The coastal waters surrounding Plymouth (United Kingdom) are located on the North Western European Continental Shelf, a dynamic environment heavily influenced by freshwater, tidal and marine currents (Figure 1). The tidal cycle of the estuary governs hydrodynamic processes in the area (Siddorn et al., 2007). In combination with biological seasonality, these influence the dynamics of the optically active constituents (total suspended matter concentration, TSM; coloured dissolved organic matter, CDOM and phytoplankton chlorophyll concentration, Chla; Doxaran et al., 2006). Offshore, in coastal waters around station L4 (water column depth 55 m, Figure 1), optically active constituents are mostly influenced by biological seasonality and river outflows. Phytoplankton dominates during spring/summer bloom events and TSM (mainly made of inorganic particles) and/or CDOM in winter (Groom et al., 2009; Martinez-Vicente et al., 2010). Low concentrations of optically active constituents are present in the area, with concentrations of chlorophyll-a around 1.5 mg m⁻³, reaching 4 mg m⁻³ during spring bloom at L4 station and TSM concentrations generally around 1 g m⁻³, reaching 9.94 g m⁻³ in winter months (Martinez-Vicente et al., 2010).

Ocean and Land Colour Instrument (OLCI) data were acquired for the period between December 2016 and March 2021. The data evaluated in this study are the fully normalised water-leaving reflectance (ρ_w) full resolution (300 m × 300 m) for 15 selected bands: 400, 412, 443, 490, 510, 560, 620, 665, 674, 681, 709, 754, 779, 865 and 885 nm. For comparison to field measurements, these were then transformed to remote sensing reflectance (R_rs) by dividing by π.

The processing stages included atmospheric correction using the POLYMER v4.15 algorithm and the IDEPIX mask generation. The POLYMER atmospheric correction algorithm was originally developed for MERIS open water products to remove Sun glint (Steinmetz et al., 2011), using a BRDF correction from Park and Ruddick (2005) and selected because of its high performance for Sentinel product (Mograne et al., 2019; Giannini et al., 2021). IDEPIX (Identification of Pixel properties, in SNAP 8) masking is used to remove land, cloud and spurious data points with high uncertainty. The processing chain follows that developed in Warren et al., 2019.

For the creation of a set of optical water types (OWTs), a spatial/temporal subset of the processed satellite data acted as a training dataset to represent the observed variability in the reflectance spectra in the region of interest (Jackson et al., 2017). The optical water type classification first involved the standardisation of the output products by subtracting the mean and dividing by the standard deviation for each band. Principal component analysis was then performed on the standardised data whereby the resulting principal components were used in the cluster analysis to classify the coastal waters into OWTs. The cluster analysis consisted of a fuzzy c-means clustering procedure using a cluster validity function to generate clusters (Xie and Beni, 1991; Moore et al., 2001). The clustering approach resulted in 6 OWTs in the study area (Figure 2). Memberships to all OWT classes were calculated for every satellite data matchup and assigned a dominant OWT class based on highest membership value. The in situ spectra associated with that satellite matchup pixel were assigned the same OWT class.

During field campaigns in 2017, 2018 and 2020, above-water radiometric measurements were collected with the unsupervised Hyperspectral Sur-face Acquisition System (HSAS, SeaBird Inc.). The sky and water leaving radiance measurements were acquired from the bow of R/V Plymouth Quest at about 5 m above the water surface, whilst out at sea. R/V Plymouth Quest is a 21.5 m length scientific vessel that has been used in previous works (Martinez-Vicente et al., 2013; Jordan et al., 2023). The downward looking radiometer has a field of view of 6°, which corresponds to about 0.6 m at the sea surface. The downwelling irradiance sensor was located at the top of the wheelhouse, on the extreme side of the top of a T shaped mast, away from any superstructure. The sensors and the setup provide accuracy similar to other sensors and approaches (Tilstone et al., 2020). The system continuously recorded downwelling irradiance (E_d), sky radiance (L_s) and water leaving radiance (L_t) at 169 wavelengths with 3.3 nm spectral resolution between 340.0 and 897.7 nm. The scanning frequency was between 4 and 0.5 Hz, depending on the sensor optics. The optical data were converted to physical units and processed to Level3a using the manufacturer’s software (Prosoft v7.7.16) which merged the data to 1 Hz. R_rs was inferred following: R_rs(λ) = L_w+(λ)/E_d(λ) where L_w+(λ) = L_t(λ)–ρ_s L_s(λ) with L_w+ the water leaving radiance above the surface and ρ_s the surface reflectance factor (Martinez-Vicente et al., 2013). At the normal speed of R/V Plymouth Quest, the average number of R_rs spectra data per OLCI FR pixel (300 m) is 150. The spatial structure of reflectance dataset collected underway from R/V Plymouth Quest is discussed in detail in Jordan et al. (2023). Sensors were factory calibrated every year.

The data were then processed following the fingerprint approach (Simis and Olsson, 2013; QC0 in Table 1) to resolve ρ_s for each observation. Following the procedure, data flagged as valid were retained and filtered according to published recommendations. Quality control was done using characteristics of the spectral shape (Simis and Olsson, 2013; Qin et al., 2017; IOCCG, 2019; Warren et al., 2019). Quality control consisted of removing spectra affected by significantly negative values (QC1 in Table 1; Qin et al., 2017), removing spectra affected by Sun glint and whitecaps (QC2 and QC3 in Table 1; Qin et al., 2017; IOCCG, 2019), removing the effect of the oxygen absorption feature from the instrument calibration (QC4 in Table 1; Qin et al., 2017). It also consisted of accurately selecting spectra representative of the study area where dominance of mineral particles compared to coloured dissolved organic matter can be found (QC5 in Table 1; Qin et al., 2017; Warren et al., 2019). By applying those successive gating criteria, the number of observations was reduced from 174,396 to 9,065 (Table 1).

In situ spectra that passed the filtering criteria were convolved to the spectral response functions of the OLCI, instrument on Sentinel-3A.

The distance to land (including smaller features such as islands and breakwaters) for each in situ measurement was computed using the great-circle distance (Distance to Nearest Hub tool in QGIS). All data points were within 15 km of land. These were grouped into three categories (namely, 0–5 km, 5–10 km, and 10–15 km), and used to investigate the effect of land adjacency on the satellite data quality as described in Section 2.5.

The selection of spatial and temporal windows has been documented to potentially affect validation performance statistics (Concha et al., 2021). Therefore, different matchup selection routines were explored in this study. Two temporal windows (±30 min and ±3 h) and two spatial windows (1 pixel and 3 × 3 pixels) were considered (Table 2). In the case of the 3 × 3 pixel area, only matchups yielding a minimum of six pixels and a coefficient of variation <0.15 were retained (Brewin et al., 2016; Bailey and Werdell, 2006). The mean in situ spectra were then used to reduce the satellite data to unique validation points for comparison purposes. A detailed study on the variability within different spatial windows is not the subject of this paper and has been explored elsewhere (Nasiha et al., 2022; Jordan et al., 2023).

The satellite matchup data points were filtered (as shown in ‘Filtering’ step in Table 2) for significantly negative values (< − 0.0005   s r − 1 ) to allow for consistency with the in situ dataset (QC1 in Table 1; Qin et al., 2017). In the case that multiple field measurements were present in the same pixel, the median value was taken (‘Median’ step in Table 2).

For very dynamic waters or stations close to the coast, EUMETSAT (2021) states that reduction of the matchup time window to 1 hour and the spatial window to 1 pixel is acceptable. Thus, given the spatiotemporal variability in the study area, the ±30 min and 1 × 1 pixel matchup selection routine was selected. The resulting matchup in situ spectra span across 19 sampling dates, covering a transect of approximately 12 km (Respective in situ spectra retained: 0–5 km: N = 11, 5–10 km: N = 38, 10–15 km, N = 22) and spanning six optical water types. The environmental conditions for those sampling dates were clear to slightly overcast skies for all days other than 14th July 2020 showing hazy skies. Most matchups occurred in OWTs 1 and 3 with N = 23 and N = 18, respectively.

From the initial 174,396 reflectance spectra collected, 430 were preserved, and converted to 71 median spectra that were then used for the statistical matchup, coming from a temporal window of 30 min and a spatial window of 1 × 1 pixel.

A set of performance metrics was then calculated to describe the difference between the OLCI and in situ data. Root Mean Square Difference (RMSD), Median Absolute Percentage Difference (MdAPD) and Median Percentage Difference (MdPD) were computed to quantify the accuracy, dispersion and bias of the data, respectively, with x the in situ data, y the satellite data and N the number of matchups 1–3: R M S D =   1 N * ∑ i = 1 N y i − x i 2 (1) M d A P D = M e d i a n   y i − x i x i * 100 (2) M d P D = M e d i a n   y i − x i x i * 100 (3)

These metrics were chosen since they are most common in validation exercises within the Ocean Colour Radiometry community (Concha et al., 2021). In addition, the distribution of the datasets was investigated.

Data were then segregated by distance to land and by optical water types. Performance metrics were then calculated for subsets depending on the OWT and distance from land. Following a Levene’s test for heterogeneity of variance and Shapiro-Wilk test for normality of distribution, non-parametric tests were used. Mann-Whitney tests were carried out to test for statistically significant differences. A permutational analysis of variance (PERMANOVA) with 1,000 permutations based on Euclidian distances was conducted to test for differences in the performance metrics between the distinct OWTs and distance from land.

A comparison of performance metrics between the various matchup selection routines can be found in Section 3.1. Results and discussion are, however, mainly focused on the matchups with a temporal window of 30 min and a spatial window of 1 × 1 pixel, on all data and on data segregated by OWT dominance and distances to the shore. Scatterplots, histograms and descriptive statistics for the other spatial and temporal match up windows explored (as described in Section 2.4) are provided in Supplementary Materials S2–S4.

First, satellite and in situ data were compared using various matchup selection criteria using distinct temporal and spatial windows. The respective performance metrics were calculated across the spectrum and for each spectral band, and further displayed in Figure 3.

The spectral distribution in the RMSD was similar across data from all matchup selection routines with a peak at 490 nm with an overall decrease with a smaller temporal and spatial window, going from 0.0021 sr⁻¹ with a ±3 h and 3 × 3 pixel window to 0.0013 sr⁻¹ with a ±30 min and 1 × 1 pixel window (Figure 3A). Similarly, the spectral distribution of the MdAPD was consistent across all data with a higher median value of 64% for the matchup data selected within a 3-h window (Figure 3B). On the other hand, the highest median MdPD value could be observed for the most stringent dataset (±30 min 1 × 1 pixel selection window) with a peak of 200% depicted at 779 nm (Figure 3C).

Next, the most stringent matchup selection criteria (±30 min and 1 pixel) were delved into further. Overall, when comparing the satellite reflectance estimates with in situ measurements, small differences were observed, yet a positive bias could be seen over most of the spectral range. To illustrate any discrepancies between satellite and in situ R_rs for the whole matchup dataset (N = 71), scatter plots and histograms for each of the S3-OLCI bands are shown in Figures 4, 5. The distribution of the data can further be observed in relation to optical water types and distance to land as depicted in Figure 4. Performance metrics were then calculated for each spectral band (Table 3; Figure 6).

In general, there was a large scatter across the wavelengths, but a positive slope of the regression between satellite and in situ R_rs can be seen for all wavelengths but 754 nm and 779 nm (Figure 4). A positive bias of the satellite estimates can be observed in relation to in situ measurements for most wavelengths, except in the range from 490 to 560 nm, Figure 4. It is, however, essential to notice that the R ² values depicted are very small (≤0.08, Figure 4). It is also worth noticing, that when looking at the distribution of the data points in relation to optical water types, measurements in waters dominated by OWT 6 were particularly out of range (Figure 4).

As the range of observations is limited, another way to visualise the data is by means of histograms. The histograms show a maximum range of 0.007 sr⁻¹ at 490 nm (Figure 5). Figure 5 shows the non-normal distribution of R_rs at most wavebands (Shapiro-Wilk normality test, p < 0.05, N = 71) except 412 nm (W = 0.98, p = 0.25) and 443 nm (W = 0.97, p = 0.11) in the satellite dataset and 665 nm (W = 0.98, p = 0.21) and 674 nm (W = 0.97, p = 0.08) in the in situ dataset. The data also showed overall positive skewness, particularly for the satellite dataset where outliers were present in three of the bands, namely, 560, 620 and 665 nm.

Testing for the differences between performance metrics, statistically, we found that significant differences between the median R_rs were shown for some of the bands (Mann-Whitney test, p < 0.05, N = 71) where there was a general shift towards higher reflectance values in the satellite dataset as suggested by the bias metric (Figure 6). This was clear at 674 nm, 681 nm, 754 nm and 779 nm where the median satellite reflectance values were at least double the in situ measurements (e.g., Median R_{rs_ satellite} (674) = 4.57 × 10⁻⁴ sr⁻¹ vs. Median R_{rs_ in situ} (674) = 2.10 × 10⁻⁴ sr⁻¹).

When comparing the satellite with the in situ measured reflectance, we used a set of metrics commonly used in the field. RMSD (a measure of accuracy) increased from 0.0001 to 0.001 sr⁻¹ from the red to the blue parts of the spectrum (Figure 6A), which is small compared to the range in R_rs in the blue (i.e. 0.001 sr⁻¹ at 490 nm; Table 3). Conversely, the MdAPD (a measure of dispersion), was lower in the green (∼30%) than in the red bands (MdAPD(779) = 204%; Table 3). This was similar to strongly negative bias at 400 nm (MdPD (400) = −4%) and 865 nm (MdPD(865) = −16%; Table 3). The rest of the wavelengths exhibited a positive bias ranging between 2.59% and 201% (Table 3). A gradual increase was observed between 510 nm and 779 nm, both in the dispersion and the bias metrics (Figures 6B, C).

Overall, this validation exercise at relatively low reflectance values in coastal waters depicted that full resolution OLCI R_rs using POLYMER produced lower accuracy and higher dispersion at shorter wavelengths than at higher wavelengths compared to underway above water R_rs. To try to disentangle the causes affecting this performance, the matchup datasets were then separated in terms of optical water types and distance to the coast.

The discrepancy metrics between satellite and field measurements of R_rs were investigated separating the matchups by their distance to land. Despite no statistically significant relationship (PERMANOVA, F2 = −0.049, R ² = −0.001, p = 0.969), the influence of varying distance from land on discrepancies was observed and summarised through performance metrics (Figure 11).

The RMSD metric had a higher mean value for measurements in the nearshore waters (RMSD = 0.001 sr⁻¹ for 0–5 km to land) compared to more offshore waters (RMSD = 0.0006 sr⁻¹ for 10–15 km to land; Figure 11A). A clear peak was associated with the 560 nm for the 0–5 km to land group (RMSD (560) = 0.002 sr⁻¹; Figure 11D) coinciding with the peak observed in OWT 6 (Figure 10A). From the location of the matchups (Figure 8), OWT 6 appeared both within the <5 km to land and 10–15 km to land regions. The averaged MdAPD (dispersion) was similar for the three groups, and varied between 64% and 72% across all wavelengths (Figure 11B), with a peak at the 709 nm for waters <5 km to land (MdAPD(709) = 142%) and at 779 nm for the two groups further offshore (MdAPD(779) = 210–221%; Figure 11E). Spectrally, the minimum for this metric was found at 443 nm for 0–5 km (MdAPD(443) = 15%), at 510 nm for 5–10 km (MdAPD(510) = 18%) and at 560 nm for 10–15 km (MdAPD(560) = 12%; Figure 11E). The MdPD metric (Bias) was overall positive (MdPD = 35–46%). This indicates a positive bias of satellite estimates in relation to in situ observations across all distance partitions considered (Figure 11C). When looking at the spectral variation of this metric, however, matchups at 5–10 km from the shore were the only ones revealing a positive bias of the satellite derived reflectance (Figure 11F). At other distances (0–5 km and 10–15 km), bias was variable and sometimes negative.

Previous studies assessing the performance of atmospheric correction processors have highlighted that disparities in results between study areas arise from differences in reflectance ranges, in part due to varying degrees of turbidity (Vanhellemont and Ruddick, 2021).

The coastal areas assessed in the present study can be classed as moderately turbid, with total suspended matter concentrations between 1 and 10 g/m³ (Martinez-Vicente et al., 2010). These turbidity conditions differ substantially from those of recent studies (Mograne et al., 2019; Alikas et al., 2020; Giannini et al., 2021), which included moderately turbid to very turbid water bodies with reflectance values 10 times greater in magnitude (0.02–0.05 sr⁻¹) than the values observed in the present study (0.003 sr⁻¹, Figure 4). The study area, characterised by a narrow range of relatively low reflectances (Figure 4), revealed a positive bias of satellite estimates in relation to in situ observations. Although the present observations are low in magnitude, they are in line with previous measurements in the area (Martinez-Vicente et al., 2013; Jordan et al., 2023). They align with the lower end of values recorded in coastal waters by Giannini et al. (2021) (at 443 nm they report in situ R_rs less than 0.005 sr⁻¹). They are also similar to more offshore waters in the Canadian coast (Vishnu and Costa, 2023).

When comparing the performance metrics for the whole dataset of this study to other studies, similarities arise with regards to the wavelength dependency. The dispersion and bias metrics indicated higher dispersion and bias at the edges of the spectral band range (<490 nm and >674 nm) with a peak in the red region (Figure 6C). The spectral differences in the dispersion and bias metrics are consistent with the metrics obtained by selecting different spatial or temporal windows explored in this study (Figures 3B, C). Lower performance at 400 nm and 865 nm may result from the low R_rs signal, given that there is a relationship between the dispersion metric and the intensity of the signal (Vanhellemont and Ruddick, 2021), and less likely to come from high absorption and backscattering by optically active constituents (Mograne et al., 2019). Larger uncertainties at shorter wavelengths could also be attributed to difficulties with aerosol correction (Soppa et al., 2021). Larger uncertainties at longer wavelengths, however, may instead be attributed to the small R_rs signal from increasing water absorption (Soppa et al., 2021). The positive bias of the satellite derived R_rs at 865 nm in relation to the in situ observations (Figure 6C) was also indicated by the MPD (Mean Percentage Difference) metric in Alikas et al. (2020) despite larger bias being observed (MdPD_present = −16%, MPD_Alikas = −50%).

Conversely, the RMSD metric suggested lower accuracy at shorter wavelengths than other studies (Figure 6A; Mograne et al., 2019; Vanhellemont and Ruddick, 2021). Also, the RMSD depicted strong differences with changes in the spatial and temporal matchup windows (Figure 3). An increase in RMSD (decrease in accuracy) was observed, mainly, at shorter wavelengths (<600 nm) between the strictest matchup selection routine (30 min and 1 × 1 pixel) and the most relaxed one (3 h and 3 × 3 pixels). Different causes have been proposed in the literature. Giannini et al. (2021) raise the fact that POLYMER does not consider the CDOM contribution to the optical signal adequately, leading to uncertainties in shorter wavelengths. This is particularly important in regions where the optical signal is significantly affected by CDOM absorption, e.g., British Columbia a_CDOM (440) 0.23–4.91 m^-1 (Giannini et al., 2021). The more offshore part of our study area is not affected by high CDOM, with a_CDOM(440) 0.05–0.20 m^-1 levels around L4 station (10–15 km from shore, Groom et al., 2009). Higher concentrations of CDOM are, however, expected in nearshore waters that could result in error in the shorter wavelengths. Warren et al. (2019) point out that the curve follows the wavelength dependency of the aerosol optical thickness with particular difficulty in removing aerosol contributions in shorter wavelengths, specifically at 443 nm (Pahlevan et al., 2021). Vanhellemont and Ruddick (2021) propose the under correction of the increasing atmospheric path reflectance between the NIR and the UV to be a potential cause.

Disparity between the present results and other studies can also be explained by the differences in the matchup selection criteria and in the data processing chain, including atmospheric correction methods and in situ data filtering. Not only have previous studies assessed more turbid types of coastal waters, different matchup selection routines and data filtering procedures have been used resulting in possible mismatch between satellite and field measurements. In this study we have selected the matchups very strictly (30 min difference between satellite and in situ observations), however for sun-synchronous satellites such as MERIS or Sentinel-3 OLCI, tidal dynamics are a known cause of bias in ocean colour products, for instance TSM concentrations in estuaries (Eleveld et al., 2014). Appropriate matchup selection criteria are therefore critical according to the spatiotemporal variability of the study area (Concha et al., 2021). It is also worth pointing out that the present study and previous studies cited have used different methods of atmospheric correction, with various versions of POLYMER used as well as different flags or correction methods to target bidirectional effects (Table 4).

Whilst both matchup selection and quality control are crucial for high quality data to support satellite ocean colour validation activities (IOCCG, 2019), strict criteria cause a large proportion of the acquired data to be discarded. Particular attention should be given to the sampling strategy and the validation protocol to obtain high quality ground measurements for accurate measures of performance. Users should follow accepted protocols (IOCCG, 2019; EUMETSAT, 2021).

Few sources in the literature describe the performance of atmospheric correction models while considering optical water types or distance from land (as proxy for adjacency effects) as potential contamination factors (Pereira-Sandoval et al., 2019; Warren et al., 2019; Pahlevan et al., 2021; Hieronymi et al., 2023).

When considering nearshore waters (0–5 km from land), the large RMSD values were present at the 560 nm band and could be a result of either nearshore enhanced TSM scattering (OWT 6) and/or adjacency from neighbouring land (Groom et al., 2009). The latter was suggested by Warren et al. (2019) when assessing the performance of atmospheric correction models for the Sentinel-2 MSI data. In their study, the RMSD at 560 nm increased by 25%–55% between the coastal and inland matchup datasets. With distance from land ranging between 0.2 and 1.2 km, this was associated with the contribution of land adjacency to the atmospheric path reflectance. The under correction of the land adjacency effect could explain the positive bias of the satellite derived reflectance in such waters. This is supported by the decreasing tendency observed in RMSD with increasing distance to land. When considering the RMSD metric, the results of the current study therefore point to the adjacency effect from neighbouring land as being a potential issue for Sentinel-3 OLCI data up to approximately 5 km from land. The dispersion (MdAPD) and bias (MdPD) metrics do, however, not show similar tendencies. This could be the result of other factors such as environmental conditions at the time of sampling such as aerosol optical depth, Sun and observation geometry. Increased efforts in the in situ R_rs sampling collocated with aerosol optical properties measurements within that distance to the shoreline would help to reduce uncertainties and improve confidence on Sentinel-3 OLCI to increase its use in water quality monitoring applications. Concerning the relationship between OWT and performance statistics, the accuracy (RMSD) and the dispersion (MdPD) measures provide the worse results, pointing to worse R_rs retrievals in more optically complex waters. This points to an application of OWT to flag coastal areas where satellite retrievals should be used with caution.

Finally, several limitations could be attributed to this analysis, in particular related to in situ data collection and analysis of radiometric data from unsupervised equipment. Non-ship superstructure, uncertainties associated with instruments and possible environmental conditions affect all measurements collected in this manner and have been discussed in specifically designed intercomparisons (e.g., Tilstone et al., 2020). In addition, some variations could come from the processing approach, which has been discussed extensively elsewhere (Groetsch et al., 2017). Ideally, optical instruments should be collected in tandem to provide closure, however, this is not always possible in the context of satellite validation from opportunistic, unsupervised platforms, which, in contrast, provide large amounts of data with the possibility to filter them. Progress could be made in the direction of improving the automatic orientation of the platform. In this way, data would be collected in the best conditions.