Section 4, including cross-validation of the reflectance and SPM concentration obtained from CZI images with OLI images, validation of in situ reflectance data and SPM concentration data with CZI images. The advantages, atmospheric correction results and application prospects of CZI images are discussed in Section 5. Finally, the conclusions are summarized in Section 6.

2. Materials

2.1. Study Area

The study area mainly includes the Subei Shoal (SBS) and the Yangtze River Estuary (YRE) (Figure 1). The SBS is located along the southeast coast of Jiangsu Province, and is a large intertidal flats and radial sand ridges. It is the largest shallow water area in China with most of the coastal waters locating within the 20 m isobath. The strong tidal force in this region makes the water well-mixed and highly turbid, resulting in permanent turbidity maxima [31,32]. Since the shallow water depth and proximity to the mainland, the seawater in its north is mainly influenced by the inlet runoff from the Guan river, Sheyang river, and the irrigation main canal in northern Jiangsu. On the other hand, the Yangtze River Diluted Water is able to influence the seawater in south of 33°N in the southern SBS [33].

The YRE is a large mesotidal subtropical estuary, where a substantial volume of sediment discharge from the Yangtze River flows into the sea [8]. The developed economy and large population of the YRE region imply its fundamental socioeconomic importance, as a tidal wetland, it is also extremely important ecologically. The SPM within the YRE is highly dynamic due to complex hydrological processes.

2.2. In Situ Data

As for the apparent optical properties (AOPs), we use Hyperspectral Surface Acquisition System (HyperSAS) to in situ measure spectral radiance parameters for estimating remote-sensing reflectance (R_rs). Two radiance sensors were set away from the sun with an optimal zenith angle of 40° and an optimal azimuth angle of 135° to minimize the impact of wind speed and reduce solar glitter effects [34]. The field-measured R_rs is calculated as follows:

$$R_{\mathrm{rs}}\left(\lambda \right)=\frac{L_t\left(\lambda \right)-\rho \left(\lambda \right)\times L_S\left(\lambda \right)}{E_d\left(\lambda \right)}$$

(1)

where L_t(λ), L_s(λ) and E_d(λ) represent the total upwelling spectral radiance, incident spectral radiance and downwelling spectral irradiance, respectively. Here, ρ(λ) is a ratio of spectral reflected sky radiance.

Water samples are collected simultaneously during field sampling, and then obtain SPM concentration after filtering, drying, weighing and other operations. The specific experimental operation was detailed in the study of Luo et al. [8].

2.3. Satellite Data

Landsat-8 and Landsat-9 satellites were launched in February 2013 and September 2021, respectively. They carried the OLI sensor with a spatial resolution of 30 m and revisit period of 16 days, the joint network operation of Landsat-9 and Landsat-8 has enabled the observation of the Earth every 8 days. OLI has a higher signal-to-noise ratios (SNR) than the thematic mapper (TM) on-board Landsat-4/5 and the enhanced thematic mapper plus (ETM+) on-board Landsat-7 [35]. Level-1 Landsat-8/9 OLI data of the SBS and YRE were downloaded from the United States Geological Survey (USGS, https://earthexplorer.usgs.gov/ (last accessed on 16 June 2022)).

HY-1C and HY-1D satellites were launched in September 2018 and June 2020, respectively, they both carried a CZI sensor with a wide swath width of about 950 km, a high spatial resolution better than 50 m, and high timeliness that can achieve twice observations in three days. Level-1 HY-1C/D CZI data of the SBS and YRE were downloaded from the National Satellite Ocean Application Service (https://osdds.nsoas.org.cn/OceanColor/ (last accessed on 12 June 2022)). The CZI sensor has four bands from visible to near-infrared, with wavelengths ranging from 0.42-0.89 μm and center wavelengths of 460, 560, 650, and 825 nm, and a quantization level of 12 bits. It is mainly used for coastal zone monitoring, understanding the distribution and change pattern of suspended matter, and to conduct real-time monitoring and early warning of marine environmental disasters such as ice disaster, red tide, green tide and pollutants.

In summary, the advantageous characteristics of HY-1C/D CZI compared with Landsat-8/9 OLI are shown in

Table 1. The advantageous characteristics of the coastal zone imager (CZI) on-board HY-1C/D compared with the operational land imager (OLI) on-board Landsat-8/9.

Satellite/Sensor	Landsat-8/9: OLI	HY-1C/D: CZI
Revisit (days)	Single: 16 Networking: 8	Single: 3 Networking: twice every 3 days
Scene Size (km)	185	950
Path time	10:00 a.m. ± 15 min local time	C: 10:30 a.m. ± 30 min local timeD: 1:30 p.m. ± 30 min local time

. The spectral response curves of CZI and OLI are shown in Figure S1.

3. Methods

In order to evaluate the potential of HY-1C/D-CZI images for quantifying SPM, it is necessary to establish accurate atmospheric correction and SPM inversion algorithms, which are described in detail below. In addition, considering that the band settings and spatial resolution of CZI and OLI are similar, this study will cross-validate the products of CZI and OLI at all levels.

3.1. Reflectance at Top of Atmosphere

The raw data received by the remote sensing sensors are stored in digital numerical (DN) format, and the downloaded DN data is converted into the zenith radiance (L_TOA) through radiometric calibration:

$$L_{\mathrm{TOA}}=\mathrm{gain}\times \mathrm{DN}+\mathrm{offset}$$

(2)

where the DN and the corresponding gain and offset for each band are provided by the downloadable Level 1 products of CZI and OLI.

In order to eliminate the influence of solar geometric differences and solar spectral radiation differences of corresponding bands, before the comparison of OLI and CZI, the L_TOA should be converted into the top-of-atmosphere reflectance (ρ_TOA) by Equation (3):

$${\rho }_{\mathrm{TOA}}\left(\lambda \right)=\frac{\pi \times L_{\mathrm{TOA}}\left(\lambda \right)}{E_0\left(\lambda \right)\times COS\left(\theta \right)/d^2}$$

(3)

where E₀(λ) is the average solar spectral irradiance at λ, the angle at which sunlight incident to the earth’s surface is called the solar zenith angle (θ) and d is the earth-sun distance in astronomical units (AU).

In addition, to maximize the elimination of effects such as solar zenith angle and spectral response function between OLI and CZI, the relative relationship between L_TOA of CZI and OLI images was obtained using the linear relationship between ρ_TOA of CZI and OLI images combined with Equation (3).

$${\rho }_{\mathrm{TOA}-\mathrm{OLI}}\left(\lambda \right)=a\times {\rho }_{\mathrm{TOA}-CZI}\left(\lambda \right)+b$$

(4)

$$\begin{gathered} L_{\mathrm{TOA}-\mathrm{OLI}}\left(\lambda \right)=a\times \frac{E_{0-OLI}\left(\lambda \right)\times cos\left({\theta }_{OLI}\right)}{E_{0-CZI}\left(\lambda \right)\times cos\left({\theta }_{CZI}\right)}\times L_{TOA-CZI}\left(\lambda \right) \\ +b\times \frac{E_{0-OLI}\left(\lambda \right)\times cos\left({\theta }_{OLI}\right)}{\pi } \end{gathered}$$

(5)

where E_0−OLI and E_0−CZI were the average solar spectral irradiance at the equivalent wavelength of OLI and CZI sensors, θ_OLI and θ_CZI were the observed solar zenith angle of OLI and CZI sensors.

3.2. Atmospheric Correction

The total reflectance, at a wavelength λ, measured at the top of atmosphere can be expressed as [21]:

$${\rho }_{\mathrm{TOA}}={\rho }_{\mathrm{rc}}+{\rho }_a+t\times {\rho }_{\omega }$$

(6)

where ρ_rc is the Rayleigh reflectance, ρ_a is the aerosol reflectance, t is the diffuse transmittance, ρ_w is the water-leaving reflectance.

The atmospheric correction mainly includes Rayleigh scattering correction and aerosol scattering correction. Currently, the Rayleigh scattering correction can be accurately calculated through LUTs. However, in general, specific LUTs need to be generated for different sensors. The generalized Rayleigh scattering correction LUTs generated based on a vector radiative transfer model has good universality, and can meet the Rayleigh scattering correction of different sensors [36,37], which is convenient and fast to use.

The aerosol scattering correction is mainly based on NIR or SWIR dark pixel method. For the atmospheric correction of turbid water in the absence of SWIR band, ESOA provides a better solution by coupling the semi-empirical radiative transfer model applicable to highly turbid waters with the AMHAD2010 aerosol model, to solve the aerosol parameters of the image for aerosol scattering correction. Since CZI lacks the band required for the dark pixel atmospheric correction method, this study refers to the ESOA method [24], constructs a water-atmosphere coupling model suitable for CZI images, and develops the ESOA-CZI atmospheric correction method.

In short, for atmospheric correction of CZI images, this study first uses the generalized Rayleigh scattering correction LUTs to complete the Rayleigh scattering correction of CZI images, and then uses ESOA-CZI to perform aerosol scattering correction of CZI images, which will be described in detail in the following sections.

3.2.1. CZI Rayleigh Correction

Current accurate Rayleigh scattering calculations for ocean color remote sensing are usually performed using look-up tables (LUTs), but since these LUTs are generated for specific remote sensing sensors, they are difficult to be directly applied to new ocean color remote sensing sensors. For this reason, He et al. [36,37] independently-developed a vector radiative transfer model termed PCOART for the coupled ocean–atmosphere system with rough sea-surface, which solves the vector radiative transfer equation using the matrix-operator or adding-doubling method.

A comparison of the results with the SeaDAS accurate Rayleigh scattering LUTs show that the calculation accuracy of the PCOART universal LUTs were better than 0.5%, which can be used for accurate Rayleigh scattering calculations for all ocean color remote sensing sensors. Therefore, Rayleigh correction of CZI was performed using PCOART in this study. The LUTs has been operated and could be downloaded from https://www.satco2.com/index.php?m=content&c=index&a=lists&catid=399/ (last accessed on 14 July 2022).

3.2.2. CZI Aerosol Correction

Semi-Empirical Radiative Transfer Model

For turbid estuarine and coastal waters, a semi-empirical radiative transfer (SERT) model based on a two-stream approximate radiative transfer theory [38] was proposed by Shen et al. [39] for remote sensing estimation of SPM, assuming that the ratio of backscattering coefficient (b_b) and absorption coefficient (a) is positively correlated with SPM, taking into account that suspended particulate matter in the water mainly has a sensitive effect on scattering, while it has little sensitive effect on absorption. The specific form of the SERT model is as follows:

$$R_{\mathrm{rs}}\left(\lambda \right)=\frac{u\left(\lambda \right)\times v\left(\lambda \right)\times SPM}{1+v\left(\lambda \right)\times SPM+\sqrt{1+2\times v\left(\lambda \right)\times SPM}}$$

(7)

where R_rs is the atmospherically corrected remote sensing reflectance in sr⁻¹, u and v are empirical and wavelength-dependent coefficients, and SPM is the suspended particulate matter concentration in g/L.

In order to reduce the discrepancy between the CZI atmospherically corrected R_rs and the in situ measured R_rs. In this study, based on the spectral response function (SRF) of CZI, the in situ measured R_rs(λ) was converted into R_rs(λ) measured at the equivalent wavelength of CZI using Equation (8).

$$R_{\mathrm{rs}}\left(ban{d}_i\right)=\frac{{{\displaystyle \int }}_{{\lambda }_1}^{{\lambda }_2}{R}_{\mathrm{rs}}\left(\lambda \right)SRF\left(\lambda \right)d\left(\lambda \right)}{{{\displaystyle \int }}_{{\lambda }_1}^{{\lambda }_2}SRF\left(\lambda \right)d\left(\lambda \right)}$$

(8)

where R_rs(band_i) is the equivalent R_rs(sr⁻¹) measured by the CZI for the i-th band; λ₁ and λ₂ are the lower and upper limits in the wavelength range for the i-th band, respectively; SRF(λ) is the spectral response function of the CZI at λ; R_rs(λ) is the in situ measured R_rs at λ.

The simultaneously simulated CZI-based derived R_rs (calculated by Equation (8)) and SPM datasets (716 samples) from multiple cruises in the estuarine and coastal waters were available for the determination of fitting parameters u and v of the SERT model based on the least-squares method according to Equation (7). In this study, the SERT model after fitting the sensor equivalent wavelength and re-optimizing the calibration parameters is referred to as the SERT_c model. The wavelength-dependent u and v are listed in

Table 2. u and v coefficients of the SERT_c model for CZI data.

Center band for CZI (nm)	u	v	RMSE (sr−1)	MAE (sr−1)	MAPE (%)	Correlation (R2, N = 716)
460	0.0246	419.1596	0.0048	0.0038	44.16	0.6300
560	0.0466	146.1654	0.0060	0.0047	36.75	0.8126
650	0.0699	32.5096	0.0053	0.0039	41.03	0.9121
825	0.0984	3.8635	0.0034	0.0024	72.46	0.9392

, the results show that the SERT_c model has higher accuracy in the red and near-infrared bands.

Implementation of ESOA_CZI Based on a Genetic Algorithm

The sum of the aerosol reflectance ρ_a(λ) and the normalized water-leaving reflectance [ρ_w(λ)]_N at the zenith is expressed as ρ_aw(λ):

$${\rho }_{\mathrm{a}\mathsf{\omega }}\left(\lambda \right)={\rho }_a\left(\lambda \right)+t_v\left(\lambda \right)\times {[{\rho }_{\omega }\left(\lambda \right)]}_N$$

(9)

where t_s(λ) and t_v(λ) represent the diffuse transmittance in the solar direction and the sensor observation direction, respectively.

Based on the definition of normalized water-leaving reflectance [ρ_w(λ)]_N and remote sensing reflectance R_rs(λ), it is easy to derive Equation (10):

$$R_{\mathrm{rs}}\left(\lambda \right)=\pi \times {[{\rho }_{\omega }\left(\lambda \right)]}_N$$

(10)

The AHMAD2010 aerosol model [40] and the SERT water radiative transfer model were used to calculate ρ_a(λ), t_s(λ), t_v(λ) and R_rs(λ), respectively, and then ρ_aw(λ) can be calculated by coupling Equations (9) and (10).

In addition, the top-of-atmosphere reflectance ρ_TOA can be linearly expressed as the sum of the reflectivity of each part. In this study, the ρ_TOA(λ) after Rayleigh scattering (ρ_rc(λ)), white cap reflection ρ_wc(λ), and flare reflection ρ_glint(λ) correction is expressed by ρ_n(λ):

$${\rho }_n\left(\lambda \right)={\rho }_{\mathrm{TOA}}\left(\lambda \right)-{\rho }_{\mathrm{rc}}\left(\lambda \right)-t_v\left(\lambda \right)\times {\rho }_{\mathsf{\omega }\mathrm{c}}\left(\lambda \right)-T\left(\lambda \right)\times {\rho }_{\mathrm{glint}}\left(\lambda \right)$$

(11)

where T(λ) represent the solar direct transmittance. Note that in the Equation (11), we ignore the sea surface reflectance from whitecaps and sun-glint [41].

In theory, ρ_n(λ) is equal to ρ_aw(λ) at all wavelengths, so combined with the above sea- atmosphere coupling model, we developed a system of nonlinear equations for CZI sensors:

$$\begin{array}{c}{\rho }_n\left(460\right)={\rho }_{\mathrm{aw}}\left(460, RH, FMF,{\tau }_a\left({\lambda }_0\right), SPM \right)\\ {\rho }_n\left(560\right)={\rho }_{\mathrm{aw}}\left(560, RH, FMF,{\tau }_a\left({\lambda }_0\right), SPM \right)\\ {\rho }_n\left(650\right)={\rho }_{\mathrm{aw}}\left(460, RH, FMF,{\tau }_a\left({\lambda }_0\right), SPM \right)\\ {\rho }_n\left(825\right)={\rho }_{\mathrm{aw}}\left(825, RH, FMF,{\tau }_a\left({\lambda }_0\right), SPM \right)\end{array}$$

(12)

where RH and FMF represent relative humidity and fine-mode fraction, respectively.

The unknown variables which are relative humidity (RH) and fine-mode fraction (FMF) that determine the aerosol type, the aerosol optical thickness at the reference wavelength τa(λ₀) and SPM, are estimated from nonlinear equations (Equation (12)).

Due to the large spatial and temporal variability of SPM and aerosol optical properties in the SBS and YRE, it is difficult for SOA2009 [42] and its previous spectral optimization algorithms to accurately estimate the initial values of the four unknown variables in Equation (12). Genetic Algorithm (GA) is a global optimization algorithm that has been widely adopted and does not depend on the initial estimation of the unknown variables. For the solution of a specific problem, GA requires the determination of three main aspects: the ranges and encodings of the variables, the determination of the objective function, and the selection of the genetic operator algorithm and important parameters. The main parameters of GA model for CZI are the same as those set by Pan et al. [24] for GOCI. Therefore, in this study, GA is used to optimally solve for the variables of Equation (12).

3.2.3. OLI Atmospheric Correction

The ACOLITE bundles the atmospheric correction algorithms and processing software for aquatic applications of high-resolution satellite data, especially suited for processing of turbid waters, The ACOLITE performs well for atmospheric correction over turbid estuarine and coastal waters [8,43]. The Level-1 Landsat-8/9 OLI data were processed using the open-source ACOLITE software (version 20220222.0), in combination with the dark spectrum fitting (DSF) algorithm (with glint correction), to obtain ρ_TOA and remote sensing reflectance (R_rs) [44].

3.3. Inversion of Suspended Particulate Matter Concentration

The study area of this paper is the Subei Shoal and the Yangtze River Estuary, which are located in the estuarine and coastal areas, where SPM in the waters is dominant, leading to high turbidity of the waters, and drastic changes in SPM dynamics. Therefore, it is of great significance to monitor SPM in this region using remote sensing data with high spatial resolution. Meanwhile, for turbid waters, R_rs is mainly affected by SPM, and the inversion algorithm of SPM is relatively mature. In addition, CZI only sets four bands of 460 nm, 560 nm, 650 nm and 825 nm, which makes it difficult and low accuracy to invert other ocean color parameters (such as chlorophyll a, CDOM, etc.).

The suspended particulate matter concentration in the SBS and YRE was retrieved by implementing the SERT_c model (Section 3.2.2) to the atmospherically corrected R_rs images of CZI and OLI. The 716 sets of R_rs and SPM matched datasets were collected simultaneously in the estuarine and coastal areas in recent years and divided into modeling and validation datasets according to 8:2 for the SERT_c model (Figure 2). The results show a high coefficient of determination of 0.91 for the R_rs and SPM modeling datasets (N = 572), while the mean absolute error of the validation dataset (N = 144) is below 0.0731 g/L.

In addition, the original SERT model was validated using the same validation dataset (Figure S2), and the comparison results showed that the accuracy of the SERT_c model to retrieve the SPM concentration was better than that of the SERT model. In the practical application of remote sensing images, the inverse form of SERT_c model is shown as follows:

$$SPM=\frac{2\times u\left(\lambda \right)\times R_{rs}\left(\lambda \right)}{v\left(\lambda \right)\times {(u\left(\lambda \right)-R_{rs}\left(\lambda \right))}^2{}_{}}$$

(13)

where for the red band of the CZI sensor, u is 0.0697 and v is 32.7876 (Figure 2a); for the red band of the OLI sensor, u is 0.0709 and v is 31.1277 (Figure 2c).

3.4. Accuracy Evaluation

The values of CZI images and OLI images can be compared based on the numerous methods or indicators available to assess the agreement between two data sets. The International Ocean Color Coordination Group (IOCCG) recommends the mean absolute percentage error (MAPE), the root mean square error (RMSE), the coefficient of determination R², defined as (Pearson coefficient), or the slope and intercept of linear regression [45]. MAPE and RMSE were obtained according to Equations (14) and (15), respectively.

$$\mathrm{MAPE}=\frac{1}{n}{\displaystyle \sum }_{i=1}^n\left|\frac{X_i^{estimated}-X_i^{measured}}{X_i^{measured}}\right|\times 100\%$$

(14)

$$\mathrm{RMSE}=\sqrt{\frac{1}{n}{\displaystyle \sum }_{i=1}^n{(X_i^{estimated}-X_i^{measured})}^2}$$

(15)

where X represents a variable (e.g., ρ_TOA, R_rs, SPM), n is the number of samples, and the superscripts estimated and measured represent the estimated and measured values, respectively.

4. Results

To evaluate the effect of ESOA_CZI atmospheric correction and the potential of CZI to estimate SPM, this study uses OLI data to cross-validate with CZI data products at all levels, and assists in verification with in situ data. Considering the difference in transit time (about 30 min), as well as the difference in spectral response functions of each band, combined with the difference of some other factors between CZI and OLI, these are not conducive to the direct comparison between them. Therefore, before the cross-comparison between CZI and OLI, CZI was relatively corrected using the linear relationship between the original L_TOA of CZI and OLI, and the angular information such as solar zenith angle is also consistent with OLI, so that the input atmospheric correction parameters are consistent as much as possible to better test the effect of ESOA_CZI atmospheric correction. Meanwhile, CZI uses the nearest-neighbor image element method to resample to a spatial resolution of 30 m × 30 m in order to generate a uniform resolution dimension with OLI images. In addition, the blue band is not considered in this study because of its atmospheric correction instability.

4.1. Comparison of TOA Reflectance between CZI and OLI

In this study, cloud-free CZI and OLI images covering the SBS and YRE were collected on 27 February 2022. Figure 3 shows that the ρ_TOA of CZI and OLI have strong correlation, specifically, from low to high reflectance in all three bands (green, red, NIR) showing a good linear relationship. Therefore, the CZI image can satisfy the need for satellite R_rs observations in waters with different turbidity by using suitable atmospheric correction process. Among them, the best results were obtained for ρ_TOA in the red band, the slope of their linear regression fitting line was closer to 1, the intercept was smaller, and the distribution of its results was closer to the 1:1 line. In addition, the coefficient of determination was as high as 0.9803 (n = 38331813), with the smaller RMSE and MAPE of 0.0050 W m⁻² μm⁻¹ and 5.01%, respectively. The overall results of ρ_TOA in the green and near-infrared bands are good, although the ρ_TOA of CZI is a little high in the green band and a little low in the high reflectance part of ρ_TOA in the NIR band.

4.2. R_rs Comparison and Validation

The R_rs of the waters were obtained after atmospheric correction of ρ_TOA. Figure 4 shows the good linear relationship of R_rs in green, red and NIR bands of CZI and OLI images. The results show that the R_rs between CZI and OLI in the red band have good agreement (R² = 0.9761), low RMSE (0.0020 sr⁻¹) and low MAPE (18.10%), the R_rs values derived from the red band in CZI and OLI data were basically equal (Figure 4b). Large MAPE values are observed in the NIR band (Figure 4c), where the reason could be that the R_rs is close to zero in the low value part of R_rs, while the R_rs of OLI is large in the high value part of R_rs. For the green band, the low R_rs value part of CZI is lower than that of OLI (Figure 4a).

In order to further test the atmospheric correction performance of the ESOA_CZI algorithm, we collected in situ measured quasi-synchronous R_rs data from two calibration stations (Muping station (121°42′00″E, 37°40′51.6″N) and Dongtou station (121°21′18″E, 27°40′30″N)) of the National Satellite Ocean Application Service for verification. First, all the in situ R_rs data of the two stations obtained from September 2019 to June 2022 were matched with the CZI images, and we select the R_rs data whose time window between the sampling time of in situ measurement and the transit time of the HY-1C/D is less than half an hour. Next, the CZI-derived R_rs values were obtained by taking the location of the calibration station as the central pixel and combining 8 surrounding pixels to form a 3 × 3 grid, and calculating the average value of R_rs within the grid. Quality control was performed before processing according to the rule that the effective pixels in the grid were more than 8 and the coefficient of variation was less than 15%.

Since the wavelengths of the in situ R_rs at the calibration stations do not exactly coincide with the central wavelengths of CZI at 460, 560, 650 and 825 nm, which the closest corresponding wavelengths are 490, 560, 667 and 779 nm. Therefore, we used 716 hyperspectral R_rs data measured on multiple cruises in estuarine and coastal areas, and also to reduce the error between the measured hyperspectral R_rs data and CZI multispectral R_rs data, the hyperspectral R_rs data were fitted to the corresponding center wavelength of CZI band based on the spectral response function of CZI. Next, the linear relationships between the CZI-based fitted R_rs at 460, 560, 650 and 825 nm and the corresponding in situ measured at 490, 560, 667 and 779 nm were established, respectively, and the in situ R_rs from two calibration stations at 490, 560, 667 and 779 nm were corrected and migrated to R_rs at 460, 560, 650 and 825 nm using these linear relationships. Finally, 28 matchups of satellite data and simultaneous in situ data were successfully obtained, and the mean values of the in situ R_rs from two calibration stations at 460, 560, 650 and 825 nm were 0.0135 sr⁻¹, 0.0201 sr⁻¹, 0.0107 sr⁻¹ and 0.0015 sr⁻¹, respectively, with the corresponding coefficients of variation were 30.72%, 39.85%, 76.54% and 143.53%, respectively.

Finally, the scatter plots of the in situ measured R_rs and CZI-derived R_rs at 560, 650 and 825 nm was shown in Figure 5. As mentioned before, due to the poor atmospheric correction effect of the blue band (460 nm), it is not involved in the comparison here. The results are as follows: the R² values of the CZI-derived R_rs and the in situ measured R_rs at green (560 nm), red (650 nm) and NIR (825 nm) bands were 0.8040, 0.8810 and 0.5295, respectively. In addition, the RMSE were 0.0036 sr⁻¹, 0.0035 sr⁻¹ and 0.0045 sr⁻¹, respectively. Furthermore, the MAPE were 0.1844%, 0.6239% and 91.1752%, respectively. Most of the data are concentrated around the 1:1 line and no negative values were obtained, which indicates that ESOA_CZI algorithm performs well in these bands.

These results show that ESOA_CZI algorithm has the ability to accurately realize the atmospheric correction of CZI, even in estuarine and coastal waters with large spatio-temporal dynamic changes, coexistence of clear water and turbid water, and complex optical properties. More importantly, the red band (650 nm) used to estimate SPM in this study has the highest R² (0.8810) and the smallest RMSE (0.0035 sr⁻¹), and its results are also basically distributed along the 1:1 line. Combining the SERT_c algorithm with strong robustness to estimate the SPM of estuarine and coastal waters has great development potential and broad application prospects.

4.3. Comparison and Validation of CZI and OLI for Quantifying SPM

According to the results of R_rs after CZI atmospheric correction, it can be seen that R_rs in the red band has the best effect. Moreover, the R_rs in the green band is not sensitive to the change of SPM and is easy to saturate. However, the R_rs in the NIR band is less stable and likely to have negative values in the clear water area. Therefore, to illustrate the accuracy of CZI and OLI images inversion of SPM, only red band was selected to input into SERT_c model to invert SPM. The cross-comparison scatter plot of CZI and OLI image inversions of SPM are shown in Figure 6. The results show that the SPM concentration of CZI inversion was highly consistent with that of OLI inversion (R² = 0.9481), and the slope of their linear regression fitting line was very close to 1 (slope = 0.9870), the intercept was as low as 0.0006 g/L, the RMSE was only 0.0167 g/L, and the MAPE was 26.39%. This shows that it is feasible and reliable to estimate SPM of SBS and YRE by SERT_c model using ESOA_CZI atmospheric corrected R_rs, and the future application of CZI images to long time series and large-scale SPM observation of estuarine and coastal waters has great value and potential.

The SPM results of SBS and YRE estimated from CZI and OLI images on 27 February 2022 are shown in Figure 7. The spatial distribution of SPM estimated from CZI and OLI images has good consistency, and both reflect the distribution pattern of SPM in the SBS and YRE waters. Compared with the medium-resolution images, the high-resolution images can observe the dynamic changes of SPM around the nearshore, islands and tidal flats, as well as details such as fine eddies in the water and the wake of ships.

To further test the performance of the CZI image quantifying SPM, we selected the SPM data quasi-synchronous (within one hour) with the local transit time of CZI images from the in situ measured SPM datasets for verification, and a total of 16 sets of SPM data were matched (Figure 8). The verification results show that the R² values was 0.86, the RMSE was 0.0362 g/L, and the MAPE was 33.21%. Most of the data are concentrated around the 1:1 line, which indicates that the CZI images are capable of estimating the SPM distribution and variability in estuarine and coastal waters by ESOA_CZI atmospheric correction and SERT_c inversion SPM algorithm.

4.4. CZI Images for Quantifying SPM

Taking advantage of the short revisit period of CZI images that can achieve 6 observations in 7 days for the SBS and YRE regions (Figure 9). Combine with the advantages of 50 m resolution and wide width, CZI images has great advantage and potential in estimating the SPM of estuarine and coastal areas. As shown in Figure 9, the spatial and temporal dynamics of SPM in SBS and YRE vary greatly, and the SPM generally shows a distribution pattern of decreasing from west (coastal) to east (open sea) and decreasing from north (SBS) to south (YRE). The SPM concentration in the coastal area of Yancheng and the waters adjacent to the radial sand ridge group of SBS is the highest, generally greater than 0.7 g/L. The SPM concentration in the coastal area of Nantong to the south decreases to about 0.3–0.7 g/L, and the SPM concentration in the maximum turbidity zone of the YRE in the further south is only about 0.2–0.4 g/L, and the SPM concentration in the upper section of YRE is as low as 0.05–0.2 g/L. The SPM concentration in the outer sea is even lower, roughly below 0.05 g/L.

Interestingly, the SPM plume with distinctive features often appears in the sea area east of the radial sand ridge group in the SBS, which are particularly spectacular and beautiful natural scenery, which may be formed by the action of underwater topography and tidal current. The water level caused by flood and ebb tides directly determines the size of the exposed tidal flats in the radial sand ridge group. Generally, the tidal flats have a large, exposed area and low SPM concentration at low tide level (Figure 9c), while the tidal flats are flooded and have high SPM concentration at high tide level (Figure 9e), which may be caused by the resuspension of the substrate in the shallows due to strong tidal power.

4.5. Advantages of Fine Spatial Resolution

Spatial resolution has a significant effect on the imaging accuracy of objects with high spatial variability, and to quantify the differences in feature observations by sensors with different spatial resolutions, the coefficient of variation (CV) is used to measure the amount of information contained within the waters observed by the sensor:

$$CV=\frac{SD}{MN}\times 100\%$$

(16)

where SD is the standard deviation, and MN is the mean value.

An HY-1C/CZI image of SBS local area in 27 February 2022 was selected. Details such as groin, vegetation, clear water areas and turbid water zones with different turbidity levels can be clearly seen from the RGB true color image with a spatial resolution of 50 m (Figure 10a). Figure 10b–d showed the original image of red band (650 nm) with a spatial resolution of 50 m, and resized image with spatial resolution of 100 m and 500 m. The cross-comparison of the L_TOA in this region of different spatial resolution is shown in Figure 10e,f.

50 m spatial resolution image are sufficient to distinguish small-scale objects such as groin, sea reclamation boundary, etc. (Figure 10b), while 500 m resolution image cannot observe the details of these features (Figure 10d). Some studies have shown that the optimal scale observation of estuarine and coastal areas should be less than 100 m, or even less than 50 m in some cases [46]. The results of cross-comparison of images with different spatial resolutions show that the reduction in spatial resolution results in significant errors (Figure 10e,f). In addition, when the resolution decreases, the information and details of remote sensing data are gradually lost, and the CV values in the region gradually decreases from 22.13% for 50 m resolution images to 16.11% for 500 m resolution images (Figure 10b–d). Therefore, compared with the spatial resolution of water color sensors at the level of hundred meters, the spatial resolution of CZI at 50 m has an incomparable advantage in observing information and details of estuarine and coastal areas.

5. Discussion

Compared with other remote sensing sensors, HY-1C/D CZI combines the advantages of high spatial resolution (50 m), high temporal resolution (2 observations in 3 days) and wide swath width (950 km), which has unparalleled advantages for monitoring estuarine and coastal waters with large dynamic changes and complex optical properties [12]. Specifically, only 9 CZI images are needed to cover the whole estuarine and coastal waters of China, while at least 46 OLI images or more MSI images are needed. The combined HY-1C/D network is capable of 10 or 12 observations in 16 days, with observations in both morning and afternoon, while the combined Landsat-8/9 network is capable of only 2 observations in 16 days, with no observation in the afternoon.

The premise of CZI application in ocean color observation is the development of a reliable atmospheric correction algorithm [12,13], so the ESOA_CZI algorithm was proposed in this study. The reason why selecting OLI data for cross-validation is that OLI and CZI have similar band settings and similar spatial resolution (Table 1), and previous studies have shown that OLI can be applied to retrieve ocean color parameters with good results [8]. However, it is difficult to achieve complete synchronization between CZI and OLI, i.e., the difference in observation time between the 2 sensors is roughly half an hour, which will inevitably cause angular differences such as the solar zenith angle, and the estuarine and coastal waters also have certain variations within half an hour under the influence of ocean currents. At the same time, the spectral response functions of the two sensors are also different Figure S1. Therefore, if a systematic correction of the L_TOA of CZI is required, it is necessary to make a linear correction by the ρ_TOA of CZI and OLI (Equation (4)) and decompose the effect of the solar zenith angle and the spectral response function (Equation (5)). In theory, the linear relationship between the ρ_TOA of CZI and OLI can be used to recalibrate the L_TOA to give specific and fixed calibration coefficients. Unfortunately, in practice, the L_TOA correction coefficients between CZI and OLI sensors are found to be somewhat different after comparing multiple sets of images at different times, which may be that the sensors have some performance degradation in different periods or the atmospheric environment parameters such as aerosols are different when imaging in different seasons. Therefore, this study uses the linear relationship between the original L_TOA of CZI and OLI to make a relative L_TOA correction of CZI, and ensures that the angle information such as solar zenith angle was also consistent with OLI, so that the input atmospheric correction parameters are consistent. The results show that the R_rs obtained by CZI using ESOA_CZI maintain good agreement with those obtained by OLI using Acolite software (Figure 4), which proves that ESOA_CZI atmospheric correction algorithm is feasible and has high accuracy, especially in the green and red bands. The main reason for the large difference of R_rs in the NIR band may be that the wavelength difference of the CZI and OLI in the NIR band is 40 nm, which leads to a certain degree of unsatisfactory effect of L_TOA linear correction.

Usually in the blue band, more than 80% of the total signal received by the satellite is due to the confounding influences of the atmosphere [21], this may be the reason for the poor atmospheric correction effect of the ESOA_CZI algorithm at blue wavelengths. The large dynamic variation of SPM in estuarine and coastal waters is the combined result of the stochastic processes of sediment movement, the heterogeneity of underwater topography, the seasonal variations of runoff and wind, and the periodic movements of tidal currents and waves, making it challenging to observe SPM in estuarine and coastal waters [8,43,47]. Therefore, we have developed atmospheric correction methods and SPM estimation models suitable for CZI images. The atmospheric correction results at green, red and NIR bands (Figure 4 and Figure 5) and the effect of SPM inversion (Figure 6, Figure 7 and Figure 8) show that the CZI has the potential to provide high-quality ocean color products for scientific research and society-level applications.

6. Conclusions

In order to take advantage of the high spatial resolution, high temporal resolution and large scene size of HY-1C/D CZI, and to exploit the potential of CZI quantitative inversion of ocean color products, this paper proposes a new atmospheric correction algorithm ESOA_CZI for CZI images applicable to turbid water in estuarine and coastal zone. The algorithm was firstly based on the vector radiative transfer model termed PCOART to calculate CZI Rayleigh scattering reflectance. Next, a semi-empirical radiative transfer model with suspended particle concentration as the parameter is used to model the water-atmosphere coupling. Finally, the parameters of the coupling model are solved by combining a global optimization approach that is based on a genetic algorithm.

The accuracy of the ESOA_CZI algorithm was verified using the quasi-synchronous OLI data and the in situ data from the calibration station. Comparison with the OLI-derived R_rs showed that the CZI-derived R_rs have good consistency in the green and red bands, especially in the red band, which maintains high consistency (R² = 0.9761), low RMSE (0.0020 sr⁻¹) and low MAPE (18.10%). Compared with the in situ data, the RMSE were 0.0036 sr⁻¹ and 0.0035 sr⁻¹, and the MAPE were 0.1844% and 0.6239%, respectively, for CZI-derived R_rs in the green and red bands.

The values and spatial distributions of CZI-estimated SPM and OLI-estimated SPM in YRE and SBS are basically consistent, and the validation using in situ data revealed that the inversion of SPM concentration by CZI was effective (R²=0.86, RMSE=0.0362 g/L). CZI has higher revisit period and wider swath width than OLI, which can better reflect the rapid spatial change characteristics of ocean color in the estuarine and coastal waters. In the future, CZI will have broad application prospects in monitoring ocean color constituents in estuarine and coastal waters in China and even the world, such as obtaining the SPM distribution of estuarine and coastal waters, and estimating the suspended sediment flux and carbon flux, which will provide data support and scientific basis for formulating and implementing the development, utilization and protection strategy of estuarine and coastal zone.