Estimation of diffuse attenuation of ultraviolet light in optically shallow Florida Keys waters from MODIS measurements

Remote Sensing of Environment 140 (2014) 519–532

Contents lists available at ScienceDirect

Remote Sensing of Environment journal homepage: www.elsevier.com/locate/rse

Estimation of diffuse attenuation of ultraviolet light in optically shallow Florida Keys waters from MODIS measurements Brian B. Barnes a,⁎, Chuanmin Hu a, Jennifer P. Cannizzaro a, Susanne E. Craig b, Pamela Hallock a, David L. Jones a, John C. Lehrter c, Nelson Melo d,e, Blake A. Schaeffer c, Richard Zepp f a

College of Marine Science, University of South Florida, 140 7th Avenue South, St Petersburg, FL, 33701, USA Dalhousie University, Department of Oceanography, Halifax, Nova Scotia, Canada National Health and Environmental Effects Research Laboratory, Gulf Ecology Division, United States Environmental Protection Agency, Gulf Breeze, FL, USA d Cooperative Institute for Marine and Atmospheric Studies, University of Miami, Miami, FL, USA e National Oceanic and Atmospheric Association Atlantic Oceanographic and Meteorological Laboratory, Miami, FL, USA f National Exposure Research Laboratory, United States Environmental Protection Agency, Athens, GA, USA b c

a r t i c l e

i n f o

Article history: Received 29 May 2013 Received in revised form 13 September 2013 Accepted 25 September 2013 Available online 22 October 2013 Keywords: Water clarity Light penetration Ultra-violet light Remote sensing Shallow water Coral reef

a b s t r a c t Diffuse attenuation of solar light (Kd, m−1) determines the percentage of light penetrating the water column and available for benthic organisms. Therefore, Kd can be used as an index of water quality for coastal ecosystems that are dependent on photosynthesis, such as the coral reef environments of the Florida Reef Tract. Ultraviolet (UV) light reaching corals can lead to reductions in photosynthetic capacity as well as DNA damage. Unfortunately, field measurements of Kd(UV) lack sufficient spatial and temporal coverage to derive statistically meaningful patterns, and it has been notoriously difficult to derive Kd in optically shallow waters from remote sensing due to bottom contamination. Here we describe an approach to derive Kd(UV) in optically shallow waters of the Florida Keys using variations in the spectral shape of MODIS-derived surface reflectance. The approach used a principal component analysis and stepwise multiple regression to parsimoniously select modes of variance in MODISderived reflectance data that best explained variance in concurrent in situ Kd(UV) measurements. The resulting models for Kd(UV) retrievals in waters 1–30 m deep showed strong positive relationships between derived and measured parameters [e.g., for Kd(305) ranging from 0.28 to 3.27 m−1; N = 29; R2 = 0.94]. The predictive capabilities of these models were further tested, also showing acceptable performance [for Kd(305), R2 = 0.92; bias = − 0.02 m− 1; URMS = 23%]. The same approach worked reasonably well in deriving the absorption coefficient of colored dissolved organic matter (CDOM) in UV wavelengths [ag(UV), m− 1], as Kd(UV) is dominated by ag(UV). Application of the approach to MODIS data showed different spatial and temporal Kd(305) patterns than the Kd(488) patterns derived from a recently validated semi-analytical approach, suggesting that different mechanisms are controlling Kd in the UV and in the visible. Given the importance of water clarity and light availability to shallow-water flora and fauna, the new Kd(UV) and ag(UV) data products provide unprecedented information for assessing and monitoring of coral reef health, and could further assist ongoing regional protection efforts. © 2013 Elsevier Inc. All rights reserved.

1. Introduction The necessity of light for photosynthetic energy creation restricts zooxanthellate corals to shallow, oligotrophic environments which can present potential exposure to ultraviolet (UV) radiation. While light is critical for photosynthesis, exposure to intense UV radiation can cause oxidative stress (Foyer, Lelandais, & Kunert, 1994; Jokiel, 1980) in corals and other reef flora (Fisk & Done, 1985; HoeghGuldberg & Jones, 1999; Lesser, Stochaj, Tapley, & Shick, 1990). Tolerance to incident UV radiation may depend on the concentration of colored dissolved organic matter (CDOM, aka gelbstoff or yellow ⁎ Corresponding author. E-mail address: [email protected] (B.B. Barnes). 0034-4257/$ – see front matter © 2013 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.rse.2013.09.024

matter) within the surrounding water column (West & Salm, 2003; Zepp et al., 2008). CDOM is a byproduct of terrestrial or marine plant matter decay (Shank, Lee, Vähätalo, Zepp, & Bartels, 2010). Spectrally, the CDOM absorption coefficient (ag) is the highest in the UV wavelengths, and it decreases with increasing wavelength with an exponential spectral slope (S; Bricaud, Morel, & Prieur, 1981). Exposure to UV light, especially in stratified surface waters, will cause photobleaching and oxidation of CDOM (Fichot & Benner, 2012; Häder et al., 1998; Moran & Zepp, 1997; Shank, Zepp, Vähätalo, Lee, & Bartels, 2010; Zepp et al., 2008), which will subsequently reduce the CDOMinduced light attenuation and increase the UV light potentially reaching coral reef environments. Despite the wealth of scientific literature describing the relationship between light and coral health, there are surprisingly few implemented

520

B.B. Barnes et al. / Remote Sensing of Environment 140 (2014) 519–532

programs to monitor light availability at coral reefs, primarily due to logistical and financial constraints. To overcome this difficulty, the U.S. NOAA's Coral Watch Program has implemented an experimental light stress product over the global ocean (http://coralreefwatch.noaa. gov/satellite/lsd/index.html). However, the product is only based on satellite-estimated surface photosynthetically available radiation in the visible wavelengths (PAR) without accounting for spatial variation in light attenuation or for depth-dependent changes in radiation reaching the benthos, noted by Yentsch et al. (2002) as crucial for all coral studies. This omission is due to a lack of reliable algorithms to derive light attenuation (or water clarity) from satellite data in optically shallow environments (see Barnes et al., 2013; Zhao, Barnes, et al., 2013). ‘Optically shallow’ is defined here as waters where the bottom is visible from an above-water instrument, which includes the environments inhabited by many zooxanthellate corals. This lack of satellite water clarity data stands in direct contrast to sea surface temperature (SST) data from satellites, which are not adversely affected in optically shallow environments (e.g., Hu et al., 2009). As a result, data products based on satellite-derived SST are currently used almost exclusively to predict coral bleaching (see Strong, Liu, Meyer, Hendee, & Sasko, 2004; Strong, Liu, Skirving, & Eakin, 2011). Estimation of water clarity or diffuse light attenuation coefficient (Kd, m−1, see Table 1 for list of symbol definitions) in optically shallow environments from satellite measurements has been problematic in the past due to bottom contamination of the satellite-derived remote sensing reflectance (RRS). Most inversion algorithms are developed and validated for waters without the benthic contribution to the RRS. For example, Johannessen, Miller, and Cullen (2003) described an empirical relationship between Kd(UV) and a reflectance ratio between the blue and green wavelengths for optically deep waters ranging from turbid estuarine to offshore oligotrophic. However, benthic albedo varies with wavelength for disparate bottom types (see Hochberg, Atkinson, & Andréfouët, 2003, Werdell & Roesler, 2003). As a consequence, application of the Johannessen et al. (2003) algorithm to optically shallow waters is likely to produce large errors in the retrieved Kd(UV). Fig. 1 shows performance of this algorithm in optically shallow waters, using data collected for this study. Benthic contamination has similarly been shown to cause errors in retrievals of water-column chlorophyll-a concentrations (Cannizzaro & Carder, 2006; Cannizzaro et al., in press; Carder, Cannizzaro, & Lee, 2005; Hu, 2008; Schaeffer, Hagy, Conmy, Lehrter, & Stumpf, 2012) and Kd in visible wavelengths (Zhao, Barnes, et al., 2013). Similar errors are likely to occur for Kd(UV) retrievals using other empirical algorithms (e.g., Fichot, Sathyendranath, & Miller, 2008; Smyth, 2011) because of the bottom interference. Nevertheless, instruments such as the Moderate Resolution Imaging Spectroradiometer (MODIS) onboard the U.S. National Aeronautics and Space Administration (NASA) satellites Aqua (2002–present) and Terra (2000–present) provide medium-resolution (1 km) RRS data for all of

Table 1 Description of symbols. Symbol

Description

Units

Kd

Diffuse attenuation coefficient for downwelling irradiance Downwelling irradiance Subsurface downwelling irradiance Radiance Reflectance Remote sensing reflectance Total absorption coefficient Absorption coefficient of gelbstoff (CDOM) Absorption coefficient of suspended particles Absorption coefficient of phytoplankton pigments Absorption coefficient of particulate detritus Spectral slope of ag Wavelength Depth

m−1

Ed Ed(0–) L R RRS at ag ap aph ad S λ z

W nm−1 m−2 W nm−1 m−2 W nm−1 m−2 sr−1 Dimensionless sr−1 m−1 m−1 m−1 m−1 m−1 nm−1 nm m

n = 54 R2 = 0.39

Fig. 1. Relationship between in situ measured Kd(380) and that derived from Johannessen et al. (2003) algorithm using the MODIS/A RRS as the algorithm inputs. Solid line is 1:1 reference, while dotted lines show ±50% error.

the world's optically shallow regions once every 1–2 days. In the Florida Keys region, Barnes et al. (2013) described a modified quasi-analytical algorithm (QAA, Lee, Carder, & Arnone, 2002) to remove bottom contamination and subsequently derive Kd in the visible from MODIS Aqua (MODIS/A) measurements. However, since MODIS does not collect RRS data in UV wavelengths, estimation of Kd(UV) requires an approach which is both resilient to variable bottom contamination and which can extrapolate beyond the measured visible satellite data. While Kd(UV) is linearly proportional to Kd in the visible bands [Kd(VIS)] in oligotrophic open-ocean waters (Lee et al., 2013), in optically complex environments, spatiotemporal differences between the relative concentrations of various water constituents (see Bricaud et al., 1981; DeGrandpre, Vodacek, Nelson, Bruce, & Blough, 1996) generally preclude use of Kd(VIS) to derive Kd(UV). Algorithms based on neural networks (Ioannou, Gilerson, Gross, Moshary & Ahmed, 2011, 2013; Jamet, Loisel, & Dessailly, 2012), matrix inversion models (Brando & Dekker, 2003; Phinn, Dekker, Brando, & Roelfsema, 2005), or empirical orthogonal functions (EOF; aka principal components analysis; Craig et al., 2012; Fichot et al., 2008; Fischer, Doerffer, & Grassl, 1986; Gower, Lin, & Borstad, 1984, Mueller, 1976; Sathyendranath, Hoge, Platt, & Swift, 1989, 1994; Sathyendranath, Prieur, & Morel, 1989; Toole & Siegel, 2001) have the potential to overcome the difficulty of deriving Kd in optically shallow waters, provided that data from these environments are included in the training datasets. The latter approach, which entails partitioning variations in the measured reflectance through the use of an EOF, has previously been used to estimate water properties in the UV (Fichot et al., 2008). This general approach, including its published variants discussed below, will herein be termed the ‘EOF method.’ The EOF method is used to reduce hyperor multispectral reflectance data to a few uncorrelated variables (EOF modes, aka eigenvectors), which retain most of the variance in the original data. Assuming that a particular water constituent affects the measured reflectance, varying concentrations of that constituent should be correlated to one or more of the EOF modes. Although the methodology for application of the EOF method has varied in the past, results have typically shown success of this approach in modeling water parameters. The inputs to the EOF method included measured reflectance (R; Gower et al., 1984; Mueller, 1976; Sathyendranath et al., 1994), simulated reflectance (Sathyendranath et al., 1989), simulated radiance (L; Fischer et al., 1986), and measured RRS (Craig et al., 2012; Fichot et al., 2008; Toole & Siegel, 2001), from regions ranging from coastal to open ocean. Water parameters derived from the EOF method included secchi disk depth (Mueller, 1976), chlorophyll concentration (Craig et al., 2012), fluorescence (Sathyendranath

B.B. Barnes et al. / Remote Sensing of Environment 140 (2014) 519–532

et al., 1989), inherent optical properties (IOPs) including absorption coefficients (Craig et al., 2012) and scattering coefficients (Gower et al., 1984), salinity and nutrient concentrations (Toole & Siegel, 2001), Kd for both visible and ultraviolet wavebands (Fichot et al., 2008), and others. Fichot et al. (2008) and Craig et al. (2012) further demonstrated the effectiveness of this approach using in situ measured RRS, which had been resampled to wavelengths of satellite instruments. To our knowledge, however, validation of the EOF method using satellite-derived data as an input (e.g., MODIS RRS) has yet to be reported. In the present study, our approach is distinct from other published EOF methods through 1) use of satellite-derived data as input, 2) parsimonious model selection via stepwise forward addition, and 3) measure of the predictive ability of models through leave-one-out-cross validation (see Sections 3.3 and 3.4 for details). Thus, given the pressing need for synoptic information on the UV light field reaching coral reef environments, the objective of this study was to develop and validate a method to estimate Kd and ag in UV wavelengths from MODIS/Aqua measurements over optically shallow waters. The method is based on a new approach to implement the EOF method, as well as extensive field data collected from the Florida Keys region for algorithm tuning and validation. 2. Study area — Florida Keys At the southern tip of the Florida peninsula are the Florida Keys (Fig. 2), a string of limestone islands at 24 to 26°N and 80 to 83°W. The Florida Reef Tract (FRT) is a series of bank and patch reefs located to the south and east of the Keys and extending west to the Dry Tortugas. The patch reefs are generally nearshore (3–5 km from land), while the larger reef tract is located approximately 8–10 km from the shore (abutting the shelf break) with depths typically ranging from 1 to 30 m. For the purposes of this research, the FRT region is delineated as all waters with depths from 0 to 30 m immediately south and east of the Florida Keys, and extending west to the Dry Tortugas. This designation of 30m matches the classification of “shallow water” in the satellite ocean color community (see Level-2 processing flag ‘COASTZ’; Patt et al., 2003). This threshold also is an appropriate delineation for the bounds of optically shallow

a)

521

waters in the FRT, as errors in chlorophyll retrievals due to bottom contamination of satellite RRS signals in this region have been found for depths shallower than 25 m (Cannizzaro et al., in press; Schaeffer et al., 2012). 3. Methods 3.1. In situ data collection To include a wide spatial and temporal scope for development of an inversion model applicable to all possible scenarios, in situ data were compiled from a number of different sources. As such, different instruments, data collection protocols, and quality assurance procedures were applied to all in situ data. All data were collected in the southwest Florida region, with an emphasis on the FRT. Overall, 292 different locations (211 in the FRT; see Fig. 2) were visited at least once, and up to 16 times per station, for a total of 540 station visits. Data collected at each station typically included vertical profiles of the downwelling irradiance (Ed) and water samples for the analyses of IOPs (Table 2). Much of these data, along with the collection and processing methods, have been included in previous publications (Ayoub, 2009; Barnes et al., 2013; Cannizzaro & Carder, 2006; Mueller & Fargion, 2002; Schaeffer, Conmy, Aukamp, Craven, & Ferer, 2011; Zepp et al., 2008, Zhao, Hu, et al., 2013). As such, the following is a summary of in situ data collection, where differences in protocol and instrumentation between the data sources are highlighted. For simplicity, wavelength (λ) and depth (z) notation for IOP and Ed descriptions are omitted except where necessary. Between January 2003 and May 2006, Ed and IOP data with a focus on the UV light environment were collected in the Florida Keys and Dry Tortugas regions by the US Environmental Protection Agency (EPA), the methodology for which has been detailed by Zepp et al. (2008). Briefly, a Satlantic (Halifax, NS, Canada) MicroPro profiler, configured to measure multispectral Ed primarily in UV wavelengths (at 305, 325, 340, 380, 412, and 443 nm) was deployed for 5–6 casts per station. Ed(z) measurements were discarded when instrument tilt or roll was greater than 5°. Water samples (approx 5 L) were collected at approximately 1 m depth and subsequently filtered to separate the

b) SOUTHWEST FLORIDA BIGHT

FLORIDA BAY Upper Keys

Middle Keys Dry Tortugas

Lower Keys STRAITS OF FLORIDA

Fig. 2. MODIS/A RGB image on 9 January 2009 (18:50 GMT) showing in situ measurement locations within the FRT. Diamonds indicate locations with concurrent satellite/in situ matchups, with color indicating the source of the in situ data. White crosses are locations with in situ data but without satellite RRS matchup. Inlaid charts show histograms of sample location a) depths and b) bottom types (from Florida Fish and Wildlife Conservation Commission habitat maps). BS = bare substrate, CR = coral reef, PS = patchy seagrass, CS = continuous seagrass, U = unmappable/uninterpretable, ND = no data.

522

B.B. Barnes et al. / Remote Sensing of Environment 140 (2014) 519–532

Table 2 Description of in situ data. Source

Dates (month/year)

Samples

MODIS/A Matchups

Kd instrument: wavelengths (nm)

ag wavelengths (nm)

ap wavelengths (nm)

ad wavelengths (nm)

EPA

1/03, 6/04, 8/05, 5/04, 8/09, 8/11,

191

28

Satlantic MicroPro: 305, 325, 340, 380, 412, 443

305, 325, 340, 380, 440

305, 325, 340, 380, 440

ND

39 34

4 8

200–800 200–800

400–800 400–800

400–800 400–800

144

34

ND Biospherical PRR-2600 & PUV-2500: PAR, 305, 330, 380, 443, 490, 555, 589, 625, 683 Satlantic HyperPro: 349–803

200–750

350–800

350–800

USF NOAA/ AOML

EPA/ USF

6/03, 8/03, 9/03, 2/04, 11/04, 1/05, 2/05, 6/05, 11/05, 3/06, 5/06 7/04, 9/04, 6/05, 7/05 12/09, 5/10, 12/10, 5/11, 10/11, 12/11, 2/12

4/11, 7/12, 8/12

dissolved (0.2 μm membrane filters) and particulate (0.7 μm nominal pore size glass-fiber filters — GF/F) water constituents, respectively. The GF/Fs were processed using the quantitative filter technique (Kiefer & SooHoo, 1982; Pegau et al., 2003; Roesler, 1998; Yentsch, 1962). A Perkin Elmer Model Lambda 35 UV–visible spectrophotometer was used for measurement of absorption coefficients due to particulates (ap) and gelbstoff (ag). Additional water samples were collected by the University of South Florida (USF) over the course of five cruises spanning May 2004 and July 2005. Water samples were collected from the surface using Niskin bottles as described by Ayoub (2009) and processed in a manner detailed by Cannizzaro and Carder (2006), Mueller and Fargion (2002), and Ayoub (2009). Samples were filtered on GF/F filters using the quantitative filter technique, from which ap spectra were measured using a custom-made, 512-channel spectral radiometer (350–850 nm; Bissett, Patch, Carder, & Lee, 1997). Hot methanol extraction (Kishino, Takahashi, Okami, & Ichimura, 1985; Roesler, Perry, & Carder, 1989) was used to remove pigments from these GF/Fs, which were re-analyzed to measure detrital absorption (ad). Phytoplankton absorption spectra were then calculated as aφ = ap − ad. Water samples were further filtered through a 0.2μm filter, and processed to measure ag spectra using a Perkin Elmer Lambda 18 UV–visible spectrophotometer. A pathlength of 10 cm was used for measurement of ag spectra from this, and all other, data sources. The NOAA Atlantic Oceanographic and Meteorological Laboratory (AOML) has conducted approximately bi-monthly cruise surveys aboard the R/V F.G. Walton Smith in the Florida Keys region as part of its South Florida Program (SFP) since 1999 (see Zhao, Hu, et al., 2013). A subset of those cruises included collection of Ed profiles and water samples (using a Niskin rosette at approximately 1 m depth) for IOP analysis. For the former, Biospherical Instruments, Inc. (San Diego, CA) Profiling Reflectance Radiometer (PRR-2600) and Profiling Ultraviolet Radiometer (PUV-2500) were used to collect Ed at 305, 330, and 380 nm (PUV-2500) and 380, 443, 490, 555, 589, 658, and 683 nm (PRR-2600). Both instruments also recorded photosynthetically available radiation (PAR) profiles with depth. Typically 2 to 6 casts were collected for each instrument at each station visited within two hours of local solar noon. Ed(z) data for which the instrument tilt or roll was greater than 10° were excluded from further analysis. Water samples were processed using the same methods and instrumentation as the USF dataset. In April 2011 and July/August 2012, in situ IOP and Ed data were collected through a joint EPA/USF effort. At each depth, hyperspectral (400–735 nm, interpolated to every 1 nm) Ed profiles were collected using a Satlantic HyperPRO (for details see Barnes et al., 2013). Ed(z) measurements were discarded when instrument tilt or roll was greater than 5°. Water samples were collected from just below the surface for measurement of both ag and ap. Similar to the methods used for the USF and NOAA/AOML datasets, water samples were filtered on GF/F filters using the quantitative filter technique. A Shimadzu UV1700 dual beam spectrophotometer was used for measurement of ap, ad (after warm methanol extraction), and ag spectra (see Schaeffer et al.,

2011 for details). Residual scattering was removed from these spectra by subtracting from each wavelength the mean of each measured value between 790 and 800 nm (Pegau et al., 2003). All Ed profiles were processed to derive Kd via Lambert–Beer's law (Gordon, 1989; Smith & Baker, 1981), using linear regression between depth and
ln

 Ed ðz; λÞ :  Ed ð0 ; λÞ

ð1Þ

For the EPA/USF and NOAA/AOML data sources, ln(Ed) at several wavelengths for each profile was plotted against measurement depth, and the longest sequence of approximately log-linear decay in Ed was manually selected. Only data within this portion of the profile were used in the derivation of Kd via the linear regression. Also, for the EPA/USF and NOAA/AOML data sources, Kd derived from each station was further assessed for quality control by testing the similarity of replicate profiles (i.e., measurement repeatability). Specifically, Kd(λ) from any station was only considered reliable if the difference in Kd(λ) from two or more replicate profiles was less than 0.01 m−1 and/or the coefficient of variation for two or more replicate profiles was less than 10%. The individual Kd(λ) measurements which fit either of these conditions were averaged to yield the final Kd(λ) used in subsequent analyses. In the NOAA/AOML dataset, both the PRR and PUV instruments collected profiles for certain wavebands (PAR and 380 nm). For stations where both instruments were deployed, data for these wavebands from the different instruments were considered replicates. As there were no instances of concurrent in situ data from any of the four data sources, no attempts were made to identify or reconcile any differences between instruments as they pertain to calibration or performance. 3.2. Satellite data collection All MODIS/A Level-1A data from the Florida Keys region and within the years 2002–2012 (4566 files) were downloaded from the NASA Goddard Space Flight Center (GSFC) ocean color website http:// oceancolor.gsfc.nasa.gov/. Only MODIS/A data were used due to the striping and scan mirror damage on MODIS/Terra, which limit its utility for ocean color research (Franz, Kwiatkowska, Meister, & McClain, 2008). Using default processing parameters and SeaDAS version 6.4, level-2 RRS at all visible bands (412, 443, 469, 488, 531, 547, 555, 645, 667, and 678 nm), as well as the level-2 processing flags, were created at 250 m resolution. For most MODIS bands, this spatial resolution was achieved by SeaDAS via interpolation from native 1 km or 500 m resolution. RRS data from each band were subsequently mapped to a cylindrical equidistant projection with bounds 24 to 26 N and 80 to 83 W, and stored in Hierarchical Data Format 4 (HDF4) files. RRS data concurrent (i.e., same day and collocated) with in situ samples were extracted from these HDF4 files and stored in ASCII text

B.B. Barnes et al. / Remote Sensing of Environment 140 (2014) 519–532

files. RRS data were removed from analysis if they were flagged by any of the standard level-2 processing flags (Patt et al., 2003). Also, any MODIS/A-derived spectrum with negative RRS at any wavelength was discarded to exclude any data with potential atmospheric correction failures. All image processing and data extraction were performed using IDL version 8.0 (Interactive Data Language; Exelis Visual Information Solutions). The concurrent satellite RRS and in situ measured water parameters that passed all quality control procedures are hereafter termed ‘matchups,’ while the matchup data from all parameters is the ‘training dataset.’ In total, concurrent MODIS/A RRS spectra were found for 74 of the 408 (18%) stations sampled, which represents the maximum possible number of matchups for any given parameter (number of matchups varies by parameter due to sampling differences between the four in situ data sources). Once validated, the algorithm described below was applied to the entire time series of MODIS/A RRS in the Florida Keys region. Two locations, A (24.548°N, 81.397°W; Lower Keys) and B (24.651°N, 81.098°W; Middle Keys), were selected from within the FRT according to their repeated and long term in situ sampling. Time series of derived water parameters were extracted for these two stations and binned by month. Image-wide monthly mean climatologies (2002–2012) of Kd(305) were also created for July and January to show representative summer and winter conditions, respectively. MODIS/A RRS from these two months were also processed to derive Kd(488) using the algorithm described by Barnes et al. (2013), from which monthly mean climatologies were calculated. Percent penetration of light to the benthos was calculated as: ð−K d zÞ

Penetrationð% Þ ¼ 100e

;

principal components regression) to model water parameters. Our approach expands on this, while ensuring parsimony, through the forward addition of EOF modes (i.e., stepwise multiple regression). As such, this process increases the statistical power of the multiple regression by including only independent variables that are significantly contributing to the explained variance of the dependent variable (Gauch, 1993). Inclusion of superfluous predictor variables will always increase the R2 and Fisher's test statistic (F) of a multiple regression (Cohen, Cohen, West, & Aiken, 2003). Although an EOF mode may explain a tiny portion of the total variance in RRS, it may still be relevant in deriving water parameters (see von Storch & Zwiers, 1999). Each water parameter was modeled independently using a modification of the EOF method. The term ‘parameter’ is used to describe Kd, ag, aφ, or ad at a specific wavelength [for example, Kd(305) or ag(490)], and there were 1930 parameters included in the training dataset. For each water parameter, only matchups with valid in situ data were included in the analyses. The number of matchups (N) varied by water parameter due to differences in sampling between the various projects from which data were collected and because of quality-control procedures. All water parameters were log transformed to account for their approximately log-normal distribution (Campbell, 1995). For simplicity, we will describe algorithm implementation for Kd(305), although all measured water parameters with matchup satellite RRS were analyzed in the same manner. All statistical analyses were performed using Matlab R2011a (Mathworks). First, to focus on variability in spectral shape (as opposed to variability in amplitude), MODIS/A RRS spectra from Kd(305) matchups were normalized by:

ð2Þ

where z is the bottom depth from a bathymetry database provided by the United States Geological Survey (USGS). 3.3. Algorithm development Previously published variants of the EOF method often involved a cursory investigation of the shape of individual EOF modes to visually interpret how various water constituents might contribute to each mode (Craig et al., 2012; Fichot et al., 2008; Fischer et al., 1986; Toole & Siegel, 2001). Explaining the first two modes is typically straightforward, while subsequent modes require more conjecture. Despite the lack of apparent physical meaning, these subsequent modes may be most effective at representing the variance in the model (see von Storch & Zwiers, 1999). Measures of correlation to a concurrently measured water parameter have also been used to determine the contribution of water constituents to individual EOF modes (Fischer et al., 1986; Sathyendranath et al., 1989, 1994; Toole & Siegel, 2001). When empirically modeling measured parameters, mode selection has previously included statistical justification using R2adj (Fichot et al., 2008) or confidence intervals of multiple regression coefficients (Craig et al., 2012; Mueller, 1976). Mueller (1976) utilized different modes for estimation of different parameters. In contrast, while noting that certain EOF modes were negligibly contributing to the models for different parameters, Fichot et al. (2008) and Craig et al. (2012) used EOF modes 1–4 in the final models for all parameters to maintain consistency. In all of these works, only the first several (typically 4) modes were considered, owing to the high percentage of total variance (generally ~99%) explained by these modes. In this paper, we demonstrate the implementation of a stepwise, forward-addition procedure for selection of EOF modes to be retained for use in modeling water parameters. This approach removes the guesswork and ambiguity from the EOF mode selection process. As such, statistically defensible mode selection is straightforward for model application to derive water parameters. Mueller (1976) and Craig et al. (2012) utilized multiple regression of EOF modes (aka

523

hRRS ðλÞi ¼ Z

RRS ðλÞ 678 412

;

ð3Þ

RRS ðλÞdλ

where b RRS(λ)N is the normalized spectra (Craig et al., 2012). An EOF was then performed on the matchup MODIS/A b RRSN, which partitioned the variance in the bRRSN along 10 modes. Since the 10 visible MODIS bands were used in this analysis, these 10 modes explain 100% of the variance in the bRRSN. The resulting scores for each of these EOF modes were used as the independent variables in a stepwise multiple regression (i.e., stepwise principal components regression), with the dependent variable being the matchup in situ log10[Kd(305)]. The significance of this global (i.e., all inclusive) model was tested using 1000 permutations. If the significance level of the global model was greater than alpha (0.05), no further analysis was performed to model or estimate this water parameter from MODIS/A data. However, if the global model was significant, stepwise multiple regression was further used to test the individual explanatory effect of each EOF mode on log10[Kd(305)]. A final model was constructed by sequentially adding the EOF modes to the model which yielded the largest partial F. Addition of EOF modes to the model was stopped if addition of the next EOF mode raised either 1) the significance level over alpha or 2) the adjusted coefficient of multiple determination (R2adj) over that of the global model (see Blanchet, Legendre, & Borcard, 2008). This approach allows parsimonious determination of the EOF modes that are contributing to the variance in Kd(305). Coefficients of determination (R2) in log-space are primarily used to describe the statistical relationships between these parsimoniously selected EOF modes and the matching in situ log10[Kd(305)], hereafter termed ‘model’ results. 3.4. Predictive validation statistical methods To test the ‘predictive’ ability of the approach, leave-one-out-crossvalidation was performed for each significant model result. To do so, multiple linear regression was performed, using only the significant EOF modes as independent variables and excluding data from one of the matchups. The EOF scores for the excluded matchup bRRSN were

524

B.B. Barnes et al. / Remote Sensing of Environment 140 (2014) 519–532

then combined with the multiple linear regression coefficients to predict log10[Kd(305)] for that matchup. This process was repeated N times, with each iteration excluding a different matchup. The predicted log10 [Kd(305)] values were compared to the in situ measured log10[Kd(305)] values using simple linear regression. Unbiased root mean squared percent error (URMS; Hooker, Lazin, Zibordi, & McLean, 2002) and bias of this predictive model were calculated on the non-transformed data as: bias ¼

N 1X ðpredi −measi Þ; N i¼1

ð4Þ

and vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u N
2 u1 X ðpredi −measi Þ ; URMS ¼ t N i¼1 ððpredi þ measi Þ=2Þ

ð5Þ

where pred and meas represent predicted and measured values at i sample number. As a benchmark for URMS, the NASA mission goal for satellite-derived chlorophyll-a retrievals is 35% accuracy (Hooker, Esaias, Feldman, Gregg, & McClain, 1992). Gregg and Casey (2004) found Sea-viewing Wide Field-of-view Sensor (SeaWiFS) default algorithms applied over global ocean waters show RMS greater than 0.2 for log transformed chlorophyll-a, which corresponds to relative RMS uncertainties over 58%. 4. Results 4.1. Model development and validation A total of 1930 in situ parameters [e.g., Kd(305)] were tested using this approach, of which 1319 showed significant global models. Table 3 presents a summary of performance for selected parameters. In the following results and discussion, only results from Kd and ag at 305 and 340 nm are described in detail. The rationale for the focus on

these specific parameters is that they represent application of the approach in both the UVA (315–400 nm) and UVB (280–315 nm) wavebands, which can lead to photoinhibition and DNA damage, respectively (Lesser & Lewis, 1996; Zepp et al., 2008). While Kd (and depth) determines the light exposure to the benthos, ag is the dominant component of water absorption for blue and UV wavelengths in the FRT region (Zepp et al., 2008). As an example, Fig. 3 shows the percentage contributions of particulate and dissolved matter absorption coefficients to the total non-water absorption coefficient at 440 nm. For most water samples, especially for those collected from the FRT, total non-water absorption is dominated by ag, with minimal contributions from non-living detrital particles (ad). Kd and ag at 305 and 340 nm all showed highly significant positive relationships between the in situ measurements and model results (Fig. 4). For Kd(305), 29 matchups were included in the training dataset with in situ values ranging from 0.28 to 3.27 m−1. Regression of the model results and measured values (in log-space) showed an R2 of 0.94, F of 438 and p-value ≪ 0.01. Kd(340) showed very similar results across all statistical indices (N = 28; range = 0.17–2.27 m−1; R2 = 0.96; F = 577; p ≪ 0.01). Strong positive relationships were also seen for ag(305) (N = 65; range = 0.10–3.19 m−1; R2 = 0.87; F = 433; p ≪ 0.01), and ag(340) (N = 65; range = 0.02–1.50 m−1; R2 = 0.81; F = 271; p ≪ 0.01), which had triple the number of matchups as Kd(UV), yet suffered little decrease in R2. The predictive ability of this approach, using leave-one-out-crossvalidation, showed slightly lower statistical performance than the model results (Fig. 5). Nevertheless, the regression of the prediction results with in situ values again showed strong positive relationships for Kd(305) (R2 = 0.92; F = 314; p ≪ 0.01; bias = −0.02 m−1; URMS = 23%), Kd(340) (R2 = 0.94; F = 386; p ≪ 0.01; bias = −0.01 m−1; URMS = 20%). Slightly lower R2 and URMS were seen for ag(305) (R2 = 0.84; F = 320; p ≪ 0.01; bias = −0.02 m−1; URMS = 30%), and ag(340) (R2 = 0.74; F = 183; p ≪ 0.01; bias = −0.01 m−1; URMS = 40%), yet the increased number of matchup data points greatly decreased the p-values for these parameters.

Table 3 Model and prediction performance for selected parameters. Parameter

MODIS/A matchups

Model R2

Kd(305) Kd(325) Kd(340) Kd(380) Kd(410) Kd(440) Kd(490) Kd(555) Kd(670) ag(280) ag(305) ag(325) ag(340) ag(380) ag(410) ag(440) ag(490) ag(555) ag(670) aφ(410) aφ(440) aφ(490) aφ(555) aφ(670) ad(410) ad(440) ad(490) ad(555) ad(670)

29 28 28 54 20 22 23 23 18 40 65 65 65 63 37 60 32 32 35 21 21 21 21 21 21 21 21 21 21

0.94 0.97 0.96 0.89 0.84 0.79 0.74 Model 0.74 0.81 0.87 0.84 0.81 0.68 0.31 0.26 Model Model Model 0.81 0.84 0.83 0.73 0.63 0.59 0.59 0.74 0.63 0.73

not

not not not

Prediction R2 0.92 0.95 0.94 0.86 0.77 0.68 0.65 significant 0.37 0.76 0.84 0.79 0.74 0.62 0.25 0.19 significant significant significant 0.58 0.66 0.64 0.44 0.07 0.46 0.47 0.56 0.47 0.53

Prediction F

Prediction p-value

Prediction Bias (m−1)

Prediction URMS (%)

EOF modes in model

313.9 526.7 386.0 331.0 60.7 42.8 39.6

≪0.01 ≪0.01 ≪0.01 ≪0.01 ≪0.01 ≪0.01 ≪0.01

−0.02 −0.01 −0.01 −0.01 −0.01 0.00 0.00

23 17 20 27 32 33 26

1, 1, 1, 1, 1, 1, 1,

3, 2, 3, 4, 2, 3, 3,

2, 3, 2, 2, 3 7 2

9 7, 9 9 6, 3

9.5 118.6 320.4 235.2 183.1 100.7 11.4 13.8

b0.01 ≪0.01 ≪0.01 ≪0.01 ≪0.01 ≪0.01 b0.01 ≪0.01

0.00 −0.06 −0.02 −0.02 −0.01 −0.01 −0.02 −0.02

13 29 30 35 40 46 60 70

4, 1, 1, 1, 1, 1, 3 1,

1, 3, 3, 3, 3, 3,

2, 7, 6, 6, 6, 6

7 6, 5, 5, 5,

26.8 36.7 33.5 15.2 1.5 16.2 17.1 24.2 17.0 21.6

≪0.01 ≪0.01 ≪0.01 ≪0.01 0.236 ≪0.01 ≪0.01 ≪0.01 ≪0.01 ≪0.01

0.00 0.00 −0.00 0.00 −0.00 −0.00 −0.00 0.00 −0.00 0.00

52 45 50 94 82 103 106 94 102 99

2, 2, 2, 1, 2, 4, 4, 4, 4, 4,

1, 1, 1, 4, 4, 9 9 9, 9, 2,

6, 6, 4, 9, 1,

9, 9, 9, 2, 9

4 4 4 2

3

1, 2 1 9, 1

4, 8 4 6 5

B.B. Barnes et al. / Remote Sensing of Environment 140 (2014) 519–532

0

100

50 50

100

0

0

100

50 ad (440)

Fig. 3. Ternary plot showing percent contributions of water constituents CDOM, detritus, and phytoplankton pigments to the total non-water absorption coefficient at 440 nm for the FRT (red circles) and adjacent waters.

4.2. Application within the Florida Keys region

Modeled (m-1)

The application of any empirical algorithm should generally be restricted to the same spatiotemporal bounds and range of environmental conditions as the training dataset. To accentuate this limitation, we

have applied this approach to estimate Kd(305) from satellite/in situ matchups (from the same data sources listed above) for waters adjacent to the FRT. Such retrievals yielded poor predictive relationships for both Straits of Florida (Fig. 6, open squares; N = 9; URMS = 126%) and Florida Bay (Fig. 6, open triangles; N=4; URMS=65%) waters. Incorporating the Straits of Florida matchups into the training dataset greatly improved the predictive performance of this approach (Fig. 6, closed squares; URMS = 36%). However, much less improvement was seen when Florida Bay waters were included in the training dataset (Fig. 6, closed triangles; URMS = 60%). The latter is likely due to the small number of matchups in this large and diverse region, but again highlights the need to limit application of this approach to the range of environmental conditions (e.g., water properties and bottom types) represented within the training dataset. As such we have strived to incorporate in situ data spanning multiple years, with inclusion of data from all seasons and a variety of bottom types into our FRT training dataset. Further, the training dataset included a range of in situ values of at least one order of magnitude for most of the parameters being modeled and predicted. Nevertheless, the same logistical and financial limitations that plague established long-term monitoring programs hindered our ability to compile a training dataset that included all possible environments and water conditions (as well as combinations thereof) in the study area. Such effect is also clearly seen when the developed model is applied to MODIS/A data to create a Kd(305) map covering the entire Florida Keys region (Fig. 7, top panel). Even within the FRT, there appear to be locations for which Kd(305) has not been properly derived. For example, in the shallow reefs (depths as shallow as intertidal; Dustan & Halas, 1987) of the Upper Keys region, the spatial pattern in Kd(305) appears to be associated with changes in depth. In the FRT, Kd (on the scale of monthly climatologies) should instead decrease with distance to shore

Kd305

Kd340

n = 29

n = 28

R2 = 0.94

R2 = 0.96

a)

b)

ag305

ag340

n = 65

n = 65

R2

R2 = 0.81

= 0.87

525

c)

d)

Measured (m-1) Fig. 4. Relationships between measured and modeled a) Kd(305), b) Kd(340), c) ag(305), d) ag(340). Solid line is 1:1 reference, while dotted lines show ±50% error.

B.B. Barnes et al. / Remote Sensing of Environment 140 (2014) 519–532

Predicted (m-1)

526

Kd305

Kd340

URMS = 23%

URMS = 20%

Bias =-0.02

Bias = -0.01

a)

b)

ag305

ag340

URMS = 30%

URMS = 40%

Bias = -0.02

Bias = -0.01

c)

d)

Measured (m-1) Fig. 5. Relationships between measured and predicted (via leave-one-out-cross-validation) a) Kd(305), b) Kd(340), c) ag(305), d) ag(340). Solid line is 1:1 reference, while dotted lines show ±50% error.

Predicted Kd(305) (m-1)

and according to inundation from Florida Bay waters through channels between the islands (see Barnes et al., 2013). This effect of bottom contamination appears to be limited to waters shallower than 4.5 m. The training dataset used for this study includes locations just over 1 m deep, but such stations were primarily inshore as opposed to offshore, and the waters from these inshore stations were relatively turbid (i.e., less bottom contamination than for offshore clear water with the same water depth). This exclusion of extremely shallow, offshore,

clear-water reef locations from the training dataset was primarily logistical (access by boat is limited due to the depth), but also due to quality control of the measured Kd (reliable Kd is difficult to measure in clear shallow waters due to the surface wave-focusing effect). Thus, we suggest limiting the application of the developed model to sites deeper than 4.5 m. This limit, however, does not represent a deficiency of the overall approach, and could be remedied with a more inclusive training dataset.

, ,

Straits of Florida Florida Bay

Training Dataset , FRT only FRT and Straits of Florida FRT and Florida Bay

Measured Kd(305) (m-1) Fig. 6. Predictive performance of the EOF approach to derive Kd(305) (m−1) for waters in the Straits of Florida (squares) and Florida Bay (triangles). Open symbols show performance using only FRT data in the training dataset, while filled symbols represent retrievals when local samples are also included in the training dataset. Solid line is 1:1 reference, while dotted lines show ±50% error.

B.B. Barnes et al. / Remote Sensing of Environment 140 (2014) 519–532

527

Fig. 7. (Top) Image showing Kd(305) (m−1) derived from MODIS/A measurements on 9 January 2009 (same as Fig. 2) using the EOF model developed in this study. Land is masked in black. White indicates coastline, clouds or algorithm failure. (Bottom) Monthly mean time series of MODIS/A derived Kd(305) for two stations A (black) and B (gray), with error bars showing standard errors. In situ measured Kd(305) are also shown for stations A and B. White arrow shows example location of shallow waters (b4.5 m) for which Kd(305) retrievals are questionable.

5. Discussion 5.1. Spatio-temporal variability of Kd in the FRT Within the FRT, excluding these noted regions where retrievals from the EOF model may be erroneous, the spatial changes in derived Kd(305) appear to be reasonable. Specifically, a strong onshore-offshore gradient in Kd(305) is apparent, which has also been described for Kd(488) from MODIS/A data (Barnes et al., 2013), and is consistent with previous reports (Ayoub, 2009; Lapointe & Clark, 1992, Szmant & Forrester, 1996). Kd(305) of the Upper, Middle, and Lower Keys regions also shows spatial patterns similar to the Kd(488) patterns demonstrated by Barnes et al. (2013). However, there is a conspicuous difference in the seasonality of Kd(305) and Kd(488) (see below). The temporal patterns at individual locations have been validated by in situ measurements, suggesting the validity of the spatial Kd(305) patterns and its temporal contrast with Kd(488). For example, Kd(305) time series at stations A and B derived for the entire span of MODIS/A measurements (Fig. 7, bottom graph) show Kd(305) fluctuations that are verified by the occasional in situ data from these locations (note that these in situ data were not used in the training dataset due to lack of concurrent MODIS/A RRS). It is also worth noting that the largest peak in Kd(305) for one of the time series (Middle Keys, Station B; Fig. 7) coincides with the January 2010 cold event in Florida (see Barnes & Hu, 2013; Barnes, Hu, & Muller-Karger, 2011; Lirman et al., 2011), yet a similar peak was not seen in the Lower Keys time series (Station A). The Kd(305) peak at Station B may be potentially attributed to cold event related die-offs of the Everglades fauna and subsequent CDOM release. Although the spatial patterns of Kd(305) and Kd(488) may be similar, their temporal changes can be drastically different. Fig. 8 shows monthly mean climatology (2002–2012) images of Kd(305) and Kd(488) for July and January to represent conditions for summer and winter, respectively.

In these images, data from locations shallower than 4.5 m have been masked. If the shallow areas were completely surrounded by deeper water, Kd data for the shallow areas were filled by interpolation from the adjacent deeper data. In the FRT, Kd(488) is generally higher in the winter than in the summer (Fig. 8) via a regular seasonal cycle (Barnes et al., 2013). In contrast, Kd(305) shows reversed temporal changes for much of the FRT, as seen in the July/January ratio images in Fig. 8. The higher Kd(305) in July than in January is possibly driven by summer precipitation, which can lead to excessive CDOM-rich terrestrial runoff. Increased CDOM production from seagrass degradation is also expected in warmer waters (Stabenau, Zepp, Bartels, & Zika, 2004). Finally, even in dry summers, Everglades waters are occasionally released into Florida Bay to minimize adverse hypersalinity effects on seagrasses, providing another potential source of CDOM to downstream FRT environments. The contrast between the July/January ratios for Kd(305) and Kd(488) suggests that, unlike oligotrophic open-ocean waters where Kd(305) is linearly proportional to Kd(488) (Lee et al., 2013), the shallow Florida Keys waters are more optically complex and one cannot rely solely on Kd(VIS) to infer information on Kd(UV). This finding is not an artifact of differences between algorithms being used, as there was a great deal of scatter between concurrent Kd(UV) and Kd(VIS) retrievals when both were derived using the EOF approach. Fig. 8 also highlights the much higher Kd(305) than Kd(488) throughout the FRT region. Consequently, the percent penetration of UV light to the benthos is much lower than that of blue light. In fact, the only locations with appreciable 305nm light reaching the benthos are the offshore shallow coral reef environments, especially in the Upper Keys region. 5.2. Potential use in coral reef assessment In the context of coral reef research and monitoring, the potential uses of this long-term (2002–present), spatially synoptic dataset of UV

528

B.B. Barnes et al. / Remote Sensing of Environment 140 (2014) 519–532

July K d(305)

January

July / January K d(305)

K d(305)

Kd(488) (Barnes et al., 2013)

K d(488) (Barnes et al., 2013)

K d(488) (Barnes et al., 2013)

Kd(305) % Penetration

K d(305) % Penetration

K d(305) % Penetration

K d(488) % Penetration

K d(488) % Penetration

K d(488) % Penetration

(Barnes et al., 2013)

(Barnes et al., 2013)

(Barnes et al., 2013)

Fig. 8. Mean Kd(305) and Kd(488) for July 2004 and January 2005 derived from MODIS/A measurements using the EOF model developed in this study and Barnes et al. (2013). Also shown are their July/January ratios. The bottom panels show the corresponding percent light reaching the benthos and their July/January ratios.

light penetration are numerous. Although the satellite measurement frequency is often semi-weekly due to cloud cover, the spatial and temporal coverage of this dataset is much higher than any in situ research and monitoring programs (e.g., 2 month repeat sampling for the NOAA AOML South Florida Program; see Kelble & Boyer, 2007). Thus, this new dataset could provide plentiful information with which to research the individual and combined effects of light and temperature on coral reefs in the FRT. In particular, a better understanding of the variation in UV light availability may help explain the distribution of corals away from tidal channels through the Keys (Ginsburg & Shinn, 1964, 1993), but seemingly contradictory success of nearshore reefs relative to offshore reefs (Lirman & Fong, 2007). Indeed, these offshore reefs are the only areas in the FRT for which appreciable levels of UV light reach the benthos. Thus, the offshore reefs may be more susceptible to the combined effects of UV and thermal oxidative stresses. However, there are complex interactions between water-column light absorption and temperature (e.g., radiative transfer) that confound a simple ‘either/ or’ explanation of UV and thermal stress on coral communities. Further, as reefs are often net exporters of CDOM (Boss & Zaneveld, 2003), these new products could be used to investigate the change in Kd(UV) or ag(UV) for waters flowing over a reef, potentially providing insight into reef function. A thorough investigation of the direct and long-term impacts of satellite-derived SST and light availability on coral reef health in the Florida Keys is beyond the scope of this work. The availability of long-term UV and visible light data in the Florida Keys allows for statistical assessment of the spatio-temporal patterns in light conditions within the FRT waters (see Section 5.1; Barnes et al., 2013). Such analyses can differentiate coral reef locations with distinct Kd(UV) and Kd(VIS) regimes, either according to mean condition or by annual or seasonal variability. The Florida Keys National Marine Sanctuary (FKNMS) is currently undergoing re-zonation of specific protection areas, in an effort to more effectively manage and protect the marine resources (NOAA 2007, 2012). As with the Kd(VIS) patterns,

synoptic accounting of the UV light availability in the FRT and its effect on corals could be valuable in informing this rezoning process. Beyond the FRT, the range of in situ values used in the training dataset and subsequently modeled using this approach are similar to other reported coral environments. Using Kd(UV) as an example, our training fits within measurements reported by Dunne and Brown (1996), and Shick, Lesser, and Jokiel (1996), for a Maldives atoll lagoon [Kd(UVB) = 0.39 m−1], Thailand coastal island [Kd(UVB) = 0.78 m−1], Thailand inshore fringing reef [Kd(UVB) = 1.61 m−1], Belize Barrier Reef [Kd(UVB) = 0.21 m−1], Hawaii estuary [Kd(320) approximately 1 m−1], and offshore Hawaii reef [Kd(320) = 0.2 m−1]. Also, the range of Kd (at 305, 320, 340, and 380 nm) observed at coral reef locations surrounding Malaysia (Kuwahara, Nakajima, Othman, Kushairi, & Toda, 2010) almost exactly matches the spread of the training dataset used in this FRT study. As such, beyond the general applicability of the approach to other environments, the particular parameterization of the models presented here may be directly applicable to coral reef environments worldwide. 5.3. Algorithm novelty — importance of stepwise model selection While the use of EOF methods to derive water quality parameters such as Kd is not new, the method presented here shows improvement over previously published EOF methods primarily through its implementation of the stepwise, forward addition procedure for selection of the EOF modes to be retained in the models. This approach allows for parsimonious and statistically defensible mode selection for model application to the various measured parameters discussed in this paper as well as for application of this EOF method to any other region. When applied to each of the in situ measured parameters, another important consequence of this new approach in implementing the EOF method is a ranking of the relative importance of the EOF modes

B.B. Barnes et al. / Remote Sensing of Environment 140 (2014) 519–532

in explaining variations in the parameter of interest. As seen in previous works (Craig et al., 2012; Fichot et al., 2008; Mueller, 1976), the mode selection varied not only by IOP or Kd, but also by wavelength within an IOP or Kd (Fig. 9). In contrast to visual interpretation of the modes, this listing gives a more definitive indication of the meaning (in terms of effect on water parameters) of various portions of the eigenvectors (loadings by wavelength) for each mode (Fig. 10). For example, this analysis allows definitive assessment of mode 4 as representing changes in bRRSN according to ad. Above approximately 550 nm, these changes are also concomitant with variability due to aφ. Further, low detrital contributions (Fig. 3) are confirmed by the modes most often selected for ad retrievals (4 and 9) which account for a total of 1.46% of the total variance in bRRSN. Not surprisingly, Kd retrievals do not generally include these modes, further highlighting the relatively small contributions of ad to total attenuation in the FRT. It is important to note that most of the EOF modes discussed here are not ‘significant’ according to common selection rules (e.g., Rule N; Preisendorfer, Zwiers, & Barnett, 1982; von Storch & Zwiers, 1999). Using the Kd305 matchups (n = 29) as an example, we performed 1000 EOF analyses on randomly generated noise (simulating 10 bRRSN bands for those 29 matchups). From these EOFs, we calculated the 95th percentile for the variance explained by each mode. As such, only mode 1 in our model explains more variance in bRRSN than would be predicted for random noise, and is thus ‘significant’ at α = 0.05. However, this does not mean that subsequent modes are noise as predictors in the stepwise multiple regression. The forward selection process only includes independent variables that are significantly contributing to the variance in the dependent variable (in this case, log[Kd(305)]), and the stopping procedures ensure parsimony. To demonstrate this, we have performed the stepwise multiple regression model for Kd(305) data 1000 times, each with a different random (noise) predictor added to the 10 EOF modes. The noise variable was only selected in 53 cases (5.3%), very close to the acceptable significance level of the procedure.

529

In contrast, EOF mode 3 (a ‘non-significant’ mode that explains only 1.77% of the variance in bRRSN) was selected in 99.3% of iterations.

5.4. Algorithm performance — resilience to instrument calibration uncertainties As mentioned in the methods section, no attempts were made to quantify differences between performance of the various instruments used to collect the in situ data. The reason for this lack of intercalibration stems mostly from the fact that, due to the various data sources, instruments were not used concurrently. However, given the diversity of instrumentation used in the creation of the training dataset, the performance of the EOF model for prediction of many parameters indicates acceptable intercalibration errors between instruments. The EOF approach described here is also likely resilient to certain calibration errors or drift of the satellite instrument, mostly stemming from the stepwise procedure for model selection. EOF modes are included in the final model only if they parsimoniously explain a significant portion of the variance in the dependent variable. Although the few MODIS/A spectra with negative RRS (indicating failure of the atmospheric correction) were excluded from this analysis in effort to have the most accurate possible training dataset, performance of the models did not generally suffer if such data were not excluded. As the normalization procedure (see Section 3.3; Craig et al., 2012) focuses on the spectral shape, negative RRS (which have no physical meaning) may still carry some information useful in modeling water parameters. More importantly, a slow drift of calibration errors in one or more of the satellite RRS bands could potentially be partitioned into one particular EOF mode. This is especially true if the matchups extend beyond the start of the instrument drift (i.e., some of the matchup RRS data are unaffected by the drift). Since the variation in the dependent variable is not likely to be explained by an EOF mode which has partitioned

Kd

ag

ad

aφ

Fig. 9. Parsimoniously selected EOF modes for all parameters (by wavelength) tested. Thickness of the lines indicate relative contribution to the model, as determined by partial F-statistic (thickest = highest F-statistic, medium thickness = second highest, thinnest = third or lower, yet still significant).

530

B.B. Barnes et al. / Remote Sensing of Environment 140 (2014) 519–532

Fig. 10. EOF modes for Kd(305) matchups, showing percent contribution to the total variance in bRRSN. Modes did not greatly differ for the various parameters tested.

the instrument drift, it is unlikely that the calibration error would be represented in the final model. Actual conditions of the MODIS/A instrument blue bands (412 and 443 nm) allowed for testing of such resilience of this approach. Starting in early 2011, deterioration in these bands required a change in the radiometric calibration (Bryan Franz, NASA/GSFC, personal comm.). This change was implemented in SeaDAS version 6.4 but not in SeaDAS version 6.2. As such, by processing all MODIS/A data (especially from 2011 to 2012) with the outdated SeaDAS version (6.2), we were able to extract RRS which included calibration drift in the 412 and 443 nm bands. These RRS data were analyzed in the same manner as described above. Interestingly, most of the resulting EOF modes (and subsequent model performances) were quite similar to those from the SeaDAS version 6.4 processing (Fig. 10), with one exception being EOF mode 6, which explained 0.21% of the total variance in Kd(305) b RRSN matchups. This mode showed indications of seemingly random variance in all of the MODIS/A bands except those at 412 and 443 nm. For this mode, near-zero loadings in the blue bands could reflect stability relative to each other (i.e., improper calibration) independent of the variance in all other bands. Also, only 14 of the 1158 (1.2%) significant models (one model for each parameter) included EOF mode 6 as parsimoniously

explaining variance in that parameter. Given that EOF mode 6 showed a spectral shape that fits expectations for the instrument drift and generally does not explain in situ parameters, we conclude that EOF mode 6 likely partitioned the variance in RRS associated with the instrument drift. The resilience of the EOF model to instrument degradation or calibration uncertainties is particularly important to construct seamless environmental data records (EDRs) of Kd and ag using more than one satellite instrument. For example, at the time of this writing, MODIS/A exceeded its 5-year mission design by 6 years. In case of unexpected malfunction, the more recently launched Visible Infrared Imaging Radiometer Suite (VIIRS; 2011–present) and other future sensors (which may extend measurements to 340 or 350nm; e.g. Fishman et al., 2012) may be used to continue the MODIS/A observations to form multi-decadal Kd and ag EDRs to address environmental problems associated with changing climate and increased anthropogenic activities. 6. Conclusions The new approach in implementing the EOF method developed in this study improves over its predecessors in its requirements for parsimonious model selection. The method has been applied directly

B.B. Barnes et al. / Remote Sensing of Environment 140 (2014) 519–532

to satellite-derived RRS data, with validation procedures designed to estimate performance of its prediction capacity. The approach should be applicable to other regions, although performance is limited beyond the bounds of the training dataset. The success of the approach to derive Kd(UV) and ag(UV) for a large dynamic range in the study region shows its applicability in optically shallow environments and outside the range of the measured RRS. The contrasting seasonal changes in Kd(UV) and Kd(VIS) suggest that the former cannot be derived from the latter through spectral extrapolation or empirical regression, as they may be controlled by different forces. The contrast also highlights the need for more research and monitoring of the long-term and large-scale effects of UV light reaching the shallow-water benthos. The established MODIS/A based long-term Kd(UV) EDR makes such research possible, especially when combined with other environmental variables such as SST and surface radiation. Acknowledgements This work was funded by the U.S. NASA through its Gulf of Mexico program, Ocean Biology and Biogeochemistry program, Decision Support program, and Water and Energy Cycle program. The authors wish to thank NOAA National Geophysical Data Center (Wessel & Smith, 1996) for providing coastline shapefiles, Florida Fish and Wildlife Conservation Commission's Fish and Wildlife Research Institute for providing benthic habitat maps, USGS for providing bathymetry data, and NASA/GSFC Ocean Biology Processing Group (OBPG) for providing MODIS data and SeaDAS software, as well as two anonymous reviewers whose comments helped to greatly enhance this manuscript. Collection of in situ data was funded, in part, by NOAA-NURC subcontract No. 2004-19B and the U.S. Environmental Protection Agency Gulf Ecology Division Grant No. X796465607-0. Several coauthors work for the U.S. Environmental Protection Agency, but the views expressed in this article do not necessarily represent the views or policies of EPA. Mention of trade names, products, or services does not convey, and should not be interpreted as conveying, official EPA approval, endorsement or recommendation. Support for N. Melo was provided by NOAA's AOML. References Ayoub, L. M. (2009). Can colored dissolved organic material protect coral reefs by reducing exposure to ultraviolet radiation? (PhD dissertation). : University of South Florida. Barnes, B. B., & Hu, C. (2013). A hybrid cloud detection algorithm to improve MODIS sea surface temperature data quality and coverage over the Eastern Gulf of Mexico. IEEE Transactions on Geoscience and Remote Sensing, 51, 3273–3285. Barnes, B. B., Hu, C., & Muller-Karger, F. (2011). An improved high-resolution SST climatology to assess cold water events off Florida. Geoscience and Remote Sensing Letters, IEEE, 8, 769–773. Barnes, B. B., Hu, C., Schaeffer, B.A., Lee, Z., Palandro, D. A., & Lehrter, J. C. (2013). MODIS-derived spatiotemporal water clarity patterns in optically shallow Florida Keys waters: A new approach to remove bottom contamination. Remote Sensing of Environment, 134, 377–391. Bissett, W. P., Patch, J. S., Carder, K. L., & Lee, Z. P. (1997). Pigment packaging and Chl a-specific absorption in high-light oceanic waters. Limnology and Oceanography, 42, 961–968. Blanchet, F. G., Legendre, P., & Borcard, D. (2008). Forward selection of explanatory variables. Ecology, 89, 2623–2632. Boss, E., & Zaneveld, J. R. V. (2003). The effect of bottom substrate on inherent optical properties: evidence of biogeochemical processes. Limnology and Oceanography, 48, 346–354. Brando, V. E., & Dekker, A. G. (2003). Satellite hyperspectral remote sensing for estimating estuarine and coastal water quality. IEEE Transactions on Geoscience and Remote Sensing, 41, 1378–1387. Bricaud, A., Morel, A., & Prieur, L. (1981). Absorption by dissolved organic matter of the sea (yellow substance) in the UV and visible domains. Limnology and Oceanography, 26, 43–53. Campbell, J. W. (1995). The lognormal distribution as a model for bio-optical variability in the sea. Journal of Geophysical Research, Oceans, 100, 13237–13254. Cannizzaro, J. P., & Carder, K. L. (2006). Estimating chlorophyll a concentrations from remote-sensing reflectance in optically shallow waters. Remote Sensing of Environment, 101, 13–24. Cannizzaro, Carder, K. L., Kelble, C. R., Melo, N., Johns, E. M., Vargo, G. A., et al. (in press). On the accuracy of SeaWiFS ocean color data products on the West Florida Shelf. Journal of Coastal Research. http://dx.doi.org/10.2112/JCOASTRES-D-12-00223.1 (in press).

531

Carder, K., Cannizzaro, J. P., & Lee, Z. (2005). Ocean color algorithms in optically shallow waters: Limitations and improvements. In R. J. Frouin, M. Babin, & S. Sathyendranath (Eds.), Proc. SPIE 5885, Remote Sensing of the Coastal Environment, 588506. Bellingham, WA. Cohen, J., Cohen, P., West, S. G., & Aiken, L. S. (2003). Applied multiple regression/correlation analysis for the behavioral sciences. Mahwah, N.J.: L. Erlbaum Associates. Craig, S. E., Jones, C. T., Li, W. K. W., Lazin, G., Horne, E., Caverhill, C., et al. (2012). Deriving optical metrics of coastal phytoplankton biomass from ocean colour. Remote Sensing of Environment, 119, 72–83. DeGrandpre, M.D., Vodacek, A., Nelson, R. K., Bruce, E. J., & Blough, N. V. (1996). Seasonal seawater optical properties of the U.S. Middle Atlantic Bight. Journal of Geophysical Research, Oceans, 101, 22727–22736. Dunne, R. P., & Brown, B. E. (1996). Penetration of solar UVB radiation in shallow tropical waters and its potential biological effects on coral reefs; Results from the central Indian Ocean and Andaman Sea. Marine Ecology-Progress Series, 144, 109–118. Dustan, P., & Halas, J. (1987). Changes in the reef-coral community of Carysfort reef, Key Largo, Florida: 1974 to 1982. Coral Reefs, 6, 91–106. Fichot, C. G., & Benner, R. (2012). The spectral slope coefficient of chromophoric dissolved organic matter (S275–295) as a tracer of terrigenous dissolved organic carbon in river-influenced ocean margins. Limnology and Oceanography, 57, 1453–1466. Fichot, C. G., Sathyendranath, S., & Miller, W. L. (2008). SeaUV and SeaUVC: Algorithms for the retrieval of UV/Visible diffuse attenuation coefficients from ocean color. Remote Sensing of Environment, 112, 1584–1602. Fischer, J., Doerffer, R., & Grassl, H. (1986). Factor analysis of multispectral radiances over coastal and open ocean water based on radiative transfer calculations. Applied Optics, 25, 448–456. Fishman, J., Iraci, L. T., Al-Saadi, J., Change, K., Chavez, F., Chin, M., et al. (2012). The United States' next generation of atmospheric composition and coastal ecosystem measurement: NASA's Geostationary Coastal and Air Pollution Events (GEO-CAPE) Mission. BAMS, 1547–1566. http://dx.doi.org/10.1175/BAMS-D-11-00201.1. Fisk, D. A., & Done, T. J. (1985). Taxonomic and bathymetric patterns of bleaching in corals, Myrmidon Reef (Queensland). Proc 5th Intl Coral Reef Congr, 6. (pp. 149–154). Foyer, C. H., Lelandais, M., & Kunert, K. J. (1994). Photooxidative stress in plants. Physiologia Plantarum, 92, 696–717. Franz, B.A., Kwiatkowska, E. J., Meister, G., & McClain, C. R. (2008). Moderate resolution imaging spectroradiometer on terra: Limitations for ocean color applications. Journal of Applied Remote Sensing, 2, 023525. http://dx.doi.org/10.1117/1.2957964. Gauch, H. G. J. (1993). Prediction, parsimony and noise. American Scientist, 81, 468–478. Ginsburg, R. N., & Shinn, E. A. (1964). Distribution of reef-building community in Florida and the Bahamas. American Association of Petroleum Geologists Bulletin, 48, 527. Ginsburg, R. N., & Shinn, E. A. (1993). Preferential distribution of reefs in the Florida Reef Tract: the past is key to the present. In R. N. Ginsburg (Ed.), Global Aspects of Coral Reefs. Health, Hazards, and History (pp. 21–26). Miami, FL: University of Miami. Gordon, H. R. (1989). Can the Lambert–Beer Law Be applied to the diffuse attenuation coefficient of ocean water? Limnology and Oceanography, 34, 1389–1409. Gower, J. F. R., Lin, S., & Borstad, G. A. (1984). The information content of different optical spectral ranges for remote chlorophyll estimation in coastal waters. International Journal of Remote Sensing, 5, 349–364. Gregg, W. W., & Casey, N. W. (2004). Global and regional evaluation of the SeaWiFS chlorophyll dataset. Remote Sensing of Environment, 93, 463–479. Häder, D. P., Kumar, H. D., Smith, R. C., & Worrest, R. C. (1998). Effects on aquatic ecosystems. Journal of Photochemistry and Photobiology B: Biology, 46, 53–68. Hochberg, E. J., Atkinson, M. J., & Andréfouët, S. (2003). Spectral reflectance of coral reef bottom-types worldwide and implications for coral reef remote sensing. Remote Sensing of Environment, 85, 159–173. Hoegh-Guldberg, O., & Jones, R. J. (1999). Photoinhibition and photoprotection in symbiotic dinoflagellates from reef-building corals. Marine Ecology Progress Series, 183, 73–86. Hooker, S. B., Esaias, W. E., Feldman, G. C., Gregg, W. W., & McClain, C. R. (1992). An overview of SeaWiFS and ocean color. NASA Tech. Memo., Vol. 104566, Goddard Space Flight Center Greenbelt, MD: National Aeronautics and Space Administration. Hooker, S. B., Lazin, G., Zibordi, G., & McLean, S. (2002). An evaluation of above- and in-water methods for determining water-leaving radiances. Journal of Atmospheric and Oceanic Technology, 19, 486–515. Hu, C. (2008). Ocean color reveals sand ridge morphology on the West Florida Shelf. IEEE Geoscience and Remote Sensing Letters, 5, 443–447. Hu, C., Muller-Karger, F. E., Murch, B., Myhre, D., Taylor, J., Luerssen, R., et al. (2009). Building an automated integrated observing system to detect sea surface temperature anomaly events in the Florida Keys. IEEE Transactions on Geoscience and Remote Sensing, 47, 1607–1620. Ioannou, I., Gilerson, A., Gross, B., Moshary, F., & Ahmed, S. (2011). Neural network approach to retrieve the inherent optical properties of the ocean from observations of MODIS. Applied Optics, 50, 3168–3186. Ioannou, I., Gilerson, A., Gross, B., Moshary, F., & Ahmed, S. (2013). Deriving ocean color products using neural networks. Remote Sensing of Environment, 134, 78–91. Jamet, C., Loisel, H., & Dessailly, D. (2012). Retrieval of the spectral diffuse attenuation coefficient Kd(λ) in open and coastal ocean waters using a neural network inversion. Journal of Geophysical Research, Oceans, 117, C10023. Johannessen, S.C., Miller, W. L., & Cullen, J. J. (2003). Calculation of UV attenuation and colored dissolved organic matter absorption spectra from measurements of ocean color. Journal of Geophysical Research, Oceans, 108, 3301. Jokiel, P. L. (1980). Solar ultraviolet radiation and coral reef epifauna. Science, 207, 1069–1071. Kelble, C., & Boyer, J. N. (2007). Southern estuaries hypothesis cluster: Water quality. Comprehensive Everglades Restoration Plan Assessment Team Final 2007 System Status Report.

532

B.B. Barnes et al. / Remote Sensing of Environment 140 (2014) 519–532

Kiefer, D. A., & SooHoo, J. B. (1982). Spectral absorption by marine particles of coastal waters of Baja California. Limnology and Oceanography, 27, 492–499. Kishino, M., Takahashi, M., Okami, N., & Ichimura, S. (1985). Estimation of the spectral absorption coefficients of phytoplankton in the sea. Bulletin of Marine Science, 37, 634–642. Kuwahara, V. S., Nakajima, R., Othman, B. H. R., Kushairi, M. R. M., & Toda, T. (2010). Spatial variability of UVR attenuation and bio-optical factors in shallow coral-reef waters of Malaysia. Coral Reefs, 29, 693–704. Lapointe, B., & Clark, M. (1992). Nutrient inputs from the watershed and coastal eutrophication in the Florida keys. Estuaries and Coasts, 15, 465–476. Lee, Z. P., Carder, K. L., & Arnone, R. (2002). Deriving inherent optical properties from water color: A multi-band quasi-analytical algorithm for optically deep waters. Applied Optics, 41, 5755–5772. Lee, Z., Hu, C., Shang, S., Du, K., Lewis, M., Arnone, R., & Brewin, R. (2013). Penetration of UV-visible solar radiation in the global oceans: Insights from ocean color remote sensing. J. Geophys. Res. Oceans, 118. Lesser, M. P., & Lewis, S. (1996). Action spectrum for the effects of UV radiation on photosynthesis in the hermatypic coral Pocillopora damicornis. Marine Ecology Progress Series, 134, 171–177. Lesser, M. P., Stochaj, W. R., Tapley, D. W., & Shick, J. M. (1990). Bleaching in coral reef anthozoans: Effects of irradiance, ultraviolet radiation, and temperature on the activities of protective enzymes against active oxygen. Coral Reefs, 8, 225–232. Lirman, D., & Fong, P. (2007). Is proximity to land-based sources of coral stressors an appropriate measure of risk to coral reefs? An example from the Florida Reef Tract. Marine Pollution Bulletin, 54, 779–791. Lirman, D., Schopmeyer, S., Manzello, D., Gramer, L. J., Precht, W. F., Muller-Karger, F., et al. (2011). Severe 2010 cold-water event caused unprecedented mortality to corals of the Florida Reef tract and reversed previous survivorship patterns. PLoS ONE, 6, e23047. Moran, M.A., & Zepp, R. G. (1997). Role of photoreactions in the formation of biologically labile compounds from dissolved organic matter. Limnology and Oceanography, 42, 1307–1316. Mueller, J. L. (1976). Ocean color spectra measured off the Oregon coast: Characteristic vectors. Applied Optics, 15, 394–402. Mueller, J. L., & Fargion, G. S. (2002). Ocean optics protocols for satellite ocean color sensor validation, revision 3. Greenbelt, MD: NASA Goddard Space Flight Center Center. National Oceanic and Atmospheric Administration (2012). Florida Keys National Marine Sanctuary marine zoning and regulatory review, summary of scoping comments. Office of National Marine Sanctuaries, National Oceanic and Atmospheric Administration. National Oceanic and Atmospheric Association (2007). Florida Keys National Marine Sanctuary revised management plan: National Ocean Service, US Department of Commerce. Patt, F. S., Barnes, R. A., Eplee, R. E., Jr., Franz, B.A., Robinson, W. D., Feldman, G. C., et al. (2003). Algorithm updates for the fourth SeaWiFS data reprocessing. In S. B. Hooker, & E. R. Firestone (Eds.), NASA Technical Memorandum 2003–206892. Pegau, W. S., Zaneveld, J. R. V., Mitchell, B. G., Kahru, M., Wieland, J., & Stramska, M. (2003). Inherent optical properties: Instruments, characterization, field measurements, and data analysis protocols. In J. L. Mueller, G. S. Fargion, & C. R. McClain (Eds.), Ocean optics protocols for satellite ocean color sensor validation, revision 4 (pp. 1–76). Greenbelt, MD: NASA Goddard Space Flight Center. Phinn, S. R., Dekker, A. G., Brando, V. E., & Roelfsema, C. M. (2005). Mapping water quality and substrate cover in optically shallow coastal and reef waters: An integrated approach. Marine Pollution Bulletin, 51, 459–469. Preisendorfer, R. W., Zwiers, F. W., & Barnett, T. P. (1982). Foundations of principal component selection rules. Scripps Institution of Oceanography Reference Series, 81–4, La Jolla, CA: Scripps Institution of Oceanography (192 pp.). Roesler, C. S. (1998). Theoretical and experimental approaches to improve the accuracy of particulate absorption coefficients derived from the quantitative filter technique. Limnology and Oceanography, 43, 1649–1660. Roesler, C. S., Perry, M. J., & Carder, K. L. (1989). Modeling in situ phytoplankton absorption from total absorption spectra in productive inland marine waters. Limnology and Oceanography, 34, 1510–1523.

Sathyendranath, S., Hoge, F. E., Platt, T., & Swift, R. N. (1994). Detection of phytoplankton pigments from ocean color: Improved algorithms. Applied Optics, 33, 1081–1089. Sathyendranath, S., Prieur, L., & Morel, A. (1989). A three-component model of ocean colour and its application to remote sensing of phytoplankton pigments in coastal waters. International Journal of Remote Sensing, 10, 1373–1394. Schaeffer, B.A., Conmy, R. N., Aukamp, J., Craven, G., & Ferer, E. J. (2011). Organic and inorganic matter in Louisiana coastal waters: Vermilion, Atchafalaya, Terrebonne, Barataria, and Mississippi regions. Marine Pollution Bulletin, 62, 415–422. Schaeffer, B.A., Hagy, J.D., Conmy, R. N., Lehrter, J. C., & Stumpf, R. P. (2012). An approach to developing numeric water quality criteria for coastal waters using the SeaWiFS satellite data record. Environmental Science & Technology, 46, 916–922. Shank, G. C., Lee, R., Vähätalo, A., Zepp, R. G., & Bartels, E. (2010). Production of chromophoric dissolved organic matter from mangrove leaf litter and floating Sargassum colonies. Marine Chemistry, 119, 172–181. Shank, G. C., Zepp, R. G., Vähätalo, A., Lee, R., & Bartels, E. (2010). Photobleaching kinetics of chromophoric dissolved organic matter derived from mangrove leaf litter and floating Sargassum colonies. Marine Chemistry, 119, 162–171. Shick, J. M., Lesser, M. P., & Jokiel, P. L. (1996). Effects of ultraviolet radiation on corals and other coral reef organisms. Global Change Biology, 2, 527–545. Smith, R. C., & Baker, K. S. (1981). Optical properties of the clearest natural waters (200–800 nm). Applied Optics, 20, 177–184. Smyth, T. J. (2011). Penetration of UV irradiance into the global ocean. Journal of Geophysical Research, Oceans, 116, C11020. Stabenau, E. R., Zepp, R. G., Bartels, E., & Zika, R. G. (2004). Role of the seagrass Thalassia testudinum as a source of chromophoric dissolved organic matter in coastal south Florida. Marine Ecology Progress Series, 282, 59–72. Strong, A. E., Liu, G., Meyer, J., Hendee, J. C., & Sasko, D. (2004). Coral Reef Watch 2002. Bulletin of Marine Science, 75, 259–268. Strong, A. E., Liu, G., Skirving, W., & Eakin, C. M. (2011). NOAA's Coral Reef Watch program from satellite observations. Annals of GIS, 17, 83–92. Szmant, A.M., & Forrester, A. (1996). Water column and sediment nitrogen and phosphorus distribution patterns in the Florida Keys, USA. Coral Reefs, 15, 21–41. Toole, D. A., & Siegel, D. A. (2001). Modes and mechanisms of ocean color variability in the Santa Barbara Channel. Journal of Geophysical Research, Oceans, 106, 26985–27000. von Storch, H., & Zwiers, F. W. (1999). Statistical analysis in climate research. Cambridge, New York: Cambridge University Press. Werdell, P. J., & Roesler, C. S. (2003). Remote assessment of benthic substrate composition in shallow waters using multispectral reflectance. Limnology and Oceanography, 48, 557–567. Wessel, P., & Smith, W. H. F. (1996). A global, self-consistent, hierarchical, high-resolution shoreline database. Journal of Geophysical Research, Solid Earth, 101, 8741–8743. West, J. M., & Salm, R. V. (2003). Resistance and resilience to coral bleaching: Implications for coral reef conservation and management. Conservation Biology, 17, 956–967. Yentsch, C. S. (1962). Measurement of visible light absorption by particulate matter in the ocean. Limnology and Oceanography, 7, 207–217. Yentsch, C. S., Yentsch, C. M., Cullen, J. J., Lapointe, B., Phinney, D. A.M., & Yentsch, S. W. (2002). Sunlight and water transparency: Cornerstones in coral research. Journal of Experimental Marine Biology and Ecology, 268, 171–183. Zepp, R. G., Shank, G. C., Stabenau, E., Patterson, K. W., Cyterski, M., Fisher, W., et al. (2008). Spatial and temporal variability of solar ultraviolet exposure of coral assemblages in the Florida Keys: Importance of colored dissolved organic matter. Limnology and Oceanography, 53, 1909–1922. Zhao, J., Barnes, B. B., Nelo, M., English, D., Lapointe, B., Muller-Karger, F., et al. (2013). Assessment of satellite-derived diffuse attenuation coefficients and euphotic depths in south Florida coastal waters. Remote Sensing of Environment, 131, 38–50. Zhao, J., Hu, C., Lapointe, B., Melo, N., Johns, E., & Smith, R. (2013). Satellite-observed black water events off Southwest Florida: Implications for coral reef health in the Florida Keys National Marine Sanctuary. Remote Sensing, 5, 415–431.