Salinity is a critical factor in understanding and predicting physical, chemical, and biological processes in the coastal ocean, where these processes vary considerably in time and space.1 Salinity is a major contributor in determining the water density structure, which in turn affects the circulation and stratification.2 Biologically, salinity influences the distribution, survival, and growth of several commercially important organisms in Chesapeake Bay, such as oysters3,4 and blue crabs.5,6 It is also employed as a variable in all the habitat suitability models that have been developed to forecast organisms in the Chesapeake Bay, i.e., sea nettles,7 harmful algae, and water-borne pathogens.8,9 Consequently, estimating salinity synoptically and accurately in near real-time would improve the ecological forecasts of these organisms.
Unfortunately, estimating salinity of sufficient quality and resolution in coastal waters is difficult. Hydrodynamic models of the coastal ocean, such as the National Oceanic and Atmospheric Administration (NOAA) Chesapeake Bay Operational Forecast System (CBOFS), currently provide routine predictions of salinity on a timely basis,10 but these simulations can have inaccuracies, requiring additional sources of data to constrain the models and their output. The spatial coverage of in situ salinity measurements in these waters is inadequate to fully capture the complexity of coastal ocean processes. Neither aerial surveys11 nor satellite microwave measurements, e.g., from the Aquarius satellite,12 supply salinity estimates at temporal or spatial resolutions sufficient to resolve the distribution of salinity in dynamic coastal regions.
Satellite ocean color radiometry may offer a method to provide estimates of sea surface salinity (SSS) at a medium spatial resolution (250 m to 1 km) in coastal waters for direct application and for assimilation into medium resolution hydrodynamic models. The technique utilizes the inverse linear relationship between salinity and the absorption of colored dissolved organic matter (CDOM), which absorbs light at ultraviolet to blue wavelengths and can be estimated from satellite ocean color radiometric measurements. The salinity–CDOM relationship arises from the mixing of seawater with freshwater rich in terrestrial plant-derived compounds.13,14 These seawater and freshwater end members are assumed to mix conservatively, i.e., without internal loss or gain of CDOM or salt.15 Because the hydrodynamics and biogeochemistry of these end members are regulated by local processes, the salinity–CDOM relationship is locally dependent and must be developed and applied on a region-by-region basis.16 In addition, some factors may disrupt the conservative mixing assumption: nonfreshwater CDOM sources may include coastal phytoplankton production, zooplankton grazing (i.e., excretion), and microbial decomposition,14,17 and CDOM sinks may include in situ microbial consumption18 and photo-oxidation in stratified surface waters.19 However, the salinity–CDOM relationship has been well documented for estuarine and coastal waters, and its coefficient of determination can be very high.13,20
The technique to estimate SSS from the salinity–CDOM relationship varies. In one approach, an explicit statistical relationship between CDOM absorption and SSS is developed empirically from in situ measurements, usually a simple linear or multiple linear relationship. This empirical algorithm is then applied to CDOM absorption estimated from satellite ocean color measurements to calculate SSS. This approach has been used in many estuarine and coastal environments: the Clyde Sea,21 Florida Bay,22 Columbia River plume,23 and East China Sea,16,24,25 among others. Reference 16 points out that the biggest challenge in salinity retrieval using this approach is the accurate estimation of satellite-derived CDOM absorption.
Another more recent approach uses remote sensing reflectance, , rather than satellite-derived CDOM absorption, where represents wavelength. is defined as the ratio of the radiance exiting the water surface that originated from underwater light to the downwelling irradiance incident onto the water surface.26 The approach to use avoids the error introduced by estimating satellite CDOM absorption. Furthermore, utilizing a statistical combination of from multiple ocean color bands both inside and outside of the ultraviolet-to-blue spectrum may bring additional information to the statistical relationship. These approaches have been used for the U.S. mid-Atlantic coast,27 Chesapeake Bay,28 and Bohai Sea29 to retrieve SSS.
In Ref. 27, Geiger et al. constructed an artificial neural network (ANN) to establish a relationship between ocean color and SSS for several regions within the U.S. mid-Atlantic coast. In Ref. 28, Urquhart et al. built upon the work of Geiger et al. of Ref. 27 by comparing eight different statistical approaches, including an ANN, to determine whether the ANN approach or a different statistical approach is better suited for calculating SSS from for Chesapeake Bay. In Ref. 28, Urquhart et al. found that the generalized additive model (GAM) provided SSS estimates with the least error in comparison with in situ salinity measurements. Both approaches use estimates from a variety of ocean color bands from the Moderate Resolution Imaging Spectroradiometer (MODIS) on NASA’s Aqua spacecraft. For a comparison of the ANN and GAM statistical techniques, see Ref. 28. Hereinafter, Ref. 27 of Geiger et al.’s approach will be referred to as the ANN algorithm and Ref. 28 of Urquhart et al.’s approach will be referred to as the GAM algorithm.
Our study evaluates the error of these two satellite algorithms for Chesapeake Bay by comparing their satellite SSS retrievals to in situ buoy SSS observations from a period outside the algorithms’ training and testing period in order to independently assess each algorithm’s performance. We also evaluate SSS predicted by the CBOFS hydrodynamic model against the in situ buoy observations. Based on these results, we assess if the satellite retrievals could improve hydrodynamic model predictions of SSS through data assimilation. If the error of satellite ocean-color-derived SSS is significantly lower than the hydrodynamic model-predicted SSS error, then these satellite retrievals could be assimilated by the hydrodynamic model to improve its predictions of salinity. Given the effort of the remote sensing community in developing ocean color radiometric SSS algorithms for coastal areas, which includes recent advancements in methodology to emphasize rather than derived CDOM, assessments of these SSS retrievals for coastal modeling applications may lead to better physical and ecological forecasts and improved understanding and decision-making in coastal environments.
Methods and Materials
The Chesapeake Bay is a partially mixed estuary with southward flowing lower-density freshwater and northward flowing higher-density seawater separated by the pycnocline.30 This circulation pattern results in a strong salinity gradient increasing from north to south with a salinity range of 0 to 35 (unitless). The Susquehanna River provides the largest freshwater supply to the Bay, contributing of the Bay’s freshwater.31
Satellite ocean color radiometric measurements
Level 2 swath files of MODIS-Aqua over Chesapeake Bay for the period of June 2011 to February 2013 were obtained from NASA’s Goddard Space Flight Center (GSFC).32 Reprocessing R2013.0 was used, which incorporated the latest retrospective instrument calibration. The satellite swath files were then converted to Mercator-projection mapped files in NOAA’s CoastWatch HDF format. MODIS-Aqua sea surface temperature, an input to the ANN algorithm, was also obtained from NASA GSFC for the same Level 2 swaths.
In situ buoy salinity
In situ salinity observations of the Chesapeake Bay Interpretive Buoy System (CBIBS) for the period of June 2011 to February 2013 were obtained from the NOAA Chesapeake Bay Office.33 CBIBS is a network of 11 buoys in the Chesapeake Bay, which collect meteorological, oceanographic, and water-quality data continuously. Five buoys were selected in the bay main stem (Fig. 1) to represent a range of estuarine salinities that fall within the geographic location of the data used to train the ANN and GAM algorithms. The salinity sensor is mounted on the buoy at a depth of 0.5 m. This depth of measurement is within a mean optical depth of 0.89 m for Chesapeake Bay,28 allowing comparison of the in-water salinity measurements with SSS estimations calculated from satellite . CBIBS salinity data are available as hourly averages. Due to the potential for biofouling of the buoys’ in-water instrumentation, the buoys’ sensors are routinely swapped at monthly intervals.
Hydrodynamically modeled salinity
The CBOFS is the current operational hydrodynamic model for the Chesapeake Bay developed by NOAA’s National Ocean Service (NOS) Coast Survey Development Laboratory10 and based on Rutgers University’s Regional Ocean Modeling System.34 Grid resolution ranges from 34 to 4895 m in the -direction and from 29 to 3380 m in the -direction, with finer resolution near the coast. The vertical grid follows the terrain and consists of 20 model levels. Hourly nowcast fields were retrieved from NOS’s six-hourly files35 for the period of June 2011 to February 2013, and the surface salinity was extracted for the five buoy locations.
Satellite Sea Surface Salinity Retrieval Algorithms
The ANN algorithm was formulated specifically for this project by Geiger (unpublished) following the methodology of Ref. 27. The predictors in the ANN algorithm include or ratios of at wavelengths of 412, 443, 412/547, 443/547, and 488/547 nm, as well as longitude, latitude, and sea surface temperature (SST). The eight ANN predictors were chosen based on principle component analysis, in which correlation was determined between in situ surface salinity and a possible set of 17 predictors, including satellite radiometry, position, and environmental parameters (SST, chlorophyll, water depth, river discharge, tides, and year-day). The parameters most correlated with surface salinity were chosen for the neural network training.27 From a matchup data set of 769 matchups between coincident MODIS-Aqua data and in situ surface salinity data, 399 randomly selected matchups were used to train the ANN. The in situ salinity data were collected by underway ships over a 6-year period (2003 to 2008) from Chesapeake Bay waters with salinities ranging from 9.58 to 32.71. The algorithm was evaluated with the remaining 370 matchups not used in the training, by comparing these withheld in situ surface salinities with the coincident algorithm-estimated SSS, which resulted in a mean absolute error (MAE) of 0.82, a root-mean-square error (RMSE) of 1.12, and a correlation coefficient of 0.968.
The GAM algorithm includes predictors of at 412-, 443-, 469-, 488-, 531-, 547-, 555-, 645-, 667-, and 678-nm wavelengths, as well as longitude and latitude.28 In Ref. 28, Urquhart et al. discuss the significance of these predictors using statistical significance tests. Training of the GAM algorithm used matchups between coincident MODIS-Aqua data and in situ surface salinity data from the Environmental Protection Agency’s Chesapeake Bay Monitoring Program’s water-quality data set for an 8-year period (2003 to 2010) with a salinity range of 0.0 to 31.65. Thus, the in situ salinity range used to train the GAM algorithm extended significantly further into freshwater than that of the ANN algorithm. In Ref. 28, Urquhart et al. also evaluated their GAM algorithm by withholding matchups from the training. Of a total of 620 matchups for the 8-year period, 496 were used to train the GAM algorithm, and 124 were used to validate the satellite-derived SSS against the in situ salinity. Validation resulted in an MAE of 1.82 and an RMSE of 2.38 between the in situ and satellite-derived SSS.
The applicability of the algorithms is limited to within the ranges of the training data of , latitude, longitude, and, in the case of the ANN algorithm, SST. Extrapolation of SSS outside of the training ranges results in poor quality SSS retrievals. Therefore, implementation of the two algorithms was very conservative. The ANN’s training ranges for , SST, latitude, and longitude were used to exclude satellite inputs (, SST, latitude, and longitude) that fell outside of these training ranges from the data processing. Effects of this exclusion were to limit the geographic domain of the satellite retrievals to the Bay’s main stem and to remove values that generally represent poor water quality. Because of these exclusions and cloud coverage, daily spatial coverage of viable satellite-derived SSS was quite low. Utilizing the same conservative exclusions in the GAM implementation as in the ANN implementation generated two output data sets with the same spatial coverage.
Evaluation of Sea Surface Salinity
Validation of satellite-retrieved sea surface salinity
The satellite-derived estimates of SSS were compared against contemporaneous surface salinity measurements at the five buoys (Fig. 1) from June 2011 to February 2013. Matchups of coincident satellite and in situ buoy data were created from satellite overpasses acquired within 2 h of buoy measurement and within 1 km of the buoy location. Accuracy and precision of the satellite retrievals were quantitatively described using the metrics of bias and RMSE, respectively.
By evaluating both algorithms’ performance outside of the temporal domain over which the algorithms were trained, we test the temporal robustness of the algorithms, which is crucial for their potential application to ongoing near real-time data assimilation for hydrodynamic model forecasting.
Validation of model-predicted sea surface salinity
Error estimates of the CBOFS model salinity were derived in the same way as the error estimates for the satellite retrievals. For each of the five buoys, the CBOFS model grid cell containing the buoy’s location was selected. Hourly nowcast surface salinity was selected for the CBIBS hourly average salinity value. Bias and RMSE between CBOFS model salinity and in situ buoy salinity were then calculated. The mean surface layer within CBOFS is deep, and we assume that the satellite-derived SSS is equivalent to the salinity of the model’s surface layer.
Satellite Sea Surface Salinity Retrievals in Chesapeake Bay
The study time period from June 2011 through February 2013 included an extended wet period with low salinities in Chesapeake Bay from September 2011 until approximately April 2012, when salinities gradually returned to higher values for the remainder of the study period. For the purposes of this study, we describe these two periods as wet and dry periods, to distinguish these periods’ relatively low and high salinities.
Monthly mean maps derived from the ANN and GAM algorithms illustrate the estuarine dynamics for one month within the wet, low-salinity period (October 2011) and dry, high-salinity period (October 2012) (Fig. 2). Both algorithms show the expected salinity gradient from fresher water in the north, where the Susquehanna River enters the Bay, to saltier water in the south at the Bay’s mouth to the Atlantic Ocean. In addition, both algorithms demonstrate the difference between wet and dry periods, with a northward shift to higher saline water from the wet period to the dry period. The shift is evident throughout the entire gradient for both algorithms.
However, the algorithms differ in the magnitude of their salinity gradients, with the ANN algorithm depicting a weaker gradient from to 25, whereas the GAM algorithm depicts a stronger gradient from to 32. Furthermore, the GAM algorithm produces a west-to-east gradient of increasing salinities in the southern half of the Bay. This west-east gradient is absent in the ANN retrievals. Missing data in Fig. 2 are due to clouds and the removal of inputs (, SST, latitude, and longitude) that fall outside of the data ranges used to train the ANN algorithm, as discussed in Sec. 2.3 above. Also, the images in Fig. 2 represent the monthly averaged SSS derived from daily SSS images, and a blurring of the gradients in these images would be expected due to averaging.
Error of Satellite-Retrieved Sea Surface Salinity and Chesapeake Bay Operational Forecast System Modeled Sea Surface Salinity
Scatter plots of coincident matchups of the ANN algorithm salinity with the CBIBS buoy salinity are shown in Fig. 3 for all five buoys. Bias ranges from to 6.37 and RMSE ranges from 3.88 to 7.31. The ANN algorithm greatly overestimates the salinity at the northernmost (Patapsco) buoy and underestimates the salinity at the southernmost (First Landing) buoy. At the three middle buoys, the ANN algorithm overestimates salinity, but to a lesser degree than at the northernmost buoy.
Scatter plots of the GAM algorithm salinity versus the CBIBS buoy salinity are shown in Fig. 4. Bias of the GAM salinity is for the three southernmost buoys but is larger (1.43) at the northernmost buoy. There is a negative bias of at the Gooses Reef buoy. RMSE ranges from 2.58 to 4.83.
Error of the CBOFS modeled salinity versus the CBIBS buoy salinity is shown in the scatter plots in Fig. 5. The CBOFS modeled salinity closely matches buoy salinity at the most northern, freshest buoy with a bias , but exhibits a positive bias at the remaining buoys with a bias as high as 2.85 at the Potomac buoy. RMSE ranges from 1.21 to 3.05.
Table 1 compares the satellite-to-buoy and model-to-buoy bias and RMSE at the five buoys.
Bias and RMSE of MODIS-Aqua satellite SSS retrieved with the ANN algorithm, the GAM algorithm, or predicted SSS from the CBOFS hydrodynamic model, as compared to observed buoy SSS at each of the five CBIBS buoys used in this study. Statistics of bias and RMSE are calculated as satellite or model salinity minus observed salinity.
|ANN versus CBIBS||GAM versus CBIBS||CBOFS versus CBIBS|
Table 2 presents the values of tests of significance between pairs of the three methods, based on two-sample (pairwise) student’s -tests. Values that are statistically significant with at least 95% confidence () are highlighted in bold. Overall, the GAM salinity is significantly lower in bias and thus more accurate than the ANN salinity at four of the five buoys (Tables 1 and 2). Only bias at the Gooses Reef buoy is significantly greater for the GAM salinity than the ANN salinity, where the GAM algorithm underestimates salinity. The GAM salinity is also lower in RMSE at the same four buoys, yet is only significantly lower at the Patapsco and Potomac buoys. There are no statistically significant differences in RMSE between the GAM and ANN salinities at the remaining three CBIBS buoys.
Significance of bias and RMSE between pairs of the three SSS calculation methods: ANN, GAM, and CBOFS. When the pair is statistically significant (p<0.05), then the pair member with the lower error (bias or RMSE in Table 1) outperforms the other pair member.
|ANN versus GAM||CBOFS versus ANN||CBOFS versus GAM|
|p values for bias comparison|
|p values for RMSE comparison|
Statistically significant pairs are indicated in bold, and nonsignificant pairs are indicated in italics.
The CBOFS model salinity is significantly less accurate, i.e., higher bias, than the GAM salinity at the southernmost three of the five buoys (Tables 1 and 2). In contrast, the CBOFS model salinity is significantly more precise, i.e., lower RMSE, than the GAM salinity at three buoys (Patapsco, Gooses Reef, and First Landing).
Time Series of Satellite, Model, and Buoy Sea Surface Salinity
Time series of all four parameters, ANN salinity, GAM salinity, CBOFS model salinity, and CBIBS buoy salinity, are shown in Fig. 6. The ANN algorithm consistently overestimates salinity at the northernmost buoy and primarily underestimates salinity at the southernmost buoy. The GAM algorithm, on the other hand, consistently overestimates salinity in the wet period when buoy salinities are low and consistently underestimates salinity in the dry period when buoy salinities are high. This wet–dry distinction in the GAM algorithm is apparent at the four northern buoys but is not apparent at the tidally dominated southernmost First Landing buoy.
The CBOFS model salinity follows the trend and variability of the buoy salinity extremely closely, but is positively biased at all but the northernmost freshest buoy (Patapsco).
The large decrease in salinity at the Patapsco, Gooses Reef, and Potomac buoys in September 2011 (day 253 on -axis) is due to the extremely high river discharge from the Susquehanna River associated with Tropical Storm Lee, which inundated a large portion of the Chesapeake Bay watershed with heavy rain September 9 to 10, 2011.
The lack of satellite-to-buoy matchups at Gooses Reef during the wet period is due to clouds or masks applied during data processing, such as atmospheric correction failure or sediment-laden water causing high surface-water reflectance. The small number of satellite-buoy matchups at the Potomac buoy may be due to degraded water quality (e.g., high turbidity) entering the main stem of the bay from the Potomac River. Poor quality, highly turbid water may result in values that are outside the range of the salinity algorithms’ training data and thus excluded during SSS retrieval.
The lack of continuous buoy data at the First Landing buoy until December 2011 (day 350 on -axis) is due to an instrument failure at that buoy resulting from the impact of Hurricane Irene on August 28, 2011.
This work represents the first assessment of ocean color radiometric SSS retrievals, an advancing algorithm development field, for a coastal model application. While Urquhart et al. in Ref. 28 evaluated the ANN and GAM approaches on a single data set, they neither evaluated the specific algorithm of Geiger et al. in Ref. 27 nor his reformulated algorithm for Chesapeake Bay (unpublished). By retrieving SSS for both Urquhart et al.’s and Geiger’s (unpublished) approaches for the 21 months of this study, we extend both of these earlier assessments. Furthermore, we compare their accuracy and precision to those of CBOFS modeled SSS during the same period to evaluate the potential of the satellite SSS retrievals for near real-time data assimilation, which occurs outside the retrieval algorithms’ training and testing periods, to improve CBOFS predictions.
We found the GAM algorithm performs better than the ANN algorithm at retrieving SSS in the Chesapeake Bay, in agreement with Urquhart et al. in Ref. 28. Both ANN and GAM satellite algorithms replicate the north-to-south increasing estuarine salinity gradient for the Chesapeake Bay and show differences between wet and dry periods, shifting to higher salinities northward in drier periods (Fig. 2). However, the GAM algorithm more realistically represents the bay’s range of salinity values with better accuracy at four of the five buoys (Table 1) and reproduces the west-to-east salinity gradient typical of the southern Chesapeake Bay. This gradient is due to both the Coriolis effect that deflects the northward flowing, higher saline waters toward the east36 and the greater freshwater discharge on the bay’s western side.37 The GAM algorithm’s training data, which possessed more samples, contained a greater number of input variables, represented a longer time period, and exhibited a broader range of salinity values, particularly that of lower salinity at the head of the bay, likely accounts for its more accurate estimation of salinity. Similarly, the lack of low salinity values in the ANN algorithm’s training data likely cause it to overestimate salinity in the fresher waters of the northern bay due to extrapolation outside its trained variable space.
The GAM algorithm’s near-consistent overestimation of salinity in the wet period and underestimation of salinity in the dry period (Fig. 6) may be due to the trained algorithm representing an 8-year average condition of the bay. As actual environmental conditions depart from the average, i.e., during wetter or drier periods, the algorithm may still calculate values closer to the average, thus saltier estimations in wet periods and fresher estimations in dry periods. Perhaps an algorithm formulation that includes precipitation or river discharge into the bay could introduce a dependency on freshwater supply when calculating the salinity. If the GAM algorithm’s overestimations and underestimations could be resolved by such a means, then its precision would markedly improve and, perhaps, also its accuracy.
The significantly greater bias of the GAM algorithm than the ANN algorithm at the Gooses Reef buoy is due to the lack of satellite-to-buoy matchups during the wet period. The GAM algorithm’s underestimations in the dry period cause the GAM’s bias to be negative overall for the study period. Had there been matchups in the wet period, a balance between overestimations and underestimations would have caused the GAM salinity bias to be closer to zero. The ANN algorithm in contrast, not subject to consistent underestimation in the dry period, exhibits a better bias for this buoy due to the presence of both overestimations and underestimations.
While Geiger (unpublished) and Urquhart et al. in Ref. 28 calculate very promising MAE and RMSE statistics for their algorithms based on randomly withheld matchups from their training data, our results indicate that confidence in these algorithms is reduced when applying the algorithms outside of the time period of the training data. For instance, Geiger (unpublished) and Urquhart et al. of Ref. 28 calculate RMSE of 1.12 and 2.38 for their algorithms, respectively, whereas our study indicates RMSE of 5.01 and 3.71 for each algorithm, respectively, when applied to our study’s time period. Temporally changing environmental conditions will alter performance when applying such algorithms outside of their training periods.
The positive bias of the CBOFS modeled salinity at all but the northernmost buoy (Fig. 5) indicates the CBOFS model may be improved. Not all freshwater inputs to the bay are accounted for in the model, e.g., ground water flow, resulting in the positive bias at locations without a strong freshwater input (Ref. 38). Assimilating a more accurate and precise salinity may help to constrain the model and improve the CBOFS model’s salinity estimations. However, while the GAM satellite algorithm offers smaller bias than the CBOFS model at three of the five buoys for the period of this study (Table 1), its overestimations in wet periods and underestimations in dry periods make it unsuited for reducing the model’s bias through ongoing near real-time data assimilation. Likewise, the GAM algorithm is statistically less precise than the CBOFS model at three of the five buoys (Table 1), so the GAM algorithm’s salinities would not improve the CBOFS modeled salinity. While the GAM is a robust algorithm for representing spatial features of the Chesapeake Bay’s estuarine dynamics, improvements in its accuracy and precision, such as adding a variable indicative of wet versus dry periods, would be needed before data assimilation for CBOFS model improvements could be considered.
Use of satellite ocean color variables, such as , in coastal areas can be prone to error due to the presence of multiple water constituents not found in open-ocean waters, such as CDOM and biogenic and terrigenous particles, as well as the presence of phytoplankton. In addition, land-based aerosols may not be fully accounted for in the atmospheric correction. Even though the algorithms assessed here are trained with in situ data and thereby implicitly account for the error in , this error is an assumed general limitation to the overall SSS retrieval error. Given everything else equal, the accuracy of the algorithms’ SSS retrieval will depend upon the error characteristics of retrieval during the period when the algorithms were trained and when they are applied.
To better characterize spatial and temporal variability of salinity in coastal waters, it may be possible to use the CDOM–salinity inverse relationship, since CDOM can be estimated from routine, medium-resolution satellite radiometric measurements. Although recent ocean color radiometric SSS research has utilized combinations of values in statistical models to estimate salinity, the SSS retrieved using the two statistical algorithms evaluated in this study is less precise than SSS predicted by NOAA’s coastal hydrodynamic model of Chesapeake Bay and therefore would likely not improve the SSS forecasts with their assimilation. This study indicates reduced confidence in these statistical algorithms when applying them outside their trained variable space, in this case to time periods not included in the algorithm training. Yet these statistical algorithms may still provide reasonable salinity accuracy and precision over space and time scales needed for assimilation by coastal models, as long as careful consideration is given to including relevant environmental variables, such as precipitation, in future retrieval algorithm design. With better estimations of salinity, forecast skill of coastal hydrodynamic models such as NOAA’s CBOFS would increase, which in turn would improve our understanding and applications of biological, chemical, and physical processes in coastal environments.
The authors thank Erick Geiger and Erin Urquhart for extensive help implementing and verifying their satellite retrieval algorithms and for important discussion. Specific acknowledgment goes to Erick Geiger for reformulating the ANN algorithm for Chesapeake Bay for use in this study. The authors thank Lyon Lanerolle, Jiangtao Xu, Phillip Keegstra, Peter Bergstrom, Doug Wilson, and Charles Pellerin for their valuable discussions and for help with data, formats, and buoy data quality. This research was funded by the NOAA Ocean Remote Sensing Program. The views, opinions, and findings contained in this report are those of the authors and should not be construed as an official NOAA or U.S. Government position, policy, or decision.
B. H. Johnson et al., Development and Verification of a Three-Dimensional Numerical Hydrodynamic, Salinity, and Temperature Model of Chesapeake Bay, p. 43, U.S. Army Engineering Waterways Experimental Station, Vicksburg, Mississippi (1991).Google Scholar
G. W. Calvo et al., “A comparative field study of Crassostrea ariakensis (Fujita 1913) and Crassostrea virginica (Gmelin 1791) in relation to salinity in Virginia,” J. Shellfish Res. 20(1), 221–229 (2001).JSHRDAGoogle Scholar
R. S. Fulford et al., “Effects of oyster population restoration strategies on phytoplankton biomass in Chesapeake Bay: a flexible modeling approach,” Mar. Ecol. Prog. Ser. 336, 43–61 (2007).MESEDT0171-8630http://dx.doi.org/10.3354/meps336043Google Scholar
R. S. King et al., “Regional, watershed and local correlates of blue crab and bivalve abundances in subestuaries of Chesapeake Bay, USA,” J. Exp. Mar. Biol. Ecol. 319(1–2), 101–116 (2005).JEMBAM0022-0981http://dx.doi.org/10.1016/j.jembe.2004.05.022Google Scholar
M. Sandoz and R. Rogers, “The effect of environmental factors on hatching, moulting, and survival of zoea larvae of the blue crab callinectes sapidus rathbun,” Ecology 25(2), 216–228 (1944).ECGYAQ0094-6621http://dx.doi.org/10.2307/1930693Google Scholar
M. B. Decker et al., “Predicting the distribution of the scyphomedusa Chrysaora quinquecirrha in Chesapeake Bay,” Mar. Ecol. Prog. Ser. 329, 99–113 (2007).MESEDT0171-8630http://dx.doi.org/10.3354/meps329099Google Scholar
C. W. Brown et al., “Ecological forecasting in Chesapeake Bay: using a mechanistic-empirical modeling approach,” J. Mar. Syst. 125, 113–125 (2013).JMASE50924-7963http://dx.doi.org/10.1016/j.jmarsys.2012.12.007Google Scholar
G. Constantin de Magny et al., “Predicting the distribution of Vibrio spp. in the Chesapeake Bay: Vibrio cholerae case study,” Ecohealth 6(6), 378–389 (2010).http://dx.doi.org/10.1007/s10393-009-0273-6Google Scholar
L. W. J. Lanerolle, R. C. Patchen and F. I. Aikman, “The second generation Chesapeake Bay Operational Forecast System (CBOFS2): model development and skill assessment,” NOAA Technical Report, Vol. 29, p. 77, NOS CS (2011).Google Scholar
J. L. Miller, M. A. Goodberlet and J. B. Zaitzeff, “Airborne salinity mapper makes debut in coastal zone,” EOS Trans. Am. Geophys. Union 79(14), 173–184 (1998).EOSMAWhttp://dx.doi.org/10.1029/98EO00126Google Scholar
G. Lagerloef et al., “The Aquarius/SAC-D mission: designed to meet the salinity remote-sensing challenge,” Oceanography 21(1), 68–81 (2008).1042-8275http://dx.doi.org/10.5670/oceanog.2008.68Google Scholar
D. G. Bowers and H. L. Brett, “The relationship between CDOM and salinity in estuaries: an analytical and graphical solution,” J. Mar. Syst. 73(1–2), 1–7 (2008).JMASE50924-7963http://dx.doi.org/10.1016/j.jmarsys.2007.07.001Google Scholar
P. Coble et al., “Colored dissolved organic matter in the coastal ocean: an optical tool for coastal zone environmental assessment and management,” Oceanography 17(2), 50–59 (2004).1042-8275http://dx.doi.org/10.5670/oceanog.2004.47Google Scholar
N. V. Blough, R. Del Vecchio, “Chromophoric DOM in the coastal environment,” in Biogeochemistry of Marine Dissolved Organic Matter, , D. Hansell and C. Carlson, Eds., pp. 509–546, Academic Press, San Diego, California (2002).Google Scholar
Y. Bai et al., “Remote sensing of salinity from satellite-derived CDOM in the Changjiang River dominated East China Sea,” J. Geophys. Res. Oceans 118(1), 227–243 (2013).http://dx.doi.org/10.1029/2012JC008467Google Scholar
M. S. Twardowski and P. L. Donaghay, “Separating in situ and terrigenous sources of absorption by dissolved materials in coastal waters,” J. Geophys. Res. Oceans 106(C2), 2545–2560 (2001).http://dx.doi.org/10.1029/1999JC000039Google Scholar
R. Del Vecchio and N. V. Blough, “Spatial and seasonal distribution of chromophoric dissolved organic matter and dissolved organic carbon in the Middle Atlantic Bight,” Mar. Chem. 89(1–4), 169–187 (2004).MRCHBD0304-4203http://dx.doi.org/10.1016/j.marchem.2004.02.027Google Scholar
A. Vodacek et al., “Seasonal variation of CDOM and DOC in the Middle Atlantic Bight: terrestrial inputs and photooxidation,” Limnol. Oceanogr. 42(4), 674–686 (1997).LIOCAH0024-3590http://dx.doi.org/10.4319/lo.1997.42.4.0674Google Scholar
C. E. Del Castillo et al., “Analysis of the optical properties of the Orinoco River plume by absorption and fluorescence spectroscopy,” Mar. Chem. 66(1–2), 35–51 (1999).MRCHBD0304-4203http://dx.doi.org/10.1016/S0304-4203(99)00023-7Google Scholar
C. E. Binding and D. G. Bowers, “Measuring the salinity of the Clyde Sea from remotely sensed ocean colour,” Estuarine Coastal Shelf Sci. 57(4), 605–611 (2003).ECSSD30272-7714http://dx.doi.org/10.1016/S0272-7714(02)00399-2Google Scholar
E. J. D’sa et al., “Estimation of colored dissolved organic matter and salinity fields in case 2 waters using SeaWiFS: examples from Florida Bay and Florida Shelf,” J. Earth Syst. Sci. 111(3), 197–207 (2002).http://dx.doi.org/10.1007/BF02701966Google Scholar
S. L. Palacios, T. D. Peterson and R. M. Kudela, “Development of synthetic salinity from remote sensing for the Columbia River plume,” J. Geophys. Res. Oceans 114(C2), C00B05 (2009).http://dx.doi.org/10.1029/2008JC004895Google Scholar
Y. H. Ahn et al., “Satellite remote sensing of a low-salinity water plume in the East China Sea,” Ann. Geophys. 26, 2019–2035 (2008).ANNGEA1432-0576http://dx.doi.org/10.5194/angeo-26-2019-2008Google Scholar
H. Sasaki et al., “Mapping the low salinity Changjiang diluted water using satellite-retrieved colored dissolved organic matter (CDOM) in the East China Sea during high river flow season,” Geophys. Res. Lett. 35(4), L04604 (2008).GPRLAJ0094-8276http://dx.doi.org/10.1029/2007GL032637Google Scholar
E. F. Geiger et al., “Satellite-derived coastal ocean and estuarine salinity in the Mid-Atlantic,” Cont. Shelf Res. 63(Suppl.), S235–S242 (2013).http://dx.doi.org/10.1016/j.csr.2011.12.001Google Scholar
E. A. Urquhart et al., “Remotely sensed estimates of surface salinity in the Chesapeake Bay: a statistical approach,” Remote Sens. Environ. 123, 522–531 (2012).RSEEA70034-4257http://dx.doi.org/10.1016/j.rse.2012.04.008Google Scholar
S. Qing et al., “Retrieval of sea surface salinity with MERIS and MODIS data in the Bohai Sea,” Remote Sens. Environ. 136, 117–125 (2013).RSEEA70034-4257http://dx.doi.org/10.1016/j.rse.2013.04.016Google Scholar
J. R. Schubel and D. W. Pritchard, “Responses of upper Chesapeake Bay to variations in discharge of the Susquehanna River,” Estuaries 9(4), 236–249 (1986).ESTUDO0160-8347http://dx.doi.org/10.2307/1352096Google Scholar
A. F. Shchepetkin and J. C. McWilliams, “The regional oceanic modeling system (ROMS): a split-explicit, free-surface, topography-following-coordinate oceanic model,” Ocean Model. 9(4), 347–404 (2005).1463-5003http://dx.doi.org/10.1016/j.ocemod.2004.08.002Google Scholar
D. W. Pritchard, “Observations of circulation in coastal plain estuaries,” in Estuaries, and G. H. Lauff, Ed., pp. 37–44, American Association for the Advancement of Science, Washington, DC (1967).Google Scholar
J. A. Austin, “Estimating effective longitudinal dispersion in the Chesapeake Bay,” Estuarine Coastal Shelf Sci. 60(3), 359–368 (2004).ECSSD30272-7714http://dx.doi.org/10.1016/j.ecss.2004.01.012Google Scholar
L. Lanerolle, Personal communication (2013).Google Scholar
Ronald L. Vogel received his BA degree in environmental science with honors from Wesleyan University and his MS degree in marine science from the University of South Carolina, specializing in coastal and estuarine processes. He is a satellite oceanographer, supporting the application of remote sensing to water quality, fisheries, and ecological modeling at the National Oceanic and Atmospheric Administration. His expertise is in ocean and land visible and infrared satellite imaging and creating environmental decision-support tools.
Christopher W. Brown received his BS degree in biological sciences from Cornell University in 1982 and his MS and PhD degrees in oceanography from the Graduate School of Oceanography at the University of Rhode Island in 1987 and 1993, respectively. After his doctorate, he held a postdoctoral fellowship at the NASA Goddard Space Flight Center and joined NOAA/NESDIS in November 1995. His primary research interests include the remote detection, characterization and prediction of marine organisms.