Challenges and New Advances in Ocean Color Remote Sensing of Coastal Waters

Knowing that coastal areas concentrate about 60% of the world's population (within 100 km from the coast), that 75-90% of the global sink of suspended river load takes place in coastal waters in which about 15% of the primary production occurs, the ecological, societal and economical value of these areas are obvious (fish resources, aquaculture, water quality information, recreation areas management, global carbon budget, etc). In that context, precise assessment of suspended particulate matter (SPM) concentrations and of the phenomena controlling its temporal variability is a key objective for many research fields in coastal areas. SPM which encompasses organic (living and non-living) and inorganic matter controls the penetration of light into the water and brings new nutrients into the system, both key parameters influencing phytoplankton primary production. Concentrations and availability of SPM are also known to control rates of food intake, growth and reproduction for various filter feeder organisms. Phytoplankton is highly sensitive to environmental perturbations (such as nutrient inputs, light, and turbulence). The abundance, biomass and dynamics of phytoplankton in coastal areas therefore reflect the prevailing environmental conditions and represent key parameters for assessing information on the ecological conditions, as well as on the coastal water quality. Because phytoplankton is highly sensitive to environmental perturbations [1], its distribution patterns and temporal variability represent good indicators of the ecological conditions of a defined region [2, 3]. Coastal waters also host complex ecosystems and represent important fishery areas that support industry and provide livelihood to coastal settlements. The food chain in the coastal ocean is generally short (especially in upwelling systems, having as low as three trophic levels) whereas the open ocean food web presents up to six trophic levels [4]. As a result, when compared to the open ocean, a relative lower fraction of the primary production gets respired in the coastal ocean while a higher fraction reaches the uppermost trophic level (fish) [5] or is exported to adjacent areas (coastal or open sea)...

of the primary production gets respired in the coastal ocean while a higher fraction reaches the uppermost trophic level (fish) [5] or is exported to adjacent areas (coastal or open sea).
The potential large fluxes of carbon in the coastal ocean underscore its significance to the global carbon cycle (around 14% of total global ocean production, along with 80-90% of new production). However, the extreme heterogeneity of coastal environments in terms of ecosystems structure and functioning, and the current large uncertainties remaining at global scale on the latter features makes difficult general assessment of net autotrophic or net heterotrophic character of these marine waters. The results of the Shelf Edge Exchange Processes Study raised the question: "Do continental shelves export organic matter?". This question still remains without any definitive answer. Bauer and Druffel [6] suggested that dissolved organic carbon (DOC) and particulate organic carbon (POC) inputs from ocean margins to the open ocean interior may be greater than inputs of recently produced organic matter derived from the open surface ocean by more than one order of magnitude. While particles from terrestrial origin are primarily deposited in the coastal region, DOC is considered as the main path for transporting terrestrial and phytoplankton derived carbon into the deep ocean [7]. The accurate assessment of the temporal variability of SPM, POC, DOC, and Chl in coastal areas over long time periods and along with physical forcing parameters may therefore provide valuable insights for improving our knowledge on the biogeochemical cycle prevailing in coastal ecosystems.
Due to the high variability of the physical and biogeochemical processes occurring in coastal areas, traditional approaches based on oceanographic cruises and in situ time series, although essential, are very time-consuming, expensive and sometimes uncertain to yield meaningful results on the studied phenomena, especially at a large synoptic scale. In this context, remote sensing of biological and physical parameters is a very powerful tool for performing large scale studies. Satellite data are not as accurate as in situ measurements and are limited to the surface layer. However, the latter limitations are largely compensated by the spatial and temporal coverage offered by the satellite observations. In situ data remain obviously necessary to validate the satellite products, in terms of absolute value, but also in terms of temporal variability in areas where long in situ time series are available. After a short introduction of the main and critical aspects of ocean color radiometry, the specific issues relative to the satellite observation of ocean color in complex natural systems such as coastal marine waters will be specifically addressed. Illustrations of current algorithms development will be then provided as well as examples of the different end-users products currently available from ocean color remote sensing in coastal waters. At last, the new challenges and concepts allowing for a better observation of coastal ecosystems from satellite ocean color observation will be discussed.

General concepts of the ocean color radiometry
The interaction of light field within the visible part of the spectrum (i.e. 400-700 nm) with the different optically significant constituents of sea water (water molecules, salt, particulate and organic dissolved matters) modifies the color of the water. These spectral variations, which bring qualitative and quantitative information about the water constituents, can be recorded by a passive radiometric sensor onboard a satellite platform. However, the conversion of the radiometric signal measured by the sensor at the top of the atmosphere (TOA) to the end-user parameters is not straightforward. The contribution of the reflected photons at the air-sea interface as well as the contribution of the atmosphere should first be removed from the top of atmosphere measured signal to assess the water-leaving radiance, L w (λ), which is the only radiometric parameter encompassing useful information on the water masses composition (λ represents the wavelength of light in nanometer, nm). The removal of the atmospheric path radiance represents the most challenging part of the ocean color atmospheric correction procedure [8]. This signal, which includes Rayleigh and aerosols components, can contribute to up to 90% of the TOA signal depending on λ, the geometry of illumination and observation, the aerosol optical thickness, and the water leaving signal [9]. The importance of the latter task is particularly crucial when considering the level of accuracy required for being able to derive accurate estimates of the desired water components from L w (λ). As a matter of fact, the radiometric accuracy required for the Sea-viewing Wide Field-of-view Sensor (SeaWiFS) is, for instance, of 5% for the visible spectral domain for absolute radiance values, and of 2% for relative values (i.e. reflectances) [10].
The retrieval of the different inherent optical properties (IOPs) and biogeochemical components of the water from the water-leaving radiance spectral values are performed through biooptical algorithms. Considering that about 90% of the L w (λ) signal originates from the upper layer of the water column (the so-called first attenuation layer), the various biogeochemical or optical variables which can be retrieved from L w (λ) are assumed to be weighted averaged parameters within this upper layer [11]. The thickness of this oceanic layer in the visible part of the spectrum typically varies from less than one meter (as in turbid waters, or in the red part of the spectrum) to about 60 meters (as for oligotrophic waters in the green), depending the amount of optically significant constituents within the water mass and the measured light wavelength [12]. Since the proof-of-concept Coastal Zone Color Sensor (CZCS) mission in 1978, satellite-derived ocean color has been routinely interpreted, with increasing accuracy, in terms of upper-ocean chlorophyll concentration, Chl [13]. The great availability of ocean color satellite data, as well as the need of complementary product (additionally to Chl) for validating global ocean biogeochemical models have greatly stimulated the rapid development of numerous inverse methods over the last decade for retrieving a large variety of information ( Figure 1). For instance, empirical [14] and semi-analytical methods [15,16] have been developed for specifically assessing the particulate organic carbon (POC) over the global ocean. In the same way, the various inherent optical properties are now routinely retrieved from space [17]. Among these different IOPs one may cite the coefficients of absorption by phytoplankton (a phy ) and colored detrital matter (a cdm ), as well as the particulate backscattering coefficient (b bp ). The spectral shape of b bp [18] and a cdom [19], which, respectively, provide relevant information on the particulate and dissolved pools, are also currently accessible over open ocean waters. Note that the retrieval of the bio-optical parameters can also be performed simultaneously to the atmospheric corrections using, for instance, neural network approaches [20]. Additionally, information about phytoplankton community composition [21][22][23] and size distribution of marine particles [24], essential for a better understanding of the oceanic carbon cycle, are also available from satellite imagery. With the ability to measure the Sun-stimulated phytoplankton Chl fluorescence, information on the physiological state of algal populations could also be derived from space as illustrated from the Moderate Resolution Imaging Spectroradiometer (MODIS) and MEdium Resolution Imaging Spectrometer (MERIS) ocean color missions. Detailed description of the different atmospheric and bio-optical algorithms for the observation of ocean color from space can be found in the extensive literature provided in the International Ocean Colour Coordinating Group web site [25].  [15] and [26], respectively.
Thanks to its continuous recording of 13-years of data, the Sea-viewing Wide Field-of-view-Sensor ocean color mission (SeaWiFS) has allowed great scientific advances in our understanding of the open ocean biological productivity [13]. In contrast, while ocean color observations allowed the achievement of numerous researches and new discoveries over open ocean waters, the algorithm development over coastal waters is still not mature enough to be applied routinely as it is currently performed for offshore waters, despites numerous recent advances. Some of the most important challenges remaining for developing accurate ocean color products in coastal waters are described in the following sections.

Challenges to overcome for ocean color radiometry in coastal areas
Remote sensing of ocean color in coastal areas is impacted by their intrinsic environmental features: vicinity of land over which photons can be reflected back to the sensor (the so-called adjacency effects), land inputs from rivers discharges and coastal washing, shallow waters which promote resuspension of the unconsolidated sediments as well as bottom reflected photons captured by the sensor in clear shallow waters, breaking waves which generate bubbles, etc. All these phenomena make difficult and challenging the assessment of bio-optical marine components from space by ocean color sensors.
The main problems to overcome in order to derive accurate observations of coastal areas from ocean colour remote sensing are associated with the presence of highly diffusing suspended mineral particles as well as high concentrations of particulate organic matter (phytoplankton, detritus) which may bias atmospheric corrections and impact bio-optical algorithms [17,27]. Another issue largely encountered for the exploitation of ocean colour data in coastal areas is the presence of clouds. In presence of highly turbid waters, cloud-free pixels are sometimes erroneously classed as clouds, leading to a loss of data. At last, but not least, crucial and mandatory validation exercises are greatly complicated to perform due to the extreme spatial heterogeneity of these areas. These different issues will be addressed in the following sections.

Cloud masking
Ocean color retrieval by satellite borne sensors is in principle only possible for a clear (cloud free) atmosphere, and cloud-contaminated pixels have to be removed from the images before any ocean color processing takes place. The cloud masks developed in the frame of ocean color missions are based on the assumption that the marine reflectance in the near-infrared (NIR) is equal or close to zero. While open ocean waters can effectively be considered as black in the NIR, except in presence of highly scattered material (due for instance to offshore river plumes, or coccolithophore blooms), this is not the case for coastal waters. Indeed, the level of signal observed at the top of the atmosphere over clouds and coastal areas may be similar due to the presence of suspended sediments in the water surface ( Figure 2). Cloud-free pixels are then sometimes classed as clouds and excluded from further processing, leading to a loss of data in these areas. Specific cloud masking algorithms have to then be developed over coastal waters.

Atmospheric corrections a. General description of the atmospheric correction
The use of satellites to monitor the color of the ocean requires effective removal of the contribution of the atmosphere (due to absorption by gasses and aerosols, and scattering by air molecules and aerosols) to the total signal measured by the remote sensor at the top of the atmosphere, L toa : the so called "atmospheric correction" process. As shown in Figure 3, the signal measured by the remote sensor is the sum of different components. The top-ofatmosphere radiance, L toa , can be expressed as [8]: where L r is the radiance due to the scattering of the air molecules (Rayleigh scattering), L ra is the multiple interaction term between molecules and aerosols, L a , the radiance due to the scattering by aerosols, L wc , the radiance due to the whitecaps on the sea surface, and L g the specular reflection of direct sunlight off the sea surface. t(λ) and t 0 (λ) are the diffuse transmittances of the atmosphere from the sun to the surface and from the surface to the sensor, respectively, T (λ) is the direct transmittance from the surface to the sensor, and L w (λ) the waterleaving radiance [28].
The terms L wc , L g , and L r can be determined during a pre-processing. So the atmospheric correction algorithm needs to solve the following equation: The goal of the atmospheric correction process is to accurately determine L w (λ) from the spectral measurements of L cor (λ). For that purpose L A (λ) has to be quantified.
The classic methods for removing the contribution of the atmosphere to the total measured signal exploit the high absorption by seawater in the red and near-infrared (NIR) spectral regions. In open seawater, i.e. where generally chlorophyll-a concentration and related pigments and co-varying materials (like detritus) determine the optical properties of the ocean, seawater can be considered to absorb all light in the NIR so that the signal observed by the satellite sensor in this spectral domain is assumed to be entirely due to the atmospheric path radiance (L A ) and sea surface reflectance [8]. This is not always the case when considering turbid waters (generally coastal optically complex waters). In these waters phytoplankton pigment and detritus, as well as suspended sediment, contribute to NIR backscatter. The resulting NIR water-leaving radiances introduce two sources of error into the standard procedure to remove the aerosols. First, the total aerosol reflectance is overestimated as some of the total radiance (L toa ) at two NIR bands (λ 1 and λ 2 ) comes from the seawater. Second, as the absorption and scattering properties of seawater change from λ 1 to λ 2 , the selection of the appropriate atmospheric model is affected, causing an error in the extrapolation of L A toward shorter wavelengths. As a result, L A is overestimated at all bands with increasing values at shorter wavelengths, even possibly leading to negative water-leaving radiances in the blue bands [30].
The difficulty of the atmospheric correction is that the atmosphere contributes to 80-90% of the total top-of-atmosphere signal at the blue-green wavelengths (400-550 nm) and the atmospheric path radiance L A significantly varies and cannot be easily approximated. Moreover, the assumption of having a non-zero L w in the NIR bands is not valid for turbid waters Figure 3. Illustration of several different light pathways in the atmosphere: a) The light path of the water-leaving radiance, b) the attenuation of the water-leaving radiance, c) scattering of the water-leaving radiance out of the sensor's field of view (FOV), d) sun glint (reflection from the water surface), e) sky glint (scattered light reflecting from the surface), f) scattering of reflected light out of the sensor's FOV, g) reflected light is also attenuated towards the sensor, h) scattered light from the sun which is directed toward the sensor, i) light which has already been scattered by the atmosphere which is then scattered toward the sensor, j) water-leaving radiance originating out of the sensor FOV, but scattered toward the sensor, k) surface reflection out of the sensor FOV which is then scattered toward the sensor. L w Total water-leaving radiance. L r Radiance above the sea surface due to all surface reflection effects within the FOV. L p Atmospheric path radiance. This Figure is adapted from [29].

Description of existing algorithms
The algorithms previously cited can be grouped into four different categories:

Assignment of hypothesis on the NIR aerosols or water contributions
For these algorithms, assumptions are made on the aerosol models and/or the bio-optical models. Hu et al. [35] estimate the aerosol model in clear waters on the studied image and extrapolate this information over the coastal waters of the image. Ruddick et al. [40] impose a constant value of the ratio of L A and L w in the NIR. The aerosol constant NIR ratio value is determined for each image of interest taking into account only the clear waters as in [35]. The constant L w (NIR) ratio is defined from radiative transfer modeling and is assumed to be constant for any location for a given satellite sensor. This ratio is valid for low and moderate turbid waters. Using other wavelengths (SWIR, see next subsection), Wang et al. [46] developed a similar scheme than the one proposed by [35] for the Lake Taihu (China). The aerosol properties are obtained at the pixel-by-pixel level. The derived aerosol model is averaged in the middle of the lake, where the waters are not turbid. Using this mean aerosol model, the SWIR atmospheric correction algorithm use a single SWIR band and can be carried out for the entire lake for estimating L w . These assumptions allow to use the scheme defined in [8] for the dark pixel hypothesis. These constant ratios only allow to determine L A (NIR) and to remove this contribution to L cor .

Use of the shortwave infrared bands
For moderate and very turbid waters, it is possible to use the shortwave infrared bands (from the MODIS-AQUA sensor). The principle of this type of algorithm is very similar than the one of Gordon and Wang [8], except that the aerosols models are determined for wavelengths between 1200 and 2100 nm instead of NIR wavelengths. For this wavelengths range, L w can be considered negligible and the dark pixel atmospheric correction processing can be applied.
The difficulty with SWIR algorithms is the low signal to noise ratio for the SWIR bands of the MODIS-AQUA sensor. The algorithm shows limitations for low and moderate turbid waters, and therefore needs to be coupled with a NIR atmospheric correction algorithm [47]. More specifically, the latter authors have proposed to use the standard NIR-based atmospheric correction unless a turbidity index [48] exceeds a predefined threshold and, thereby, triggers the use of a SWIR-based correction where two SWIR bands are used instead of two NIR bands. The relevant MODIS-Aqua SWIR band pair is 1240 and 2130 nm. Shi and Wang [48] concluded that these bands satisfy the black pixel assumption in moderately (1240 nm) to extremely (2130 nm) turbid waters. This combined "NIR-SWIR" atmospheric correction approach has been evaluated extensively in several geographic locations [34,47,48] and validated against an independent, globally-distributed in situ data set [45]. Moreover, the difficulty of using the SWIR bands relies on the selection of the aerosol model. Analysis of the Ångström exponent retrievals suggests that the SWIR approach cannot retrieve the correct aerosol spectral dependence when the atmospheric path radiance is low. The latter feature tends to produce spectrally flat models, thus underestimating the aerosol reflectance in the shorter wavelengths.
In fact, at low signals the SWIR approach often predicts aerosol spectral dependence that is beyond the range of the current aerosol models, as it would be associated with unrealistically large aerosol size distributions. Conversely, the SWIR approach tends to overestimate aerosol reflectance at higher aerosol loads, thus resulting in negative L w (λ) [49].

Correction/modelling of the non-negligible L w (NIR)
This type of algorithm aims to correct the non-negligible L w (NIR) signal using iterative methods coupled with a bio-optical algorithm [30,31,38,43,46]. These algorithms need to define a first value of L w (NIR) which can be directly fixed, or calculated from a first guess of Chl or IOPs. This can be done by using either default values of the parameters or a dark pixel atmospheric correction procedure such as [8]. Then, a NIR water-leaving reflectance model, developed from in situ measurements, is used to calculate L w (NIR). The problem encountered with these algorithms is that the empirical bio-optical models limit their applicability to waters that are similar to those over which these empirical models were developed.

Coupled ocean-atmosphere inversion
As the ocean and the atmosphere cannot be decoupled in turbid waters (L w (NIR) ≠ 0), an appropriate solution for developing atmospheric correction scheme would be to couple the two systems and to inverse them together. In practice, this can be done using either artificial neural networks (NN) [41] or optimization techniques [20,31,36,41,[50][51][52][53][54].
The more common type of NN is the Multi-Layer Perceptron (MLP) [55]. Thanks to this association of elementary tasks, an MLP is able to solve complex inverse problems. The specificity of an MLP depends on the topology of the neurons (number of layers, numbers of neurons on each layer) and on the connection weights w ij from a neuron j of a layer to the neuron i of the next layer. The w ij weight values are computed through a training phase, using a training dataset. Once the training phase is finished, the MLP will only do algebraic operations, which leads to faster computations, which is very convenient for satellite applications. When using a neural network for the atmospheric correction phase, the top-of-atmosphere radiance is usually directly inverted for estimating the water-leaving radiance in the visible as well as the aerosol optical properties. Another option is to use optimization techniques. The principle of those algorithms is based on the iterative minimization of a dedicated cost function by adjusting relevant atmospheric (aerosol optical thickness, and Ångström coefficient) and oceanic (Chlorophyll concentration, IOPs) parameters which are the control parameters [20,56]. There are some difficulties that one encounters when using this type of techniques: the parameters initialization [57], definition/use of a direct model, calculation of the adjoint code and the method for the parameters adjustment [56].

2.
Are these algorithms accurate enough?
As shown previously, there are several methods to estimate L w from L toa in complex coastal waters. Few round-robin comparisons have been made to evaluate different atmospheric correction algorithms from in situ data [58,59]. These studies showed that the different algorithms highly over-estimated L w in the blue and red and were quite accurate for the intermediate wavelengths providing similar estimates of L w ( Figure 5). But the retrieval accuracy in the blue is still too low, with negative L w values in some cases, which prevent to use these bands for bio-optical applications.  [58] and [59], respectively).

Other issues
Besides the presence of water leaving radiance signal in the near infrared, other issues can be encountered during the atmospheric correction processing. The three major problems are the bottom albedo for shallow waters, the adjacency effects and the presence of absorbing aerosols.
The bottom effect corresponds to the light reflected off the bottom of a water body, providing the water is sufficiently shallow and clear. The influence of the seafloor on the colour of water depends on the depth of the water body, the clarity of the water masses, the type of optical substances present in the water, as well as on the type of substrate composing the seafloor. The bottom may be rocky or sandy, and may be covered, partially or fully, by a variety of benthic organisms (e.g., algae, mollusks). All of these factors will influence the manner in which bottom effects will act on the colour of the water, as seen by a remote sensor [27]. Figure  6 shows the particulate backscattering coefficient estimated at 490 nm using [60]. The area in light pink and white correspond to negative R rs (and then b bp ) values induced by this bottom effect. If this effect is not corrected before or during the atmospheric correction, the estimation of the R rs (b bp ) will be biased and will lead often to obtain negative values. This is particularly true when studying clear shallow waters such as lagoons or the Coral Reefs Bareer in the east coast of Australia [61][62][63]. The so called "adjacency effect" refers to the process by which a photon, reflected from a surface adjacent to a targeted pixel, is scattered by the atmosphere between the sensor and the target, blurring the sharp boundary between the land and the coastal water [64][65][66]. In coastal areas, a fraction of the light reflected by the land can reach the sensor. Modeling the reduction of image contrast when the atmospheric turbidity increases requires large amounts of computing time. The adjacency effects impart a significant bias to the model for retrieval of aerosol values because of the high contrast between land and ocean in the near infrared spectral region. First, the aerosol optical thickness is overestimated by classic algorithm because of the increase of the atmospheric signal. Then the wavelength dependence is modified, leading to a wrong extrapolation of the aerosol optical thickness in the blue-green region. The variation of the aerosol optical thickness between the near infrared and the red on a transect from the coast can be a good indicator of contamination by adjacency effects of the aerosol product over land and an indicator of this effect on satellite imagery [66].
The third major problem is the presence of absorbing aerosols from urban and desert origin in coastal regions. The aerosol models used in the classic atmospheric correction are all nonor weakly absorbing [8,9,67]. Absorbing aerosols have a lower aerosol radiance at the shorter wavelengths than any of the models used in the current atmospheric correction process. The presence of absorbing aerosols would, thus, have a similar effect to the aforementioned incorrect model selection, though likely to a more significant degree in the blue portion of the spectrum. The presence of such strongly-absorbing aerosol can only be inferred in the visible, where multiple scattering is high. In such situations, one can no longer derive water properties by a two-step process -atmospheric correction followed by a bio-optical algorithm to estimate water properties. To solve this issue, only few methods have been currently proposed [20,32,51,53,68,69]. A promising way to deal with the presence of absorbing aerosols would be to use spectral optimization method which allows the atmospheric and oceanic properties to be retrieved simultaneously.

Bio-optical algorithms
Satellite remote sensing of ocean colour is a very powerful tool for the management of resources and activities of continental shelf waters. Besides estimates of Chl, which represents the historical parameter investigated from optical remote sensing technique, more recent developments have allowed the retrieval of a variety of bio-optical (e.g. phytoplankton and detrital matter absorption or particulate backscattering coefficients) and biogeochemical (e.g. POC, phytoplankton size distribution and community composition) parameters with now a satisfying accuracy at global scale [70]. The latter developments have been supported by the improvement of optical sensors, from the first (CZCS launched in 1978) to the most recent ones (e.g. SeaWiFS, MODIS or MERIS), both regarding the radiometric data quality and spatial resolution.
An accurate assessment in the coastal marine domain of the various optical and biogeochemical parameters previously cited and now available for oceanic waters still represents an important challenge since it would provide relevant and innovative insights on the dynamics and functioning of these complex and highly diverse ecosystems. The development of innovative bio-optical algorithms is for instance cruelly needed for precisely identifying the occurrence of specific phytoplankton species. As a matter of fact, a precise monitoring of harmful algal bloom (HAB) events which strongly impact the functioning of the concerned coastal ecosystems and which dynamics are currently relatively poorly constrained, represents a crucial field of investigation. Various algorithms have, in that sense been, proposed in the recent years for computing indices specifically dedicated to the identification of red tides events [71][72][73]. Among the various important challenges to overcome in coastal waters, one particularly crucial concerns the estimation of dissolved and particulate organic carbon stocks and dynamics which would strongly condition our ability to quantify the impact of coastal ecosystems domain on the global carbon cycle. The assessment of POC and DOC concentrations in coastal waters however still represents an open field of investigation even though some empirical regional algorithms have been already proposed for estimating DOC content from CDOM absorption coefficient retrieval in some coastal regions [74][75][76].
The assessment of the latter bio-optical and biogeochemical parameters from space depends on the relevance of the bio-optical algorithm required to infer marine water IOPs or biogeochemical products from satellite L w (λ) measurements. This key step for interpreting the satellite signal in these marine regions is complicated by various specific issues. Indeed, coastal waters are highly dynamic systems at both temporal and spatial scales being subjected to a variety of physical (tides, current, fronts, turbulence...) and environmental forcings (interactions with terrestrial ecosystems, phytoplankton blooms, re-suspension...). This high hydro-biological variability of these waters logically induces a strong optical complexity ( Figure 7) making more difficult the interpretation of the 'color' in these environments. Specifically, unlike in case 1 waters, substances significantly active from an optical point of view (i.e. phytoplankton, nonalgal particles and colored dissolved material) usually vary independently in time and space and have specific optical properties which may vary over wide ranges [77][78][79]. These specificities of coastal waters create different issues. First, a strong dispersion around the average general laws generally used in open ocean waters for linking the radiometric measurements to bio-optical (absorption and backscattering coefficients) or biogeochemical parameters (e.g. chlorophyll a). The latter feature prevents the use of Case 1-like general relationships and emphasizes the crucial need to develop specific inversion schemes allowing to take into account for the peculiar optical characteristics of the coastal marine domain [27,80,81]. Second, standard algorithms used for estimating ocean colour such as Chla concentration in oceanic waters are based on reflectance ratios in the blue and green spectral domain [27]. In coastal waters, the significant contribution of dissolved and particulate non-algal material to the optical budget at these latter wavelengths complicates our ability to detect phytoplankton pigment optical signal and often induces a strong failure in the accuracy of these classical formulations ( Figure 8).
Accurate assessment of the different in water bio-optical components from ocean colour measurements in coastal areas are therefore largely controlled by: (i) our ability to understand and to account for the origin of the variability in the radiometric signal and (ii) the realism of the parameterizations used between the inherent optical properties (IOPs) and the biogeochemical component (BC). Numerous challenges still remain [17,27] in that sense, however some progresses have been recently performed by developing inversion approaches specifically dedicated to face the various issues encountered in coastal waters (see sections 3.3 and 3.4).

Validation
Validation of ocean color products (i.e. L w (λ), IOPs, and biogeochemical parameters) should theoretically be performed from in situ measurements acquired simultaneously to the satellite overpass over the same location. However, these ideal conditions are rarely encountered and specific protocols should be applied [82,83]. For instance, it is in practice recommended to use a 2-hour time window applied to the satellite overpass time at the measurement site to select  satellite imagery. Second, the match-up procedure extracts a 3-by-3 pixel satellite image box with the middle pixel closest to the measurement site. The mean value of the desired parameter over the box is calculated for each image. Lastly, an atmospheric spatial uniformity criterion is applied, based on a prescribed coefficient of spatial variation, defined as the ratio of the standard deviation to the mean pixel value within the selected image box. This match-up exercise is made more complicated for coastal waters where typical high frequency physical and biological processes cause high temporal variability and strong spatial heterogeneity ( Figure 9). In situ measurements are generally performed from oceanographic vessel or instrumented fixed buoy from which the spatial heterogeneity of the satellite pixel can not be necessarily taken into account. Typical pixel size of ocean color missions is about 1x1 km 2 at nadir, even if higher spatial resolution data (about 250x250 m) are now accessible for some specific wavelengths (i.e. MODIS), or over the whole spectrum but only on request, that is not routinely (i.e. MERIS-FR). Thanks to the very recent development of autonomous underwater vehicles such as gliders, equipped with optical and radiometric sensors, the spatial heterogeneity of a given pixel could now be sampled in situ. However, much more researches are needed to reach the level of accuracy needed for ocean color validation, from measurements collected using these new platforms.

New and futures developments of ocean color remote sensing of coastal areas 4.1. Cloud masking
Prior to any application of atmospheric correction and bio-optical algorithms a cloud masking algorithm should be applied over the satellite observed area. For that purpose, different algorithms which differ by their inherent assumptions and the wavelengths used have been developed [85][86][87][88][89]. The standard algorithms to process the image acquired by SeaWiFS [85] and MODIS [89] use a constant threshold value at 865 nm (SeaWiFS) and 869 nm (MODIS) over which the pixel is considered as cloud. Wang and Shi [86] present a more sophisticated algorithm based on the fact that water is almost totally black in the shortwave infrared (SWIR) wavelengths. The latter authors have also proposed an alternative algorithm for ocean color sensors which do not allow for measurements in the SWIR bands (such as SeaWiFS and MERIS). Basically, this algorithm is based on the assumption that the spectral variability in the NIR is lower for clouds than for water. Based on the observation that water is generally spatially more homogeneous than clouds, the cloud mask for the POLarization and Directionality of the Earth's Reflectances (POLDER) sensor uses a threshold on the spatial variability in 865 nm [87]. To be able to automatically process and exploit the long time series of SeaWiFS ocean color images over coastal areas, a recent algorithm has been developed for the SeaWiFS sensor [88]. This algorithm is mainly based on the spectral differences of the Rayleigh free TOA signal between cloud and turbid waters. Performance of different algorithms for the very turbid waters of the Amazon delta is provided in Figure 10.

Atmospheric correction
As shown in Figure 5, the accuracy of the existing atmospheric correction algorithms is still not enough satisfying, especially when considering the blue and red spectral bands. It is therefore still necessary to develop innovative schemes to decrease the errors at these bands. As mentioned previously, one of the most promising algorithm development suggests the use of the shortwave infra-red bands (SWIR) for estimating the atmospheric path radiance [44,47]. However, because of the low signal-to-noise ratio of the MODIS-AQUA SWIR bands, which induces uncertainties in the extrapolation of the aerosol models from the SWIR to the NIR and visible, and also due to the fact that none of the future ocean color sensors (OLCI, GOCI-II, OCAPI, S-GLI) will possess the SWIR bands, it is necessary to find other ways to improve the atmospheric correction procedure. Another promising option would be to consider the shorter wavelengths, i.e. in the ultraviolet bands [90]. For highly productive waters with high amount of colored dissolved organic matter, the water-leaving reflectance at 412 nm can be used to constrain the aerosol model retrieval considering that L w (412) is relatively low being negligible when compared to L w in the visible and NIR bands. Considering the latter feature, UV bands could be used to determine the aerosol model. The major disadvantage of using the UV for the atmospheric correction is the assumption of a negligible L w for these bands which are important for the retrieval of CDOM. This assumption needs to be further validated by increasing the amount of dedicated in situ data in the future.
Another way for improving atmospheric correction schemes in coastal waters would be to further develop the use of the formulations which consider a definition of the NIR bio-optical algorithms, by adding constrains on the relationships between L w at several wavelengths. Several relationships already exist. For instance, Wang et al. [46] proposed a new atmospheric correction algorithms for the GOCI sensor using a relationship between L w (765) and L w (865). However, this algorithm needs in a preliminary step to calculate the diffuse attenuation coefficient, K d (490) for estimating L w (765) from the MODIS-AQUA sensor using the SWIR atmospheric correction algorithm. This step might add errors and complexity for developing such algorithms. This emphasizes also the need for developing more direct approaches. One solution would be to mix a new formulation of the NIR similarity spectrum approach [40,91] such as defined in [59,92] with the algorithm developed by [31] which has been demonstrated to be robust in low but also very turbid waters. Figure 11 presents evaluation of the similarity spectrum [40,91] and formulation of Doron et al. [92] against in-situ data for moderately and highly turbid waters. We can see that the constant R rs (NIR) ratios is not valid for those waters, notably for highly turbid waters. These ratios could be integrated inside a current atmospheric correction algorithm for constraining the inversion. Another way is to find linear or polynomial relationships between two R rs ( Figure 12).

Bio-optical algorithms: The classification approach vs. regional approaches
Different approaches have been considered in the recent years for developing inversion algorithms in order to face the issue represented by the strong optical complexity and heterogeneity of coastal waters (see section 3.3). Two contrasted approaches based on a geographical or optical partition of the coastal waters will be considered here, both aiming to constrain the dispersion around bio-optical R rs (λ)-IOPs-BC relationships and presenting their own advantages and requirements.
The first approach, which is the more usual one, consists in focusing on the range of optical variability specific to a defined coastal area by developing local or regional, usually empirical inversion algorithms [93][94][95]. This way of dealing with the optical complexity of coastal waters, although relatively convenient to implement (as soon as a reasonable amount of in situ measurements are available for a defined region) might present different limitations in its application. Indeed, while such regional algorithms may reduce the variability in the IOPs-BC relationships, they are highly dependent on the dataset used for their development. In other terms, the accuracy of regional relationships mostly depends on their ability to account for the natural variability of the water masses optical properties occurring in a defined region. However, a relationship defined for a given area is likely to vary over time (i.e. at the daily scales, seasonal or inter-annual) since it depends strongly on the combined action of various environmental forcing (phytoplankton blooms, in suspension, leaching of the sides, riverine inputs and advection of the offshore waters....). Furthermore, considering the high dynamics of coastal ecosystems it might be presumably challenging to capture, even for a defined region, the many high-frequency processes affecting bio-optical relationships. Finally, in the scope of applications of remote sensing techniques in coastal waters over large spatial scales (i.e. global), the use of a regional approach (by definition non-exportable geographically) would consist in performing a collection of algorithms developed on a mosaic of coastal areas. This seems both difficult to implement in practice and would inevitably lead to ignore some coastal areas currently not covered by ad hoc in situ measurements.
Another way to deal with the optical heterogeneity of the coastal domain is to specifically consider the optical diversity of these waters within the algorithm development procedure. In practice, this alternative approach to regional methods is based on optical classifications, which aims at grouping waters with similar optical features and develops optically adapted algorithms for each water class. Such a definition of optical water types implicitly assumes that different coastal regions can present similar optical characteristics of the marine components Figure 11. Reflectance ratio at 765 and 865 nm according to [91] (dashed line) and [92] (grey polygon) for (a) moderately and (a) highly turbid waters [59]. Circles and triangles are for data collected in the Eastern English Channel/North Sea and the French Guiana, respectively.
which translate into similar reflectances, at least during some moments of the year. Classification-based approaches are therefore intrinsically independent of the location and the time,  [59]. For this latter figure, the dotted line represents the polynomial regression line over the in-situ dataset. The grey thick line represents the polynomial relationship defined by [59].
and thus present a stronger universal character being presumably more suitable for large scale applications than classical regional approaches. Previous studies have treated the partition of the marine waters into optical classes. They were based on sets of bio-optical parameters diversely including the following: the diffuse attenuation coefficient, the Secchi depth or inherent optical properties and biogeochemical parameters [12,[96][97][98]. Other studies, in particularly, at global scale, were dedicated to optical water type definition based on the information provided by the remote sensing reflectance spectra [99,100]. These studies have emphasized the potential of using ocean typology based on ocean color radiometry for characterizing the uncertainties related to ocean color products [101] and improving the performance of the inversion procedure [102]. Furthermore, the advantage of the optical classification approach for providing relevant insights into ocean water masses optical dynamics and therefore studying its biogeochemical quality has also been illustrated for the open ocean waters [103]. Similar applications of optical classification dedicated to the coastal ocean are currently limited. Few studies have specifically focused on a characterization of the optical diversity of coastal water masses from in situ measurements [83,105].
Satellite applications of coastal ocean optical typology for monitoring coastal water quality [106] or improving ocean color product inversion [107,108] are also still relatively scarce. A recent study [109] performed from an in situ data set gathered in contrasted coastal waters (i.e. eastern English Channel, southern North Sea and French Guiana) has emphasized the applicability and the advantages of this approach (Figure 13). Specifically the main results of this study have emphasized that (i) the ubiquitous character of R rs spectra optical shape at global scale (ii) the need of a reasonable amount of optical classes for describing coastal waters optical diversity (iii) the interest of the optical classification for dynamically monitoring the coastal waters masses quality and (iv) the potential for this approach for improving estimates of satellite products (with preliminary results on the SPM retrieval). The potential of the classification-based approach should be however explored more in detail through the establishment of large data set coupling optical and biogeochemical measurements gathered in a large variety of coastal waters.

The need of other "tools"
a. Towards the use of other spectral domains The past and present ocean color sensors are generally characterized by a common set of spectral channels. The blue (412, 443, 490 nm) and green (510, 550/560 nm) wavelengths are used for bio-optical algorithms purposes whereas the red (670) and near infrared (765, 865 nm) wavebands are used for atmospheric corrections. To these standard ocean color channels, some sensors also have the ability to measure the chlorophyll-a fluorescence peak using a band centered at about 676 nm in combination with two surrounding bands (around 665 and 746 nm) used for the baseline [110]. To overcome the challenges of ocean color remote sensing of coastal areas, the use of visible wavelengths other than the standard ones should also be examined. For instance, the maximum reflectance signal observed in coastal areas is, in many cases, observed around 590 nm, which is far from the 550/560 standard bands which can not adequately capture this maximum ( Figure 14). Because of the generally low absorption coefficient at this wavelength, it could also be used to improve the determination of the spectral slope of the particulate backscattering coefficient [18]. Figure 14. Example of normalized remote sensing spectra, R rs , collected in various coastal environments (English Channel, North Sea, and French Guyana). The description of this data set can be found in [105]. The normalization has been performed by dividing raw R rs spectrum by its integrated value in order to emphasize changes in the shape of the reflectance found in coastal waters. The recent use of the ultraviolet (UV) bands in the frame of the development of ocean color algorithms over coastal waters has shown some promising results. For instance, the exponential increase of detritus and colored dissolved organic matter absorption coefficient from the long to the short wavelengths makes the UV domain an ideal candidate for atmospheric correction purposes in turbid waters [90]. This new method, which based on a low and relatively stable water reflectance signal in the UV, does therefore not require any assumption on the water inherent optical properties. Based on the same observations, the 412 nm channel has previously been used with success in an atmospheric correction scheme applied over the Chesapeake region [39].
The ultraviolet spectral domain is also essential for assessing information about the colored dissolved organic matter distribution as already illustrated by recent works [111,112]. The advantage of using nLw(325)/nLw(565) have been for instance emphasized by the latter authors for specifically assessing a CDOM coefficient. Specifically, the interest of using of the UV reflectance signal stands in its ability for minimizing the potential issues represented by the overlapping of CDOM and phytoplankton absorption spectra occurring for the classically used blue and green bands. Indeed, in the UV domain CDOM is expected to dominate the absorption budget of the non-water compounds while the contribution of phytoplankton photoprotective compounds such as microsporine-like amino acids [113], although still uncertain, is expected to be relatively restricted in this spectral region. In addition to a gain in CDOM retrieval accuracy, the use of successive UV bands for deriving CDOM absorption from the marine reflectance signal would provide the opportunity to derive information on CDOM spectral slope which might provide innovative information on the nature of CDOM [114,112].
Besides, reflectances in the near infrared domain have also been demonstrated to be of particular interest for estimating the SPM concentration from satellite imagery [e.g. AVHRR: [115]; MODIS, MERIS SeaWiFS: [116]]. In this spectral domain, direct relationships between NIR-reflectance and SPM values appear indeed to be robust due to the very low influence of CDOM absorption at these wavelengths. The latter reason has also led to the development of a variety of Chla inversion algorithms based on red and NIR wavelengths [117][118][119][120][121][122] allowing to avoid issues related to the use of classical blue/green reflectance ratios and taking in some cases advantages of the Chla fluorescence properties in the latter spectral domains.

Directional polarized measurements
While the UV and NIR spectral domains are excellent candidates to overcome the challenges of ocean color remote sensing of coastal areas, the use of directional polarized radiometric measurements in the visible should also be examined. The inverse methods used to estimate the optical (IOPs) and biogeochemical parameters (i.e. Chl, POC, etc) information from space are based on different assumptions and mathematical developments. However, they all use the total remote sensing reflectance, R rs (λ), as input parameters (or a similar radiometric quantity). While theoretical and field measurements have highlighted that the polarization of the underwater light field is sensitive to the nature of the suspended marine particles (for example phytoplankton vs. mineral), none of these models exploits the linear polarization of the upwelling light field from the ocean surface. This is due to the fact that the marine polarized signal only represents a small fraction of the signal measured from space (about 10% of the total marine signal) which makes this measurement very challenging. Recent studies have however showed that space measurement of the polarized light field in different directions could provide useful information for both advanced atmospheric correction and bio-optical algorithms [123,124]. For instance, the signal measured from the POLarization and Directionality of the Earth's Reflectances sensor (POLDER-2) over turbid areas has been found in excellent agreement with theory with regards to its variability with the bulk particulate matter [123]. However, the exploitation of such measurements in the frame of algorithms development for the futures ocean color missions requires an excellent radiometric resolution.

Hyperspectral measurements
Hyperspectral sensors cover the visible and near infrared spectra with a much more complete spectral resolution, compared to multispectral ocean color sensors which have about 8 (i.e. MODIS) to 15 (i.e. MERIS) spectral bands in this spectral domain. For instance, the Hyperspectral Imager for the Coastal Ocean (HICO), operating on the international Space Station (ISS), specifically samples the coastal ocean every 5.7 nm [125]. Such high spectral resolution, associated with a high signal-to-noise ratio, allow the spectral features of the observed scene to be resolved. The interest of hyperspectral measurements compared to multispectral measurements for spatial ocean color applications is still under debate. However, more and more studies emphasize the importance of hyperspectral remote sensing data for phytoplankton detection, optically significant water constituents assessment, as well as bottom characterization [126][127][128][129][130]. Most of these studies are based on the derivative spectroscopy technique which enhances subtle spectral features present over hyperspectral data. While the second derivative of the radiometric signal allows specific spectral features to be detected, it is also much less sensitive to the presence of additional material such as colored dissolved organic matter which generally biases the standard algorithms based on a restricted number of wavelengths [128]. Futures ocean color sensors such as the HYPERSPECTRAL IMAGER (HSI) on board EnMAP will offer the opportunity to go further in the exploitation of hyperspectral data for ocean color applications.

Advanced statistical methods
Due to the complexity of the coastal waters, the estimation of the water contents are more and more difficult to obtain with a high accuracy as shown in the previous sections. Therefore, there is a need in developing innovative inversion methods. Advanced statistic methods that can be very promising for the monitoring of the ocean color are the machine learning. These methods, developed initially for computational/artificial intelligence, are now more and more used in environmental sciences. As they have shown their utility in climate, meteorology, atmospheric sciences, satellite data processing, analysis and modeling of environmental data, weather prediction, they can be very helpful for studying the ocean color. Several studies already exist for the application of artificial neural networks for the atmospheric correction processing [50-52, 20, 57, 41], the estimation of the chlorophyll-a concentration [54,[131][132][133][134][135][136], the inherent optical properties retrieval [33,[137][138] or the diffuse attenuation coefficient [139]. But there exists other types of machine learning that the community could use such as the support vector machine [140,141], the unsupervised competitive learning [142], regression trees [143], non-linear principal component [144], canonical correlation analysis [145]. These methods can be used for regression analysis for the direct estimation of water constituents, for classification (water types, phytoplankton functional groups) and for analysis of time series.
For instance, Figure 15 shows a comparison of different algorithms for the estimation of the spectral diffuse attenuation coefficient K d (490) [139]. The use of a NN (bottom right Figure) allows to decrease the error for the estimation of moderate and high values of K d (490) (K d (490)> 0.5 m -1 ).

High spatial and temporal resolutions satellite data
A high spatial resolution is required for ocean color remote sensing of coastal waters to catch the high spatial variability generally occurring in these areas, as well as for validation purposes. The recent development of geostationary ocean colour sensors will increase the precision of the remote sensing measurements and will provide relevant insights for the study of marine biogeochemical cycles [146,147]. Geostationary satellites continuously view the same region of the earth's surface. It thus allows obtaining high quality and frequent observations of a defined area. Such an instrument is therefore particularly useful in order to follow the response of marine ecosystems to short-term variations in the environmental conditions. In particularly, it is of interest for monitoring the effects of rivers plumes and tidal front and mixing on the biotic and abiotic material present in coastal areas or assessing the effects of exceptional events (storms, red tides, dissemination of sediments or pollutants). It will also provide relevant information for biogeochemical modelling purposes as well as for research activities related to the biogeochemical cycles at daily scales. South Korea's instrument on board the COMS-1 satellite (GOCI, Geostationary Ocean Color Imager), which has been recently launched (2010), is the first ocean-colour sensor in a geostationary orbit. The target area of GOCI covers a large region (2500x2500 km) around the Korean peninsula. Its resolution is 500 m while it acquires data at a 1 hour frequency. The other ocean color geostationary missions which are currently planned (OCAPI-CNES, GeoCape-NASA) will increase the spatial coverage and the number of information delivered by such sensors.

Conclusions
The use of ocean color remote sensing in coastal waters offers a capability to routinely monitor the surface water constituents over large areas. Thanks to our better understanding of the remote sensing signal, as well as to the improvement sensor technology, new innovative products, compared to the historical chlorophyll concentration, are now available, at least at regional scales. One may cite for instance the colored dissolved organic matter, and concentration and nature of suspended particulate matter. Such information coupled with other data coming from other satellite observations (sea surface temperature, wind speed and direction, sea surface elevations) and physical and ecological modeling provide essential inputs to understand the complex bio-physical coupling occurring in coastal waters. For instance, the coupling between bottom sediment resuspension and observed surface suspended particulate matter has been stressed using satellite and physical modeling [148]. In the same way, based on the synergic use of satellite products deriving from ocean color observations, thermal measurements, and sea surface height, tight coupling bio-physical processes can now be examined [149]. While numerous ocean color products are now available in coastal areas, their assessment is not as accurate as for open ocean waters. Note that there is still no consensus of the community for a common algorithm to assess the chlorophyll concentration in coastal areas. One of the major issues is the retrieval of the marine radiometric signal from the top of atmosphere (i.e. the atmospheric corrections). Numerous new methods, listed in the present chapter, are now developed to address this particular point. A second issue is related to the bio-optical algorithms development. In the frame of algorithms development, new paths, compared to the ones used for open ocean waters, are still in the experimental phase. Among them, the simultaneous retrieval of atmospheric and oceanic components, the algorithms based on the classification approaches, as well as the exploitation of geostationary ocean color sensors should open new ways in a recent future.