Evaluation of Rayleigh-Corrected Reflectance on Remote Detection of Algal Blooms in Optically Complex Coasts of East China Sea

Abstract

This study used GOCI-II data to systematically evaluate the feasibility of Rayleigh-corrected reflectance (R_rc) to detect algal blooms in the complex optical environment of the East China Sea (ECS). Based on long-term in situ remote sensing reflectance (R_rs), R_rc spectra demonstrated the similar capability of reflecting the water condition under various atmospheric conditions, and the baseline indices (BLIs) derived from R_rc and R_rs showed good consistency (R² > 0.98). The effectiveness of five R_rc-based BLIs (SS490, CI, DI, FLH, and MCI) for algal bloom detection was assessed, among which SS490 and MCI showed better performances. A synthetic bloom detection algorithm based on the BLIs of R_rc was then developed to avoid the impact of turbid water. The validation of the BLI algorithm was carried out based on the in situ algal abundance data from 2021 to 2023. Specifically, SS490 showed the best bloom detection result (F-measure coefficient, FM = 0.97), followed by MCI (FM = 0.88). Since the 709 nm bands used in MCI were missing in many ocean color satellites, the SS490 algorithm was more useful in application. Compared to R_rs based bloom detection algorithms, synthetical R_rc BLI proposed in this paper provides more effective observation results and even better algal bloom detection performance. In conclusion, the study confirmed the feasibility of utilizing R_rc for algal bloom detection in the coastal areas of the ECS, and recognized the satisfactory performance of synthetical SS490 by comparing with the other BLIs.

1. Introduction

Harmful algal blooms (HABs) are a well-known ecological anomaly characterized by rapid proliferation and microplankton aggregation, causing increasingly serious ecological problems along coastal areas worldwide [1,2]. Over the past decade, substantial progress has been achieved in enhancing the spatial and temporal resolution of ocean color satellite remote sensing, enabling the synoptic detection and characterization of the location and extent of HABs from space [3,4,5,6,7,8,9,10]. Several approaches have been developed to detect algal blooms, such as the chlorophyll a concentration (Chla)-based approach [7,8,11,12] and the reflectance-based indices method [13,14]. Furthermore, over the past decade, studies have also advanced in identifying the types or causative species of phytoplankton blooms [4,15,16,17,18,19,20,21,22,23]. For example, Cannizzaro et al. [4] utilized the satellite-derived backscattering/Chla ratio to distinguish Karenia brevis blooms from other blooms in the coastal waters of Florida. Kim et al. [24] explored the potential for optically distinguishing Cochlodinium polykrikoides blooms from non-dinoflagellate blooms based on two band ratios in Korean coastal areas. Notably, all these outcomes rely on the fundamental product of remote sensing reflectance (R_rs) derived from ocean color sensors.

However, the reliable retrieving R_rs over optically complex coastal or estuarine regions still remains challenging [25,26,27,28]. For example, despite the near-daily coverage of moderate resolution imaging spectroradiometer (MODIS) measurements and a ~25–30% cloud free probability over global oceans [29], the probability of obtaining high-quality R_rs or Chla retrievals at the pixel scale is only ~5% [30]. This limited spatial and temporal coverage of MODIS observations provides timely information on phytoplankton bloom dynamics, but other ocean color sensors also face similar challenges. The main causes of such problems are significant uncertainties that occurred in atmospheric correction schemes, including incorrect presumptions such as the assumption of a negligible water signal in the near-infrared and the relationship between different spectral bands used in the iterative atmospheric correction.

Instead of using R_rs data, the Rayleigh-corrected reflectance (R_rc), which is obtained by correcting the atmospheric path radiance received by a sensor at the top-of-atmosphere (TOA) to remove the effects of ozone absorption and molecular (Rayleigh) scattering, proved to be useful in the monitoring of harmful algal blooms (HABs) [29,31,32,33,34,35]. The benefits of using R_rc include (1), the R_rc can be accurately derived since the Rayleigh component can be accurately computed, even when considering polarization effects [25] and (2), R_rc is less susceptible to atmospheric correction failures so that it can then significantly improve data coverage without the need for complex aerosol removal steps [36,37]. Additionally, the coverage of R_rc can even extend to the severe sun glint regions, allowing for the successful derivation of color patterns [33]. Due to the advantages of R_rc, Hu (2010) [32] introduced a MODIS floating algae index (FAI), defined as the difference between R_rc in the near-infrared (NIR) and a baseline formed by the red and short-wave infrared (SWIR) bands, to effectively detect phytoplankton in the ocean. Additionally, Hu (2011) also proposed another R_rc-based color index (CI) to successfully derive the color patterns even under severe sun glint conditions, and recently, this index has been converted into a high spatial coverage Chla product through machine learning [33]. Another example is the maximum peak height (MPH) algorithm, which quantifies the phytoplankton biomass by measuring the height above a spectral baseline established linearly between R_rc at 664 nm and 885 nm, specifically where the maximal R_rc occurs at 681, 709, or 753 nm [38,39]. Recently, R_rc has also shown some potential in the classification of algae species, especially for some macroalgal blooms such as Sargassum and Ulva [40,41]. Cannizzaro et al. (2019) [42] demonstrated that Cyanobacteria index and a spectral shape around 488 nm (SS488) derived from R_rc showed strong correlations with those derived from R_rs, and the estimated uncertainty of SS488 (R_rc) will minimally impact the accuracies of classifying the picocyanobacteria blooms from diatom ones in an optically shallow lagoonal estuary of Florida Bay. However, there is a lack comprehensive analysis on the ability of these methods to meet the needs of operational algal bloom monitoring in the ECS.

Similarly, in the ECS, microalgal blooms of diatoms or dinoflagellate [43,44,45,46] occur frequently every year. Til present R_rs data which still remain pivotal in the microalgal bloom detection. Shang et al. (2014) [44] initially proposed a novel method using the Bloom Index (BI), FLH, and a(443) to distinguish blooms. Tao et al. (2015) [47] employed two MODIS land bands to devise two indices: Prorocentrum donghaiense (P. donghaiense) index (PDI) and diatom index (DI), aimed at discriminating between the two species. Although a modified algal bloom index (RI) derived from the R_rc data, as proposed by Lou and Hu (2014) [43], can effectively delineate the P. donghaiense bloom in sediment-rich waters, other indices (e.g., fluorescence line height based on R_rc and CIE color represented by R_rc) were also proved to be effective [35,48,49]. Nevertheless, only a few studies on the comprehensive evaluation of R_rc data in identifying various species-dominated algal bloom can be found, as optical properties among harmful bloom species can vary, limiting the applicability of most bio-optical methods to specific species [50].

To address the lack of evaluation of R_rc in identifying microalgal blooms in the ECS, this study aims to (1) evaluate the spectral features of R_rc for various optical water types, (2) develop a synthetical R_rc-based baseline algorithm for identifying microalgal blooms in the coastal areas of the ECS, and (3) validate different R_rc baseline indices in bloom detections. By comparing R_rc and R_rs in identifying algal blooms in the ECS, this study comprehensively examines the strengths and weaknesses of the R_rc-based approach, ultimately aiming to recommend the most effective R_rc baseline index for accurately identifying microalgal blooms in the ECS.

2. Data and Methods

2.1. Study Area

This study centers around the ECS, specifically within the coordinates of 26–29°N and 119.5–122.5°E, encompassing most of the coastlines of Zhejiang Province as depicted in Figure 1. Given that the majority of the seabed in this area is less than 50 m deep, the region is greatly influenced by dynamic and intricate physical and biogeochemical oceanographic processes [51,52]. The ECS is biologically productive, with frequent annual occurrences of algal blooms, encompassing both harmful and non-harmful micro- and macro-algae [45,53,54,55]. As a result of its optically complex waters, atmospheric correction failures often leads to substantial data gaps. Improving data coverage by using R_rc is therefore important in assessing long-term bloom monitoring.

2.2. GOCI-II Data

The Geostationary Ocean Color Imager-II (GOCI-II), a maritime sensor located on the geostationary satellite Geo-Kompsat-2B (GK-2B), replaces the previous GOCI and provides comprehensive monitoring coverage of the entire sector, including Northeast Asia. GOCI-II captures images at a spatial resolution of 250 m, capturing up to 10 images daily with hourly intervals from 23 UTC to 8 UTC. Its local modes encompass 12 slots, collectively covering an extensive area of 2500 km × 2500 km across Northeast Asia, among which slot 9 encompassing our study area as depicted in Figure 1. Equipped with 12 spectral bands, GOCI-II facilitates with 1 near-ultraviolet (UV) band (380 nm), 8 VIS bands (412, 443, 490, 510, 555, 620, 660, and 680 nm), and 3 NIR bands (709, 745, and 865 nm) for ocean environmental monitoring. Particularly, GOCI-II features three new bands centered at 510 nm, 620 nm, and 709 nm, providing enhanced wavelength information to discern the optical properties difference among various types of blooms [46,50,56,57]. For the purpose of this study, the GOCI-II Level-2 data, obtained from the Korea Hydrographic and Oceanographic Agency (KHOA), were employed in bloom detection. These data include both Rayleigh-corrected reflectance and fully atmospherically corrected remote sensing reflectance data across 12 spectral bands from 380 to 865 nm. Briefly, R_rc was derived as Equation (1):

$$R_{rc,\lambda }=\frac{\pi L_{TOA,\lambda }^{Corr}}{F_{0,\lambda }\cos {\theta }_s}-R_{r,\lambda }$$

(1)

where $L_{TOA,\lambda }^{Corr}$ represents the calibrated top-of-atmosphere radiance at band λ after accounting for gaseous absorption and whitecap correction, F_0,λ is the extraterrestrial solar irradiance at the data acquisition time, θ_s is the solar zenith angle, and R_r,λ is the reflectance due to Rayleigh (molecular) scattering. GOCI-II R_rc and R_rs data were generated using the GOCI-II atmospheric correction algorithm (G2AC) [58,59].

For cloud mask, an empirical method was employed to flag cloud pixels by utilizing the relationship between neighboring spectral bands [33]. A pixel is classified as cloud if:

R_rc,₈₆₅ > 0.1, or
R_rc,₈₆₅ > 0.06 and R(R_rc,_745, R_rc,₈₆₅) < 1.15, or
R_rc,₈₆₅ > 0.027 and R(R_rc,_745, R_rc,₈₆₅) < 1.15 and S(R_rc,_745, R_rc,₈₆₅) > 0.01.

(2)

where R(R_rc,745, R_rc,865) = R_rc,745/R_rc,865, and S(R_rc,745, R_rc,865) = R_rc,745 − R_rc,865. Pixels around a cloud pixel marked by this method are also flagged as cloud to minimize the influence caused by cloud shadows.

2.3. Field Data

The in situ R_rs data were derived from two distinct sources. Firstly, a stationary oceanographic platform named Dongou, situated at 27.675°N and 121.358°E, approximately 80 km southeast of Wenzhou, China, served as a reliable source for collecting long-term radiometric measurements. These measurements encompassed remote sensing reflectance (R_rs) and normalized water-leaving radiance (L_wn). The shallow water depth around the platform minimized the impact of bottom effects on L_wn measurements. Since 2017, the platform has been equipped with a Sea-Viewing Wide Field-of-View

Sensor Photometer Revision for Incident Surface Measurements (SeaPRISM, CE318H1590TV12, Cimel Electronique, Paris, Frence) autonomous radiometer system (depicted in Figure 2), which collects sea-viewing data every 30 min, within a four-hour window centered at 12:00 PM local time. This system is capable of performing ocean color measurements across 11 wavelengths ranging from 400 to 1020 nm. The SeaPRISM data used in this study, collected between April 2021 and October 2022, were manually inspected to ensure data integrity and the absence of corrupted spectra. For the purpose of match-up comparison analysis, L_wn and R_rs data were selected at specific center wavelengths: 412, 442, 490, 510, 560, 620, 667, 779, and 865 nm. To leverage the high-frequency measurements provided by SeaPRISM, the in situ data used for the quantitative match-up comparison analysis were carefully chosen, ensuring they corresponded to the measurements made within a narrow (±0.5 h) time window surrounding the satellite overpass time at the Dongou site [60]. This approach minimized potential biases introduced by the rapid dynamic changes observed in coastal waters.

The second source of R_rs data originated from the underway observations conducted during the LORCE cruise in the ECS in August 2021. For this cruise, a shipboard fully automated hyperspectral radiometric measurement system named CrusieAOP (CAOP-H2301, Anhua Ocean Intelligent Equipment Ltd., Guangzhou, China, depicted in Figure 2d) was utilized. This system simultaneously measured the total radiance leaving the sea surface, sky radiance, and surface incident irradiance across wavelengths ranging from 350 to 900 nm with a 1 nm interval. Consequently, the in situ R_rs values were derived from these three radiometric parameters, adhering to the NASA Ocean Optics Protocols.

The records of algal blooms, encompassing the occurring time, location, and causative species, were retrieved from the WMC. These records were collected during field measurements conducted on a series of cruise in the ECS, covering the period from 2021 to 2023. At each visited bloom station during these cruises, we collected 1-L or 500-mL water samples from the surface (at a depth of 2–3 m) using Niskin bottles and preserved them in 2% glutaraldehyde [61]. Additionally, surface water samples were collected in 500 mL bottles and stored with 2% formalin. Later, the phytoplankton taxa were identified and enumerated through sedimentation on a scaled slide, employing a light microscope like the Leica DM2500. For each sample, a minimum of 300 units were counted, adhering to the morphological classification criteria [61,62]. The detailed species identified as the dominant algal species in the bloom are listed in

Table 1. Information of algal blooms in the ECS during 2021–2023.

Algal Bloom Type	Date	Center Longitude	Center Latitude	Cell Abundance (×106 cell/L)
P. donghaiense	27 April–17 May 2021	121.0866	28.0246	0.01–41.6
	4–7 June 2021	120.8045	27.3679	3.43–4.59
	7–26 May 2022	120.9105	27.4302	0.15–28.4
	2–13 June 2022	120.7438	27.3579	0.02–74.6
	18 August 2022	121.3583	27.6750	0.72
	29 April–28 May 2023	120.9378	27.4363	0.07–9.85
	30 August 2023	120.5241	27.2764	0.95
A. sanguinea	17–28 September 2021	121.1573	27.8136	0.55–6.10
Diatom	1–8 September 2021	121.2145	27.9651	5.02–6.71
	24 August 2022	120.9358	27.9675	4.00
	3 July 2023	120.5435	27.3049	0.08
	27 August 2023	120.5241	26.4276	11.48
	6 July 2023	121.0812	27.7982	0.08

.

2.4. Baseline Subtraction Algorithm

R_rc-based bloom detection algorithms typically employ a baseline subtraction method to discern spectral signatures exhibited by blooms across various wavebands. This technique has also found widespread application in diverse ocean color analyses. [31,32,38]. The popularity of the algorithms stems from the their lower sensitivity, compared to band-ratio approaches widely used in algorithms based on remote-sensing reflectance, towards various errors caused by instrument noise and imperfect atmospheric correction, including sun glint and whitecap corrections [33]. More importantly, the band ratio of R_rc cannot be directly compared between images captured at different time and under different atmosphere conditions due to the large variation in the aerosol scattering contributions.

A series of spectral indices have been developed using the baseline subtraction algorithm. These indices are generally used to quantify the height of reflectance at the specific waveband (λ) in comparison to a baseline that is linearly established between two neighboring wavebands (λ⁺ and λ⁻) as follows:

$$SS(\lambda )=R(\lambda )-R({\lambda }^{-})-\{R({\lambda }^{+})-R({\lambda }^{-})\}\times \frac{\lambda -{\lambda }^{-}}{{\lambda }^{+}-{\lambda }^{-}}$$

(3)

where R was either satellite-derived R_rc(λ) (dimensionless) or π∗R_rs(λ). In this study, five spectral indices—SS490, CI, DI, FLH, and MCI—are applied for algal bloom detection. The specification of λ⁻, λ, and λ⁺ for the five indices are listed in

Table 2. Five BLIs used to detect algal blooms and their formulas.

Index	Algorithm	GOCI-II	Reference
SS490	$R_{rc}({\lambda }^{-})+\frac{R_{rc}({\lambda }^{+})-R_{rc}({\lambda }^{-})}{({\lambda }^{+}-{\lambda }^{-})}\ast (\lambda -{\lambda }^{-})-R_{rc}(\lambda )$	λ− = 443, λ = 490, λ+ = 555	Cannizzaro et al., 2019 [42]
CI	$R_{rc}(\lambda )-\left[R_{rc}({\lambda }^{-})+\frac{R_{rc}({\lambda }^{+})-R_{rc}({\lambda }^{-})}{({\lambda }^{+}-{\lambda }^{-})}\ast (\lambda -{\lambda }^{-})\right]$	λ− = 490, λ = 555, λ+ = 620	Hu et al., 2011 [33]
DI		λ− = 555, λ = 620, λ+ = 660	Tao et al., 2015 [47]
FLH		λ− = 660, λ = 680, λ+ = 745	Hu et al., 2005 [13]
MCI		λ− = 660, λ = 709, λ+ = 745	Gower et al., 2005 [31]

.

2.5. Accuracy Assessment

In this paper, we use the confusion matrix to calculate the algal bloom detection accuracy of different BLIs. The measured stations are categorized into algal bloom (b) stations and non-bloom stations (nb) using Chinese Marine Monitoring Specification [63]. Similarly, the remote sensing detection results can be categorized into algal bloom (B) and non-bloom areas (NB), and scenarios will appear after matching ((A) b-B, (B) b-NB, (C) nb-B, (D) nb-NB).

The overall performance of the algorithm was assessed using the F-measure coefficient (FM), which represents the harmonic mean between precision and sensitivity, and was computed as follows [42]:

$$\begin{gathered} FM=\frac{({\beta }^2+1)\times precision\times sensitivity}{{\beta }^2\times precision+sensitivity} \\ precision=\mathrm{A}/(\mathrm{A} + \mathrm{C}) \\ sensitivity=\mathrm{A}/(\mathrm{A} + \mathrm{B}) \end{gathered}$$

(4)

when β is set to 1, precision and sensitivity are weighted equally. However, for this study, a value of β = 0.5 was chosen to prioritize precision over sensitivity.

3. Results and Discussion

3.1. Spectral Characteristics of R_rc

Figure 3 illustrates typical R_rc spectra of clear, medium turbid, turbid and algal bloom waters, as well as the normalized R_rc spectra using the maximum and minimum values of spectral reflectance. The spectral characteristics of algal bloom water differ from various water types. The clear water spectra show peak reflectance at blue bands, while those of algal bloom indicate a negative spectral shape around 490 nm (positive SS490). Although both are medium turbid and algal bloom water peak at 555 nm (positive CI), the bloom water showed a relatively lower reflectance and peaks up at 680 nm (positive FLH) or 709 nm bands (positive MCI) owing to the integrated effect of chlorophyll fluorescence and absorption at red band. Turbid water R_rc spectra exhibit significantly high values compared to other water types, and their reflectance peaks also shift to the red band around 620 nm (DI). These results suggest that R_rc and its baseline indices can effectively characterize the spectral differences between algal bloom and non-bloom water.

To demonstrate the stability of the R_rc BLIs under different aerosol conditions, we calculated the baseline indices using the in situ measured R_rs from SeaPRISM on the Dongou platform, and compared them with those calculated from GOCI-II R_rc. On top of that, we conducted more direct comparisons between GOCI-II R_rc and R_rs indices. Specifically, CI derived from in situ R_rs and CI derived from GOCI-II R_rc showed strong correlations (R² > 0.91, N = 204), and correlations for SS490 and DI are still high (Figure 4a–c). Most importantly, the indices derived from GOCI-II R_rs and R_rc appear to be stable and consistent (Figure 4d–f). Since the SeaPRISM does not contain the 680 and 709 nm bands, we also used the hyperspectral R_rs measurement from the shipboard CrusieAOP system to calculate the baseline indices of FLH and MCI to evaluate the synchronous GOCI-II R_rc results, along with the other three indices. Figure 5 depicted a section of underway R_rs measurements that traverse the coastal area, encompassing a range of water bodies from turbid to clean, including turbid, moderately turbid, bloom, and clean water. To minimize the uncertainties caused by different observation times of GOCI-II and shipboard measurements, we used three GOCI-II image data from the different time slots on the same day to match the different shipboard trajectories according to the measurement time of the in situ R_rs. The GOCI-II calculated baseline indices show good agreement with those from the underway R_rs measurements. Notably, SS488 and MCI appear to perform better on identifying algal blooms water, and these two indices exhibit significant increases in the algal bloom water and can be distinguished from other water types.

Additionally, we further evaluate the time series BLIs (including FLH and MCI) from both GOCI-II R_rc and R_rs at the Dongou platform in 2021 (Figure 6). The results of all indices from R_rc and R_rs have shown good consistency. More importantly, the effective data volume of R_rc is approximately 20% much greater than that of R_rs. It can be observed that the points representing only R_rc data mainly occur in January and February. During this season, sediment-dominated waters disturbed near the Dongou Platform, in which the atmospheric aerosol correction was usually poor and thus R_rs data could not be successfully retrieved. Similarly, during algal blooms in May, bloom water with extremely high chlorophyll concentrations also achieves good atmospheric aerosol correction result. Nevertheless, despite these scenarios, the strong stability of R_rc indices is effective in characterizing the spectral features of various optical water types.

3.2. Comparison of the Effectiveness of Different BLIs for Algal Bloom Detection

To compare the algal bloom detection effects of different BLIs, GOCI-II remote sensing images of several typical algal bloom and non-bloom events were selected and matched with the in situ measurements as shown in Figure 7. Each column shows, from left to right, the RGB true-color composite map and the distribution of the corresponding BLIs (SS490, CI, DI, FLH, MCI). The yellow circles indicate the measured bloom stations, with the corresponding algal cell abundance values labeled nearby (the reference concentration of algal bloom is set as 5 × 10⁵ cell/L [63]). There were significant differences in the validation results of different BLIs. The first and second rows showed the remote sensing images of 30 April 2021 and 1 May 2023, both depicting algal blooms caused by P. donghaiense and occurring in similar periods of the year. Among the five BLIs, SS490 showed the best validation for the P. donhaiense blooms, with the measured bloom stations corresponding to areas of high SS490 values. In contrast, CI had errors in indicating algal bloom area, with bloom stations corresponding to the boundary between the high and low-value areas. The stability of DI was poor, with matches in 2021 and 2023 completely reversed, as the measured bloom stations in 2021 fell in the high-value area of the DI, while those in 2023 were situated in the low-value area. The validation effect of FLH was also unstable. Although the measured points fell in areas of relatively high value in both cases, the differentiation between bloom water and non-bloom water was obviously reduced (Figure 7k). The validation effect of MCI was much better than that of CI, DI, and FLH, as the measured bloom points were located in the area of high value.

For diatom bloom (Figure 7m–r), SS490 and MCI also performed well, with both measured bloom sites located in the high-value area, while DI, CI, and FLH correspond to the boundary between low and high value areas or low-value area. The matching of DI, CI, and FLH with the remote sensing imagery of A. sanguinea bloom was still poor. MCI was slightly better, but the southern measured points were matched to the low-value area. Only SS490 had the best validation performance, with all measured bloom stations located in the high- value area. Additionally, based on the winter typical non-bloom images on 22 December 2022, it can be observed that, when blooms did not occur, either type of BLI may also show high values in turbid waters. This result suggests that any single index cannot accurately identify algal blooms, and further analysis for the differences between turbid and bloom waters is needed.

3.3. Algorithm Development of HABs Detection

Affected by turbid waters, no single BLI can accurately identify algal blooms. Therefore, we constructed a Turbidity Index TI (TI = R_rc,660 − R_rc,745) using the spectral properties in the red to near-infrared bands to indicate water turbidity. Figure 2 illustrates the scatter distribution between different BLIs and TI for the boxed area in Figure 7. When plotted with TI, each BLI showed certain distribution characteristics.

For SS490, the scattering points in the region of the two P. donhaiense blooms (Figure 8a,d) always present a shape resembling the letter “S”, corresponding to the mixing of clean water, algal bloom water, medium turbid water and turbid water in the region. Different water types were labeled in Figure 8 with blue, red, green, and yellow circles, respectively. Clean waters close to the open sea had low values of SS490 and TI, which gradually increased during mixing with algal bloom water. Then, bloom water mixed with medium turbid water with increasing TI but decreasing SS490. Finally, SS490 and TI increased together again during the transition from medium turbid to turbid waters. As a result, the scatterplot shows a clear “S” shape, indicating the mixing trends.

A similar phenomenon was observed in the scatterplots of DI, FLH, and MCI with TI. However, there were some changes in the shape compared to SS490. These three BLIs were much higher in the turbid waters, resulting in the elongation of the upper part of the “S”, reducing the ability to distinguish between clear water and algal bloom water. Also, this mixing demonstrated by the scatter of DI, FLH and MCI was not stable. In Figure 8h–j, the scatter showed almost consistent mixing trends, making it difficult to distinguish algal bloom waters from the other types of waters. The mixing of medium turbid and high turbid waters was less pronounced in the area surrounding the diatom bloom event (Figure 8k–o), where there was a mixing of algal bloom water and high turbid water. Fortunately, all BLIs but CI were able to differentiate the algal bloom scatters clearly. In the scatterplot of the A. sanguinea bloom (Figure 8p–t), the difference was more pronounced. SS490 scatter displayed a distinct “S” shape, demonstrating excellent capability in distinguishing the algal bloom water, yet the scatters of DI, FLH, and MCI showed very consistent scatter variation characteristics as in the second row. Additionally, a typical non-bloom dataset from 22 December 2022 was selected and compared with the scattering points during algal bloom events. When there was no algal bloom, the mixing occurred only between clean, medium turbid, and turbid waters, so BLIs increased as TI increased and showed an upward trend and this trend was similar in the scatter plots of all five BLIs but CI. Overall, among the four algal blooms, SS490 had the best stability and responsiveness, and basically identified the algal bloom points in each bloom event, while DI, FLH, and MCI identified only one P. donghaiense bloom and one diatom bloom, and CI was almost incapable of detecting algal blooms.

Based on the comparisons conducted above, an algorithm for detecting algal blooms was developed, focusing on the correction of turbid water effects (Figure 9a). To demonstrate the efficacy of each step of the algorithm, we take synthetical SS490 as an example, during a P. donghaiense bloom on 31 May 2023 (Figure 9b–f).

First, the distribution of SS490, calculated directly from R_rc data obtained from GOCI-II, is shown in Figure 9c. Subsequently, cloud pixels are identified based on the spectral features of the R_rc data (as described in Section 2.2), resulting in the distribution of the SS490 after cloud masking, as shown in Figure 9d. Meanwhile, it was also observed that, while SS490 captures different algal blooms, it also exhibits similar characteristics in turbid water, so it is imperative to differentiate between various water types. In this case, we empirically set the TI determination threshold for turbid water bodies at 0.02. Since algal blooms usually do not occur in turbid water in the sediment-rich ECS, this study does not conduct further processing. Water with TI values between 0.012 and 0.02, pointing out the medium turbid water or the mixed turbid and algal bloom water, requires identification from true algal bloom area due to the similar high SS490 with bloom area. To address this issue, the cubic deflation of SS490 was carried out for this type of water, so that the reduced SS490 values will be compared to those in the algal bloom area, shown as in Figure 9e. For non-turbid water, SS490 effectively distinguishes algal bloom water from clean water, obviating the need for further processing. Lastly, a synthetical SS490 is obtained by combining SS490 values for non-turbid and corrected SS490 values for turbid waters. A threshold of 0.002, determined from field measurements, is applied to detect areas above this threshold as algal bloom areas (Figure 9f).

3.4. Validation during the 2021–2023 Bloom Events

In order to compare the algal bloom detection accuracy of different BLIs, we collected field-measured data on dominant algal species and the corresponding cell abundance during algal blooms from 2021 to 2023. A total of 75 pairs of data were collected, including the sampling time, latitude, longitude, dominant algal species, and algal abundance. According to the Chinese National Ocean Monitoring Code [63], stations with an algal abundance exceeding 5 × 10⁵ cells/L were determined as algal bloom stations. Furthermore, the field-measured data were matched with the GOCI-II within a 2 km × 2 km area around the sampling points on the same day [64]. After excluding cloudy or partially missing R_rc data, a total of 35 data pairs were successfully matched. In addition to SS490, the other four BLIs were also corrected for turbid water to minimize their influences. The scatter plots of different BLIs against measured algal density are shown in Figure 10, with green dots representing non-algal bloom points and red dots representing algal bloom points recognized by the measured data. Thresholds were also set for each BLI based on previous research and experience, with the points above the threshold determined to be algal bloom points. Clearly, the synthetical SS490 was the best for agal bloom detection, showing a distinct difference between algal bloom and non-bloom scatters. By calculating confusion matrices, the quantitative accuracy of algal bloom detection by different BLIs was obtained, as detailed in

Table 3. Pixel-based statistics determined by employing confusion matrixes for comparing various algal bloom detection approaches.

Algorithms	Threshold(s)	A (b-B)	B (b-NB)	C (nb-B)	D (nb-NB)	Sensitivity A/(A + B)	Precision A/(A + C)	False Neg. % B/(A + B)	False Pos. % C/(C + D)	FM
SS490	0.002	23	3	0	9	0.88	1	0.12	0	0.97
CI	0.005	18	8	5	4	0.69	0.78	0.31	0.56	0.76
DI	0.000	18	8	7	2	0.69	0.72	0.31	0.78	0.71
FLH	0.001	10	16	3	6	0.38	0.77	0.62	0.33	0.64
MCI	0.000	20	6	2	7	0.77	0.91	0.23	0.22	0.88

.

According to the results demonstrated in Table 3, SS490 presented the best detection performance, with an overall F-measure coefficient (FM) value of 0.97, surpassing other indices in both algal bloom and non-algal bloom point discrimination accuracy. MCI ranked the second with an FM value of 0.88, indicating relatively high accuracy. However, MCI required data from the 709 nm band, which was missing in some commonly used sensors such as SNPP/VIIRS, limiting its application to some extent. In comparison, DI, CI, and FLH showed lower accuracy in bloom detections. Specifically, DI had a lower accuracy in recognizing non-algal bloom points, limiting the overall accuracy obtained from the confusion matrix calculation. FLH exhibited lower accuracy in identifying algal bloom points, resulting in insufficient sensitivity and reducing the overall FM value.

We also analyzed the results of R_rs of GOCI-II to detect algal bloom using several algorithms based on R_rs, including LHR [46], RAB [47], RI [43], and SS510 [65]. Scatter plots of algal cell abundance and baseline indices are shown in Figure 11, where green dots represent non-algal bloom points, red dots represent algal bloom points, and black dots represent points where R_rc data were valid, but R_rs data were missing. Both R_rc and R_rs data show good potential in terms of detecting algal blooms, but there were signficantly less GOCI-II R_rs data than there were R_rc data. Specifically, the number of data pairs successfully matched with measured data for R_rc data was 35, while the number of data pairs successfully matched with spectral data for R_rs was 20, resulting in a missing rate of 42.86%. value.

Meanwhile, we found that the values of these indices did not show a significant difference between non-bloom points (green) and algal bloom points (red) (Figure 11). Therefore, we marked the positions and corresponding measured algal cell abundance of these three points on RGB images, and plotted their R_rc and pi∗R_rs spectra (as shown in Figure 12, in which R_rs is multiplied by pi to have the same measure as R_rc). Clearly, these three points were all located in turbid waters, lacking algal bloom spectral characteristics, indicating that these points were likely influenced by turbid water and cannot be distinguished as non-algal bloom points. The algorithm proposed for calculating BLIs based on R_rc spectral data includes a turbid correction step, which can effectively eliminate the influence of turbid water on detecting algal blooms. Overall, the accuracy of algal bloom detection using different BLIs calculated based on R_rc spectra varies greatly, with synthetical SS490 achieving the highest detection accuracy and demonstrating good versatility and practicality. Additionally, the accuracy of R_rc-derived BLIs was not weaker than that of R_rs, but there was a significant increase in the amount of valid R_rc data compared to R_rs.

4. Conclusions

In this paper, we developed a synthetical SS490 algorithm to identify algal blooms in the ECS using the R_rc data of GOCI-II and validated the feasibility of R_rc for algal bloom detection. The R_rc-derived BLIs can reflect the water condition well and show a strong agreement (R² > 0.98) with the R_rs-derived BLIs. Among the listed BLIs, SS490 is optimal for algal bloom detection in the measured bloom events, followed by MCI. However, a single index for algal bloom detection is susceptible to the interference from turbid waters. Thus, we developed an algal bloom detection algorithm based on the synthetical R_rc derived BLI, incorporating a correction for turbid water. The synthetical SS490 performs optimally in the matching validation with the measured algal abundance data (FM = 0.97), and can stably recognize regardless of whether the blooms are P. donghaiense, diatom, or A. sanguinea bloom. We also presented the results of GOCI-II-derived R_rs for algal bloom detection and found that R_rc-derived BLIs were not weaker than the algorithms of R_rs for algal bloom detection, with an obvious increase in valid data for the R_rc-derived BLI. In conclusion, algal bloom detection using R_rc data is feasible, and the synthetical SS490 algorithm analyzed in this study can be widely applied.