To assess the accuracy of GOCI-II remote sensing products in eutrophic waters, in situ measurements were conducted in Lake Taihu. Lake Taihu was segmented into 10 subareas, with the lake center area exhibiting a better water quality and less disruption from human activities, whereas the southwestern area is more significantly influenced by human activities, causing rapid changes in Chl-a (Xu et al., 2019; Zhang et al., 2019; Duan et al., 2009). Consequently, time-series measurements within a day were conducted in both the lake center and southwestern areas, along with cruise-based sampling throughout the entire Lake Taihu region. The distribution of the voyage stations is shown in Figure 1A , where the red dots denote the two time-series stations operating on August 29 and September 1, 2022 (in a total of 20 measurements). The black dots denote the cruise-based sampling stations operating on September 6, 2022 (in a total of 16 measurements). Specific measurement times were aligned with GOCI-II observation times, which included 10 satellite measurements per day from 7:15 AM to 4:15 PM local time, encompassing parameters such as Rrs, TSM concentration, Chl-a concentration, and a_ph.

Rrs was measured via a portable spectrometer manufactured by the American ASD company (spectral range: 350–2,500 nm; spectral resolution: 3 nm). The implementation of the spectrometer measurement method followed the specifications outlined in the Ocean Optics Protocol established by the National Aeronautics and Space Administration (NASA) (Mueller and Fargion, 2023). Each day, the upward radiance at the water surface and the downward sky radiance, along with the upward radiance from a reference gray panel, were simultaneously measured. During these measurements, the viewing azimuth angle of the ASD probe was adjusted to form a 135-degree angle with the direction of direct sunlight, while the zenith angle was set at 45 degrees to minimize the influence of sun glint. The Rrs can be calculated as follows:

where L t denotes the upward radiance above the water surface, L s denotes the downwelling sky radiance, r denotes the diffuse reflectance factor, L r denotes the upward radiance from the reference gray panel, and ρ denotes the water surface reflectance, which was set to 0.028 as the wind speed during the spectral measurements was below 4 m/s (He et al., 2013). Figure 1B displays the spectral curves of the Rrs values obtained during the Lake Taihu cruise, while Table 1 lists the measurement times.

During the cruise, the Chl-a concentration in water was measured via a Turner fluorometer manufactured by Turner Biosystems. Chl-a samples were obtained by filter membrane samples, thereby employing Whatman GF/F glass fiber filter membranes, with a diameter of 25 mm and a pore size of 0.7 µm. Each sample was measured three times, and the average value was determined (Holm-Hansen et al., 1965). Moreover, a_ph was measured via glass fiber filter membranes following the same filtration method as that used for Chl-a determination. The filter membranes were subsequently collected and stored in a freezer at -17°C, and measurements were conducted in the laboratory via an Ultraviolet-visible spectrophotometer (Lambda750S). The TSM concentration was measured via filter membrane samples. Before sampling, the sample acetate fiber membranes (diameter: 47 mm; pore size: 0.7 μm) were weighed in the laboratory. Then, they were transported to the experimental site for onsite filtration of water samples at each station. The filter membranes were then collected and stored in a freezer at -17°C, and measurements were conducted in the laboratory. Via repeated drying in a muffle furnace until a stable weight was reached, the weight of the total suspended matter membrane was subtracted from the weight of the corresponding blank membrane to obtain the weight and concentration of suspended particulate matter. In total, 128 sets of data were obtained, including TSM, Chl-a, and a_ph data and Rrs data.

Hourly GOCI-II Level-1B data for August 29, September 1, and September 6, 2022, were obtained from the Korea National Ocean Satellite Center (https://www.nosc.go.kr/eng/main.do). The data comprised 10 observations per day. Two types of atmospheric correction algorithms were subsequently applied to derive the Rrs in various GOCI-II bands, including an ultraviolet (UV)-based atmospheric correction algorithm and the official NIR-based atmospheric correction algorithm used by the GOCI-II.

The Moderate Resolution Imaging Spectroradiometer (MODIS)-Aqua data used in this study were acquired from NASA’s Ocean Color Web. The MODIS-Aqua Level-1 data were downloaded from https://oceancolor.gsfc.nasa.gov/data/aqua/. The MODIS-Aqua satellite passes at approximately 1:30 PM, and it provides data with a spatial resolution of 1 km in the visible light spectrum. The visible band Rrs products used in this study were derived from Level-1 data via a NIR atmospheric correction model, with reference bands of 748 and 869 nm for atmospheric correction.

The inversion algorithm used for Chl-a is the OCx algorithm (O’reilly et al., 1998), which is derived from the ratio of the fourth-order polynomial to the maximum Rrs value of the green band at 555 nm and the blue band. The expression is as follows:

where Chl oc 4 is the Chl-a concentration, and a₀ to a₄ are fitting coefficients specific to GOCI-II, with values of 0.3272, -2.9940, 2.7218, -1.2259, and -0.5683.

TSM can be estimated via the YOC (Yellow Sea Large Marine Ecosystem Ocean Color Project.) algorithm of Siswanto et al. (2011) as follows:

where c0 to c2 are fitting coefficients specific to GOCI-II, with values of 0.649, 25.623, and 0.646.

In this work, the absorption coefficient of phytoplankton pigments was calculated via a semianalytical method based on a bio-optical model, namely, the quasi-analytical algorithm (QAA), which was proposed by Lee et al. (2002). The QAA can be expressed as:

where a ( λ ) is the total absorption coefficient, a d g ( λ ) is the absorption coefficient of gelbstoff and detritus, and a w ( λ ) is the absorption coefficient of pure seawater. Where a d g can be calculated as follows:

Notably, a g ( 443 ) , ζ, ξ, and the spectral slope S can be obtained as follows:

And a ( λ )   can be calculated as follows:

The in situ and satellite data were matched spatiotemporally by averaging the valid pixels within a 3×3 pixels box centered on the in situ point, with a time window of available data of ±1 hour (Bailey and Werdell, 2006). Additionally, the following conditions must be met:

By employing these selection criteria, we aimed to ensure that the matched satellite data were both spatially representative and temporally close to the in situ measurements, minimizing the impact of spatial heterogeneity and temporal mismatch on the accuracy assessment.

The accuracy of remote sensing products was assessed via various statistical parameters, namely, the root mean square deviation (RMSD), absolute percentage deviation (APD), and relative percentage deviation (RPD), which can be calculated as:

where Y_i, X_i, and N are the algorithm inversion value, the measured value, and the sample size, respectively. RPD is a predictive indicator of the systematic error or average relative deviation, and APD reflects the absolute accuracy of the inversion product relative to the measured data.

Figure 2 shows the distribution of Rrs(443 nm) in Lake Taihu on September 6, 2022, obtained via the NIR atmospheric correction algorithm with the GOCI-II data. Owing to the clear weather conditions on that day, a large amount of valid data was obtained. As shown in the figure, the value of Rrs(443 nm) varied significantly during the observation period, showing characteristics of higher Rrs(443 nm) values around noon and abnormally lower values close to dawn and dusk.

Table 2 provides the statistical results for the comparison of Rrs values across various bands between the NIR algorithm-retrieved and in situ data. The APD across the 6 bands ranged from 20.96% to 31.52%, with the smallest APD value occurring at 555 nm and the largest APD value occurring at 680 nm. This finding is similar to the results obtained by Kim et al. (2016), who validated inversion values against measured data for low-turbidity waters via voyage data from 2010 to 2011, with APD values ranging from 18% to 35%.

Figure 3 shows a comparison between the Rrs values for various bands derived from the GOCI-II data via the NIR algorithm and the in situ data. The scatter plot is relatively dispersed and the inverted values are slightly lower than the measured values, especially at higher Rrs values, suggesting underestimation by the NIR algorithm.

In the Rrs results inverted via the NIR-correct atmospheric correction algorithm, the APD values for the 555 nm band Rrs data were relatively small, indicating high accuracy. However, for the Rrs results of the 412 nm and 443 nm bands, the APD values were relatively large. This difference may be attributed to the use of the NIR-correct atmospheric correction method for extrapolating aerosol scattering from the NIR spectrum to the visible light spectrum, yielding larger inversion errors in the blue light spectrum, which is farther from the NIR spectrum (Abbott et al., 1999). In addition, Lake Taihu is a moderately turbid water body, thus, the water-leaving radiance in the NIR spectrum cannot be neglected, and the assumption of dark pixels in the NIR used by the NIR atmospheric correction algorithm is not appropriate for the turbid waters. Considering the high absorption capacity for UV wavelengths of eutrophic water bodies, it can assume that the water-leaving radiance signal at UV wavelengths could be neglected. Leveraging the advantage of the GOCI-II sensor with a 380 nm UV band, UV atmospheric correction algorithms were employed to improve the inversion performance.

Figures 4 , 5 show the distributions of the retrieved Rrs (443 nm) data for Lake Taihu on September 6, 2022, obtained via the use of the UV_380nm-correct (taking 380 nm band as reference band) and UV_412nm-correct (taking 412nm band as reference band) atmospheric correction algorithms, respectively. Compared with the Rrs (443 nm) results inverted via the NIR-correct atmospheric correction algorithm (as shown in Figure 2 ), the Rrs (443 nm) values inverted via the UV atmospheric correction algorithms were generally greater. The temporal variation in Rrs across the ten images captured throughout the day was minimal, but there was clear algorithm failure in the last two images of the day.

In this study, Rrs data were obtained from 33 in situ measurement stations matched with Rrs data inverted via UV atmospheric correction algorithms. Figure 6 shows scatter plots that provide a comparison of the Rrs values across various wavelengths obtained via the UV_380nm-correct and UV_412nm-correct atmospheric correction algorithms with in situ data. Figure 7 shows the comparison of remote sensing reflectance products derived from various inversion algorithms with measured values at certain stations. It is evident that the results from the UV_412nm-correct atmospheric correction algorithm align closely with the trend of the measured spectral curves. Table 3 provides statistical results for the comparison of the inverted Rrs values across the different wavelengths with the Rrs values inverted via the two UV atmospheric correction algorithms. Compared with those of the NIR-correct atmospheric correction algorithm, the Rrs results inverted via the two UV atmospheric correction algorithms for shorter wavelengths of 412 nm, 443 nm, and 490 nm were closer to the 1:1 line, indicating better performance. As indicated in Table 3 , the APD values for the Rrs results for the NIR wavelengths inverted via both UV atmospheric correction algorithms were higher, exceeding 30%. The APDs for the Rrs results at the 660 nm and 680 nm wavelengths inverted via the UV_380nm-correct atmospheric correction algorithm were 42.47% and 51.56%, respectively. For the UV_412nm-correct atmospheric correction algorithm, the APDs for the 660 nm and 680 nm wavelengths were 36.83% and 33.00%, respectively, indicating greater inversion performance in the NIR wavelengths. The statistical parameters revealed that the UV_412nm-correct atmospheric correction algorithm performed better in correcting data for Lake Taihu waters, with the accuracy in the shorter wavelength bands surpassing that of the NIR-correct atmospheric correction algorithm. Both the UV_380nm-correct and UV_412nm-correct atmospheric correction algorithms could reproduce the spectral distribution of the measured Rrs data. Compared with the UV_380nm-correct atmospheric correction algorithm, the UV_412nm-correct algorithm slightly overestimated the aerosol scattering reflectance at NIR wavelengths because of the significant contribution of the water-radiance of water at 412 nm, but the overestimation of the aerosol scattering reflectance at 865 nm could compensate for the minor overestimation by the UV atmospheric correction algorithms due to the white aerosol assumption at short visible wavelengths. The lower accuracy of the UV_380nm-correct algorithm may result from lower signal-to-noise ratios or calibration inaccuracies in the data obtained at the 380 nm wavelength by the GOCI-II. Lake Taihu is characterized by high aerosol optical depth (AOD) and varying water vapor content, which can influence the accuracy of the NIR algorithm. While the NIR-based atmospheric correction method is widely used, its performance is limited in turbid waters like Lake Taihu, where the high concentration of suspended matter can interfere with the near-infrared signal. In contrast, the UV algorithm may perform better under such conditions, as it is less affected by the water’s optical complexity.

Table 4 provides the statistical results for the comparison of the Rrs values inverted via the three algorithms with the measured Rrs values. Overall, the results inverted via the UV_412nm-correct atmospheric correction algorithm were the best. This occurred because UV atmospheric correction algorithms are designed primarily for highly turbid or eutrophic waters. Therefore, in Lake Taihu, UV atmospheric correction algorithms have better performance than the NIR atmospheric correction algorithm.

Figure 8 shows a comparison between the Chl-a, TSM, and a_ph values inverted via the NIR-correct atmospheric correction algorithm for the GOCI-II data and in situ values, with the corresponding statistical parameters also shown in Figure 8 . There were 30 corresponding measured station data points matching the inverted Chl-a data, as shown in Figure 8A . At lower measured Chl-a values, the inverted GOCI-II results were greater than the measured values, whereas at higher measured Chl-a values, the inverted GOCI-II results were lower than the measured values. The overall RMSD was 16.301 mg/m³, with the APD of 45.76% and the RPD of 16.79%, indicating a significantly lower accuracy than that of the results for general open ocean waters. TSM inversion exhibited 25 corresponding measured station data matches, as shown in Figure 8B . When the TSM was greater than 55 mg/L, the inverted TSM values from the GOCI-II data were generally lower than the in situ values, and the scatter points were more dispersed. Owing to the Lake Taihu cruise being conducted only at two time-series stations (T-A and T-B) for sampling a_ph, there were fewer station data that could match the satellite observations, ultimately yielding 18 matching samples for the a_ph products. As shown in Figure 8C , similar to the scatter plot evaluation for the Chl-a product, when the measured value of a_ph was low (<0.7 m^-¹), the scatter points occurred near the 1:1 line, whereas at higher measured values, the a_ph product inverted from the GOCI-II sensor data was significantly underestimated.

The Rrs products inverted via the UV_412nm-correct atmospheric correction algorithm were used to estimate the Chl-a and TSM concentrations and a_ph. Figure 9 provide scatter plots and statistical parameters used to compare the inversion results of these three parameters with the in situ values. Compared with the inverted Chl-a values by the NIR-correct atmospheric correction algorithm, the inverted Chl-a values with the UV_412nm-correct algorithm were more accurate ( Figure 9A ), with an APD of 28.84%. However, the underestimation issue at high Chl-a concentrations was not mitigated when this algorithm was used, primarily because of spectral band saturation attributed to high Chl-a concentrations, rendering the Chl-a inversion algorithm unsuitable. The accuracy of the TSM concentration calculated from the Rrs data inverted with the UV_412nm-correct atmospheric correction algorithm also improved ( Figure 9B ), with results closer to the 1:1 line and a smaller APD value of 9.94%, which is lower than the value of 13.26% achieved with the NIR-correct atmospheric correction algorithm. Similar to Chl-a, the underestimation issue of a_ph at high values was not addressed. For eutrophic waters, there is a need to develop more suitable models for Chl-a and a_ph inversion algorithms.

The GOCI-II obtains one observation per hour from 7:15 to 16:15 local time, providing a total of 10 observation images per day. Owing to the relatively high solar zenith angles during the first and second observations, as well as the ninth and tenth observations of the GOCI-II satellite, which can exceed 70° during the autumn and winter seasons, the stability of remote sensing products at different observation times must be examined. The comparison of the two observation periods depicted in Figure 10 reveals that the Rrs values inverted by the GOCI-II are closer to the measured Rrs values from 9:15 to 14:15, when the solar zenith angle is relatively low. The results for this period are clustered around the 1:1 line, whereas the results for other periods are more dispersed. The agreement between the Rrs values observed from 9:15 to 14:15 and inverted via the UV_412nm-correct atmospheric correction algorithm was the best ( Figure 10C ). However, there was underestimation in the high-value areas for the inversion results of the 660 nm and 680 nm bands, whereas the results for the other bands were better. The Rrs results for the various bands from 9:15 to 14:15, as inverted via the UV_380nm-correct atmospheric correction algorithm, were slightly worse than those inverted via the UV_412nm-correct atmospheric correction algorithm. Among these bands, the results for the 660 nm and 680 nm bands were more scattered than those for the other bands were, but the results for the other bands were more accurate ( Figure 10B ). Similarly, the Rrs results for the various bands inverted via the NIR-correct algorithm between 9:15 and 14:15 showed that the results for the 660 nm and 680 nm bands were more scattered than those for the other bands were, and underestimation occurred when the Rrs value was high ( Figure 10 ). Outside the period from 9:15 to 14:15, the inversion accuracies of the three atmospheric correction algorithms were poor, with very scattered data points.

Tables 5 , 6 provide the error statistics of the GOCI-II derived Rrs results for each band obtained via the different atmospheric correction algorithms across the two observation periods. The resutls indicate that for the NIR-correct atmospheric correction algorithm, the APD ranges from 20.80% to 31.37% at midday and from 20.68% to 41.63% at other times. The UV_380nm-correct atmospheric correction algorithm yielded APD ranging from 11.12% to 25.08% at midday and from 32.66% to 56.73% at other times. Similarly, the UV_412nm-correct atmospheric correction algorithm yielded APD ranging from 14.95% to 19.56% at midday and from 26.19% to 45.27% at other times. Overall, the APD range of the three algorithms at midday was much smaller than that at other times, indicating significantly better retrieval results at midday than during the dawn and dusk periods. This finding agrees with the validation results of ocean color remote sensing products from geostationary satellites reported by Lamquin et al. (2012) and Li et al. (2020).

In addition to evaluating the accuracy of GOCI-II products using in situ data from Lake Taihu, we also conducted a cross-comparison with MODIS-Aqua products. The primary reason for this is that MODIS-Aqua products are widely recognized as a standard for ocean color remote sensing accuracy. This comparison highlights potential discrepancies between satellite products in eutrophic waters. The time window for matching MODIS and GOCI-II data is consistent with the matchup protocol, and spatially, we projected the data from both satellites onto a 0.04-degree resolution grid for comparison. Figure 11 shows the comparison of the Rrs products from the GOCI-II and MODIS satellite data in Lake Taihu on September 6, 2022. The figure reveals a notable difference in the Rrs products between the two satellites, despite a generally similar overall distribution. This observed difference could be attributed to variations in the spatial resolution, spectral bands, and atmospheric correction methodologies employed by the respective satellites. Previous research (Xiao et al., 2016) has emphasized a high level of agreement between the MODIS and GOCI products in oceanic environments, as well as between the GOCI and GOCI-II products. However, as shown in Figure 11 , there was a significant lack of consistency between the satellite products in eutrophic waters, indicating a certain bias in the NIR atmospheric correction models when applied to such aquatic environments.