Preliminary Assessment of GOCI-II Band-Ratio Algorithms for Retrieving Particulate Organic Carbon in the Optically Complex Waters of Jeju Island

Abstract

Particulate organic carbon (POC) monitoring in dynamic coastal environments is essential for estimating Blue Carbon fluxes and understanding local biogeochemical cycles. Conventional satellite algorithms developed for open-ocean (Case-1) waters frequently underperform in optically complex coastal environments because of competing optical signals from Colored Dissolved Organic Matter (CDOM) and suspended particles. To address this gap, we present a preliminary effort to evaluate and locally optimize POC retrieval algorithms for the Geostationary Ocean Color Imager-II (GOCI-II) in the coastal waters of Jeju Island, Korea. Previously published GOCI-II POC products in Korean waters have applied global open-ocean models without regional tuning. Using a structured diagnostic methodology, we first evaluated three standard blue-to-green band ratio algorithms (412/555, 443/555, and 490/555 nm) using in-situ radiometric and biogeochemical data from 23 stations in the Korea Strait. All three algorithms performed poorly (R² ≤ 0.093), indicating that high CDOM concentrations decouple blue reflectance from particulate scattering. A cross-sensor comparison with Visible Infrared Imaging Radiometer Suite (VIIRS) (R² = 0.617, N = 2,675 coincident pixels) provided evidence of GOCI-II's radiometric integrity, suggesting that the failures were algorithmic rather than instrumental. We then evaluated all 15 GOCI-II band-ratio combinations spanning 412-745 nm under the same Stramski et al. (2008) power-law framework. The analysis revealed a clear spectral hierarchy: Red-near-infrared (NIR) ratios consistently outperformed all shorter-wavelength alternatives, with the 680/745 nm combination yielding the best performance (R² = 0.2422, root mean square error (RMSE) = 31.69 mg m^-3, mean bias error (MBE) = -2.82 mg m^-3, p = 0.0171). Although the explained variance of the best-performing model remains modest, the substantial and consistent improvement over blue-green algorithms across multiple Red-NIR pairs supports the conclusion that the Red-NIR domain is more appropriate for POC retrieval in this CDOM-rich environment. The proposed 680/745 nm relationship should be regarded as a preliminary regional candidate that requires further validation against an expanded multi-seasonal in-situ database before operational deployment.

1. Introduction

Particulate organic carbon (POC) is an essential component of the global carbon cycle, connecting the atmosphere and the deep ocean. Through the biological pump, phytoplankton fix atmospheric carbon dioxide (CO₂) into organic matter, which is then transferred to the deep ocean as these organisms die, produce fecal pellets, or aggregate into marine snow (Volk & Hoffert, 1985; Siegel et al., 2014; Legendre et al., 2015). This export flux governs atmospheric CO₂ levels and, in turn, the Earth’s climate system (Sarmiento & Siegenthaler, 1992). Although the open ocean covers the vast majority of the world’s ocean surface, coastal edges and continental shelves play a disproportionately large role in carbon sequestration. Muller-Karger et al. (2005) estimated that continental margins, which cover only 7–10% of the ocean area, could account for up to 40% of total organic carbon sequestration, emphasizing their critical significance in the global carbon budget. Shelf-sea POC is increasingly recognized as a mobile Blue Carbon pool within the Blue Carbon paradigm, which has typically focused on vegetated coastal ecosystems such as mangroves and seagrasses (Krause-Jensen & Duarte, 2016).

The dynamics of this carbon pool in coastal seas are complex because of the interaction of land inputs, tidal resuspension, and high biological productivity driven by riverine nutrient delivery (Stramski et al., 2008; Bauer et al., 2013). The Korea Strait and the waters surrounding Jeju Island exemplify this complexity, serving as a dynamic mixing zone between the sediment-laden, Colored Dissolved Organic Matter (CDOM)-rich waters of the East China Sea and the more oligotrophic Kuroshio Branch Current. Unlike the open ocean, where phytoplankton and their derivatives dominate the optical signal, these coastal waters are influenced by a diverse mixture of autotrophic biomass, detrital matter, and suspended inorganic minerals (Chen et al., 2009). Accurately estimating the mobile POC component is critical to understanding the link between coastal ecosystems and deepocean carbon sinks (Najjar et al., 2018).

Traditional POC monitoring relies on ship-based in-situ sampling, which, although it delivers excellent analytical accuracy, is inherently limited in spatial and temporal coverage (Gardner et al., 2006). The dynamic nature of coastal Blue Carbon ecosystems, characterized by tides, currents, and episodic riverine discharge, requires observation frequencies that ship surveys cannot sustain economically. Satellite ocean color remote sensing provides a synoptic view of the ocean surface, enabling the retrieval of bio-optical properties on large spatial scales and offering the only viable method for assessing regional and global POC stocks (McClain, 2009; Frouin et al., 2019). Algorithms for estimating POC from space have typically relied on apparent optical properties (AOPs) such as the remote sensing reflectance (R_rs) ratios in the blue and green spectral bands (Stramski et al., 1999; Loisel et al., 2018). The theoretical basis for these methods is that POC concentration co-varies with suspended particle concentration, which scatters light. In Case-1 waters, where phytoplankton and their degradation products dominate the optical properties, chlorophyll-a absorption in the blue band and particle backscattering allow for strong empirical relationships (Morel, 1988; O’Reilly et al., 1998; Gordon & Morel, 1983). Stramski and Kiefer (1991) demonstrated that the backscattering efficiency of marine particles depends on particle size distribution and refractive index, establishing a physical link between remote sensing signals andorganic carbonpools. However, this relationship is strongly predicated on the assumption that blue light absorption is primarily driven by phytoplankton, an assumption that routinely fails in coastal zones.

Standard algorithms designed for open-ocean (Case-1) waters frequently underperform in optically complex coastal (Case-2) waters. CDOM and non-algal particles (NAP) typically dominate the optical signal in the blue domain in areas such as the coastal waters of Jeju Island and the wider Korea Strait (Son et al., 2009; Kim et al., 2016). The high absorption of CDOM at blue wavelengths (particularly 412 and 443 nm) can mimic the optical signature of high chlorophyll or POC concentrations, leading to substantial overestimation of carbon stocks (Mannino et al., 2008; Siswanto et al., 2011). Despite the clear need for region-specific POC algorithms in this environment, only limited development efforts have been directed toward POC retrieval for the GOCI-II sensor in Korean coastal waters. As a result, the research challenge calls for a systematic investigation to identify the root cause of retrieval errors and the most appropriate algorithm for this environment. Our investigation followed a logical chain: first, we determined whether standard blue-green algorithms (412/555, 443/555, and 490/555 nm) are viable in these waters; when they failed as expected because of CDOM interference, we then assessed whether the failure was attributable to the algorithm itself or to the radiometric performance of the GOCI-II sensor. To address this, a cross-comparison with the well-established polar-orbiting Visible Infrared Imaging Radiometer Suite (VIIRS) was performed to assess the observational capacity of GOCI-II. Finally, after confirming the radiometric integrity of the sensor, the investigation shifted to longer wavelengths in the Red-near-infrared (NIR) domain, where CDOM absorption is negligible, in order to develop a regionally tuned algorithm.

This study addresses the gap in satellite-based POC monitoring for the Korean Peninsula by following the investigative approach outlined above. The specific objectives are: (1) to evaluate the performance of standard blue-to-green band ratio algorithms for POC retrieval in the coastal waters of Jeju Island and to quantify their limitations under optically complex conditions; (2) to determine whether retrieval failures are caused by sensor anomalies or by algorithmic physics by comparing GOCI-II radiometric products against established VIIRS data, thereby assessing the suitability of GOCI-II for coastal monitoring; and (3) to evaluate and locally optimize a set of empirical POC algorithms in the Red-NIR spectral domain, applied systematically across all 15 available GOCI-II band ratio combinations spanning 412–745 nm. By achieving these objectives, this work provides a region-specific, GOCI-II-tuned POC retrieval candidate for the Korean coastal ocean. The proposed relationship offers a preliminary framework for exploring hourly POC monitoring with the geostationary observation capability of GOCI-II, contributing to ongoing efforts to resolve tidal-frequency carbon dynamics and to quantify Blue Carbon fluxes in the Korea Strait (Intergovernmental Panel on Climate Change, 2022).

2. Materials and Methods

2.1. Study Area and Field Sampling

The coastal waters surrounding Jeju Island, South Korea (approximately 33.2°N, 126.5°E to 33.6°N, 126.8°E)were selected as the study area (Fig. 1). The Yellow Sea Cold Water Mass, the Tsushima Warm Current, and local coastal processes combine to create complex hydrodynamic conditions in the study area, which encompasses a variety of marine ecosystems within the Korea Strait. With water depths ranging from shallow nearshore areas to deeper offshore seas, the sampling domain was chosen to represent the typical coastal waters of the East China Sea margin. Oceanographic conditions in the study area change seasonally, with moderate sea surface temperatures and fluctuating levels of primary productivity characteristic of temperate coastal ecosystems during the sampling period.

Fig. 1. Map of the study area surrounding Jeju Island, South Korea. The locations of the 23 sampling stations used for in-situ radiometric and biogeochemical data collection are shown, distinguishing between the June 2024 (J: Jeju June 2024 campaign series, pink squares) and May 2025 (JS: Jeju Spring 2025 campaign series, green circles) field campaigns. The bottom-right inset shows the full Jeju Island with the sampled coastal zone outlined in red, and the top-right inset provides a magnified view of the sampling cluster along the northeastern coast.

A total of 23 sampling stations were established across the study region, of which 18 stations (JS-01 to JS-18) were sampled during May 26–29, 2025, and five additional reference stations (J-01 to J-05) were sampled in June 2024. A small research vessel was used to collect water samples and optical measurements, enabling flexible deployment in coastal and nearshore settings. Daily sampling was conducted between 09:00 and 17:00 KST (Korea Standard Time) to optimize coincidence with GOCI-II satellite overpasses and favorable solar illumination conditions. Surface water samples were collected at each station using a clean-bucket method to minimize contamination. The full station metadata, including dates, geographic coordinates, and measured in-situ POC concentrations, is provided in Table 1.

Table 1. Metadata of the 23 in-situ stations used in this study, including station ID, sampling date, geographic coordinates, and measured POC concentrations

Water samples for POC analysis were processed immediately upon return to the field laboratory or stabilized for transport. Filtration was performed following the IOCCG (2021) protocols, using pre-combusted 25-mm GF/F glass fiber filters. A volume of 1 L of seawater was filtered under low vacuum (< 17 kPa) to prevent particle breakage. Following filtration, filters were folded with the particulate-laden side inward, wrapped in pre-combusted aluminum foil, and immediately flash-frozen in liquid nitrogen, then stored at −80°C to preserve integrity through transport and pre-analysis holding. Prior to elemental analysis, samples were exposed to hydrochloric acid (HCl) fumes (0.12 N equivalent) for 24 hours in a glass desiccator to remove inorganic carbon and then driedat 55±5°C for 24 hours, per protocol recommendations. Filters were analyzed on a carbon-hydrogen-nitrogen (CHN) elemental analyzer calibrated with certified standards; analytical and packaging blanks were measured and subtracted, and filtrate blank corrections were applied per IOCCG guidance to account for dissolved organic carbon (DOC) adsorption artifacts. The final POC concentration was computed from the blank-corrected carbon mass normalized to the filtered volume as given in Eq. (1):

\(\begin{align}P O C=\frac{C_{\text {sample }}-C_{\text {blanks }}}{V_{\text {filtered }}}\end{align}\)       (1)

where C_sample is the total organic carbon mass on the sample filter after inorganic carbon removal, C_blanks aggregates all required blank corrections, and Vfiltered is the volume of seawater filtered.

Concurrently, hyperspectral radiometric measurements were obtained to characterize the AOPs of the water column. An Analytical Spectral Devices (ASD) FieldSpec4 spectroradiometer was used for above-water radiometric observations, covering a spectral range of 350–2,500 nm with a spectral resolution of 3 nm. The measurement protocol followed established ocean optics standards (Mobley, 1999; IOCCG, 2021). Total water-leaving radiance (L_t) measurements were made at an azimuth angle of 135° relative to the sun and a viewing angle of 40° from nadir to minimize sun glint. Sky radiance (L_sky) was measured at a zenith angle of 40°, corresponding to the specular reflection angle for the water measurement. Downwelling irradiance (E_d), the total incoming solar irradiance, was measured using a remote cosine receptor (RCR) pointed at the zenith. The raw digital values from the ASD spectroradiometer were converted to physica lradiometric units (W m^–² nm^–¹ sr^–¹) using instrument-specific calibration files. The R_rs was then calculated as given in Eq. (2):

\(\begin{align}R_{r s}(\lambda)=\frac{L_{t}(\lambda)-\rho \cdot L_{s k y}(\lambda)}{E_{d}(\lambda)}\end{align}\)       (2)

where λ is the wavelength and ρ is the sea surface reflectance factor, set to 0.025 for average viewing geometries and wind speeds below 5 m s^–¹ (Mobley, 1999). Retrieved R_rs spectra were subjected to quality control, and spectra with clear anomalies such as sun glint contamination or instrument saturation were identified and removed.

GOCI-II Level-1B data corresponding to all 23 field campaign dates (May 26–29, 2025, and June 12–14, 2024) were acquired from the Korea Ocean Satellite Center (KOSC). The data were processed to Level-2 using the GOCI-II Ground Segment (G2GS) standard atmospheric correction algorithms to derive R_rs at standard bands (380, 412, 443, 490, 510, 555, 620, 660, 680, 709, 745, and 865 nm). It should be noted that the empirical regressions presented in this study (Section 2.2) were derived using the in-situ R_rs measurements as the primary radiometric input. GOCI-II satellite R_rs values were used as an independent cross-validation reference and for spatial mapping. For each in-situ station, the closest GOCI-II pixel was extracted within a time window of ±30 minutes between satellite overpass and in-situ measurement to minimize temporal mismatch errors in the dynamic coastal environment.

2.2. Algorithm Development Framework

Local empirical POC algorithms were developed using the full set of 23 in-situ POC measurements collected from the coastal waters of Jeju Island. All models followed the standard power-law formulation originally proposed by Stramski et al. (2008), which has been widely applied in regional coastal POC studies (Son et al., 2009; Wei et al., 2019). This general form is expressed as in Eq. (3):

\(\begin{align}P O C=A \cdot\left(\frac{R_{r s}\left(\lambda_{1}\right)}{R_{r s}\left(\lambda_{2}\right)}\right)^{B}\end{align}\)       (3)

where A and B are empirically determined regression coefficients, and λ₁/λ₂ represents the numerator and denominator wavelength pair of the band ratio. In all models, POC and the reflectance ratio data were log10-transformed to linearize the power-law relationship, facilitating coefficient estimation using ordinary least squares (OLS) regression, consistent with established remote sensing literature (Morel & Maritorena, 2001; Stramski et al., 2008). The log-transformed model is calculated as in Eq. (4):

\(\begin{align}\log _{10}(P O C)=\beta_{0}+\beta_{1} \cdot \log _{10}\left(\frac{R_{r s}\left(\lambda_{1}\right)}{R_{r s}\left(\lambda_{2}\right)}\right)\end{align}\)       (4)

The power-law coefficients are recovered by back-transformation: A = 10^β₀ and B = β₁ (Hu et al., 2011; Loisel et al., 2018). A comprehensive evaluation covering all 15 available GOCI-II band ratio combinations was conducted, spanning the full sensor spectral range from 412 to 745 nm: 412/555, 443/555, 490/555, 510/660, 510/680, 555/620, 555/660, 555/680, 620/660, 660/680, 660/709, 680/709, 660/745, 680/745, and 709/745 nm. This systematic spectral sweep was designed to identify the most informative retrieval window without imposing a priori spectral assumptions. Algorithm performance was assessed using several complementary statistical metrics. The coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE) are defined as follows in Eqs. (5–7):

\(\begin{align}\mathrm{R}^{2}=1-\frac{\sum\left(y_{i}-\hat{y}_{i}\right)^{2}}{\sum\left(y_{i}-\bar{y}\right)^{2}}\end{align}\)       (5)

\(\begin{align}R M S E=\sqrt{\frac{1}{N} \sum\left(y_{i}-\hat{y}_{i}\right)^{2}}\end{align}\)       (6)

\(\begin{align}M A E=\frac{1}{N} \Sigma\left|y_{i}-\hat{y}_{i}\right|\end{align}\)       (7)

where y_i is the in-situ observed POC value, \(\begin{align}\hat y_i\end{align}\) is the model-predicted POC value, \(\begin{align}\bar y\end{align}\) is the mean of observations, and N is the number of samples. The mean bias error (MBE) was also computed to characterize the direction and magnitude of any systematic over- or under-estimation:

\(\begin{align}M B E=\frac{1}{N} \Sigma (\hat {y}_{i}-y_i)\end{align}\)       (8)

Together, these metrics provide an evaluation of model precision, bias, and accuracy. In addition, log-log calibration plots and frequency distribution histograms were used to assess the structural fit of each candidate algorithm against the observed POC distribution.

3. Results

3.1. Standard Blue-Green POC Algorithms

The 23 in-situ POC measurements ranged from 100.7 to 280.3 mg m^–³ with a mean of 222.9±37.9 mg m^–³ (Table 1). These values are substantially elevated above typical open-ocean concentrations (5–50 mg m^–³), reflecting the combined influence of strong semi-diurnal tides and associated sediment resuspension, substantial terrestrial inputs from the Changjiang (Yangtze) and Korean peninsula river systems, and elevated biological productivity that together characterize the Korea Strait and the wider East China Sea margin (Siswanto et al., 2011; Yamaguchi et al., 2012; Kim et al., 2016; Wei et al., 2019). The confluence of the Yellow Sea Cold Water Mass and the Tsushima Warm Current further enhances these dynamics by driving frontal mixing and seasonal nutrient delivery to the surface layer (Son et al., 2009; Bauer et al., 2013). The concurrent R_rs spectra exhibited suppressed blue reflectance and a distinctRed-NIR shoulder, spectral features that are diagnostic of Case-2 optical conditions in which CDOM and suspended particles co-dominate the signal alongside phytoplankton. These optical characteristics foreshadow the poor performance of blue-band algorithms and the need to explore region-specific algorithms for this environment.

Three standard blue-green band ratio algorithms (412/555, 443/555, and 490/555 nm) were evaluated against the full in-situ dataset (N = 23) using the power-law regression framework. Table 2 summarizes the regression coefficients and performance statistics for all three models.

Table 2. Statistical performance of standard blue-green POC retrieval algorithms

Based on the coefficients in Table 2, the specific regression equations for the three standard blue-green algorithms are Eqs. (9–11):

\(\begin{align}P O C=205.89 \cdot\left(\frac{R_{r s}(412)}{R_{r s}(555)}\right)^{0.279}\end{align}\)       (9)

\(\begin{align}P O C=186.76 \cdot\left(\frac{R_{r s}(443)}{R_{r s}(555)}\right)^{1.024}\end{align}\)       (10)

\(\begin{align}P O C=186.20 \cdot\left(\frac{R_{r s}(490)}{R_{r s}(555)}\right)^{0.687}\end{align}\)       (11)

Fig. 2 shows the log-log relationships between in-situ POC and the three blue-green band ratios. All three algorithms exhibited poor performance. The 412/555 nm ratio showed the weakest correlation (R² = 0.008), as 412 nm is the wavelength most strongly affected by CDOM absorption, which effectively decouples the R_rs signal from particle backscattering. The 490/555 nm ratio also performed poorly (R² = 0.041), failing to reveal any meaningful trend. Although the 443/555 nm model showed marginally higher explanatory power among the blue-band options, the coefficient of determination remained negligibly low (R² = 0.093) and statistically insignificant (p = 0.158), indicating that blue-band ratios are ineffective proxies for POC in these waters.

Fig. 2. Log-log regression scatter plots of in-situ POC versus blue-to-green band ratios: (a) 412/555 nm (R² = 0.008), (b) 443/555 nm (R² = 0.093), and (c) 490/555 nm (R² = 0.041). The OLS fit (green line) with the 95% confidence interval (shading) and annotated statistics are shown for each panel. The highly scattered distributions indicate the absence of meaningful correlation in the blue spectral domain (N = 23).

Despite its marginal superiority among the blue-green models, the 443/555 nm algorithm was selected for full evaluation in order to quantify the extent of the failure. The scatter plot comparison of in-situ versus predicted POC (Fig. 3b) demonstrates broad scatter and a substantial departure from the 1:1 line, with quantitative metrics of R² = 0.132, RMSE = 34.74 mg m^–³, MBE = −3.48 mg m^–³, and MAE = 24.09 mg m^–³, indicating large prediction errors and systematic underestimation.

Fig. 3. Validation scatter plots comparing in-situ POC measurements with algorithm-predicted POC for the three blue-green band ratio models: (a) 412/555 nm, (b) 443/555 nm, and (c) 490/555 nm. The dashed orange line represents the 1:1 reference line; the solid green line is the best-fit regression line. The broad scatter and departure from the 1:1 line confirm the inapplicability of these models in the Korea Strait (N = 23).

The frequency distribution comparing in-situ and 443/555 nm-predicted POC reveals a characteristic distributional compression: model predictions are constrained to a narrow range clustered near the dataset mean, failing to reproduce the full dynamic range of the observations. The algorithm fails to capture the higher concentration values (> 250 mg m^–³), compressing predictions into a range that reflects the mean CDOM optical state rather than the true variability of particulate carbon. This is a diagnostic signature of optical decoupling, not a simple calibration offset, and it cannot be resolved by coefficient re-tuning alone.

3.2. Cross-Sensor Comparison: GOCI-II vs. VIIRS

To determine whether the poor performance of the blue-green algorithms was attributable to a sensor-level anomaly in GOCI-II rather than to an algorithmic limitation, we compared POC products derived from GOCI-II and VIIRS using the same Stramski et al. (2008) global algorithm applied to R_rs from both sensors. Spatially, both the GOCI-II and the VIIRS POC maps showed consistent large-scale patterns over the Jeju Island region (Fig. 4), with elevated concentrations in nearshore waters and lower values offshore, demonstrating that both sensors resolve the same physical gradients in surface POC. The comparison is based on the monthly mean composites for May 2025, which corresponds to the principal field campaign period of this study; we acknowledge that this represents a single seasonal snapshot rather than a multi-temporal evaluation.

Fig. 4. Spatial distribution of monthly mean POC (mg m^–³) for May 2025 derived from (a) GOCI-II and (b) VIIRS using the Stramski et al. (2008) global algorithm. Both sensors capture consistent large-scale spatial gradients with elevated concentrations in nearshore waters adjacent to Jeju Island.

Quantitative pixel-level comparison (N = 2,675 coincident pixels) yielded R² = 0.617, slope = 0.701, RMSE = 25.26 mg m^–³, MBE = −10.07 mg m^–³, and MAE = 19.27 mg m^–³ (Fig. 5). This level of spatial coherence between two independently calibrated and atmospherically corrected satellite datasets is consistent with the radiometric integrity of GOCI-II. We note, however, that this comparison is based on a single-month composite rather than a multi-season analysis, and the agreement should therefore be interpreted as evidence of consistency under the conditions of the study period rather than as a comprehensive cross-platform validation. Within these constraints, the result supports the interpretation that the algorithm failures documented in Section 3.1 are attributable to algorithmic design, specifically the use of CDOM-sensitive blue bands, ratherthan to any deficiency in the observational capacity of GOCI-II.

Fig. 5. Pixel-by-pixel comparison of GOCI-II and VIIRS POC retrievals for the May 2025 composite (N = 2,675 coincidence pixels). The hexagonally binned scatter plot iscolor-coded by pixelcount (log scale). The solid blue line is the best-fit linear regression (y = 0.70x + 35.91), and the dashed red line is the 1:1 reference. The level of coherence (R² = 0.617) supports the radiometric integrity of GOCI-II under the conditions of the study period, with the systematic positive offset attributable to inter-platform atmospheric correction differences.

A systematic positive offset of approximately 10–20 mg m^–³ was observed in GOCI-II retrievals relative to VIIRS throughout the full concentration range, as also visible in the pixel-by-pixel scatter (Fig. 5). This bias is most likely caused by differences in observation geometry between geostationary and polar-orbiting platforms, by discrepancies in atmospheric correction algorithms, and by differences in sensor spectral band response functions (Franz et al., 2007). This offset is consistent with previous inter-platform comparisons in the East Asian ocean color literature (Ryu et al., 2012) and represents a tractable calibration bias rather than a fundamental radiometric flaw.

3.3. Spectral Evaluation of 15 GOCI-II Band Ratios

After establishing the radiometric consistency of GOCI-II with VIIRS, we evaluated POC retrieval across all 15 available GOCI-II band-ratio combinations, spanning the full sensor spectral range from 412 to 745 nm. The same OLS power-law framework (Eq. 3) was applied consistently to all combinations. Table 3 presents the regression coefficients, R² values, p-values, and sample sizes for all 15 band ratios, ranked by R².

Table 3. Regression coefficients (A, B), coefficient of determination (R²), p-value, and sample size (N) for all 15 GOCI-II band ratio POC retrieval algorithms evaluated in this study, ranked by R²

Fig. 6 presents the log-log calibration scatter plots for all 15 band ratio combinations, illustrating a clear spectral performance hierarchy. Band ratios incorporating short wavelengths (412–490 nm) exhibit highly scattered distributions with negligible R² values. In contrast, the Red-NIR band ratios show markedly superio rlinearity in log-log space. The three strongest-performing combinations, namely 680/745 nm (R² = 0.2422), 680/709 nm (R² = 0.2370), and 709/745 nm (R² = 0.2182), all achieve statistical significance at p < 0.05,representing an approximately three-fold improvement in explanatory power relative to the best blue-green algorithm. A fourth Red-NIR ratio, 660/745 nm (R² = 0.1942, p = 0.0353), also reaches statistical significance, supporting the consistency of the NIR 745 nm band as a reference wavelength for POC retrieval in this environment. We emphasize that the absolute R² values remain modest (0.19–0.24), and the principal finding is therefore the relative improvement of Red-NIR over blue-green ratios rather than a high-accuracy retrieval solution.

Fig. 6. Log-log scatter plots of in-situ POC versus all 15 GOCI-II band ratio combinations (N = 23). The OLS fit (green line) and 95% confidence interval (shading) are shown for each panel with annotated statistics (A, B, R², p-value). Panels are arranged from shorter to longer wavelength ratios, illustrating the progressive improvement in model skill toward the Red-NIR domain.Asterisks (*) denote statistical significance at p < 0.05.

The three statistically significant Red-NIR algorithms (p < 0.05) with the highest explanatory power are Eqs. (12–14):

\(\begin{align}P O C=271.88 \cdot\left(\frac{R_{r s}(680)}{R_{r s}(745)}\right)^{-0.636}\end{align}\)       (12)

\(\begin{align}P O C=291.53 \cdot\left(\frac{R_{r s}(680)}{R_{r s}(709)}\right)^{-1.468}\end{align}\)       (13)

\(\begin{align}P O C=253.03 \cdot\left(\frac{R_{r s}(709)}{R_{r s}(745)}\right)^{-0.995}\end{align}\)       (14)

The consistently negative exponents (B = −0.636 to −1.468) in all three equations are physically self-consistent, as POC concentration increases, enhanced particle backscattering brightens the NIR reference band (745 or 709 nm) relative to the red band (680 nm), thereby decreasing the band ratio while POC increases. This inverse spectral relationship is a direct expression of the particle-dominated optical regime in the red-NIR domain, where CDOM absorption is negligible (Babin et al., 2003; Wang & Shi, 2007), and its consistency across three independent band ratio pairs provides cross-ratio corroboration of a coherent optical signal in these waters.

3.4. Distributional Fidelity Assessment

The frequency distribution histograms (Fig. 7) show that the Red-NIR algorithms reproduce the statistical properties of the observed POC distribution more faithfully than the shorter-wavelength alternatives. The 680/745, 680/709, and 709/745 nm algorithms reproduce the modal structure of the in-situ data (peak near 180–220 mg m^–³) and the overall spread of the distribution. Incontrast, the shorter-wavelength algorithms produce compressed or shifted distributions that fail to capture the observed POC range, particularly at the upper tail.

Fig. 7. Frequency distribution histograms comparing in-situ POC (blue) with algorithm-predicted POC (red/orange) for all 15 band ratio models (N = 23), with overlaid kernel density estimates. The Red-NIR algorithms reproduce the modal structure and overall spread of the observed POC distribution most faithfully.

Collectively, the log-log calibration (Fig. 6) and the distributional fidelity (Fig. 7) identify the 680/745 nm Red-NIR algorithm as the best-performing candidate within the 15 band ratio combinations tested for POC retrieval from GOCI-II in the coastal waters of the Korea Strait. We recognize that the absolute predictive skill (R² ≈ 0.24) remains modest and that further validation with an expanded multi-seasonal in-situ database is required before the candidate algorithm can be considered fully validated for operational use.

4. Discussion

4.1. Optical Complexity and the Failure of Blue-Green Algorithms

The poor performance of all three standard blue-green algorithms (R² ≤ 0.093) in the coastal waters of Jeju Island underscores the limitations of applying Case-1 open-ocean models to optically complex Case-2 waters. In open oceans, optical properties are dominated by phytoplankton, where absorption in the blue bands co-varies predictably with particulate scattering. However, our results showed R² values below0.1 for all blue-green combinations, consistent with previous findings in the Yellow and East China Seas (Siswanto et al., 2011; Kim et al., 2016) and with the broader literature on CDOM-dominated Case-2 environments (Bricaud et al., 1981; Mannino et al., 2008).

The primary driver of this failure is the high concentration of CDOM and NAP characteristic of the Korea Strait. CDOM absorption decreases exponentially with increasing wavelength, exerting its maximum influence at 412 nm and 443 nm (Bricaud et al., 1981; Twardowski et al., 2001). This strong absorption masks the scattering signal of POC, causing blue reflectance to remain loweven when POC concentrations are high. As a result, the blue-to-green band ratio does not decrease in response to increasing POC as expected in phytoplankton-dominated waters; instead, it varies independently based on the CDOM and NAP load (Mannino et al., 2008). This optical decoupling explains why the 412/555 nm ratio performed the worst (R² = 0.008), as 412 nm is the wavelength most heavily impacted by dissolved organic matter. The distributional compression observed in the predictions of the 443/555 nm model, where outputs cluster near the dataset mean regardless of true POC concentration, is a signature of this optical decoupling that cannot be resolved by coefficient re-tuning alone; a spectral shift to wavelengths outside the CDOM absorption window is physically necessary. This finding suggests that algorithm development in this region should move away from blue-band ratios in favor of longer wavelengths where CDOM absorption is negligible (Wei et al., 2019).

4.2. Physical Basisfor the Improved Performance of Red-NIR Algorithms

The shift to the Red-NIR spectral domain, specifically to the 680/745 nm ratio, yielded a statistically significant, approximately three-fold improvement in POC retrieval performance (from R² ≤ 0.093 to R² = 0.2422, p = 0.0171). The physical basis for this improvement lies in the inherent optical properties of water and particles at these wavelengths. At wavelengths beyond approximately 600 nm, molecular water absorption (aw) increases steeply and becomes the dominant absorbing constituent, progressively overshadowing the contributions of CDOM and phytoplankton pigments (Pope & Fry, 1997; Babin et al., 2003). In this spectral regime, R_rs is governed primarily by particle backscattering (bbp) relative to total absorption, making it a more direct and CDOM-independent proxy for suspended particle concentration and POC (Stramski & Kiefer, 1991; Stramski et al., 2008).

The 745 nm band serves as a useful reference because water absorption at this wavelength (aw ≈ 0.3 m^–¹) is sufficient to render the reflectance signal almost exclusively particle-driven, with negligible CDOM or phytoplankton pigment contributions (Wang & Shi, 2007). By normalizing the 680 nm band against this CDOM-insensitive 745 nm reference, the 680/745 ratio effectively isolates the particulate scattering signal from the dissolved phase. This approach aligns with the principles of the Generalized Ocean Color Inversion Model (GIOP; Werdell et al., 2013), which recommends exploiting longer wavelengths for more reliable scattering coefficient inversions in turbid waters. Our results indicate that, for GOCI-II in the sediment-rich coastal zones of the Korea Strait, the 745 nm band offers an advantageous reference wavelength for POC monitoring. The consistent negative exponents across all three significant Red-NIR algorithms are physically self-consistent with this mechanism, providing cross-ratio corroboration of the optical relationship.

4.3.Implicationsfor GOCI-II POC Retrieval in Korean Coastal Waters

The development of a statistically significant, physically grounded, and locally tuned POC retrieval candidate for GOCI-II in Korean coastal waters is the principal contribution of this study. Previously published ocean color products that used GOCI-II data in this region have applied the standard global Stramski et al. (2008) algorithm, which was calibrated on open-ocean datasets from the South Pacific and the Atlantic, without regional tuning. A bibliometric review of 578 GOCI and GOCI-II publications from 2005 to 2023 identified POC as among the less-studied ocean color products for the sensor series and noted the limited dedicated regional POC algorithm development effort in the Korean coastal domain (Ryu et al., 2023). The present study contributes to filling this gap.

The potential implications of a region-specific GOCI-II POC algorithm for Korean waters are noteworthy. The geostationary orbit of GOCI-II provides hourly observations, a capability that is unique among operational ocean color sensors. With further validation, this temporal resolution could enable the resolution of tidal-frequency POC variability in the Korea Strait. Tidal pumping, the mechanism by which tidal oscillations transport organic matter between coastal wetlands, shelf sediments, and the open ocean, operates on hourly time scales and has been estimated to be a significant contributor to cross-shelf organic carbon transport in macrotidal environments such as the West Sea of Korea (Yamaguchi et al., 2012; Bauer et al., 2013).

If the proposed 680/745 nm relationship is confirmed in additional seasons and locations, it could support the computation of hourly surface POC fields from GOCI-II, integration over tidal cycles, and the derivation of daily and monthly carbon flux estimates for the Korea Strait. Accurate POC retrievals from GOCI-II would also provide useful input for estimating sea-surface partial pressure of CO₂ (pCO₂) and air-sea gas exchange rates, which are tightly coupled to organic carbon cycling in the surface ocean (Takahashi et al., 2009; Wei et al., 2019).

The systematic positive bias of GOCI-II relative to VIIRS (approximately 10–20 mg m^–³), although operationally relevant, can be attributed to inter-platform differences in atmospheric correction methodology and observation geometry (Franz et al., 2007) rather than to a fundamental sensor deficiency. Future deployment of the Red-NIR algorithm should incorporate NIR-SWIR (short-wave infrared) atmospheric correction approaches (Wang & Shi, 2007; Wang et al., 2012), which perform better over aerosol-loaded coastal atmospheres in East Asia. Such corrections would also reduce systematic inter-sensor bias and improve the potential to harmonize GOCI-II and VIIRS data products into a multi-sensor coastal POC record with broader temporal coverage.

4.4. Limitations and Future Research Directions

Several limitations of this study should be acknowledged. The algorithm development and evaluation are based on 23 in-situ measurements from stations distributed around Jeju Island’s southern and southeastern coasts during spring (May 2025) and early summer (June 2024). Although this sample size is comparable to orlarger than those used in analogous preliminary algorithm development efforts in the East Asian region (Son et al., 2009; Kim et al., 2016; Wei et al., 2019), it is insufficient to fully capture the seasonal and inter-annual optical variability of the Korea Strait. The modest R² values (~0.24) obtained by the best-performing algorithms reflect both this sampling constraint and the inherent challenge of empirical POC retrieval in a highly variable coastal optical environment. Persistent cloud cover during the spring season further limited the number of successful satellite–in-situmatch-up pairs, a common challenge in temperate monsoon regions (IOCCG, 2021).

It is also important to recognize that the 680/745 nm relationship reported here has been developed, evaluated, and selected using the same set of 23 in-situ observations. The cross-sensor comparison with VIIRS provides evidence that the GOCI-II radiometric signal is consistent with that of an independent platform but does not constitute an independent validation of the proposed regional algorithm itself, which would require additional in-situ datasets collected outside of the present field campaigns. The 680/745 nm relationship should therefore be regarded as a preliminary regional candidate rather than a fully validated operational algorithm.

The geographic scope of the in-situ data is concentrated in the Korea Strait near Jeju Island; the applicability of the 680/745 nm algorithm to other optically distinct environments along the Korean coast, including the highly turbid macro-tidal flats of the West Sea (Yellow Sea) and the upwelling-influenced deeper waters of the East Sea, requires independent validation before deployment in those regions. In addition, chlorophyll fluorescence at 685 nm may contribute to R_rs (680) under highly productive conditions (Gower et al., 1999), and future algorithm development should explore fluorescence correction or the use of off-peak red bands to mitigate this effect.

Future research should prioritize three complementary directions. First, multi-season field campaigns should expand the in-situ database to encompass summer, autumn, and winter optical conditions across a wider geographic range of Korean coastal waters, enabling generalization and seasonal robustness testing of the algorithms presented here. Second, the development of semi-analytical retrieval approaches based on Inherent Optical Properties (IOPs), particularly the particle backscattering coefficient bbp derived via quasi-analytical inversion (Lee et al., 2002; Werdell et al., 2013), would provide a more physically transferable POC retrieval framework less sensitive to regional CDOM variability. Third, machine learning approaches capable of exploiting the full spectral information content of GOCI-II may offer further improvements in retrieval accuracy, particularly in transitional optical regimes where single-band-ratio algorithms are constrained by their one-dimensional spectral representation.

5. Conclusions

This study presents a preliminary evaluation and local optimization of GOCI-II band-ratio algorithms for POC retrieval in the optically complex coastal waters of Jeju Island. Standard global models that rely on blue-to-green band ratios (412/555, 443/555, and 490/555 nm) perform poorly in this environment (R² ≤ 0.093) because of CDOM-induced decoupling of POC scattering. In contrast, a systematic spectral sweep across all 15GOCI-II band ratio combinations identifies the Red-NIR domain (680/745, 680/709, 709/745, and 660/745 nm) as the only spectral region yielding statistically significant relationships (p < 0.05) with in-situ POC, with the 680/745 nm pair providing the best-performing candidate (R² = 0.2422, RMSE = 31.69 mg m^–³, MBE = −2.82 mg m^–³). Although the explained variance of the best-performing model remains modest, the substantial relative improvement over blue-green algorithms and the consistency of negative exponents across multiple Red-NIR pairs provide converging evidence that the Red-NIR spectral domain is more appropriate than the conventional blue-green domain for POC retrieval in this CDOM-rich environment.

Methodologically, the study illustrates the value of combining log-log calibration analysis with frequency-distribution comparison and cross-sensor consistency checks (against VIIRS) as a diagnostic toolkitfor ocean color algorithm selection in Case-2 waters. These diagnostics together help distinguish algorithmic limitations from sensor-related issues, which is particularly useful in environments where multiple optical constituents co-dominate the reflectance signal. The 680/745 nm relationship proposed here should be regarded as a preliminary regional candidate; broader operational deployment will require independent validation against an expanded multi-seasonal in-situ database and progressively more semi-analytical, IOP-based retrieval frameworks. With these further developments, the unique hourly observation capability of GOCI-II has the potential to contribute meaningfully to Blue Carbon flux quantification in the Korea Strait and the broader Asian marginal seas.

Author Contributions

Conceptualization: Rahman MR, Hwang DJ, Choi JK; Data curation, Formal analysis: Rahman MR; Funding acquisition: Hwang DJ, Son YB, Lee T, Choi JK; Investigation: Rahman MR, Hwang DJ, Jang E; Methodology: Rahman MR, Hwang DJ, Jang E; Project administration: Hwang DJ, Son YB, Lee T, Choi JK; Resources: Hwang DJ, Jang E, Son YB, Lee T, Choi JK; Software: Rahman MR; Supervision: Hwang DJ, Son YB, Lee T, Choi JK; Validation: All authors; Visualization: Rahman MR; Writing–original draft: Rahman MR; Writing–review and editing: All authors.

Conflicts of Interest

No potential conflict of interest relevant to this article was reported.

Funding

This research was supported by the Korea Institute of Marine Science & Technology Promotion (KIMST), funded by the Ministry of Oceans and Fisheries (RS-2022-KS221660).

Data Availability Statement

The in-situ radiometric and biogeochemical data used in this study are available from the corresponding author upon reasonable request.