Removal of the glint effects from satellite imagery for accurate retrieval of water-leaving radiances is a complicated problem since its contribution in the measured signal is dependent on many factors such as viewing geometry, sun elevation and azimuth, illumination conditions, wind speed and direction, and the water refractive index. To simplify the situation, existing glint correction models describe the extent of the glint-contaminated region and its contribution to the radiance essentially as a function of the wind speed and sea surface slope that often lead to a tremendous loss of information with a considerable scientific and financial impact. Even with the glint-tilting capability of modern sensors, glint contamination is severe on the satellite-derived ocean colour products in the equatorial and sub-tropical regions. To rescue a significant portion of data presently discarded as “glint contaminated” and improving the accuracy of water-leaving radiances in the glint contaminated regions, we developed a glint correction algorithm which is dependent only on the satellite derived Rayleigh Corrected Radiance and absorption by clear waters. The new algorithm is capable of achieving meaningful retrievals of ocean radiances from the glint-contaminated pixels unless saturated by strong glint in any of the wavebands. It takes into consideration the combination of the background absorption of radiance by water and the spectral glint function, to accurately minimize the glint contamination effects and produce robust ocean colour products. The new algorithm is implemented along with an aerosol correction method and its performance is demonstrated for many MODIS-Aqua images over the Arabian Sea, one of the regions that are heavily affected by sunglint due to their geographical location. The results with and without sunglint correction are compared indicating major improvements in the derived products with sunglint correction. When compared to the results of an existing model in the SeaDAS processing system, the new algorithm has the best performance in terms of yielding physically realistic water-leaving radiance spectra and improving the accuracy of the ocean colour products. Validation of MODIS-Aqua derived water-leaving radiances with in-situ data also corroborates the above results. Unlike the standard models, the new algorithm performs well in variable illumination and wind conditions and does not require any auxiliary data besides the Rayleigh-corrected radiance itself. Exploitation of signals observed by sensors looking within regions affected by bright white sunglint is possible with the present algorithm when the requirement of a stable response over a wide dynamical range for these sensors is fulfilled.
Ocean colour remote sensing data provided by the NASA's Sea-viewing Wide Field-of-view Sensor (SeaWiFS, on board the SeaStar satellite) and Moderate-resolution Imaging Spectroradiometer (MODIS, on board the Terra and Aqua satellites), ESA's MEdium Resolution Imaging Spectrometer (MERIS, on board the Envisat satellite), and ISRO's Ocean Colour Monitor (OCM-1 and OCM-2, on board the IRS-P4 and Oceansat-2 satellites respectively) are a vital resource for operational forecasting and oceanographic research, and related applications in the global oceans. With the advent of these sensors, the prospects of better algorithms to enable the interpretation of ocean colour in Case 2 waters have particularly improved. A few examples of some of the ways that ocean colour data have been used in various studies include monitoring the spatial and temporal variability of algal blooms (instrumental in characterizing variability of marine ecosystems and is key tool to investigate how marine ecosystems respond to climate change and anthropogenic perturbations), monitoring coastal marine pollution and river plumes, understanding global carbon budgets and climate change impacts (Shanmugam et al., 2013; Shanmugam, 2011). The largest sources of error for retrieval of ocean radiances in Case 2 waters (optically complex) are often attributed to the treatment of aerosol and sunglint radiances in the atmospheric correction procedure (Shanmugam, 2012; Rakesh Kumar and Shanmugam, 2014; Wang and Bailey, 2001). Since simultaneous in-situ measurements of atmospheric optical properties are not available in the most of the cases, atmospheric correction of ocean colour imagery usually relies on the satellite-derived data alone (Ruddick et al., 2000). The former problem has been successfully addressed in a recent study (Rakesh Kumar and Shanmugam, 2014) which uses the Rayleigh corrected radiance to estimate and extrapolate the aerosol radiance rather than using the aerosol models and relative humidity (Gordon and Wang, 1994). The problem of sunglint is particularly acute under a wind-roughened sea surface condition and one of the greatest confounding factors limiting the quality and accuracy of satellite data (Kay et al., 2009; Zhang and Wang, 2010). This often results in the periodic black portions on swaths found in images of the ocean colour products of these regions (Ottaviani et al., 2008).
Sunglint is a phenomenon caused by the specular reflection of the incident light from the sun to the sensor. The region affected by sunglint may vary from a single disk (image of sun), on a perfectly flat and clam surface, to a distinctive widespread spatial pattern (due to reflection into a wide range of angles) on the wind-roughened sea surface due to the reflection by a large number of wave facets (Zhang and Wang, 2010). This region often extends to several hundred kilometers, with associated reflectance factor greater than 0.2 (Hagolle et al., 2004). The effect of sunglint is highest at the sub-solar point due to the decrease in angle between the sun and sensor (Wang and Bailey, 2001). The sunglint pattern varies with respect to the wind speed and direction, sensor geometry and illumination conditions. The viewing geometry, relative orientation of the sensor's viewing angle, solar zenith angle, and the slope of the water surface along with the sea surface roughness governed by the wind speed and wind direction, play a significant role in determination of sunglint (Cox and Munk, 1954).
Most ocean colour sensors are designed to capture the radiances over a given dynamic range (minimal
threshold corresponding to the saturation limit) in a given band. The minimal threshold defines the
lowest intensity of radiances to be detected by the sensor in a given band, whereas the saturation
limit is the maximum radiance to which the sensor can respond. To reduce this contamination, some
sensors (e.g., SeaWiFS and OCM-2) use a steering mechanism to change their viewing angle by
20
The extent of the glint-contaminated region and its contribution to the radiance is generally computed from the Cox and Munk model (Cox and Munk, 1954) with the input of the sea-surface wind speed. Several recent studies have improved this model based on redefined sea surface statistical parameters. For instance, Shifrin utilized the Richardson number to link the stability of the atmosphere–water interface to the sea surface roughness (Shifrin, 2001). Ebuchi and Kizu re-estimated the slope statistics with a more general data set from a radiometer onboard Geostationary Meteorological Satellite (GMS) and ancillary data from space-borne scatterometers, which resulted in a narrower distribution for the glint which was similar to the Cox and Munk model (Ebuchi and Kizu, 2002). Bréon and Henriot used the Polarization and Directionality of Earth's Reflectance (POLDER) (Deschamps et al., 1994) with NASA scatterometer for wind speed and direction to quantify glint contamination by redefining the slope statistics (Bréon and Henriot, 2006). It should be noted that these methods utilized certain ancillary data, which are difficult to obtain in real time. Doerffer et al. used a Neural Network (NN) to estimate glint radiance from the radiative transfer calculations (Doerffer et al., 2008). The efficiency of this algorithm depends on the training of the NN and there is a chance of producing irrelevant glint radiance due to the synthetic output from the radiative transfer equations. Steinmetz et al. used a spectral matching approach for modelling atmosphere and sunglint using all available spectral bands and matching with the spectrum to be corrected (Steinmetz et al., 2011). Another approach was proposed by Shanmugam (2012) which empirically related the glint radiance to the Rayleigh corrected radiance. Recently, Kutser et al. (2013) attempted to correct glint contamination by fitting a power function on the measured (in-situ) reflectance values from the blue and NIR (near infrared) region. Many of the models developed for satellite applications have been reviewed and evaluated for correction of the sunglint contamination effects in satellite ocean colour data (Kay et al., 2009; Zhang and Wang, 2010).
A model for sunglint correction presently used in the SeaDAS processing system was proposed by Wang
and Bailey (Wang and Bailey, 2001), which is based on the glint radiance computation from Cox and
Munk (1954) with inputs from the solar and viewing geometries, sea-surface wind speed and direction,
and the estimated aerosol optical thickness. This method determines the normalized glint radiance
The major drawback of these models for sunglint correction is the absence of ancillary information such as wind speed, wind direction, sea surface slope and other parameters. The objective of this paper is to develop an alternative robust sunglint correction algorithm that is entirely dependent on the satellite-derived products alone. The new algorithm (hereafter referred to as “New Glint Correction – NGC” algorithm) takes into account the absorption by clear water as the ancillary data which is almost constant for a wide variety of waters. The performance of NGC algorithm is tested for several MODIS-Aqua images of the Arabian Sea and its results are compared with those of the default model available in the SeaDAS processing system (called as SeaDAS Glint Correction (SGC) algorithm for brevity). The applicability of NGC algorithm over the global oceans is further discussed.
Satellite ocean colour sensors measure the spectrum of sunlight reflected from the ocean–atmosphere
system at several visible and near-infrared (NIR) wavebands. About 80–90 % of the signal
recorded at the top-of-the-atmosphere (TOA) is contributed by the atmosphere through the process of
scattering by molecules and particles (aerosols) and the remaining signal is the desired
water-leaving radiance (
The default model used in the SeaDAS processing system computes the sunglint radiance as a function of the sea-surface wind speed, wind direction and solar and sensor geometries. The SeaDAS model is built on the Cox and Munk model which ignores the sky radiance, which raises the question of the validity of the Cox and Munk distribution for the global oceans and for all weather conditions (Shifrin, 2001). Further the sunglint correction is limited to the region at the edge of the sunglint, where the contribution of sunglint is below a predetermined threshold, and beyond this threshold it deteriorates the quality of the ocean colour products or simply creates the flag in areas with strong sunglint effects. The sunglint contamination is particularly evident in both atmospheric and ocean products (e.g., the aerosol optical thickness and water-leaving radiances) (Wang and Bailey, 2001). Since most of the ancillary information needed for sunglint correction are not measured at the time of each satellite overpass, it is necessary to develop a robust algorithm that relies only on the satellite-derived information for removal of the sunglint effects. Thus, the new algorithm does not use the ancillary data (such as wind speed and direction) but entirely depends on the Rayleigh corrected radiance itself. This way of sunglint correction is efficient and has wider applicability regardless of the water types and weather conditions. Figure 1 shows a step-by-step procedure to compute the sunglint radiance and remove its effect on satellite ocean colour imagery. The procedure is elucidated in the following sections.
The glint radiance spectra do not have any features, but are a continuously increasing function of
wavebands (Doerffer et al., 2008; Shanmugam, 2012). Addition of a strictly increasing sunglint
radiance spectrum through the wavebands would decrease the slope of two consecutive glint
contaminated radiance bands. It implies that the lower the ratio of two Rayleigh corrected
wavebands, the higher the glint contamination. This condition can be quantified by defining a glint
ratio as
The
To identify the pixels contaminated by sunglint and determine the extent of the glint-contaminated
region, a threshold value of
Due to the absence of 2130
The NGC algorithm uses the normalized absorption coefficient of water
The glint spectral function is a constant spectrum, which defines the basic behavior of glint and
its spectral shape is altered depending on the glint intensity in a given pixel. The multiplication
of other terms (defined in the next step) depending on the magnitude of glint to
In satellite imagery, water acts as a background which absorbs the radiance strongly at longer
wavelengths (Pope and Fry, 1997). The absorption by water alters the shape of the glint spectrum,
which can be determined as the difference between glint spectral function and normalized absorption
by water as given below,
The efficiency of the NGC algorithm is assessed based on the digital interpretation and assumption of spatial homogeneity of glint corrected products. To evaluate its efficiency in-situ observation data (with and without glint contamination) are used. The validation results are also compared with those of the default glint correction procedure in SeaDAS software.
Several MODIS Level-1A data (Local Area Coverage data) of the Arabian Sea (available at
The selected MODIS–Aqua L1A data are converted to the calibrated and scaled L1B (Level 1-B)
top-of-atmosphere radiance (
To analyze the performance of NGC method, concurrent in-situ data and MODIS–Aqua data are used for validation. The in-situ data used in this study is a part of the NASA bio-Optical Marine Algorithm Data set (NOMAD) (Werdell and Bailey, 2005), consisting of 35 in-situ and MODIS-Aqua matchups in regions away from the sunglint or without glint contamination and 4 matchups with glint contamination. Though the number of matchups with glint contamination is small, it is sufficient to show the validity and behavior of the present algorithm.
Saturation issue is well addressed with an example of MODIS-Aqua imagery (A2004026202500) from the
Pacific Ocean (Fig. 4a), where the Rayleigh corrected radiances (
The results of NGC algorithm are compared with those of the SGC model, the Fig. 5a is a typical
example of the Rayleigh-corrected radiance image (667
To better visualize these issues, the false colour composite images
(R–G–B
The new
Figure 7g and h shows quantitative comparisons of the derived chlorophyll concentration for cases
with and without glint correction (Fig. 7b, d and f). It is evident that when glint correction is
ignored in the atmospheric correction procedure, the chlorophyll values are abnormally high (beyond
120
To examine the consistency of the NGC algorithm, both the SGC and NGC algorithms were tested on six
other MODIS-Aqua images where the sunglint contamination was obvious. The sunglint mask was
purposely turned off before applying these algorithms to all glint contaminated
regions. Figures 8a–c and 9a–c display the Rayleigh-corrected images with different glint patterns
ranging from highly concentrated (e.g., 22 February 2013 and 5 March 2013) to wide-spread glints
(e.g., 17 February 2009 and 26 February 2013). Note that the glint corrected radiance data produced
by the SGC model still contain high level of residual glints surrounding the glint mask and in
regions of the concentrated sunglint patterns (Figs. 8d–f and 9d–f), whereas the glint corrected
radiances from the NGC algorithm appear to be much more reasonable (Figs. 8j–l and 9j–l). The
residual sunglint radiances and glints produced by aerosols and clouds, clearly seen in Figs. 8e and
9d, likely increase the aerosol radiances to be used in the subsequent atmospheric correction
procedure. The aerosol correction procedures (Gordon and Wang, 1994; Ruddick et al., 2000;
Shanmugam, 2012; Rakesh Kumar and Shanmugam, 2014) use the NIR bands which are the most affected
bands when residual glint radiances come into play (Gordon, 1978). These high radiances in the NIR
region are assumed to be due to aerosols and extrapolated to other visible bands, which ultimately
results in highly erroneous
Since the northern Arabian Sea is surrounded by Thar Desert in the east, the Rub-Al-Khali (Arabian Desert) in the west and Iranian Desert in the north, the glint effects produced during the transport of these aerosols are simply ignored by the SGC model leading to overestimating of the chlorophyll concentration. Similarly, the derived chlorophyll concentration is high for moderate bloom waters affected by the cloud-induced glint (indicated by white arrow in Fig. 10a and b). The effect of residual glint contamination due to sunglint, aerosols and clouds is already reported to bias the derived aerosol optical thickness high and to overestimate the chlorophyll (Wang and Bailey, 2001). For clear oceanic waters surrounding the glint mask in the southern part of the Arabian Sea, the chlorophyll concentration derived with the SGC model is reduced significantly compared with the results of the NGC algorithm.
To explore the possibility and see the applicability of the NGC algorithm for rescuing the discarded data, we also extended our analysis to exploit signals observed by MODIS-Terra sensor looking within regions of the Arabian Sea affected by high glint (not shown for brevity). The MODIS-Terra instrument is designed to operate over a wide dynamical range to capture low water-leaving radiance and high surface radiance from land. When examined the performance of the NGC algorithm, it was found that a large portion of such glint contaminated region is successfully recovered by the NGC algorithm. Thus, it can also be applied to similar regions affected by bright sunglint when the requirement of a stable response over a wide dynamical range for the new generation ocean colour sensors is fulfilled.
To validate the results obtained by the NGC method and to compare its results with those from the
SGC method, 4 glint contaminated matchups and 35 matchups from other regions are used. Many data
showing the highest magnitude in the water-leaving radiance signal were collected from the coastal
regions. To examine whether the water signals (high radiance) without glint contamination are
corrected for glint, the in-situ matchups are chosen from the Florida Keys and Bay of Fundy which
are dominated by suspended sediments (Fig. 11a and b). As expected, the SGC method underestimates
the water-leaving radiance data in both relatively clear waters and turbid waters. The
Figure 11e and f shows true colour composite images of
MODIS-Aqua showing the Bay of Fundy with sampling
locations affected by the glint effects. These points
are present in the low glint region and are subjected to glint correction by
both the glint correction methods. Figure 11g shows
improvement in the resulting products from the NGC over the SGC method when
related to the in-situ
Sunglint correction is an important step in atmospheric correction of satellite ocean colour imagery which minimizes or removes the sunglint effects to derive more accurate water-leaving radiances. The effects of other glint effects produced by intense aerosols and clouds are also significant especially in low- and mid-latitude regions. Existing models are largely dependent on ancillary data (e.g., sea-surface wind speed and direction, solar and viewing geometries, and aerosol optical thickness) which are either unavailable for every satellite overpass or insufficient for accurate glint correction. Further if the ancillary data such as wind speed and direction and solar and sensor geometry are not synchronized with each other, it would often lead to the incorrect approximation of sea surface slope and hence inaccurate and erroneous ocean colour products required for further analyses. Regardless of these drawbacks, there are also obvious problems with these methods ignoring other glint contributing elements in the imagery.
To overcome these problems, a new algorithm for sunglint correction (NGC) has been developed and
implemented in the SeaDAS processing system along with a recent aerosol correction method (Rakesh
Kumar and Shanmugam, 2014). The NGC algorithm is novel because it entirely depends on the
satellite-derived product (
The performance of NGC algorithm when tested on several MODIS-Aqua images acquired over Arabian Sea
waters in the presence of sunglint and complex aerosols and clouds is exceptionally good. Comparison
of the water-leaving radiances and chlorophyll products generated with and without glint correction
demonstrates the necessity of glint correction by the NGC algorithm. Further validation conducted
based on the concurrent in-situ and MODIS-Aqua data confirms that the NGC algorithm yields
significantly low errors when compared to the SGC model. The later model often leads to
significantly reduced
This work is supported by grants from the ISRO-IITM cell (Number: ICSR/ISRO-IITM/OEC/13-14/149/PSHA). We gratefully acknowledge the Ocean Biology Processing Group of NASA for the distribution of the MODIS data and NOMAD in-situ data and the development and support of the SeaDAS Software.
Error statistics for the SGC and NGC algorithms with glint and without glint conditions.
Schematic flow diagram depicting the new glint correction algorithm.
The false colour composite images
(R–G–B
The water-leaving radiance spectra (for the defined transect)
The false colour composite images
(R–G–B
The false colour composite images
(R–G–B
Comparison of the chlorophyll concentration images derived from the ABI algorithm using the SGC and NGC products.