Improving low-quality satellite remote sensing reflectance at blue bands over coastal and inland waters

Highlights

• A spectral shape based algorithm is developed to estimate Rrs(41×) and Rrs(443).

• The algorithm can greatly increase the satellite R_rs data quality and quantity.

• It allows for more accurate and increased number of valid ocean color retrievals.

Abstract

The satellite remote sensing reflectance (R_rs(λ)) at two short blue bands (410 or 412 nm and 443 nm) are prone to large uncertainties in coastal and inland waters, prohibiting algorithms from generating reliable ocean color products associated with these bands. In this study, we developed an algorithm to estimate R_rs(41×) and R_rs(443) when the satellite R_rs(λ) in blue bands suffer from large uncertainties. The algorithm first determines the R_rs(λ) spectral shape from the satellite-measured R_rs(λ) values at three wavelengths of 48× (486, 488, or 490), 55× (547, 551, or 555), and 67× (667, 670, or 671) nm. The algorithm then derives R_rs(41×) and R_rs(443) from the estimated R_rs(λ) spectral shape with algebraic formulations. We assessed the algorithm performance with satellite (SeaWiFS, MODISA, and VIIRS-SNPP) and in situ R_rs(λ) matchups from global waters. It is shown that the uncertainties of estimated R_rs(41×) and R_rs(443) are substantially smaller than the original satellite products when applicable. Besides, implementation of the algorithm contributes to a significant increase in the number of utilizable R_rs(41×) and R_rs(443) values. The algorithm is relatively stable and is best applicable to the satellite R_rs(λ) spectra for which the R_rs(48×) and R_rs(55×) measurements are subject to small uncertainties. The demonstrations support the application of the blue-band estimation algorithm to a wide range of coastal waters.

Introduction

Ocean color satellites provide a means of collecting remote sensing reflectance (R_rs(λ)) on spatial and temporal scales unattainable by conventional in situ measurements. The R_rs(λ) data allow for the derivation of important biological and biogeochemical properties of the upper oceans, such as phytoplankton chlorophyll-a concentration (Chl) (Hu et al., 2012; McClain, 2009; Wang and Son, 2016), phytoplankton light absorption (Wang et al., 2017; Wei and Lee, 2015), colored dissolved organic matter (CDOM) absorption (Mannino et al., 2014; Wei et al., 2016a; Yu et al., 2016), and primary production (Behrenfeld et al., 2005; Lee et al., 2015a). For reliable estimation of these bio-optical properties, it is vital to obtain accurate R_rs(λ) product from satellites over a wide spectral domain, particularly at the blue bands of 41× (410 or 412) nm and 443 nm.

The ocean color satellites, including the decommissioned Sea-viewing Wide Field-of-view Sensor (SeaWiFS, 1997–2010) and the operational Moderate Resolution Imaging Spectroradiometer (MODIS) onboard the Aqua satellite (MODISA, 2002–present) and Visible Infrared Imaging Radiometer Suite (VIIRS) onboard the Suomi National Polar-orbiting Partnership satellite (VIIRS-SNPP, 2011–present), retrieve R_rs(λ) from the radiance measured at the top of atmosphere (TOA) through atmospheric correction (AC). Operational AC algorithms usually start with the selection of an aerosol model by assuming null contribution of water-leaving radiance (L_w(λ)) at the near-infrared (NIR) or shortwave infrared (SWIR) bands (Antoine and Morel, 1999; Gordon and Wang, 1994; Wang, 2007; Wang and Shi, 2007). This “black-pixel” assumption works well in open oceans, but is faced with some difficulties in coastal regions using the NIR and SWIR approach. The strongly absorbing aerosols, for instance, can be prominent in the coastal waters near anthropogenic sources of fossil-burning products, soot, and smog, or under the influence of dust transport. The weakly or strongly absorbing aerosols are hardly discriminable from radiance measurements in the NIR/SWIR domain but quite distinctive at short blue wavelengths (IOCCG, 2010). That being said, the standard AC schemes cannot correctly estimate the strongly absorbing aerosols and often fail to yield robust R_rs(λ) products at the blue wavelengths in many coastal regions due to lack of aerosol vertical distribution information (IOCCG, 2010; Kahn et al., 2016). As a consequence, and expectedly, the satellite-derived R_rs(41×) and R_rs(443) products are prone to large uncertainties in the coastal waters (Antoine et al., 2008; Feng et al., 2008; Hlaing et al., 2013; Qin et al., 2017; Zibordi et al., 2009). Apart from atmospheric correction, the large uncertainty in satellite R_rs(41×) product is also related to system vicarious calibration and instrument degradation (Franz et al., 2018). The relatively large uncertainties at blue bands and the subsequent lack of utilizable data at these two bands add to the difficulties of satellite ocean color applications in coastal regions (Mouw et al., 2015).

To address the blue-band R_rs(λ) quality issues, a great deal of effort has been invested. Gordon et al. (1997) and Chomko and Gordon (2001) developed an AC procedure to quantify the strongly absorbing aerosol contribution under dust-dominating conditions. Application of their methods, however, remains impractical because the vertical distribution of absorbing aerosols in the atmosphere is not known a priori (Banzon et al., 2009; IOCCG, 2010). In another case study, Hu et al. (2000) tested a nearest-neighbor method to account for the aerosol contribution at the NIR bands, with reasonable results. Variants of the standard AC schemes also emerged. Oo et al. (2008) reported improved atmospheric correction for turbid coastal waters by placing constraints onto R_rs(412) within their AC procedures. For highly absorptive waters, He et al. (2012) proposed null water-leaving radiance at 412 nm so as to make a guess of the aerosol contribution at that band, which was then used for atmospheric correction. In analogy, Wang and Jiang (2018) forced the negative R_rs(410) values from VIIRS-SNPP to zeroes to estimate the aerosol contributions and ultimately to obtain improved R_rs(λ) products.

Besides the above contributions, there exists another line of methodology to deal with the blue-band R_rs(λ) measurements. Basically, this category of methods attempts to “correct” the problematic blue-band R_rs(λ) data. In this regard, Ransibrahmanakul and Stumpf (2006) tested a power-law function to account for the spectral artifacts caused by the absorbing aerosols in the northeast U.S. coasts. D'Alimonte et al. (2008) derived multi-linear regression coefficients from satellite and in situ R_rs(λ) matchups and further applied those coefficients to estimate R_rs(412) for the same region. Undoubtedly, all strategies mentioned above have demonstrated varying degrees of success, specific to the sensors, dataset, or environments, in improving blue-band R_rs(λ) quality. Yet, it remains an open question to implement such algorithms as a universal approach to global coastal and inland waters where the satellite R_rs(41×) and R_rs(443) data are of low quality.

The blue-band R_rs(λ) data are needed for many bio-optical retrievals in the upper water columns. Among others, R_rs(41×) and R_rs(443) are indispensable for many semi-analytical algorithms to retrieve the phytoplankton and CDOM absorption coefficients from multiband satellite R_rs(λ) (IOCCG, 2006; Lee et al., 2002; Wei et al., 2019; Werdell et al., 2013). R_rs(443) data may also be required for Chl estimation in both open and coastal oceans (Hu et al., 2012; O'Reilly et al., 1998; Wang and Son, 2016). Use of R_rs(41×) and R_rs(443) data with large errors can result in unrealistic retrievals for the water bio-optical properties (Werdell et al., 2018) and biased Chl products (Hyde et al., 2007). It is also risky that merging of such satellite products may generate spurious trends in ocean color time series. In order to best interpret the large-scale and long-term ocean color retrievals, this long-standing problem associated with the blue-band satellite R_rs(λ) measurements needs to be resolved.

In this study, we propose a spectral shape based algorithm to estimate R_rs(41×) and R_rs(443) when the satellite R_rs(λ) spectra at blue bands are subjected to large uncertainties. Such estimation will help ultimately enhance the bio-optical retrievals with the ocean color algorithms requiring R_rs(41×) and/or R_rs(443) data. In the following, we first describe the evaluation data from SeaWiFS, MODISA, and VIIRS-SNPP, the proposed algorithm, and relevant analyses (Section 2). Next, we demonstrate the algorithm performance in estimating R_rs(41×) and R_rs(443) by comparison with in situ matchup measurements (Section 3). Finally, we discuss the algorithm applicability, uncertainty, potential impacts on satellite bio-optical retrievals, and potential applications to satellite image processing (Section 4).