OLI measures TOA radiance in the visible, NIR, and short-wave infrared (SWIR) bands, at a spatial resolution of 30 m. Compared to its predecessors, OLI includes a new coastal/aerosol band (435–451 nm) in addition to the traditional blue (452–512 nm), green (533–590 nm), and red (636–673 nm) bands. The addition of a new band, together with enhanced spectral coverage and radiometric resolution, enables improved observation of water bodies from space and the ability to estimate the concentration of atmospheric aerosols for AC [16,24,25] (note that aerosol estimation for AC by the coastal band is done over land). Compared to data from existing global ocean colour missions, the higher spatial resolution has the potential to make important contributions to ocean colour remote sensing, such as separating and mapping in-water constituents in coastal waters [16,25]. Although OLI signal-to-noise ratios (SNRs) (Table 1) are generally lower than those of heritage ocean colour sensors such as SeaWiFS or the Moderate Resolution Imaging Spectroradiometer onboard Aqua (MODISA), its enhanced SNR across all bands compared to the past Landsat missions improves OLI’s ability to measure subtle variability in near-surface conditions and ultimately make OLI products a valuable source of data for ocean colour studies [16,18].

To validate the performance of the AC processors applied to the OLI data, we acquired 122 cloud-screened and fully quality-controlled Level 2.0 AERONET-OC in-situ measurements of normalized water-leaving radiance (nL_w) for 14 AERONET-OC sites, including 12 coastal sites (i.e., Galata Platform, Gloria, GOT Seaprism, Gustav Dalen Tower, Helsinki Lighthouse, Long Island Sound Coastal Observatory (LISCO), Martha’s Vineyard Coastal Observatory (MVCO), Thornton C-Power, USC Seaprism, Venise, WaveCIS Site CSI-6, Zeebrugge-MOW1) and two lake sites (Lake Erie, Palgrunden) (Figure 1). AERONET-OC, managed by NASA’s Goddard Space Flight Center (GSFC) [26], is a sub-network of the AERONET federated instrument [27,28]. Although OLI and AERONET-OC have somewhat different spectral bands, a set of comparable bands centred at 441 nm, 491 nm, 551 nm, and 667 nm can be used for cross-comparison purposes. Thus, AERONET-OC data were collected in four spectral bands centred at 441, 491, 551, and 667 nm, for comparison with OLI’s four visible bands, centred at 443, 483, 561 and 655 nm. Note that there is a slight difference, i.e., ±1 to ±3 nm, in all four bands among measurements at different sites. As R_rs is not directly available from the AERONET-OC sites, the normalized water-leaving radiances (nL_w, W m⁻² sr⁻¹) were divided by the top-of-the atmosphere (TOA) solar irradiance (F₀) [29] to obtain R_rs.

To obtain the in-situ R_rs data needed to test AC procedures for OLI, we performed a match-up exercise between the AERONET-OC measurements and OLI data as follows: (i) Using the OLI metadata database file provided by the United States Geological Survey (USGS), python code was created to automatically retrieve all Landsat 8 OLI scenes and the contemporaneous AERONET-OC data (from the AERONET-OC website) that were within a ±30-min time window of Landsat 8 overpass times, for the April 2013 to May 2017 timeframe (note that a strict time window of ±30 min, which reduces the number of match-up pairs, was used to limit the error introduced by water movement between satellite and AERONET-OC observations, and ensure the quality of match-ups). This yielded a text file containing a total of 122 match-ups with coincident satellite and in-situ data for 14 AERONET-OC sites (Figure 2), as well as information on aerosol optical thickness, solar zenith angles (SZA), and wind speed for each match-up. All corresponding OLI scenes were subsequently bulk-downloaded using Landsat-util, a tool to automatically find and download multiple Landsat 8 scenes. Some of the 122 OLI scenes visibly contained a non-negligible amount of specular reflection off the sea surface (sunglint) in the area of the AERONOET-OC site. As not all AC algorithms have the capacity for sunglint correction, to obtain realistic and comparable R_rs across all AC methods, scenes with visible specular reflection were excluded, leaving 69 of the original 122 scenes; (ii) Following the approach of Bailey and Werdell (2006) [30], a regional subset of Landsat 8 data was generated for each scene by (1) extracting a 7 × 7 pixel window centred on the location of the AERONET-OC site, and (2) removing the centre 3 × 3 pixels from that window to limit the effect of noise from the site’s superstructures and shadows cast by it; (iii) For SeaDAS, remaining low-quality pixels were then removed by employing the internal SeaDAS exclusion flags, which include flags for land, clouds, cloud-shadow, ice, stray light, low nL_w (555), high viewing zenith angle (>60°), high sunglint, and high TOA radiance. Scenes without any unflagged pixels were eliminated from the match-up exercise, leaving 54 scenes for comparison with AERONET-OC data. Similarly, internal ACOLITE exclusion flags were used to remove low-quality pixels for ACOLITE, which left 56 scenes for AERONET-OC comparison. The 54 scenes remaining after applying the SeaDAS exclusion flags also passed ACOLITE’s exclusion flags, and were therefore used for further analysis. To ensure an unbiased inter-comparison, we included pixels with negative (Table A3) and zero R_rs retrievals from all methods; (iv) We then obtained the per-band median R_rs values of the unflagged pixels from each Landsat regional subset for final comparison with in-situ data, and used the median AERONET-OC measurements collected within the ±30-min window of the Landsat 8 overpass to represent in-situ match-ups.

Scatter plots showing the estimated (OLI) and observed (AERONET-OC) R_rs values for each match-up are presented in Figure 3, and summary statistics are tabulated in Table 2. There are clear differences between the water-based and land-based AC algorithms, with SeaDAS and ACOLITE outperforming ARCSI and LaSRC in all metrics for all bands, with only one exception (slope for R_rs(482)). Between the two water-based methods, SeaDAS outperforms ACOLITE in every metric for all bands, with the exception of slopes for R_rs(482), R_rs(561), and R_rs(655). SeaDAS had RMSEs close to zero between 0.0013 and 0.0005 1/sr across all four wavelengths) demonstrating a high degree of similarity between in-situ and OLI-estimated R_rs with OLI data processed through SeaDAS (Table 2). A comparison of RMSE results of the per-band difference with and without band adjustments (Figure A1) shows that SeaDAS was the AC method most sensitive to spectral band differences, with the largest improvement of band adjustment occurring in the 655 nm channel. Spectral Angle values obtained for all algorithms showed that SeaDAS and ACOLITE have the highest similarity with in-situ R_rs spectra (ARCSI: 0.46, ACOLITE: 0.27, LaSRC: 0.53 and SeaDAS: 0.20).

The overall performance of SeaDAS reveals that the NIR-SWIR aerosol correction option can yield satisfactory results in low-to-moderately turbid waters. A possible reason for this is that the aerosol correction scheme, constructed following Ahmad et al. (2010) [42], was based on aerosol data obtained mainly from AERONET-OC sites [46]. Comparison of R² values among all methods shows that the lowest and most diverse values are in the 443 nm wavelength, with values between 0.05 and 0.84. For LaSRC in particular, the regression line for the comparison in the 443 nm wavelength deviates very much from the 1:1 line, yielding a poor R² and slope (Figure 3). The poor correlation and low RMSE (R²: 0.05, slope: 0.23) are mostly a result of the large discrepancies between the observed and estimated R_rs for the Zeebrugge-MOW1 site, where mean R_rs was underestimated by ~70%. This is the largest underestimation by any method, across all sites. For this site, which is one of the most turbid sites in the AERONET-OC network, mean observed in-situ R_rs in the red band is 0.0155 sr⁻¹, making it the only site with R_rs(655) one order of magnitude greater than the mean value of ~0.001 sr⁻¹ observed for all 14 AERONET-OC sites. This level of turbidity is common for this site, which is located only ~3.65 km from the coastline and receives sediment-rich water inputs from nearby rivers, as also noted by Vanhellemont and Ruddick (2015) [47] and was clearly visible in additional scenes excluded during the match-up exercise. Note that the TOA radiance data were used ‘as is’ without optimizing the vicarious calibration gains (as computed by Pahlevan et al. (2017a) [15]), which might further improve R_rs retrievals. Similarly, none of the AC methods with different configuration capabilities that might improve performance were optimized as there was no optimal setting that would work for all cases considered in this paper.

Comparison of mean estimated and observed R_rs at each AERONET-OC site (Figure A2) shows that all algorithms except SeaDAS generally overestimate R_rs across all wavelengths, with the largest and smallest overestimation occurring in the 443 and 665 nm wavelengths, respectively. This is further supported by the RMSE and bias results (Figure 4). The largest errors (RMSE) and overestimations (bias) are observed in the 443 nm wavelength, probably due to the strong atmospheric scattering in this band. ARCSI has the largest overall positive bias in this wavelength, and indeed the highest overestimation at each site. However, its R_rs results across all wavelengths at the Zeebrugge-MOW1 site compare well with R_rs estimates from SeaDAS and ACOLITE. This suggests that ARCSI has a low sensitivity to the high concentrations of suspended sediments that dominate this site, as reported by De Maerschalck and Vanlede (2013) [48]. This may also serve as an indication that ARCSI can better deal with turbid conditions than can LaSRC, which underestimated R_rs by ~45% at this site. The failure of LaSRC for this band is likely due to the fact that it is part of the bands used for the aerosol inversion scheme [23].

The best performance from LaSRC across all bands is at Lake Erie (with two match-ups) where all other AC algorithms except ARCSI also have the best match with in-situ R_rs. LaSRC outperforms other algorithms for the first three bands. Percentage difference values are 4.5%, −3.2%, and −0.3% for the first three bands, respectively. For ACOLITE and SeaDAS, the corresponding values are 28.8/−20.7%, 13.6/−12.8%, and −0.4/−6.4%. Indeed, at this site LaSRC has the best R_rs estimate of all methods in the 561 nm channel, while ACOLITE has the best R_rsestimate in the 655 nm channel, with a percentage difference of −1.3%, whereas SeaDAS and LaSRC are −18.2% and −30%, respectively. Similar to the estimated R_rs by LaSRC in Lake Erie in the 561 nm channel, LASRC-derived R_rs(561) also agrees well with that of in-situ at Zeebrugge-MOW1 site; the percentage difference here is −0.25%. For SeaDAS and ACOLITE, these values are 6.7% and 4.1%, respectively. LaSRC also outperforms other AC algorithms in the 443, 482, and 655 nm channels at the GOT-Seaprism site, with only one match-up. Percentage differences between estimated and in-situ R_rs are 8.8%, 11.1%, and −40.8%, respectively. For ACOLITE and SeaDAS, the corresponding values are 87.7/−79.3%, 60.9/−46.3%, and 96.4/−127.6%. This is the only site where SeaDAS uncharacteristically underestimates R_rs across all four bands. While any conclusion is tentative as GOT Seaprism (2014-026) only has one match-up, the poor performance of SeaDAS here is as a result of algorithm failure (very low R_rs in 443, 482, and 561 nm wavelengths, and negative R_rs in 655 nm wavelength), which can be attributed to conditions such as residual effects from cloud shadow or overcorrection for aerosol contribution in one or more visible band.

One possible reason for the generally poor performance of ARCSI can be the aerosol contribution removal which relies on estimates from (1) dense dark vegetated surfaces, based on the assumption that reflectance of vegetated pixels is sufficiently dark, and a linear relationship between reflectance in the SWIR and blue bands or (2) dark pixels in the blue band, based on the assumption of an invariant aerosol concentration over the entire scene. However, these assumptions can easily be violated as finding a vegetated pixel that satisfies this condition may be difficult in scenes acquired over coastal waters and AOT variations may be sufficiently large such that adjoining pixels may have significantly different AOT. For the blue bands in particular, per-scene AOT estimates may lead to erroneous retrievals. For LaSRC, the generally poor performance may be due to the use of land-based pixels for aerosol estimation. In addition, for LaSRC, retrieving accurate R_rs estimates over water requires the presence of a considerably large land area adjoining the water pixels. The majority of the AERONET-OC sites used in this study only have relatively small nearby land surfaces. This may help explain the few instances of good performance near land masses (e.g., for the Lake Erie and Zeebrugge MOW-1 sites).

To understand the impact environmental factors may have on R_rs retrieval errors from the water-based AC methods, we investigated the influence of three variables: AOT(869), SZA, and hourly wind speed. These three variables are known to influence R_rs retrievals [50,51], e.g., AOT(870) and SZA have been found to reduce the quality of water-leaving radiance derived from SeaWiFS and MODIS sensors [52]. Figure 5a–c illustrates the error (x^est − x^mea) for each match-up point as a function of each environmental parameter, for each AC method. Negative values imply that an algorithm underestimated the observed R_rs value, and vice versa. We used tests of the statistical significance (two-tailed, α = 0.05, critical value = 0.2262) of the individual Pearson correlation coefficients to guide this analysis (Pearson correlation coefficients were used to show the strength of the relationship as they have been widely used in similar studies). While ACOLITE consistently overestimated R_rs in the 443 and 482 nm bands, as also noted in [53], errors for both SeaDAS and ACOLITE were not significantly influenced by AOT (Figure 5a, no statistically significant correlations). However, SZA was significantly and positively correlated with R_rsretrieval errors from SeaDAS for all four bands (i.e., r = 0.495486743, 0.483529464, 0.253699366, and 0.427793365, respectively), and from ACOLITE for the 443 and 482 nm bands (i.e., r = 0.239717715 and 0.228792001, respectively) (Figure 5b). A similar pattern was evident for wind speed, which was significantly positively correlated with R_rs retrieval errors from SeaDAS for all bands except band 3 (for which a positive correlation was present, but not statistically significant) and from ACOLITE for the 443 and 482 nm bands (Figure 5c). We further examined the significance of the relationship between each environmental variable and errors across all wavelengths and found that SZA and AOT(869) were significant for SeaDAS in the first three and first two wavelengths, respectively, while ACOLITE was only influenced by wind speed in the 443 nm channel. These patterns, while generally causing only small errors in R_rs retrieval, may guide further developments of both AC methods to make them more robust across the range of environmental conditions.