High-quality remote sensing reflectance products over China and US coast

Remote sensing reflectance (R_rs) is fundamental for deriving bio-optical properties of global surface waters. However, accurate atmospheric correction (AC) to derive R_rs in coastal waters remains challenging due to strongly absorbing aerosols and complex water optics. To address this, we developed an improved processing framework that integrates flexible use of global gridded aerosol models better suited for coastal environments and incorporates tailored masking strategies. Based on this framework, we generated a high-quality R_rs dataset from Moderate Resolution Imaging Spectroradiometer (MODIS) Aqua observations, spanning 2003–2022 for the coastal waters of China and the United States (US) – two regions where complex water optics and frequent anthropogenic aerosols have long impeded retrieval accuracy and valid data yield. Compared to the NASA standard MODIS Aqua R_rs products, the improved framework achieved a significant reduction in systematic bias. Multiple validation metrics support this, with the restoration of the mean regression slope to unity (1.00 ± 0.08 vs. 0.90 ± 0.06) across eight bands being the most prominent. Performance gains are particularly pronounced in the 488–555 nm range, which is most relevant to water color applications. The new dataset also yields significantly more valid retrievals, with regional mean increases of 56 % in the Chinese coastal waters and 18 % in the US coastal waters at 443 nm. Furthermore, it recovers fine-scale features in turbid waters and captures long-term dynamics, demonstrating its robustness for coastal monitoring. As a complementary resource optimized for estuaries, bays, and turbid coastal regions, the new dataset fills a critical gap in the ocean color data portfolio where standard products are less reliable. The dataset is available at https://doi.org/10.5281/zenodo.16413443 (Zhao et al., 2025).

1 Introduction

Remote sensing reflectance (R_rs), representing the fraction of the top-of-atmosphere (TOA) signal attributable to water-leaving radiance (Gordon, 1997), is fundamental for deriving bio-optical properties of global surface waters, such as inherent optical properties (IOPs) (Lee et al., 2002; Werdell et al., 2013), concentrations of chlorophyll a (Chl) (O'Reilly, 2000; Hu et al., 2019; Tong et al., 2022) and particulate organic carbon (POC) (Stramski et al., 1999; Stramski et al., 2022), among many others. Accurate atmospheric correction (AC) – which removes atmospheric contributions from the total radiance measured by satellite sensors – is critical for retrieving R_rs and estimating water quality parameters (Gordon and Wang, 1994; Gordon, 1997). However, a key challenge in AC lies in accurately characterizing aerosol radiative properties, especially in the presence of absorbing aerosols which strongly absorb light at short blue bands (410 or 412 nm and 443 nm) (Bassani et al., 2015; IOCCG and and Mélin, 2019). In coastal regions where such aerosols (e.g., dust, smoke, and urban aerosols) are prevalent (Charlson et al., 1992), the aerosol models developed by Ahmad et al. (2010) (hereafter referred to as AF10), which are widely used in current operational AC, fail to account for strong aerosol absorption (Zhao et al., 2023), leading to systematic underestimation and limited reliability of R_rs (Wang and Jiang, 2018). In addition to aerosol-related issues, the cloud mask applied in current ocean color data processing pipelines often misclassifies highly reflective turbid waters (e.g., those with high suspended sediment concentrations) as clouds, resulting in the loss of usable R_rs data in coastal regions (Wang and Shi, 2006). Together, these issues severely constrain the usability of R_rs products in coastal regions and hamper the accurate retrieval of biogeochemical properties in these dynamic environments.

One of the key factors determining the accuracy of R_rs retrievals after atmospheric correction is whether the adopted aerosol models can effectively capture the microphysical and optical properties of aerosols, thereby accurately estimating their scattering contributions (Prospero et al., 1983; Gordon et al., 1997; Bassani et al., 2015). To address this, a range of aerosol models has been developed and refined by incorporating observational data and analyzing aerosol sources (Shettle and Fenn, 1979; Gordon and Wang, 1994; Ahmad et al., 2010; Zhao et al., 2023). Early work by Shettle and Fenn (1979) classified aerosols into five types (rural, urban, maritime, tropospheric, and fog) and accounted for humidity effects to produce a set of models (denoted as SF79). Building on this, Gordon and Wang (1994) formulated 12 humidity-dependent aerosol models (denoted as GW94) tailored for marine applications. With improved observational capacity, Ahmad et al. (2010) developed the AF10 models based on long-term data from 11 AErosol RObotic NETwork (AERONET) sites. The AF10 models parameterize aerosol particle size, refractive index, and fine-mode fraction as functions of relative humidity. Compared to GW94, they significantly reduced the uncertainties in aerosol optical thickness (AOT) and Ångström exponent (He et al., 2011; Mélin et al., 2013). As a result, the AF10 models have become the default aerosol model set in most current operational ocean color data processing systems (Mobley et al., 2016).

However, current operational AC based on AF10 models frequently yields poor R_rs retrievals in coastal regions, often resulting in underestimated R_rs values, particularly in the blue bands (Wang and Jiang, 2018). This limitation stems from the spatial representativeness of the AERONET sites used to build AF10 models: only three coastal stations (all in Chesapeake Bay) were included, while the remaining sites were located over open oceans. As a result, AF10 models primarily represent non- or weakly absorbing aerosols over the open ocean, and do not adequately reflect the high AOT and strong absorption features commonly present in coastal regions (Zhao et al., 2023). Although many studies have proposed aerosol model refinements (Giles et al., 2012; Hamill et al., 2016; Bru et al., 2017; Zhou et al., 2020; Montes et al., 2022), most focus on specific aerosol types or lack representation of seasonal variability, limiting their broader applicability for coastal waters.

Recognizing the limitations of the AF10 models, Zhao et al. (2023) developed a set of global gridded aerosol models tailored for inland and coastal waters. Leveraging AERONET observations from 1475 sites worldwide, they generated 3207 monthly aerosol models distributed across 310 5°×5° grids, and compiled the models into lookup tables compatible with the Sea-viewing Wide Field-of-view Sensor (SeaWiFS) Data Analysis System (SeaDAS) to allow seamless integration into operational AC workflows. Validation using Moderate Resolution Imaging Spectroradiometer (MODIS) Aqua data over six lacustrine or estuarine regions showed that the gridded models substantially improved R_rs retrieval accuracy compared to the AF10 models. Recently, Lv et al. (2024) further demonstrated the robustness of the gridded models through extended validations across multiple coastal regions and MODIS bands, with notable reductions in aerosol overcorrection errors under complex aerosol conditions.

Although such global gridded aerosol models have shown promise in characterizing the complex aerosol properties of coastal regions, several practical challenges hinder their direct application in operational coastal R_rs product generation. One key limitation lies in the intrinsic complexity of coastal environments, where both atmospheric and water optical conditions can exhibit strong spatial heterogeneity within a single satellite scene – from nearshore turbid waters dominated by strongly absorbing continental aerosols to offshore regions with clearer water and weakly absorbing marine aerosols. Applying either model type uniformly across such spatial gradients can lead to potential retrieval errors – for instance, using AF10 models nearshore often results in overestimated aerosol path radiance and underestimated R_rs, while applying the gridded models offshore may produce unrealistically high R_rs values due to overestimated aerosol absorption. These challenges highlight the necessity of developing a dynamic and integrated model selection scheme that can adaptively select between gridded and AF10 aerosol models under spatially heterogeneous conditions, to support accurate AC across the full coastal domain and generate spatially consistent, quality-assured R_rs products suitable for large-scale analysis.

In addition to aerosol model selection issues, a more suitable cloud masking strategy is also essential for improving the usability of coastal R_rs data. The default cloud mask employed in current operational processing workflows – originally designed for open-ocean applications – relies on a Rayleigh reflectance threshold at 869 nm ($R_{\mathrm{rc},\mathrm{869}}=\mathrm{0.027}$) to identify cloudy pixels (Robinson et al., 2003). However, in turbid coastal waters, this threshold can easily be exceeded due to high water-leaving radiance, leading to misclassification of valid pixels as cloudy (Wang and Shi, 2006). As a result, a substantial portion of valid data was incorrectly discarded, limiting the spatial coverage and continuity of the standard R_rs products. To address this, previous studies have proposed using alternative thresholds at 1240 or 2130 nm, as these relaxed cloud masking thresholds help retain more turbid water pixels. However, the relaxation of the cloud masking criteria may increase the risk of residual contamination from thin clouds. Therefore, targeted masking approaches must be incorporated during data processing to complement the modified cloud masking threshold, thereby ensuring data reliability.

In summary, while prior studies have addressed two major limitations of standard R_rs products in coastal waters – namely, the inadequacy of aerosol models and the excessive cloud masking – these improvements are not yet fully integrated into operational data production. This study bridges the gap by developing an improved processing framework that integrates refined aerosol model assignment and targeted masking strategies.

The framework is applied to generate a consistent, high-quality MODIS R_rs dataset (2003–2022) for two strategically selected regions: the coastal waters of China and the United States (US). These regions together provide representative cases for the optical challenges in coastal zones – from the highly turbid waters and frequent absorbing aerosols of China's coast to the diverse optical environments and significant anthropogenic influences along the US coast (Charlson et al., 1992; Wang et al., 2017), and serve as a rigorous testbed to validate the framework's robustness under the most demanding conditions.

The performance of the resulting dataset is comprehensively evaluated, with analyses presented on its radiometric accuracy, data coverage, and ability to resolve fine-scale and long-term spatiotemporal patterns. These results are then discussed to critically position the new framework within the current landscape of coastal ocean color research and outline its potential for future development.

2 Data and Methods

2.1 Datasets

2.1.1 Satellite Imagery

This study utilized two types of MODIS Aqua satellite data acquired from the NASA Ocean Biology Processing Group (https://oceandata.sci.gsfc.nasa.gov/, last access: 8 October 2026): Level-1A (L1A) and Level-2 (L2) ocean color products. L1A products contain unprocessed full-resolution data, supplemented with radiometric and geometric calibration coefficients and georeferencing parameters. L2 products provide derived geophysical variables (e.g., R_rs, Chl) at the same spatial resolution and geolocation as L1A data. A total of 136 544 L1A images acquired between 2002 and 2022 (data beyond 2022 were excluded here due to satellite orbital changes beginning in 2023 (Twedt et al., 2023; NASA Ocean Biology Processing Group, 2024) were used in this study for two main purposes:

1. Matchup validation: A subset of 4961 L1A images (2002–2022) was used to generate satellite R_rs for validation against in situ measurements, with temporal coverage aligned to field campaigns.

2. New coastal R_rs dataset generation: Most of the images (128 226 scenes, from 2003 to 2022) were used to generate the new R_rs dataset for two regions: the coastal waters of Chinese mainland (106–127° E, 17–41° N) and the US (128–65° W, 18–51° N).

NASA standard Level-2 R_rs products generated using the operational AC algorithm developed by NASA OBPG (NASA Ocean Biology Processing Group, 2022) were also obtained. This dataset was used for:

1. Comparative validation with R_rs data generated in this study using a common set of in situ measurements.

2. Evaluation of improvements in valid data yield by comparing valid observations counts (see Sect. 2.3.2) and assessing single-scene retrievals.

Both the new dataset and standard products include R_rs at 10 spectral bands: 412, 443, 469, 488, 531, 547, 555, 645, 667, and 678 nm, all at 1 km spatial resolution.

2.1.2 Ancillary Environmental Data

To support accurate AC, this study also incorporated ancillary data (Global Modeling and Assimilation Office, 2015a, b, c; Canada Meteorological Center, 2012, 2017) during the processing of MODIS imagery. These inputs – automatically acquired via the SeaDAS processing environment – include meteorological variables (e.g., wind speed, surface pressure, relative humidity) and concentrations of atmospheric gases (e.g., water vapor, ozone, nitrogen dioxide). These parameters are essential for calculating Rayleigh scattering, atmospheric transmittance, and gaseous absorption, and thus directly influence the quality of retrieved R_rs values. The best available ancillary data at the time of processing were used, consistent with standard SeaDAS ancillary data usage practices. This approach helps minimize the influence of auxiliary input differences when comparing with the standard R_rs products and improves the reliability of R_rs values derived with our improved processing framework.

2.1.3 In Situ Data

To evaluate the performance of the improved processing framework and compare its outputs with the standard products, we used in situ R_rs measurements archived in the SeaWiFS Bio-optical Archive and Storage System (SeaBASS, https://seabass.gsfc.nasa.gov/, last access: 8 October 2026) (Werdell et al., 2003; Bailey and Werdell, 2006). SeaBASS provides precompiled satellite–in situ matchup records, each containing metadata such as geographic coordinates, sampling time, cruise ID, measured R_rs values, corresponding MODIS Aqua filenames, and validation flags. In this study, the metadata was used to extract satellite-derived R_rs from outputs of our improved processing framework and standard products, rather than using the satellite R_rs values directly included in the SeaBASS records. This allowed us to apply a unified and more stringent quality control (QC) strategy (see Sect. 2.2.2) to both data, ensuring consistency and reliability of the accuracy assessment.

The in situ measurements cover eight MODIS bands: 412, 443, 488, 531, 547, 555, 667, and 678 nm. Due to a lack of in situ records at 469 and 645 nm, these two bands were excluded from all accuracy assessments (see Sect. 2.3.1).

To ensure compatibility with the coverage of the global gridded aerosol models and focus on coastal conditions, observations located outside the spatial domain of the global gridded aerosol models or over inland waters were excluded. A total of 6362 in situ records were retained for validation, distributed across a wide range of coastal environments (see Fig. 1).

Figure 1 Distribution of in situ measurements. Also shown are the study areas – the coastal waters of China and the United States – as well as four specific regions of interest: the Mississippi River Estuary, Chesapeake Bay, Pearl River Estuary, and Bohai Sea.

2.2 Processing Framework

Accurate AC is essential for deriving reliable R_rs in optically complex coastal waters, where high turbidity and variable aerosol properties present persistent challenges. In this study, we refined current operational ocean color AC algorithm (Mobley et al., 2016) to enhance its performance in such environments. All processing was conducted using SeaDAS software (version 8.3.0).

Our modifications build upon the standard near-infrared (NIR) iterative approach, which estimates aerosol contributions by assuming negligible water-leaving radiance in the NIR bands and minimizing residual signals through iteration (Gordon and Wang, 1994; Mobley et al., 2016). While effective over open ocean, this method could produce biased results in coastal waters due to sharp gradients in both water optical properties and aerosol characteristics. To address these limitations, we introduced a more flexible AC approach tailored to spatially heterogeneous coastal conditions. The full procedure is detailed in Sect. 2.2.1.

Additionally, a unified set of QC strategies was applied to both the newly generated data in this study and standard R_rs products to ensure consistency in evaluation. These procedures are described in Sect. 2.2.2.

2.2.1 Improved Atmospheric Correction

Building upon the standard NIR iterative approach, we further enhanced AC performance in coastal waters by incorporating the global gridded aerosol models (Zhao et al., 2023) and introducing a spatially adaptive aerosol model selection strategy. The overall process is illustrated in Fig. 2.

Figure 2 Improved atmospheric correction (AC) workflow. Notation: t (diffuse transmittance), L_w (water-leaving radiance), λ (wavelengths), NIR (near-infrared), N (normalization).

After preprocessing steps such as Rayleigh scattering and sun glint removal, the initial aerosol contribution was estimated using a gridded aerosol model selected based on the pixel's geographic location and date (Zhao et al., 2023), since the gridded models offer improved representation of the diverse and potentially absorbing aerosols common in coastal regions.

To account for spatial heterogeneity in coastal conditions, each pixel was preliminarily classified based on its normalized NIR water-leaving radiance (denoted as [L_w(NIR)]_N). Specifically, a pixel was classified as “turbid” if [L_w(NIR)]_N exceeded the empirical threshold defined by the TURBIDW flag in NASA Standard Level-2 Ocean Color Flags. Although this classification nominally distinguishes between clear and turbid waters, it served here primarily as a proxy for proximity to land. Turbid waters – characterized by elevated NIR reflectance due to high concentrations of suspended sediments or phytoplankton – are typically concentrated in nearshore zones, where riverine input and absorbing aerosols from anthropogenic sources are most pronounced (Husar et al., 1997). For these pixels, the gridded aerosol model originally assigned to the location was retained throughout iteration to better represent local aerosol properties. In contrast, pixels classified as “clear” were generally located offshore, where continental aerosol influence is weaker and water clarity is higher. For these pixels, an AF10 aerosol model was selected following the conventional open-ocean AC algorithm and used in subsequent iterations, as continued use of the gridded model in such areas could lead to underestimation of aerosol path radiance and overestimation of water-leaving radiance. This spatially adaptive model selection approach – guided by an indirect indicator of coastal influence – was key to improving retrieval performance across diverse coastal environments.

The iteration proceeded until the change in [L_w(NIR)]_N between successive steps fell below a set threshold, after which the final [L_w(λ)]_N was converted to R_rs(λ).

While the NIR-SWIR switching method is considered more robust for highly turbid waters – leveraging the near-zero water-leaving radiance assumption in the shortwave infrared (SWIR) bands – it is constrained by the low signal-to-noise ratios (SNR) of MODIS Aqua SWIR bands (Wang et al., 2009; Werdell et al., 2010). As demonstrated in previous studies (Shi and Wang, 2007; Wang et al., 2009) and confirmed by our tests, excessive noise in these bands can significantly compromise retrieval accuracy. Therefore, we retained the NIR iterative approach, with enhanced reliability achieved through spatially adaptive aerosol model selection.

2.2.2 Quality Control Strategies

To ensure reliability and comparability of R_rs retrievals, a unified set of QC strategies was applied to both the newly generated data in this study and standard products prior to subsequent analysis. These procedures aimed to eliminate contaminated or highly uncertain retrievals and to support robust evaluation of data accuracy and comparison of valid data yield.

NASA Standard Level-2 Ocean Color Flags (l2_flags) were used to mask pixels affected by clouds, ice, sun glint, stray light, and extreme solar or sensor viewing angles. In addition, pixels flagged as MAXAERITER, indicating that the maximum number of aerosol iterations was reached without convergence, were also excluded to minimize the influence of potential AC failure under complex water or aerosol conditions.

Among these flags, the CLDICE mask was customized for turbid coastal waters by applying a revised threshold ($R_{\mathrm{rc},\mathrm{2130}}=\mathrm{0.037}$). This adjustment reduces excessive cloud masking over areas with high AOT and turbid waters (Zhao et al., 2023; Zhang et al., 2014) and serves as the targeted first step to recover valid turbid-water pixels that are frequently misclassified. In contrast, the standard R_rs products used for comparison was originally generated with the conventional $R_{\mathrm{rc},\mathrm{869}}=\mathrm{0.027}$ threshold.

To mitigate the potential inclusion of non-water features resulting from the relaxed cloud mask, a secondary spectral filter was applied. Specifically, any TURBIDW-flagged pixel with R_rs in any of the four blue bands (412–488 nm) exceeding the value at 531 nm was excluded. This criterion effectively identifies cloud-contaminated water pixels whose spectral shapes deviate from typical turbid water profiles – characterized by a reflectance peak near the green wavelengths and decreasing reflectance toward the blue, due to strong absorption by colored dissolved organic matter (CDOM) and Chl (Kirk, 2010). In contrast, cloud-contaminated pixels often exhibit spuriously enhanced blue-band reflectance, resulting in anomalous spectral peaks uncharacteristic of true turbid waters. This filter is hereafter referred to as the “Wrong Turbid Water Mask”.

Extremely turbid waters were also excluded to reduce uncertainty in AC performance. Specifically, pixels exhibiting saturated R_rs signals in the 667 nm band – typically associated with very high suspended sediment concentrations – were filtered out, as such conditions compromise the reliability of both aerosol estimation and water-leaving radiance retrievals (Hu et al., 2012). This step is referred to as the “Saturated R_rs(667) Mask”.

By implementing this multi-step QC framework – comprising initial flagging, targeted cloud mask relaxation with subsequent spectral filtering, and exclusion of unreliable water signals – we prioritize data recovery in challenging coastal regions while proactively controlling for quality, thereby achieving a balance between coverage and reliability. Applying these QC strategies consistently across both datasets ensured that subsequent analyses were based on equivalently screened, quality-assured R_rs products.

2.3 Data Usability Assessment

2.3.1 Accuracy Evaluation

To assess the accuracy of the R_rs data derived with our processing framework and compare it with standard products, satellite–in situ matchups were compiled following NASA OBPG protocols (Bailey and Werdell, 2006). As mentioned in Sect. 2.1.3, both two kinds of satellite R_rs values were re-extracted based on the SeaBASS-provided metadata (e.g., geographic coordinates, acquisition time), with the same stricter QC strategy (Sect. 2.2.2) applied to ensure comparability.

Matchups were identified based on a ±3 h temporal window between satellite overpass and in situ measurement. To ensure spatial representativeness and minimize the impact of pixel-level variability, a 5×5 pixel window centered on the in situ location was extracted from each satellite image. Within each window, only non-land pixels were considered. A matchup was deemed valid only if more than 50 % of the pixels in the window passed QC screening and the coefficient of variation (standard deviation divided by the mean) was below 0.15 – ensuring spatial homogeneity and consistency with in situ conditions.

R_rs accuracy was evaluated using four metrics: regression slope, coefficient of determination (R²), median absolute percentage error (MdAPE), and normalized root mean square error (NRMSE). Among them, MdAPE serves as a robust indicator of relative error, being less sensitive to outliers than MAPE (mean absolute percentage error). NRMSE provides a scale-independent measure suitable for cross-band error comparison. These metrics were calculated as follows:

$$\begin{array}{}\text{(1)}& {\displaystyle }\mathrm{MdAPE}=\mathrm{median}\left(\left|{\displaystyle \frac{Y_i-X_i}{X_i}}\right|\times \mathrm{100}\mathit{\%}\right)\text{(2)}& {\displaystyle }\mathrm{NRMSE}={\displaystyle \frac{\sqrt{\frac{\mathrm{1}}{N}\sum _{i=\mathrm{1}}^N{\left(Y_i-X_i\right)}^{\mathrm{2}}}}{\left(X_{max}-X_{min}\right)}}\times \mathrm{100}\mathit{\%}\end{array}$$

where X_i and Y_i denote the in situ and satellite-derived R_rs values respectively, and N is the number of valid matchups.

2.3.2 Valid Data Yield Comparison

Because valid R_rs retrievals are often restricted by atmospheric conditions and QC thresholds (Wang and Shi, 2006; IOCCG and Wang, 2010), the frequency with which a dataset provides usable values serves as a practical indicator of its spatiotemporal utility. To quantify differences in data availability between the new dataset for coastal waters of China and US and the standard products, we computed a pixel-level metric representing the number of days with valid R_rs retrievals throughout the analysis period. For each pixel and spectral band, the number of days with valid retrievals was counted and denoted as VO (Valid Observation). A retrieval was considered valid if the R_rs value was non-negative and passed all QC filters described in Sect. 2.2.2.

The difference in valid observation counts between the two datasets was defined as:

$$\begin{array}{}\text{(3)}& \mathrm{VOD}={\mathrm{VO}}_{\mathrm{New}}-{\mathrm{VO}}_{\mathrm{Standard}}\end{array}$$

Here, VOD (Valid Observation Difference) captures the pixel-wise and band-wise difference in the number of valid retrieval days between the new and standard datasets. A positive VOD indicates that the new dataset provided valid R_rs values on more days than the standard dataset at the specific pixel and spectral band.

2.4 Analysis of Spatiotemporal Dynamics

To demonstrate the practical utility of the new R_rs dataset for long-term monitoring, we conducted a preliminary spatiotemporal analysis over the study regions. For each pixel and spectral band, a monthly time series of mean R_rs was first constructed by averaging all valid retrievals within each calendar month over the 20-year period (i.e., 240 monthly averages per pixel). The overall mean of this time series was then calculated and used to visualize the typical spatial patterns of R_rs across the study region.

To detect long-term trends, we applied the seasonal Mann–Kendall test – a non-parametric method that accounts for intra-annual variability and is robust to missing or abnormally distributed data (Mann, 1945; Kendall, 1970) – to the 240-month R_rs time series of each pixel. To ensure reliable trend estimation, only pixels with valid data in at least 50 % of the months were retained. The rate of change (% yr⁻¹) was then calculated by dividing the Mann–Kendall slope by the corresponding long-term mean R_rs for each pixel.

3 Results

3.1 In Situ Validation

The comprehensive validation (shown in Fig. 3) demonstrates that R_rs products generated with our improved processing framework achieve strong agreement with field observations. The most significant improvement is the correction of the pervasive systematic low bias present in the standard product, as evidenced by the average regression slope across the eight MODIS Aqua bands increasing from 0.90 ± 0.06 to 1.00 ± 0.08. This indicates that the aerosol overcorrection problem inherent to the standard AC algorithm in coastal waters has been substantially mitigated.

The improvement is particularly pronounced in the 488–555 nm range, where the new R_rs data consistently outperform the standard products across all metrics. These results confirm that the use of global gridded aerosol models, providing a regionally informed prior, is highly effective in enhancing retrieval accuracy for water color applications.

Despite these improvements, some challenges remain in the blue bands (412 and 443 nm), where strong absorption and high aerosol sensitivity complicate AC (IOCCG and Sathyendranath, 2014; Kirk, 2010). While the regression slopes improve relative to the standard products, other metrics show slightly reduced performance, likely due to the limited temporal resolution of monthly aerosol models in highly dynamic coastal regions. Nevertheless, the use of geographically informed aerosol priors still helps reduce the overcorrection tendencies typical of the standard products.

In the red bands (667 and 678 nm), the retrievals exhibit a slight overestimation (slopes = 1.06–1.16), attributable to the combined effects of aerosol model uncertainty and the very low water-leaving signal (IOCCG and and Mélin, 2019). Crucially, even with this residual uncertainty, our new data better preserve high R_rs values in turbid waters that are systematically underestimated by the standard product, demonstrating an effective correction of this bias.

In summary, the improved processing framework delivers a substantial advance in coastal R_rs accuracy by using global gridded aerosol models to achieve a major reduction in systematic bias. This establishes a more reliable radiometric foundation for coastal monitoring, long-term change detection, and downstream biogeochemical applications.

Figure 3 Scatter plots showing the validation of R_rs (unit: sr⁻¹) from this study and from NASA standard MODIS Aqua products against in situ measurements across eight spectral bands. Accuracy metrics and the number of match-ups (N) are annotated.

3.2 Valid Data Yield Comparison

Figure 4 presents the spatial distribution of valid observations (VO) at 443 nm for both the new R_rs dataset and standard products, along with their difference (VOD), over the coastal waters of China and the US in 2007 (region locations shown in Fig. 1). The year 2007 was selected as a reference due to its high data availability and broad spatial coverage in both study regions, providing a practical basis for consistent comparison.

In the Chinese coastal waters, the new dataset consistently yields more valid retrievals than the standard products, with notable improvements in the Bohai Sea, the Yellow Sea, and along the broader coastline. On average, the new dataset provides 50 VO pixel⁻¹ yr⁻¹, with local maxima reaching 140 d in the Bohai Sea – a region where semi-enclosed geography and sediment-laden river inflows result in persistently turbid conditions that challenge conventional AC (Xu et al., 2018). In contrast, the standard products yield an average of 32 VO, with a maximum of ∼ 110 d observed off the northern coast of the Philippines. The resulting VOD indicates an average increase of 18 d yr⁻¹ in the Chinese coastal waters, reaching up to 94 d in some areas. These improvements reflect the enhanced retrieval capacity of the new processing framework in optically complex coastal regions where the standard operational processing tends to underperform.

In the US coastal waters, the difference between the two datasets is generally smaller but still appreciable. The new dataset achieves an average of 80 VO yr⁻¹, compared to 68 of the standard products. Both datasets report their highest VO in the Gulf of California, with values of 232 (new) and 225 (standard). The average VOD in this region is 12 d yr⁻¹, with localized peaks up to 115 d, particularly around Chesapeake Bay, the western Gulf of Mexico, and Cuba.

Overall, the new dataset delivers a regional average increase of 18 d yr⁻¹ in the Chinese coastal waters and 12 d yr⁻¹ in the US coastal waters for R_rs(443), corresponding to relative gains of approximately 56 % and 18 %, respectively. Similar improvements are observed at other visible wavelengths, as illustrated in Figs. 5 and 6 for 547 and 667 nm, further confirming the broader enhancement in valid data yield provided by the new dataset across the visible spectrum.

Figure 4 Spatial distribution of valid observations for the new dataset (VO_New), the standard products (VO_Standard) and their difference (VOD, valid observation difference) for R_rs(443) in the coastal waters of China and the United States (US) in 2007.

Figure 5 Spatial distribution of VO_New, VO_Standard, and VOD for R_rs(547) in the coastal waters of China and the US in 2007.

Figure 6 Spatial distribution of VO_New, VO_Standard, and VOD for R_rs(667) in the coastal waters of China and the US in 2007.

Since cloud masking significantly affects the availability of valid satellite ocean color retrievals (Zhao et al., 2023; Zhang et al., 2014), a sensitivity analysis was conducted to isolate the effect of the updated cloud-masking strategy on data availability.

Specifically, we reprocessed the 2007 MODIS data for both study regions using the standard NIR iterative AC algorithm and the AF10 aerosol models, but replaced the default cloud-masking threshold ($R_{\mathrm{rc},\mathrm{869}}=\mathrm{0.027}$) with the revised threshold ($R_{\mathrm{rc},\mathrm{2130}}=\mathrm{0.037}$) adopted in our processing framework. All other processing steps, including the QC procedures described in Sect. 2.2.2, were kept consistent. This configuration is denoted as AF10&2130, and its VO_AF10&2130 was then compared to the new dataset's VO_New. Their difference, defined as VOD_AF10&2130 = VO_New − VO_AF10&2130, quantifies the change of VO when using the full improved framework beyond the cloud-masking update alone.

Results show that adjusting the cloud-masking threshold alone leads to substantial increases in VO relative to the standard products (VO_AF10&2130 in Fig. 7 > VO_Standard in Fig. 4), particularly along the Chinese coast. This confirms that the updated threshold more effectively preserves valid water pixels in highly turbid areas that would otherwise be incorrectly excluded.

Interestingly, Fig. 7 shows that VOD_AF10&2130 is negative in many regions, suggesting that the AF10&2130 configuration can sometimes produce more VO than our full processing framework. However, the added retrievals may not be of comparable quality. Although these additional pixels passed the same QC filters, these filters – especially those based on spectral shape and signal saturation – were optimized for the improved processing framework. Consequently, they may not fully capture retrieval errors specific to the AF10&2130 configuration, especially the known R_rs underestimation issue associated with AF10 models in coastal waters (see Fig. 3).

Overall, the new dataset achieves a more balanced outcome by integrating a spatially adaptive aerosol model selection strategy, an improved cloud mask, and tailored QC criteria. These components jointly enhance not only the data yield but also the reliability of R_rs retrievals in challenging coastal environments.

Figure 7 Spatial distribution of valid observations for the new dataset (VO_New), the AF10&2130 configuration (VO_AF10&2130) and their difference (VOD_AF10&2130 = VO_New−VO_AF10&2130) for R_rs(443) in the coastal waters of China and the US in 2007.

3.3 Single-Scene Imagery Comparison

To further assess the utility of the new R_rs dataset, we conducted single-scene comparisons across four representative coastal regions: the Pearl River Estuary, the Bohai Sea, the Chesapeake Bay, and the Mississippi River Estuary (see Fig. 1 for their locations). This analysis aimed to evaluate the dataset's ability to capture regional R_rs features under diverse optical conditions, and to delineate the respective contributions of the updated cloud mask and improved aerosol model. To this end, the comparisons include not only our new dataset and the standard product, but also the AF10&2130 configuration (defined in Sect. 3.2) that applies the standard AF10 aerosol models with our updated cloud mask ($R_{\mathrm{rc},\mathrm{2130}}=\mathrm{0.037}$).

Figures 8 and 9 present the comparative results for the Pearl River Estuary and the Bohai Sea, both of which are characterized by turbid nearshore waters (Liu et al., 2012; Xu et al., 2018). In these regions, the new dataset provides a markedly more extensive coverage of valid data than the standard product. For instance, while the standard products yield few valid pixels within the Pearl River Estuary and large portions of the Bohai Sea, the new dataset successfully recovers these missing observations. This enables the capture of fine-scale coastal structures – such as the distinct high-reflectance band along the Pearl River Estuary's western coast – that are otherwise lost. Critically, the AF10&2130 configuration achieves near-identical spatial coverage of valid pixels, confirming that the relaxed cloud mask is the primary driver for the increased data availability in our new dataset.

However, among the restored turbid water pixels, critical differences exist in the retrieved R_rs values. In specific regions characterized by intense sediment plumes, such as the western coast of the Pearl River Estuary and the central Bohai Sea, the R_rs values retrieved by the AF10&2130 configuration are systematically and significantly lower than those in our new dataset. This pattern is consistent with the known limitation of the standard AF10 models, which tend to overestimate aerosol path radiance in the presence of absorbing aerosols, thereby over-correcting the water-leaving signal and leading to a pervasive low bias. The regionally informed aerosol models integrated into our framework effectively mitigate this issue. Previously validated for strongly absorbing coastal regions like the Bohai Sea (Lv et al., 2024; Zhao et al., 2023), these models provide a more accurate representation of local aerosol properties. Consequently, within the sample windows (marked by black rectangles in Figs. 8 and 9), our new dataset yields higher and physically more plausible R_rs values. For example, the mean R_rs(547) in the Pearl River Estuary window increases from approximately 0.0233 sr⁻¹ (AF10&2130) to 0.0270 sr⁻¹, representing a more realistic spectral magnitude for these turbid waters.

We further examined two representative regions in the US coastal waters: the Chesapeake Bay and the Mississippi River Estuary. The same three-way comparisons presented in Figs. 10 and 11 (new dataset, standard product, AF10&2130) corroborate the findings from the Chinese seas. The relaxed cloud mask remains the primary driver for increased data coverage, recovering valid pixels in areas (e.g., the northern Chesapeake Bay and western Mississippi River Delta) that are typically flagged in the standard product. Meanwhile, the improved aerosol model ensures the radiometric fidelity of these recovered pixels. Within the highlighted sample windows, the new dataset yields R_rs values that are consistently higher than the AF10&2130 baseline – for example, approximately 0.0242 sr⁻¹ versus 0.0199 sr⁻¹ in the Chesapeake Bay at 547 nm – further evidencing the mitigation of systematic underestimation. These consistent results across geographically and optically diverse waters confirm that the observed improvements are not region-specific.

The analysis of the QC masks provides further insight. The missing pixels in the standard product are primarily flagged as cloud- or ice-contaminated, which is a misclassification attributable to the high reflectance of turbid water at 869 nm. Our updated cloud mask successfully recognizes these pixels as valid water observations. Furthermore, the new dataset's complementary QC filters (e.g., the “Wrong Turbid Water Mask” and “Saturated R_rs(667) Mask” introduced in Sect. 2.2.2) work in tandem with the relaxed cloud mask. These filters enhance overall reliability by screening out pixels with anomalous spectral features, ensuring that data quality is maintained even if the detection constraints are relaxed.

Nevertheless, the new dataset still fails to retrieve valid data in certain extremely turbid waters, such as portions of the western Bohai Sea, indicating remaining challenges in AC for waters with exceptionally high reflectance. In summary, the integrated framework underlying our new dataset delivers a substantial advancement: it significantly increases data availability in coastal zones through a tailored cloud mask, while concurrently ensuring the radiometric accuracy of the recovered data through optimized aerosol correction and specialized QC filters. This dual improvement enhances the dataset's suitability for detailed ocean color analysis and reliable water quality monitoring in optically complex nearshore environments.

Figure 8 Comparison of R_rs (units: sr⁻¹) derived from single-scene MODIS Aqua observations over the Pearl River Estuary (14 December 2014). The top image shows a true color composite (a), followed by R_rs at 443, 547, and 667 nm from the new dataset (b–d), from the standard dataset (e–g) and from the AF10&2130 configuration (h–j). Invalid pixels are marked with hatched patterns based on quality control masks. Black rectangles mark the sample windows where average R_rs values are calculated and compared quantitatively in the text.

Figure 9 Comparison of R_rs (units: sr⁻¹) derived from single-scene MODIS Aqua observations over the Bohai Sea (11 March 2020). The top image shows a true color composite (a), followed by R_rs at 443, 547, and 667 nm from the new dataset (b–d), from the standard dataset (e–g) and from the AF10&2130 configuration (h–j). Invalid pixels are marked with hatched patterns based on quality control masks. Black rectangles mark the sample windows where average R_rs values are calculated and compared quantitatively in the text.

Figure 10 Comparison of R_rs (units: sr⁻¹) derived from single-scene MODIS Aqua observations over the Chesapeake Bay (5 February 2019). The top image shows a true color composite (a), followed by R_rs at 443, 547, and 667 nm from the new dataset (b–d), from the standard dataset (e–g) and from the AF10&2130 configuration (h–j). Invalid pixels are marked with hatched patterns based on quality control masks. Black rectangles mark the sample windows where average R_rs values are calculated and compared quantitatively in the text.

Figure 11 Comparison of R_rs (units: sr⁻¹) derived from single-scene MODIS Aqua observations over the Mississippi River Estuary (13 January 2021). The top image shows a true color composite (a), followed by R_rs at 443, 547, and 667 nm from the new dataset (b–d), from the standard dataset (e–g) and from the AF10&2130 configuration (h–j). Invalid pixels are marked with hatched patterns based on quality control masks. Black rectangles mark the sample windows where average R_rs values are calculated and compared quantitatively in the text.

3.4 Spatiotemporal Patterns and Trends

Figure 12 illustrates the spatial distribution of long-term mean R_rs values at 443, 547, and 667 nm for the coastal waters of China and the US over the 2003–2022 period. These patterns reveal clear wavelength dependence and regional variability.

In Chinese coastal waters (Fig. 12a–c), the regional average R_rs(443) is 0.00679 sr⁻¹, rising to 0.00866 sr⁻¹ within the 30 m isobath (marked by the white dashed line), where optically complex conditions dominate. Notably high R_rs(443) values are observed in nearshore areas such as the western Bohai Sea (0.00738 sr⁻¹), the Yangtze River Estuary and Hangzhou Bay (0.01213 sr⁻¹), the Pearl River Estuary, and the eastern Beibu Gulf. These high R_rs(443) values persist despite strong absorption by high concentrations of CDOM and Chl, suggesting that backscattering by suspended sediments dominates the blue reflectance signal in these turbid waters (Kirk, 2010; Feng et al., 2014; Ye et al., 2016). In contrast, significantly low R_rs(443) values (< 0.003 sr⁻¹) occur in the central and northern Yellow Sea, where sediment loads are reduced and absorption remains high, resulting in suppressed reflectance in the blue band (Kirk, 2010; Ye et al., 2016; Ling et al., 2020). At longer wavelengths (547 and 667 nm), high R_rs values are more restricted to highly turbid nearshore and estuarine waters. Within the 30 m isobath, the mean R_rs(547) and R_rs(667) values reach 0.01551 and 0.00819 sr⁻¹, respectively, decreasing to 0.00506 and 0.00184 sr⁻¹ farther offshore. These spatial patterns reflect the varying contributions of optically active constituents to spectral reflectance.

In US coastal waters (Fig. 12d–f) the long-term mean R_rs values exhibit distinct spatial features. At 443 nm, high R_rs values (> 0.03 sr⁻¹) are found around the shallow Bahamas Banks, while low values (< 0.003 sr⁻¹) dominate the East and West Coasts of the US, indicative of strong absorption by CDOM and Chl (Mannino et al., 2008; Kirk, 2010). At 547 and 667 nm, high R_rs values are observed in regions such as the northwestern Gulf of California, the Gulf of Mexico coast, and around Chesapeake Bay. As in the Chinese coastal waters, R_rs decreases sharply with increasing distance from shore.

Figure 12 Spatial distribution of long-term mean R_rs (units: sr⁻¹) at 443, 547, and 667 nm for China (a–c) and US (d–f) coastal waters, derived from the new dataset (2003–2022). White dashed lines indicate the approximate 30 m isobath.

Figure 13 illustrates the spatial distribution of long-term R_rs trends (% yr⁻¹) for the same wavelengths, as estimated using the seasonal Mann-Kendall test over the 240-month period.

In Chinese coastal waters (Fig. 13a–c), R_rs(443) generally shows increasing trends, with a mean rate of 0.64 % yr⁻¹ and localized peaks reaching 6.85 % yr⁻¹ in the central Yellow Sea. R_rs(547) and R_rs(667) also exhibit upward trends in this region, potentially reflecting rising suspended sediment concentrations and declining Chl and CDOM levels in recent years, as previously observed by Wang et al. (2023). In contrast, most of Laizhou Bay, Bohai Bay, and the eastern Liaodong Bay show consistent negative trends across all three wavelengths, possibly indicating a reduction in suspended sediment loads. This trend aligns with earlier findings that attribute such declines to weakening wind-driven sediment resuspension (Li et al., 2022; Zhao et al., 2022). The mean trends for R_rs(547) and R_rs(667) are 0.18 % yr⁻¹ and 0.24 % yr⁻¹, respectively, although widespread declines are also evident in other areas, such as the northern South China Sea, with maximum decreases reaching −6.70 % yr⁻¹ and −9.08 % yr⁻¹.

In the US coastal waters (Fig. 13d–f), R_rs(443) exhibits a mean increase of 0.28 % yr⁻¹, with peak increases reaching 7.22 % yr⁻¹. Strong positive trends are observed along the US West Coast, near the Bahamas, and within the Gulf Stream region. In contrast, notable declines occur in areas such as southern Florida Peninsula and the western Gulf of Mexico, with rates reaching −6.06 % yr⁻¹. For R_rs(547) and R_rs(667), the overall trends are negligible (−0.004 % yr⁻¹ and 0.042 % yr⁻¹, respectively), with most coastal regions – except the western Gulf of Mexico – showing decreasing trends, with minimum values of −5.68 % yr⁻¹ and −7.51 % yr⁻¹, respectively. A particularly complex pattern emerges along the Gulf Stream: while R_rs(443) increases significantly, R_rs(547) and R_rs(667) exhibit moderate decreases. This may reflect the growing influence of oligotrophic Gulf Stream waters (Leonelli et al., 2022) and enhanced upper-ocean stratification associated with regional ocean warming (Todd and Ren, 2023). These processes are known to suppress vertical mixing and nutrient supply (Li et al., 2020), leading to reduced phytoplankton biomass (Yu et al., 2023) and lower concentrations of suspended matter (Wei et al., 2021) in the surface layer. Such reductions in optically active constituents diminish absorption in the blue bands and scattering in the green and red bands, consistent with the observed R_rs trends in this region. Additionally, the magnitude of R_rstrends is greater in the Chinese coastal waters than in the US coastal waters, likely reflecting stronger anthropogenic influences, hydrological variability, and regional-specific geomorphological settings.

The observed spatiotemporal patterns and trends in R_rs underscore the robustness of the new dataset in capturing diverse coastal optical dynamics. Their consistency with previously reported trends lends further confidence in the dataset's reliability and demonstrates its value as a foundation for future research into long-term environmental change, human impacts, and climate-related ecosystem responses in coastal waters.

Figure 13 Long-term R_rstrends (units: % yr⁻¹) from 2003 to 2022, derived using the seasonal Mann-Kendall test at 443, 547, and 667 nm for China (a–c) and US (d–f) coastal waters. Rates are computed as the Mann-Kendall slope divided by the corresponding 20-year average R_rs.

4 Discussion

4.1 Comparison Against Other AC Algorithms for Coastal Waters

To further contextualize the performance of our framework, we conducted a scene-based intercomparison with three advanced AC algorithms for complex waters: OC-SMART (Fan et al., 2021), POLYMER (Steinmetz et al., 2011), and NIR-SWIR, using the same four representative regions from Sect. 3.3. This analysis also included results from the standard NIR iterative algorithm with AF10 aerosol models and relaxed cloud mask (AF10&2130, as mentioned before). This configuration provides a reference baseline that embodies the persistent low bias, so the corrective effect of other algorithms can be assessed against it. For a fair and controlled comparison, the NIR-SWIR and AF10&2130 configurations used the same cloud mask and quality control steps as our new dataset, while OC-SMART and POLYMER were run with their default settings.

The comparative imagery (Figs. 14–17) demonstrates the superior performance of our framework. While all five methods successfully capture major spatial patterns associated with turbid water features (e.g., sediment plumes) in the 547 and 667 nm bands, our new dataset consistently retrieves the highest R_rs values in the most turbid regions across all scenes and bands (443, 547, 667 nm). In contrast, the R_rs values retrieved by OC-SMART, POLYMER, and NIR-SWIR in these challenging areas are frequently comparable to, or only marginally higher than, those of the AF10&2130 configuration. Given that the AF10&2130 configuration is quantitatively demonstrated to suffer from systematic underestimation (Sect. 3.1, Fig. 3), this indicates that the corrective effect of these other algorithms on the low bias is limited in optically extreme waters, whereas our improved framework provides the most robust correction.

Regarding data quality, NIR-SWIR results exhibit increased image noise due to the low signal-to-noise ratio of MODIS Aqua's SWIR bands. While POLYMER and OC-SMART generally perform well, isolated issues such as unmasked anomalous pixels (Bohai Sea for POLYMER) and striping artifacts (Mississippi River Estuary at 443 nm for OC-SMART) were observed. In terms of data coverage, the default masks for OC-SMART and POLYMER retained slightly more data in the Pearl River Estuary than our more conservative QC chain.

Figure 14 Comparison of R_rs (units: sr⁻¹) derived from single-scene MODIS Aqua observations over the Pearl River Estuary (14 December 2014). The figure shows R_rs at 443 nm (top row), 547 nm (middle row), and 667 nm (bottom row) as processed by the AF10&2130 configuration (first column), our new dataset (second column), the NIR-SWIR algorithm with AF10 aerosol models and $R_{\mathrm{rc},\mathrm{2130}}=\mathrm{0.037}$ as cloud mask threshold (third column), OC-SMART (fourth column), and POLYMER (fifth column).

Figure 15 Comparison of R_rs (units: sr⁻¹) derived from single-scene MODIS Aqua observations over the Bohai Sea (11 March 2020). The figure shows R_rs at 443 nm (top row), 547 nm (middle row), and 667 nm (bottom row) as processed by the AF10&2130 configuration (first column), our new dataset (second column), the NIR-SWIR algorithm with AF10 aerosol models and $R_{\mathrm{rc},\mathrm{2130}}=\mathrm{0.037}$ as cloud mask threshold (third column), OC-SMART (fourth column), and POLYMER (fifth column).

Figure 16 Comparison of R_rs (units: sr⁻¹) derived from single-scene MODIS Aqua observations over the Chesapeake Bay (5 February 2019). The figure shows R_rs at 443 nm (top row), 547 nm (middle row), and 667 nm (bottom row) as processed by the AF10&2130 configuration (first column), our new dataset (second column), the NIR-SWIR algorithm with AF10 aerosol models and $R_{\mathrm{rc},\mathrm{2130}}=\mathrm{0.037}$ as cloud mask threshold (third column), OC-SMART (fourth column), and POLYMER (fifth column).

Figure 17 Comparison of R_rs (units: sr⁻¹) derived from single-scene MODIS Aqua observations over the Mississippi River Estuary (13 January 2021). The figure shows R_rs at 443 nm (top row), 547 nm (middle row), and 667 nm (bottom row) as processed by the AF10&2130 configuration (first column), our new dataset (second column), the NIR-SWIR algorithm with AF10 aerosol models and $R_{\mathrm{rc},\mathrm{2130}}=\mathrm{0.037}$ as cloud mask threshold (third column), OC-SMART (fourth column), and POLYMER (fifth column).

Beyond these empirical results, our methodological approach itself warrants discussion. It enhances a physically based, modular component (aerosol modeling) within the radiative transfer framework. This differs from strategies that depend on comprehensively pre-trained spectral or bio-optical models (e.g., POLYMER, OC-SMART), whose performance is inherently tied to the optical domains covered during their training or optimization. Adapting our framework to a new coastal region would primarily involve updating the modular aerosol prior with local or regional climatological data. In contrast, methods like POLYMER or OC-SMART may require significant recalibration or retraining to maintain performance outside their original optimization domain. Thus, our design offers a more straightforward and transparent path for regional scalability.

4.2 Considerations on the Aerosol Model Selection Strategy

A pivotal component of our improved AC framework is the dynamic selection of aerosol models, guided by the normalized water-leaving radiance in the NIR band ([L_w(NIR)]_N). This design is predicated on the strong empirical spatial co-occurrence observed in coastal zones: waters exhibiting high turbidity (and thus elevated [L_w(NIR)]_N) are typically nearshore, and this high turbidity spatially co-occurs with prevalent continentally influenced, often absorbing aerosols. Conversely, in offshore regions, waters are generally clear (with low [L_w(NIR)]_N) and maritime aerosols prevail. We leveraged this consistent geographic pattern as a practical and effective proxy to implement a necessary transition between aerosol models across the land-ocean gradient. This spatially adaptive selection prevents the systematic underestimation that would occur if AF10 models were applied to turbid nearshore waters, while simultaneously avoiding the overestimation that would arise if gridded models representing absorbing aerosols were used in clear offshore waters. The success of this approach is evidenced by the reduction of systematic bias (Fig. 3) and the spatially coherent R_rs fields without discontinuities (e.g., Figs. 8, 9).

It is important to clarify that this strategy does not assert a direct physical causation between water optical properties and aerosol characteristics. Rather, it is a pragmatic solution that capitalizes on their correlated spatial distributions driven by common terrestrial influences. Consequently, the performance of this approach is inherently tied to the robustness of this co-occurrence relationship. In regions or under conditions where this relationship weakens or decouples (e.g., during long-range transport of dust or smoke over clear ocean), the proxy may be less effective. To overcome this limitation and extend the framework's applicability, research could focus on integrating direct, physically based aerosol indicators (e.g., from multi-angle or polarimetric sensors). Alternatively, developing regionally adapted or higher-resolution co-occurrence models could refine the proxy and enhance the robustness of the aerosol selection process.

4.3 Considerations on Quality Control Strategies

The QC strategy is integral to the reliability of any ocean color dataset. Our framework combines a relaxed cloud mask ($R_{\mathrm{rc},\mathrm{2130}}=\mathrm{0.037}$) with additional filters (e.g., the “Wrong Turbid Water Mask”), to prioritize the retention of valid turbid water pixels that are routinely discarded by standard processors, while implementing safeguards against anomalous retrievals.

However, when comparing with OC-SMART and POLYMER (Sect. 4.1, Fig. 14), we observed that their default masks retained slightly more valid pixels than our more conservative QC chain. This difference is primarily attributable to our “Wrong Turbid Water Mask,” (See Fig. 8), which excluded some pixels based on spectral criteria. This illustrates an inherent tension in QC design for coastal waters: an overly conservative spectral filter may reject some valid but optically extreme water pixels, whereas a more permissive mask risks retaining subtle cloud contamination or other artifacts. Our current strategy prioritizes per-pixel confidence over initial data yield, accepting a minor reduction in the latter for enhanced overall reliability.

Looking forward, cloud and cloud-shadow detection remains one of the most persistent challenges in global aquatic remote sensing. Although the QC framework used in this study performs robustly in the target regions, its universal optimization is non-trivial due to the vast diversity of atmospheric and aquatic conditions. The development and integration of more advanced, physically based or machine-learning-driven cloud masking algorithms, capable of regional adaptation and distinguishing between high turbidity and true cloud cover, represents a critical future direction. Such advancements could potentially obviate the need for heuristic spectral filters like the “Wrong Turbid Water Mask”, further automating the pipeline and enhancing the consistency and yield of quality-controlled data products across all coastal zones.

4.4 Considerations on Model Representativity

The performance of our AC framework is inherently linked to the quality of the aerosol prior. The representational characteristics of the gridded aerosol models – namely, their spatial (5°) and temporal (monthly) resolution – thus constitute a fundamental aspect of the framework's design. While these models have proven effective, their fixed resolution may not fully capture the high heterogeneity of aerosols in coastal areas at instantaneous and pixel scales. However, as integrated into our processing chain (Sect. 2.2.1), the model serves as an optimized starting point for iterative refinement, a design that mitigates the impact of imperfect prior matching. The validation results (Fig. 3) and scene-based comparison (e.g., Fig. 8) confirm that this approach yields a significant and effective correction of the systematic bias at the current model resolution. Future work to further enhance fine-scale coastal retrievals could therefore focus on developing regional aerosol climatology or integrating real-time assimilation products with higher spatiotemporal resolution, thereby providing an even more accurate AC starting point.

5 Data availability

The high-quality coastal R_rs dataset generated in this study is publicly available at https://doi.org/10.5281/zenodo.16413443 (Zhao et al., 2025). The dataset is provided in monthly composites and stored in GeoTIFF format at 0.01° spatial resolution under the EPSG: 4326 (WGS84) spatial reference system.

6 Conclusions

This study has developed and validated an improved processing framework for coastal ocean color retrieval, integrating and operationalizing key advancements in aerosol modeling and cloud masking. A central innovation is the spatially adaptive selection of aerosol models, guided by water turbidity as a proxy for continental aerosol influence, which effectively addresses the systematic bias across the land-ocean gradient. Coupled with tailored quality control strategies, this framework was used to generate a high-quality, consistent MODIS Aqua R_rs dataset (2003–2022) for the coastal waters of China and the United States. The resulting dataset was rigorously evaluated against NASA standard R_rs products through a series of comparative analyses under consistent and stringent quality control strategies.

Compared to the standard products, the new dataset demonstrates a decisive improvement: a substantial reduction in systematic low bias alongside a marked increase in valid retrieval coverage. This dual enhancement is most pronounced in turbid coastal regions and is evident across all scales of analysis – from pixel-level validation to the resolution of fine-scale features and long-term spatiotemporal patterns. The dataset thus provides a more reliable and complete observational foundation for characterizing coastal optical variability.

The dataset is explicitly positioned as a complementary resource optimized for turbid coastal regions, estuaries, and bays, filling a gap where standard products are less reliable. Although this study focused on MODIS Aqua and two regional cases, the methodology is scalable and transferable to other sensors and global coastal zones. Future work may involve broader in situ validation and the application of this dataset to retrieve biogeochemical parameters such as Chl, suspended sediments, and CDOM, thereby extending its utility for coastal management and climate-related research.