Integrating Landsat-7 Imagery with Physics-based ... - Ingenta Connect

12-023.qxd

10/19/12

8:14 PM

Page 1163

Integrating Landsat-7 Imagery with Physics-based Models for Quantitative Mapping of Coastal Waters near River Discharges [THIS PAPER WAS THE WINNER OF THE 2012 BAE SYSTEMS AWARD GIVEN AT THE ASPRS 2012 ANNUAL CONFERENCE] Nima Pahlevan, Alfred J. Garrett, Aaron D. Gerace, and John R. Schott

Abstract Remote sensing has traditionally been used to retrieve water constituents by establishing a relationship between in situ measured quantities and image-derived products. Motivated by the dramatically improved potential of the Landsat Data Continuity Mission (LDCM), this paper describes a different approach for water constituent retrieval where both thermal and visible spectral bands of the Enhanced Thematic Mapper Plus (ETM⫹) instrument on board Landsat-7 are utilized. In this effort, Landsat data is integrated with a 3D hydrodynamic model to obtain profiles of particles and dissolved matter in the near shore zone in the vicinity of two river discharges. The procedure is based upon performing many hydrodynamic simulations by adjusting input environmental/physical variables and generating Look-UpTables (LUTs). The best match, obtained using optimization, demonstrated an average root-mean-squared-error (RMSE) of 0.68 percent, i.e., 0.0068 reflectance units, calculated over the two river plumes. It is concluded that calibrating a physics-based model using the Landsat-7 imagery can provide a more lucid insight into the dynamics of spatially non-uniform waters.

Introduction Overview In order to monitor a highly variable environment, such as coastal/inland waters, using remotely-sensed observations, high-frequency measurements are needed. Although it provides adequate spatial details over coastal/inland waters, Landsat-7 (L7) has a 16-day revisit cycle, which may not be ideal for regular monitoring of such dynamic systems. This issue, however, can be compensated by leveraging numerical Nima Pahlevan is with the Department of Environmental, Earth, and Ocean Sciences, University of Massachusetts Boston, 100 Morrissey Blvd. Boston MA 02125, and formerly with the Chester F. Carlson Center for Imaging Science, Rochester Institute of Technology, Rochester, NY, 14623 ([email protected]). Alfred J. Garrett is with the Savannah River National Laboratory, Aiken, SC 29808. Aaron D. Gerace and John R. Schott are with the Chester F. Carlson Center for Imaging Science, Rochester Institute of Technology, 54 Lomb Memorial Drive, Rochester, NY, 14623. PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

models allowing for simulating the state of the environment at any point-in-time. The objective of this paper is to take advantage of the Landsat’s thermal and reflective imagery to calibrate a 3D hydrodynamic model to capture the dynamics of coastal waters in the vicinity of river plumes where large amounts of land-based materials are introduced to fresh lake waters. In this effort, the 3D hydrodynamic model is calibrated in the thermal domain using Landsat-derived surface temperature products. The thermally-calibrated model is then coupled with an in-water radiative transfer model to retrieve profiles of water constituents by optimizing the coupled-modeling outputs against Landsat-derived surface reflectances. The profiles of water constituent components derived from this integration include Total Suspended Solids (TSS) and dissolved matter, that control the optical regime of the coastal areas near river discharges. A unique advantage of the proposed approach is the retrieval of TSS through the water column beyond the penetration depth of optical signals in turbid waters. Motivated by the Landsat Data Continuity Mission (LDCM), this study uses L7 imagery as a surrogate for LDCM whose enhanced features allow for a more accurate retrieval of water constituents than those achieved by the conventional Landsat sensors. LDCM equipped with two independent reflective and thermal instruments will acquire imagery in eight reflective channels and two thermal bands (Gerace and Schott, 2009; Pahlevan and Schott, 2011). This paper is structured as follows. A brief background on the remote sensing of coastal waters and its application when coupled with hydrodynamic modeling is presented in the following subsections. An overview of the hydrodynamic model, its parameterization and input variables are given in the next section. The proposed approach, where the study areas, the model calibration technique, and the constituent retrieval process are detailed, is then described. This is followed by the results and error analysis. The concluding remarks and the future directions are given in the final section. Remote Sensing of Coastal Waters Remote sensing has long been used to investigate the water quality conditions in near-shore zones (Jensen, 2006). Based

Photogrammetric Engineering & Remote Sensing Vol. 78, No. 11, November 2012, pp. 1163–1174. 0099-1112/12/7811–1163/$3.00/0 © 2012 American Society for Photogrammetry and Remote Sensing N o v e m b e r 2 0 1 2 1163

12-023.qxd

10/19/12

8:14 PM

Page 1164

upon the extensive research conducted in such environments (Gould, Jr. and Arnone, 1997; Hakvoort et al., 2002; Brando and Dekker, 2003; Binding et al., 2005; Gitelson et al., 2007), two key findings are essential for grasping the complexity of water constituent retrieval through existing remote sensing data and algorithms. The first issue addresses the complex nature of coastal waters. The coastal areas are where land, water, and atmosphere meet making them physically dynamic regions to study. To the extent that chemicals, nutrients, dissolved organic, and inorganic materials are entering from rivers and streams, they add to the complexity of physical and biological processes occurring in such regions. The combined biological and physical processes contribute to the complexity of coastal waters and highlight the need for optimal monitoring techniques among which remote sensing has been proven an effective option. The requirements for appropriate mapping and monitoring of coastal waters through satellite-borne remote sensing raises the second issue, i.e., whether suitable remote sensing systems and methods are available to adequately address the dynamics of such environments. There are always trade-offs among the applications of different imaging systems regarding their capabilities for resolving spatial/spectral complexities in coastal waters. More importantly, in order to capture the temporal variability of coastal waters, we require highfrequency satellite systems acquiring imagery at least once/twice a day. The availability of the new generation of commercial satellites, including WorldView-2 and earlier QuickBird, with flexible pointing technology, has made data acquisition with very high spatial/radiometric resolution possible. However, for long-term monitoring of coastal waters at regional/global scale, it may not be cost-effective to utilize such image products. The Moderate Imaging Spectrometer (MODIS) and the Sea-view Wide-Field of View (SEAWIFS) are the two imaging systems developed and designed for monitoring global/regional waters. Having appropriate spectral bands configured for water studies along with their superior radiometric fidelity, i.e., 12-bit quantization rate and high signal-to-noise ratio (SNR), MODIS, and SEAWIFS, however, lack sufficient spatial resolution to map spatially heterogeneous waters in the near-shore regions. Although the Enhanced Thematic Mapper plus (ETM⫹) onboard L7 has been designed for monitoring land features, its 30 m pixel size and four broadband spectral channels in the Visible-Near-Infrared (VNIR) region have made it a suitable choice for water quality studies in coastal/inland waters over the past decade (Dekker et al., 2001; Olmanson et al., 2008). Using remote sensing, one can potentially retrieve optically active components (OAC) of water including concentrations of TSS and chlorophyll-a (CHL), as well as Colored Dissolved Organic Matter (CDOM) absorption, which collectively determine the optical regime of coastal waters, also known as Case II waters. Non-linear and complex interactions between these components together with optical properties of pure water contribute to the Apparent Optical Properties (AOPs) of water. The AOPs are commonly used in regression models or are supplied to bio-optical models for the retrieval of either IOPs or water constituents. In most case studies, the water constituents are retrieved solely at an instant in time for a scene. Under ideal atmospheric conditions (no cloud contaminations), when multiple scenes are available, the temporal variability of the water constituents can be monitored. In addition, the concentration retrieval in Case II waters is limited to the top surface layer, i.e., Secchi depth, beyond which optical signals are attenuated due to large absorption and scattering within water column. The above-noted issues can be compensated through combining the remotely sensed imagery with 3D hydrodynamic models capable of continuously simulating dynamics of coastal waters. 1164 N o v e m b e r 2 0 1 2

Remote Sensing Integrated with Numerical Modeling Surface physical properties derived from remotely sensed data when combined with physics-based models can assist in understanding the dynamics of water bodies. Remotely sensed data can be employed to initialize, calibrate, and validate physics-based models to achieve accurate modeling efforts. Numerical modeling has long been used in conjunction with remote sensing in coastal/inland waters to enable predicting distributions of suspended particles, dissolved matter, and other water quality parameters (Jensen et al., 1989; Miller and Cruise, 1995; Chen et al., 2010; Ouillon et al., 2010) In this effort, we intend to integrate L7 imagery with a thermally driven hydrodynamic model (ALGE), which predicts temperature, material transport, and deposition of stream plumes (Garrett et al., 2000; Schott et al., 2001; Garrett, 2002; Li et al., 2008). We apply L7 imagery in conjunction with the ALGE model to simulate dissipation and transport of two river plumes discharging into lake waters. The proposed approach aims at retrieving surface and vertical distributions of water constituents, i.e., TSS and dissolved matter, in two river plumes that discharge into southern Lake Ontario, New York. The 3D hydrodynamic model is first calibrated by refining the input environmental variables and optimizing it against L7-derived surface temperature maps. When realistic spatial distributions of the river plumes are determined, the input concentrations of particles and dissolved matter are varied to find the optimal surface/vertical distributions of water constituents by comparing the model outputs and L7 products in the surface reflectance domain. The well-validated in-water radiative transfer model (Hydrolight (Mobley, 1994)) is used to simulate water-leaving surface reflectance from multiple profiles of water constituent concentrations obtained from the model outputs. The best surface reflectance map, found through optimization, is associated with a pair of input concentrations of sediment and dissolved matter yielding the closest agreement with the L7-derived water constituent maps.

Methodology Hydrodynamic Modeling ALGE is a three-dimensional, time-dependent, hydrodynamic model that simulates movement and dissipation of stream plumes as well as transport, diffusion, and deposition of materials. ALGE solves the partial differential equations that govern conservation of mass, momentum, thermal energy, turbulent kinetic energy, suspended particulates, dissolved chemical species, and chemical species absorbed to particles (Garrett, 1995; Garrett and Hayes, 1997). ALGE computes sensible, evaporative, thermal radiation, and solar radiation energy fluxes into and out of a water body. Mass sources and sinks, surface wind stress, and density gradients drive currents in the water. For a river plume simulation, the ALGE code is initiated by a variety of user-defined input parameters. ALGE requires hourly input data, including meteorological data, insolation (including cloud effects), radiosonde inputs, river discharge, and river temperature observed in the region of interest. The meteorological data consist of wind speed, wind direction, air and dew point temperature, cloud heights, and air pressure. It should be noted that for the river plume simulations wind stress and wind direction are provided to the model in a 1D form, i.e., the wind data taken from stations nearby the discharges. Profiles of precipitable water and temperature of the upper-air atmosphere are also derived from sounding data. Time-varying river discharge and temperature are also provided to the model to incorporate the volume and the ambient temperature of the plume entering the lake, i.e., Lake PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

12-023.qxd

10/19/12

8:14 PM

Page 1165

Ontario. The initial temperature of the receiving waters can be estimated through remotely sensed data, in situ measurements, or lake-wide simulation results. In addition to the locally-derived meteorological parameters, ALGE is also able to incorporate lake-wide circulation patterns by allowing the user to optionally supply timedependent profiles of temperature as well as surface current velocities into the river plume simulations (localized simulations). Moreover, a zero-gradient boundary condition is adopted for our modeling efforts owing to the negligible tidal forcing in Lake Ontario. Of particular interest among ALGE outputs are spatial patterns of the temperature, sediment, and dissolved matter, which can be linked to remotely sensed observations. This utility arises from the nature of its outputs. In fact, the user can define the grid size for the model such that the model outputs can be directly comparable to the observed imagery. In the absence of in situ measurements of lake circulation patterns, lake-wide simulations are carried out to obtain temporally variable boundary conditions when conducting localized simulations. Lake-wide Simulations The most eastern of the Laurentian Great Lakes in North America is Lake Ontario, which serves as the outlet to the Atlantic Ocean. Lake Ontario is the smallest in surface area among the Great Lakes. It is elevated 75 m above the mean sea level and has a maximum depth of 244 m in its eastern basin (Schott et al., 2001). There are numerous rivers, creeks, and canals entering Lake Ontario with the Niagara River being the primary source of water, i.e., annual average discharge of 6,000 m3/s (Hayashida et al., 1999). ALGE is capable of performing lake-wide simulations to model the dynamics of the lake temperature and its circulation pattern. The lake-wide simulation is provided with a 2D wind field generated by interpolating the in situ-measured wind data collected from seven different land stations surrounding the

lake, as well as two buoy stations located in its eastern and western basins. The upper-air atmospheric condition (assumed constant throughout the lake) was obtained from the radiosonde station at the Buffalo International Airport, New York. In order to achieve reasonable results, ALGE simulations were allowed to spin up over more than twomonth periods with target dates/hours consistent with those of localized (plume) simulations. Figure 1 illustrates three lake-wide simulation results (surface temperature) for May 2009 (May09), July 2009 (July09), and October 2010 (Oct10), when compared to the L7- and MODIS-derived surface temperature maps (SST). The SSTs obtained from the model exhibit reasonable qualitative and quantitative agreements with the satellite observations in May and July 2009, i.e., rootmean- squared error (RMSE) less than 0.86°C.. The best model outputs were derived after a few experimental runs by slightly adjusting the initial lake temperature (vertically uniform) and the 2D wind patterns. The model outputs were compared to the satellite-derived SSTs. The May09 model result, in particular, shows the stratified, near-shore, warm waters (⬎4°C), i.e., thermal bar, against the well-mixed, cold waters in the deeper zones. The small disparities between the absolute temperature, to some extent, can be related to the relatively coarse vertical resolution of the model domain, 2 m, representing “bulk” temperature of the top layer of the model domain while the SSTs drawn from satellite observations provides sea surface “skin” temperature. Moreover, the coarse grid spacing of the simulations, i.e., 1.5 km, may not be sufficient to capture small-scale eddies leading to dissimilarities between the model outputs and the satellite-derived maps (July09). The predicted surface current velocities and temperature profiles for this timeframe, i.e., March through July, were utilized for the localized simulation at the Genesee River site (see Figure 2). For Oct10, although portions of the image products are masked out due to cloud contamination, significant discrepancies are evident in the eastern basin.

(a)

(b)

(c)

(d)

(e)

(f)

Figure 1. Surface temperature maps of Lake Ontario produced from Lake-wide simulations in May 2009, July 2009, and October 2010: (a) May09-L7, (b) July09-MODIS, (c) October10-MODIS, (d) May09-Model, (e) July09-Model, and (f) October 10-Model. The model results are compared with satellite-derived observations. Note that the L7-derived surface temperature has been downsampled to 1.5 km resolution and that the MODIS product in Oct 2010 has been masked out for the cloud contamination.

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

N o v e m b e r 2 0 1 2 1165

12-023.qxd

10/19/12

8:14 PM

Page 1166

experimental method termed a nudging technique or Newton relaxation scheme (Garrett, 1995; Wang, 2001). The nudging technique primarily used in atmospheric modeling and forecast systems is represented as: 0v ⫽ F(v,t)⫹N W (x,y) (vndg⫺vmod) 0t

Figure 2. A temperature map [°C] derived from a lake-wide simulation overlaid onto a surface current velocity vectors. The filled circles show the nodes from which the nudging vectors are extracted for each river plume simulation.

In addition to the above-noted issues causing uncertainties in the model prediction, it is believed that non-uniform spatial distribution of environmental parameters (e.g., air/dewpoint temperature, cloud height, sky fraction) across the lake yields inconsistent results as compared to the satellite observations. Although a 2D wind pattern is provided to the code, the other environmental inputs are supplied to the model in a 1D form. Such data are obtained from a station located in the western side of the lake. Moreover, the lake-wide simulation for Oct10 was initiated in early-August when the lake is stratified. Thus, initiating the model with a vertically uniform thermal structure introduces errors in the modeling effort. Although the model has failed to properly predict the lake’s eastern thermal structure, the modeled surface temperature in the western basin is consistent with the MODIS-derived SST. The modeled lake circulation patterns in this period are prescribed to the localized simulation at the Niagara River site driven by the locally observed environmental inputs. Nudging As described, the lake-wide simulation results are utilized to enhance the river plume simulations. ALGE facilitates incorporating lake-derived circulation patterns through an

(a)

(b)

(1)

where the term on the left is the velocity gradient at a grid point, F(v,t) represents all of the physical forcing components, and the multiplication factor N is the nudging term which determines the intensity of the forcing from the lake-wide circulation. The surface velocities derived from the lake-wide circulation are introduced to the localized simulations at three boundary nodes (see Figure 2). The velocities are then interpolated via a weighting function, i.e., W(x, y), where 0 ⭐ W (x, y) ⭐ 1, which alters the strength of the nudging forces as a function of distance from the shoreline. The vndg and vmod are the surface nudging and the modeled velocities for each model node at the surface level, respectively. Assimilated surface nudging forces are extended to lower nodes within the water column through a logarithmic approximation (Garrett, 1995). The nudging technique allows for enhanced modeling and interaction between the lake circulation and the plume simulations driven primarily by inputs from local observatories. In addition to the nudging process, the temperature profiles are also extracted from all of the nodes lying along the nodes associated with the localized simulation boundaries (shown as boxes in Figure 2). The vertical profiles, which are derived from a course resolution domain, i.e., 1.5 km, are then upsampled to the number of nodes on the boundaries of the localized simulations. Study Sites The Genesee River, S1 from this point on, originates from the Allegheny Plateau of Northern Pennsylvania and travels approximately 240 km northwards before emptying into Lake Ontario at the Rochester embayment. The Genesee watershed, totaling approximately 6,000 km2 (Figure 3), covers seven counties in western New York state (Makarewicz, 2010). The 58 km Niagara River, hereafter referred to as S2, which runs northward from Lake Erie, is the major source of inflow into Lake Ontario. The river is

(c)

Figure 3. Relief map of the State of New York (b) along with Niagara River draining into Lake Ontario (a) and Genesee River watershed which originates from high mountains of Pennsylvania (c) indicated. Boxes indicate domain boundaries for the two sites.

1166 N o v e m b e r 2 0 1 2


12-023.qxd

10/19/12

8:14 PM

Page 1167

the drainage outlet for the four upper Great Lakes having an aggregate basin area of more than 600,000 km2. Also, as water is carried away from Lake Erie, all of the nutrient loads and nuisance algae, from both the US and Canadian sides, are transferred into Lake Ontario through the river (Figure 3). Landsat Data This paper takes full advantage of L7 imagery to calibrate the ALGE model by which surface and vertical distributions of sediment and dissolved matter are predicted. Here, two cloud-free L7 images acquired on 14 July 2009 (path/row 16/30), and on 19 October 2010 (path/row 18/30), were used to examine the methodology for modeling the dynamics of the Genesee (S1) and the Niagara (S2) river plumes. The L7’s missing data in the form of diagonal stripes were estimated using a spatial interpolation technique that utilizes available edge pixels to fill the gaps. However, the interpolation does not achieve a perfect replacement for the missing data (particularly in the reflective domain) and care must be taken when calculating error values associated with retrieval process within gap areas. The seven spectral channels in digital numbers (DNs) were converted to the Top-of-Atmosphere (TOA) radiance [W/m2 sr um] using bandspecific calibration coefficients provided in the metadata. Remotely Sensed Surface Products The thermal and reflective solar bands of L7 were utilized to retrieve surface temperature and reflectance over coastal waters at S1 and S2. In order to estimate the L7’s waterleaving surface products, the governing remote sensing equation (Schott, 2007), expressed in terms of radiance [W/m2 sr um], is defined as: LSensor ⫽ [(LS ⫹ LdS ⫹ LdT)rd ⫹ LT]t ⫹ rLdS ⫹ LuS ⫹ LuT

(2)

where LSensor denotes the at-sensor radiance. The radiance quantities with subscript S corresponds to the solar paths (reflective bands) and subscript T indicates the thermal (emissive) components.; Ld and Lu stand for down-welled and up-welled contributions, rd is the unitless diffuse surface reflectance, ␶ is the path transmission, LS is the direct solar term reflected upward, and ␳Lds represents the portion of the down-welled sky radiance reflected off the water surface where, ␳ 0.022 is the Fresnel reflection coefficient for calm waters at the nadir-viewing angle (Doxaran et al., 2004). One should note that in Equation 2, all of the components are wavelength dependent and the effects due to adjacent targets and whitecaps are ignored. Surface Temperature Retrieval Since the emissive water-leaving radiance component is directly related to the Blackbody radiance, i.e., LT ⫽ ␧LBBT, the surface temperature can be found by applying the Planck equation and assuming a constant emissivity (␧ ⫽ 0.986) for water in the thermal region. The other three components, including transmission, down-welled, and up-welled radiances, were modeled using the MODtran resolution atmospheric TRANmission (MODTRAN) simulations (Berk et al., 1989) supplied with observed (but interpolated in time and space) radiosonde data. The estimated spectral radiance values were then convolved with the L7’s thermal Relative Spectral Response, i.e., RSR (␭), ranging from 10 to 13 um. Surface Reflectance Retrieval Producing surface reflectance products from the sensorreaching radiance consists of removing intervening atmospheric effects together with masking out pixels affected by sun glint and accounting for sky glint effects. In the absence PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

of in situ optical measurements of aerosol particles, such as dust and smoke, atmospheric compensation over coastal waters is a daunting task due to the poor knowledge of particle size and distribution of aerosol. Thus, when possible, it is preferred to employ image-derived techniques, such as dark object subtraction or empirical line method (ELM), for atmospheric correction (Karpouzli, 2003). In this paper, the ELM technique was used to remove/diminish atmospheric effects over the two sites. The beach sand, located adjacent to the Genesee River mouth, and simulated surface reflectance in the offshore areas produced from the Hydrolight code (Mobley, 2008) were utilized as bright and dark objects to calculate band-specific regression equations. While the CHL concentrations for the offshore waters were estimated from MODIS products, the IOPs were obtained from the previous measurements (Raqueno, 2003). Due to the lack of adequate numbers of bright pixels at the S2 site, it was decided to conduct a locally robust ELM to transform the TOA radiance scene to the surface reflectance domain. In this case, the bright pixels were drawn from the turbid waters, whereas, similar to the procedure followed at S1, the dark pixels were derived from the dark, deep lake waters. The in situ measured surface reflectance (rd) near the river mouth and the modeled deep waters were used to determine the linear regression against the TOA radiances. It should be noted that the L7 imagery for S1 was contaminated with spatially heterogeneous haze and thus, care was taken when analyzing the image for the retrieval process. As the measured spectrum was corrected for the sky glint and the modeled spectrum derived from Hydrolight is sky glint-free, there was no need to apply further corrections for the sky glint. The atmospheric removal process was followed by removing pixels contaminated with the sun glint. For S1, it was found that there were only a few pixels affected by sun glint; therefore, no sun glint correction was implemented. No sun-glint effects were found for the S2 site perhaps because of the low sun angle and relatively calm waters, i.e., wind speed ⬍2.5 m/s, at the time of imaging. The clear sky with high horizontal visibility ensured a high quality dataset for this site. Localized Simulations The river plume simulations were performed for two different rivers, namely the Genesee River and the Niagara River characterized with different environmental/physical conditions. The simulations are driven primarily by the locally derived environmental inputs, i.e., meteorological and riverine, together with the lake-derived time-series of temperature profiles and surface velocities prescribed at the boundaries. Model Inputs The wind data were taken from the National Oceanic and Atmospheric Administration (NOAA) stations in the vicinity of the river mouths, i.e., 43.258°N, 77.592°W and 43.261°N, 79.064°W for the S1 and S2 sites, respectively. The river discharges were obtained from the USGS and the US Army Corps for S1 and S2, respectively. The average discharge in mid fall for the Niagara River exhibits 120 times, on average, greater values than those for the Genesee River in July 2009. The other meteorological data were obtained from the nearest local airports. The daily river temperatures obtained from a nearby creek in the region were incorporated into the model for S1. The hourly temperatures measured at the eastern basin of Lake Erie by NOAA’s National Ocean Service were used for S2. The bathymetry map obtained from NOAA’s National Geophysical Data Center was used to create the spatial grids as the basis for the hydrodynamic modeling with the deepest area in the north-east and north-west for the S1 and S2 sites, respectively, i.e., 57 m and 112 m. The grid N o v e m b e r 2 0 1 2 1167

12-023.qxd

10/19/12

8:14 PM

Page 1168

spacing was resampled to 60 m and 120 m resolution for the Genesee and the Niagara plume simulations. The coarser spatial resolution for the Niagara simulations was adopted to reduce the computational time. To properly capture the entire extent of the Niagara plume, its domain size (19 km ⫻ 28 km) was selected to be approximately four times greater than that of the Genesee plume, i.e., 10 km ⫻ 11 km. The vertical resolutions were chosen to be 1 m for both sites. Model Stabilization (Long-term Simulations) In order to begin the modeling efforts, the model is allowed to run and stabilize for 10 days for both simulations. The last hour of simulations, i.e., 12:00 pm, are nearly coincident with the L7 acquisitions (11:45AM EST). The ALGE code was initiated at 14:00 EST on 04 July and on 10 October for S1 and S2, respectively. As noted, the initial boundary temperature (lake temperature) is an important factor that considerably affects the model outputs. The initial boundary temperatures were estimated from the vertically-averaged, lake-wide simulations. The corresponding modeled surface temperatures were in a reasonable agreement with the MODIS-derived SST products. Provided there is confidence in the performance of the instruments measuring environmental variables, the model should realistically predict the shape of the plume and its ambient temperature after nearly 240 hours of simulations. However, this scenario is not often the case as uncertainties associated with the in situ measurements lead to errors in the model predictions. In addition to the instrumental errors, some environmental variables, such as the river discharge and river temperature, are not observed at the proximity of the river mouths and do not represent realistic values as model inputs. Data assimilation techniques have been demonstrated to account for the above-mentioned uncertainties (Li, 2007), and, as a result, refine the model. In this effort, we take into account such uncertainties via simulating various environmental conditions and populating LUTs with the model outputs. The model is then considered calibrated when a model-derived surface temperature map closely resembles that obtained from the L7 thermal imagery. Model Calibration From the preliminary studies, it was known that environmental variables such as wind speed (WS), wind direction (WD), river discharge (RD), and river temperature (RT) are the major parameters driving the temperature and the physical shape of the plume. After completing the long-term simulations as described in the previous section, ALGE is re-started for the last 40 hours of the simulations during which WS, WD, RD, and RT are varied resulting in a 4D LUT. This was carried out by extracting one cube of ALGE’s long-term outputs and using it to initialize many shorter simulations for a 40-hour timeframe. Various combinations of input variables lead to various temperature distribution maps (Pahlevan et al., 2011). A broad range of coefficients was chosen to model a variety of sizes, shapes, and temperatures of realistic plumes. This process, which follows the gradient descent concept, is an iterative approach to achieve an optimal solution, i.e., the closest match with the L7-derived surface temperature map (reference map). The first iteration is started by choosing an initial set of coefficients for WS, WD, RD, and RT with a relatively large step sizes and a reasonable range spanning realistic environmental conditions. In each iteration, the input parameters associated with the best model output is used as the starting point for the next iteration conducted with finer step sizes of the coefficients. In order to evaluate the model outputs, we define a metric, which measures the similarity between two arbitrary raster outputs. The metric termed RMSE-Cross-Correlation 1168 N o v e m b e r 2 0 1 2

(RCC) is calculated for each model output, indexed with k, and expressed as:

a a m(i, j) 䊟 r(i, j) RCC k ⫽

i

j

2 a 1mn ⫺ rn2

(3)

n

Q

N

where m indicates the model output, r stands for the reference, and N is the total number of water pixels. The numerator represents a 2D cross-correlation operation while the denominator shows the RMSE computed on a pixel-bypixel basis indexed with n. In other words, RMSE is weighted with the cross-correlation function to enhance the separability between the errors associated with different simulation outcomes. The cross-correlation computation was performed in the Fourier domain by taking the inverse Fourier transform of the product of the Fourier transform of m(i,j) and r(i,j). After performing the short-term simulations a few times, a non-linear optimization technique (Lagarias et al., 1998) was used to search for a local maximum within the LUT-derived RCC space, which is equivalent to minimization in the 1/RCC space. The search is accomplished within a finely sampled space such that errors due to the choice of optimization technique are minimal. The goal is to find the RCC value for which the set of input parameters result in the maximum similarity with the reference images. Constituent Retrieval Determination of Inherent Optical Properties of the Sites As described, the Hydrolight code is utilized to model the outgoing optical field (surface reflectance) for each ALGE output on a per-pixel basis. Hydrolight is a one-dimensional, plane-parallel radiative transfer code, which solves the radiative transfer equations by incorporating the absorption and scattering properties of pure water and water constituents (Mobley, 2008). In addition to the profiles of concentrations, the inherent optical properties (IOPs) of water and its constituents, the imaging geometry, and the environmental conditions must be specified. The inherent optical properties (IOPs), which consist of scattering and absorption of water and its constituents, associated with S1 were determined using a combination of previous measurements and experimental procedures. The specific absorption properties of TSS and CHL as well as the CDOM absorption were obtained from the measurements made at the Genesee River in May 2010 and August 2011 (Pahlevan, 2012). The specific absorption measurements were consistent with the measurements made at the Rochester Embayment in 1999 and 2003 (Raqueno, 2003). It was found that although the shape of the absorption spectra for the above-noted constituents could be assumed identical, their magnitudes could vary according to the seasonal effects and levels of runoff. The specific scattering spectra associated with CHL and particles were estimated through fitting modeled reflectance spectra to those measured in situ. The modeled spectra were generated using the Hydrolight code. In this curve-fitting procedure, the in situ measured absorption spectra of TSS, CHL, and CDOM were held constant whereas the scattering properties were adjusted to achieve reflectance spectra consistent with the measured surface reflectances (rd). To account for uncertainties in the concentration measurements conducted in the lab, the concentrations of TSS and CHL were also simultaneously slightly varied. Based upon the previous studies in moderately turbid waters of Lake Ontario (Raqueno, 2003), the 1.8 percent analytically-derived Fournier-Forand phase functions for CHL and TSS were PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

12-023.qxd

10/19/12

8:14 PM

Page 1169

adopted (Fournier and Forand, 1994). The coefficient essentially indicates the backscattering fraction and implies that only 1.8 percent of the incident light field scatters backward in the viewing direction, i.e., 180° from the forward direction. For the solar radiation components, the in situ measured direct solar and diffuse irradiance were provided to Hydrolight for the S2 site while the total and diffuse solar irradiance spectra built into Hydrolight were used for S1 as no coincident measurements were made at the time of imaging (Mobley, 2008). The Pop-Fry absorption and scattering coefficients of pure water were adopted for the both sites (Pope and Fry, 1997). The in-water simulations, totaling ~15,000 reflectance estimates for each site, were conducted over the range of 350 to 730 nm with 10 nm spectral resolution. From this point on, the IOPs for the two sites are assumed known and the concentrations of water constituents are desired. Coupled Simulations It was evident that four ALGE input parameters, including the concentration of particles (COP) [g/cm3], concentration of dissolved matter (COD) [micro-gram/I], particle size (PS) [micro-meter], and particle density (PD) [g/cm3], contribute the most to the prediction of the distribution of water constituents. While PD and PS, which are the average values derived from previous studies (Li, 2007), remain constant during the course of the simulations, the COP and COD can be set as either constants or hourly varying components. Here, the assumption is that the shape of the plume, which has already been determined in the calibration procedure, would remain unchanged by varying the above-noted variables. Preliminary results based on a sensitivity analysis (Pahlevan, 2012) showed that the variations of PS and PD do not significantly affect the concentrations, i.e., approximately 0.01 units of concentrations, and, as a result, the optical field. Thus, the COP and COD were considered the primary factors controlling the optical regime of the waters under study. Furthermore, it was assumed that the COD represents the concentration of CHL in the immediate

(a)

(b)

vicinity of the river discharge and the hydrodynamic processes are the dominant phenomena relative to the biological processes driving the growth/decay of phytoplankton communities. In other words, CHL is modeled in a similar fashion as dissolved matter diluting into lake waters. In order to find the reflectance map that best approximates the L7-derived blue, green, and the red spectral images, ALGE was re-started several times for the last 40 hours by varying the COP and COD components. For each model output, a handful of profiles corresponding to different grid cells, i.e. approximately 35 percent of the area of interest, were supplied to the Hydrolight code to simulate the surface reflectance maps (rd). The subsampling was carried out to reduce the computational burden. Since the main objective of this study is to characterize the plume waters, the Hydrolight simulations were performed only over the plume areas. To select for the best pair of COP-COD, the RMSE and the correlation coefficients (␳␣) calculated between each model output and the L7-derived maps were used as the metrics. The metrics were calculated across solely the three visible bands after the results from the sensitivity analysis indicated minimal improvements by incorporating the NIR band. The best combination of COP-COD is specified in the same manner as in the model calibration stage. After a few iterations of the ALGE simulations, the optimization was conducted to find the optimal solution, which simultaneously satisfies a minimum RMSE and a maximum ␳␣ with respect to the L7 surface reflectance products.

Results and Analysis Model Calibration The calibration process is expected to result in realistic extents and orientations of the plumes. This was carried out by optimizing the modeled surface temperature outputs against that obtained from the remotely sensed data (Equation 2). Plate 1a and 1b show the model outputs

(c)

(d)

Plate 1. Model calibration results at the Genesee River site (a) and (b) shown with the L7-derived surface temperature map (c). The long-term simulation result (a) has improved after iterations and optimization (b). A statistical comparison between the matched model and L7 products are made in (d).


N o v e m b e r 2 0 1 2 1169

12-023.qxd

10/19/12

8:15 PM

Page 1170

associated with the model stabilization and optimization for the Genesee River plume, respectively. Plate 1c, on the other hand, illustrates the spatially smoothed (3 ⫻ 3window) surface temperature obtained from the L7 data. Following the long-term simulation (Plate 1a), it appears that the ALGE model tends to slightly overestimate the extent of the plume while correctly predicting its orientation towards northeast as seen in the L7 data (Plate 1c). The physical shape of the plume is clearly improved (Plate 1b) by refining the input variables after three series of iterations followed by the optimization. Quantitatively, the average RMSE calculated over the plume area was found to be 0.34°C by providing the code with the originally observed meteorological inputs (Plate 1a). Following the optimization, the RMSE was reduced to 0.27°C demonstrating enhanced model predictions. This model prediction resulted from the combination of a 4.9 percent increase in the WS, rotating the wind axis ⫹9.2°, a 10 percent increase of the RD, and a 5.4 percent increase in the RT. The largest degree of adjustment was made for the river discharge, which is measured ~5 km upstream. As noted earlier, the 5.4 percent boost in the river temperature was applied to the daily averaged measurements in degreesCelsius obtained from the nearest creek in the region. The availability of the hourly RT observations for the stream would improve the model performance and mitigate uncertainties related to the other variables. In other words, the input variables are non-linearly correlated and large errors associated with one variable have to be compensated with the other parameters. By more closely comparing Plate 1b and 1c, it was found that the overall temperature throughout the domain, which is driven primarily with the wind stress, appears to be cooler than that of L7. This is due to the 4.9 percent increase in the wind speed, which intensifies the vertical mixing and cooling of the surface waters. It should be emphasized that the best model output was determined by taking a subset over the plume and the spatial patterns in the non-plume areas were avoided. The statically derived plot (Plate 1d) gives a more quantitative way of comparing the model output and the L7-derived surface temperature

(a)

(b)

map over the plume area totaling 124 pixels. The tail in the higher end of the L7-derived histogram can be attributed to the adjacency effects near the pier causing an overestimation of the temperature. The 8-bit quantization of the L7 has led to a less uniform histogram when compared to that of the model. The errors due to the quantization artifacts, especially at the peak of the plume, are noticeable. Plate 2a through 2d illustrates the results for the Niagara River in a similar fashion as for the Genesee River. The iterations followed by the optimization improved the discrepancies between the model output and the reference data, i.e., the average RMSE calculated throughout the domain was reduced from 0.54°C to 0.46°C. The major difference is most notable at the boundaries where time varying temperature profiles were being prescribed from the whole-lake model during the simulation. The warm pattern on the northeast side of the domain is the extension of a large eddy in the middle of the lake, which has not been captured in the localized simulation at the Niagara River. Plate 2b exhibiting the highest consistency with the L7derived thermal map, i.e., 0.46°C, was obtained by multiplying the WS with 119.2 percent (19.2 percent increase), rotating the wind axis ⫹8.8°, and boosting the RD and RT 4.2 percent and 3.1 percent, respectively. Although spatially resampled, the L7-derived temperature map still contains relatively significant instrument induced spatial variability as shown in Plate 2c. Plate 2d illustrates the histogram derived from a subset of pixels (n ⫽ 943). The poor radiometric fidelity of L7 largely due to the quantization artifacts is more noticeable in the Niagara River than in the Genesee River due to the large extent of the plume. It is also interesting to notice the dynamic range associated with the model output and the reference data. L7 has been unable to distinguish the temperature differences in the core of the plume (Plate 2b) when the river temperature has slightly dropped (⌬T ⬍0.24°C) in the last 10 hours of the simulations (Pahlevan, 2012). This is evident from Plate 2b and the time series of the river temperature not presented here. The flow of relatively colder waters has created a relatively cold inner

(c)

(d)

Plate 2. Model calibration results at the Niagara River site shown with the L7-derived surface temperature map. The long-term simulation result (a) has improved after iterations and optimization (b). A statistical comparison between the matched model and L7 products are made in (d).

1170 N o v e m b e r 2 0 1 2


12-023.qxd

10/19/12

8:15 PM

Page 1171

side and a warm ring on the edge of the plume (Plate 2b). It should be noted that the noise equivalent difference in temperature (NE ⌬T) for L7 is about 0.28°K @ 280°K for the high gain setting used in this study (Barsi et al., 2003). In addition to the L7’s radiometric considerations, the ALGE model has not been able to accurately simulate the current velocities/temperature variations at the boundaries. These boundary conditions were extracted from the lake-wide simulations driven by identical river flow inputs (Niagara River) and meteorological data, but with 2D wind fields computed from adjacent weather stations and lake weather buoys. Moreover, due to the computational limitations, the lake-wide simulation was also conducted at a relatively coarse resolution of 1.5 km horizontal and 2 m vertical, that may have not been adequate to capture the subtle circulations/structures required for the plume simulations. Also, note that the individual turbulent eddies apparent in the MODIS image in Figure 1 are much larger than the limited computational domains used for the river plume simulations. It would have been necessary to simulate these eddies accurately in time and location to generate correct boundary conditions for the limited area river plume simulations. Constituent Retrieval Following fixing the shape of the plume in the thermal calibration phase, the sediment load (COP) [g/cm3] and the volume of the dissolved matter (COD) [micro-gram/l] were adjusted by optimizing the modeled water-leaving reflectance (rd) against the L7-derived rd products across the visible bands. As noted, in this study, it is assumed that the CHL distribution is modeled as a dissolved component similar to COD. Plate 3a through 3d illustrate the image- and the modelderived TSS and CHL surface distribution maps in the proximity of the Genesee plume area (1 ⫻ 1 km2). The model outputs shown in Plate 3c and 3d have provided the best agreement with the L7 imagery in the surface reflectance domain (rd), i.e., on average, RMSE ⬍0.0055 [unitless] calculated over the plume. The distribution maps produced from

the L7 imagery were based upon an independent, LUT-based approach where various water types are simulated with the Hydrolight code (Pahlevan and Schott, 2011). It should be noted that the L7-drived concentrations contain artifacts, such as adjacency effects and atmospheric haze, which were identified through inspecting the Short-Wave-InfraRed (SWIR) bands. The west side of the pier, for instance, is clearly affected by the haze and the reflection off the pier resulting in seemingly higher concentrations of water constituents. In addition, the re-suspension phenomenon could also strengthen the in-water scattering, and as a result, greater water-leaving signal in the near-shore area. As expected, the distribution of particles and the dissolved matter has nearly identical shape owing to the relatively low settling velocity of the particles allowing for a realistic prediction of the particle distribution. Plate 3e shows the surface distribution of the TSS and CHL along the plume centerline, i.e., the polyline in Plate 3a, compared for the model and the smoothed L7 imagery. The concentrations are very consistent close to the pier while the disparities increase up to 25 percent towards the end of the transect. The discrepancy is, in part, due to the spatially inhomogeneous atmosphere, as inferred from the analysis of the SWIR bands (not presented here for brevity), and underestimation of the constituents by the ALGE model as the plume expands northward. In addition, the lack of accurate knowledge of the particle size and density, river discharge, river temperature, hourly COP and COD, and perhaps current velocities at the boundaries introduce uncertainties in the model performance. The vertical profiles of the particles (TSS) associated with the pixels along the transect are shown in Plate 3f with increasing depth from right to left. These profiles cannot obviously be quantified through remotely sensed measurements due to the limited penetration depth of the light field through the water column. A relatively uniform vertical distribution in such turbid waters resulting from large turbulence and vertical mixing near the pier is noticeable. The uniformity tends to decrease towards the tip of the plume. As expected, the presence of the thermocline,

(a)

(b)

(e)

(c)

(d)

(f)

Plate 3. The TSS and CHL derived from the L7 (a) and (b); data and the best model output (c) and (d) for the Genesee plume. The surface distribution obtained along the line indicated from the model and L7 products are shown in (e). The vertical profiles derived along the same transect are plotted in (f).


N o v e m b e r 2 0 1 2 1171

12-023.qxd

10/19/12

8:15 PM

Page 1172

(a)

(b)

(e)

(c)

(d)

(f)

Plate 4. The TSS and CHL derived from the L7 (a) and (b); data and the best model output (c) and (d) for the Niagara plume. The surface distribution obtained along the model and L7 products are shown in (e). The vertical profiles derived along the same transect are plotted in (f).

which was observed with a similar trend in the model’s thermal output profiles, has affected the vertical distribution of the TSS in the deeper zone where the concentration are lower at the bottom (⬍3 g/cm3). Plate 4 shows similar plots to those of the Genesee River for the Niagara River plume (19 ⫻ 28 km2). The matched model outputs giving rise to the surface reflectance map that most resembles that of L7. The disparity between the model output and the image products was found to be less than 0.009 units of reflectance, on average, over the plume area. Although the optimal concentration maps exhibit, on average, the minimal disparity against the L7derived concentrations, the maximum concentrations shown on the image-derived products (note the scales) have not been achieved with the model due to inconsistencies in the spatial distribution of the constituents. The differences in the spatial distributions can be identified in Plate 4e where the surface distributions along the transect (the polyline drawn on Plate 4a) across the plume is plotted. Due to the natural co-existence of the phytoplankton with the suspended particles in the plume waters, there is relatively a high correlation (␳␣ ⬎0.62) between the CHL and the TSS concentration maps as shown in the L7-derived products. The ALGE model, however, is unable to take into account biological factors that influence the distribution of the chlorophyll-a in such a large scale. This is evident from surface distribution of CHL across the plume representing constant quantities (Plate 4e). On the other hand, the TSS distribution has been rather well simulated through the model even though incorporating the hourly measured concentrations can significantly enhance the model predictions. The vertical profiles of the TSS concentration along the transect (#1 to #16) is also shown in Plate 4f. At the depth of ~5.5 m, the thermocline causing a relatively large gradient in the concentrations can be clearly identified in most of the extracted profiles. Nonetheless, more uniform distributions across the highly concentrated, well-mixed areas (core of the plume with 1.1 g/cm3) were found.

1172 N o v e m b e r 2 0 1 2

With the calibrated model in the thermal and reflective domains, one can re-start the model over a specific timeframe (e.g., ⫾50 hours) around the hour at which the model was calibrated to either pre-cast or fore-cast the spatial and vertical distributions of the water constituents providing the meteorological/environmental variables are made available through in situ measurements. It should be emphasized that the ALGE model is currently capable of incorporating tidal oscillations and could be used in areas with tide effects. The simulation capabilities of ALGE can also be developed to incorporate biological components affecting phytoplankton communities if needed.

Conclusions This paper introduced a novel technique for the retrieval of surface and vertical distributions of water constituents in river plume environments using a combination of physicsbased models and Landsat-7 imagery. A hydrodynamic model (ALGE) was calibrated in the thermal domain through optimization against the L7-derived surface temperature maps. At this stage, the model’s environmental inputs are refined. The profiles of water constituents were then obtained through optimizing the coupled modeling outputs against the surface reflectance maps produced from the L7 reflective imagery. After finding the optimal model output in the thermal and the reflective domains, not only can the profiles of water constituents be achieved but also the calibrated model can be utilized to pre-cast or forecast the state of the environment in a certain timeframe during which the model remains stable and calibrated. This work will be further extended by improving the model performance through incorporating spatially non-uniform environmental variables, accurate knowledge of particle size-distribution, and biological factors enabling realistic distributions of phytoplankton communities. In addition, the possibility of developing the presented method for potential forecasting will be investigated. With its significantly enhanced performance, the new generation of


12-023.qxd

10/19/12

8:15 PM

Page 1173

Landsat (LDCM) enables a more accurate implementation of our proposed approach. Furthermore, LDCM, when in orbit, complements the L7 mission and reduces its revisiting cycle to eight days allowing for a more frequent calibration of the model.

Acknowledgments This work was funded by the United State Geological Survey (USGS) under the contract No. 06CRCN0024. The authors appreciate the staff at the Research Computing (RC) center of RIT for facilitating the computational burden of this study. We also would like to thank Mrs. Nina Gibson Raqueno and Dr. Rolando Raqueno with the Digital Imaging and Remote Sensing Laboratory (DIRS) for their assistance and the previous research efforts in similar studies conducted at RIT. It should also be noted that the content of this paper does not represent the views or policies of the Department of Interior, nor does mention of the trade names, commercial products or organizations imply endorsement by the US Government.

References Barsi, J.A., J.R. Schott, F.D. Palluconi, D.L. Heider, S.J. Hook, B.L. Markham, G. Chander, and E.M. O’Donnell, 2003. Landsat TM and ETM⫹ thermal band calibration, Canadian Journal of Remote Sensing 29(2):141–153. Berk, A., L. Bernstein, and D.C. Robertson, 1989. MODTRAN: A moderate resolution model for LOWTRAN 7, Spectral Sciences. Binding, C.E., D.G. Bowers, and E.G. Mitchelson-Jacob, 2005. Estimating suspended sediment concentrations from ocean colour measurements in moderately turbid waters; The impact of variable particle scattering properties, Remote Sensing of Environment, 94(3):373–383. Brando, V.E., and A.G. Dekker, 2003. Satellite hyperspectral remote sensing for estimating estuarine and coastal water quality, IEEE Transactions on Geoscience and Remote Sensing, 41(6): 1378–1387. Chen, X., J. Lu, T. Cui, W. Jiang, L. Tian, L. Chen, and W. Zhao, 2010. Coupling remote sensing retrieval with numerical simulation for SPM study-Taking Bohai Sea in China as a case, International Journal of Applied Earth Observation and Geoinformation, 12(Supplement 2):S203–S211. Dekker, A.G., R.J. Vos, and S.W.M. Peters, 2001. Comparison of remote sensing data, model results and in situ data for total suspended matter (TSM) in the southern Frisian lakes, Science of the Total Environment, 268(1-3):197–214. Doxaran, D., R.C.N. Cherukuru, and S.J. Lavender, 2004. Estimation of surface reflection effects on upwelling radiance field measurements in turbid waters, Journal of Optics A: Pure and Applied Optics 6(7):690–697. Fournier, G.R., and J.L. Forand, 1994. Analytic phase function for ocean water, Ocean Optics XII (J. S. Jaffe, editor), Proceedings of SPIE, 2258:194–201 Garrett, A.J., 1995. ALGE: A 3-D Thermal Plume Prediction Code for Lakes, Rivers and Estuaries, Savannah River Technology Center, Aiken, South Carolina 76 p. Garrett, A.J., 2002.Analyses of MTI imagery of power plant thermal discharge, Proceedings of SPIE. Garrett, A.J., and D.W. Hayes, 1997. Cooling lake simulations compared to thermal imagery and dye tracers, Journal of Hydraulic Engineering, 123(10):885–894. Garrett, A.J., J.M. Irvine, A.D. King, T.K. Evers, D.A. Levine, C. Ford, and J.L. Smyre, 2000. Application of multispectral imagery to assessment of a hydrodynamic simulation of an effluent stream entering the Clinch River, Photogrammetric Engineering & Remote Sensing, 66(3):329–335. Gerace, A., and J. Schott, 2009. The increased potential for the Landsat Data Continuity Mission to contribute to case 2 water quality studies, Proceedings of SPIE, San Diego, California.


Gitelson, A.A., J.F. Schalles, and C.M. Hladik, 2007. Remote chlorophyll-a retrieval in turbid, productive estuaries: Chesapeake Bay case study, Remote Sensing of Environment, 109(4):464–472. Gould, R.W. Jr., and R.A. Arnone, 1997. Remote sensing estimates of inherent optical properties in a coastal environment, Remote Sensing of Environment, 61(2):290–301. Hakvoort, H., J. de Haan, R. Jordans, R. Vos, S. Peters, and M. Rijkeboer, 2002. Towards airborne remote sensing of water quality in The Netherlands-Validation and error analysis, ISPRS Journal of Photogrammetry and Remote Sensing, 57(3):171–183. Hayashida, T., J.F. Atkinson, J.V. DePinto, and R.R. Rumer, 1999. A numerical study of the Niagara River discharge near-shore flow field in Lake Ontario, Journal of Great Lakes Research, 25(4):897–909. Jensen, J.R., 2006. Remote Sensing Of The Environment: An Earth Resource Perspective, Prentice Hall, Upper Saddle River, New Jersey. Jensen, J.R., B. Kjerfve, E.W. Ramsey III, K.E. Magill, C. Medeiros, and J.E. Sneed, 1989. Remote sensing and numerical modeling of suspended sediment in Laguna de Terminos, Campeche, Mexico, Remote Sensing of Environment, 28:33–44. Karpouzli, E., and T. Malthus, 2003. The empirical line method for the atmospheric correction of IKONOS imagery, International Journal of Remote Sensing, 24(5):7–16. Lagarias, J.C., J.A. Reeds, M.H. Wright, and P.E. Wright, 1998. Convergence properties of the Nelder-Mead Simplex Method in low dimensions, SIAM Journal on Optimization, 9(1):112–147. Li, Y., 2007. An Integrated Water Quality Modeling System with Dynamic Remote Sensing Feedback, Ph.D. Dissertation, Rochester Institute of Technology, College of Science, Center for Imaging Science, Rochester, New York, 2006, 172 p. Li, Y., A. Vodacek, N. Raqueño, R. Kremens, A. Garrett, I. Bosch, J. Makarewicz, and T. Lewis, 2008. Circulation and stream plume modeling in Conesus Lake, Environmental Modeling and Assessment, 13(2):275–289. Makarewicz, J.C., and M.J. Nowak, 2010. Genesee River Environmental Health Report. Rochester, The College at Brockport, State University of New York, 5 p. Miller, R.L. and J.F. Cruise, 1995. Effects of suspended sediments on coral growth: Evidence from remote sensing and hydrologic modeling, Remote Sensing of Environment, 53(3):177–187. Mobley, C.D., 1994). Light and Water: Radiative Transfer in Natural Waters, Academic Press, Inc. Mobley, C.D., L.K. Sundman, 2008. Hydrolight 5, Ecolight5 User Guide, Bellevue, Washington, Sequoia Scientific, Inc., 97 p. Olmanson, L.G., M.E. Bauer, and P.L. Brezonik, 2008. A 20-year Landsat water clarity census of Minnesota’s 10,000 lakes, Remote Sensing of Environment, 112(11):4086–4097. Ouillon, S., P. Douillet, J.P. Lefebvre, R. Le Gendre, A. Jouon, P. Bonneton, J.M. Fernandez, C. Chevillon, O. Magand, J. Lefèvre, P. Le Hir, R. Laganier, F. Dumas, P. Marchesiello, A. Bel Madani, S. Andréfouët, J.Y. Panché, and R. Fichez, 2010. Circulation and suspended sediment transport in a coral reef lagoon: The south-west lagoon of New Caledonia, Marine Pollution Bulletin 61(7-12):269–296. Pahlevan, N., 2012. An Integrated Physics-based Approach to Demonstrate the Potential of Landsat Data Continuity Mission (LDCM) for Monitoring Coastal/Inland Waters, Ph.D. Dissertation, Rochester Institute of Technology, College of Science, Center for Imaging Science, Rochester, New York, 195 p. Pahlevan, N., A.D. Gerace, and J.R. Schott, 2011.Using thermal remote sensing as a tool for calibrating a hydrodynamic model in inland waters, Proceedinhs of SPIE, Orlando, Florida. Pahlevan, N., and J.R. Schott, 2011. Investigating the potential of the Operational Land Imager (OLI) for monitoring case II waters using a look-up-table approach, Proceedings of the ASPRS Fall Conference and Pecora 18: Forty Years of Earth Observation . . . Understanding a Changing World, Herndon, Virginia, ASPRS. Pope, R.M., and E.S. Fry, 1997. Absorption spectrum (380-700 nm) of pure water. II - Integrating cavity measurements, Applied Optics, 36(33):8710–8723.

N o v e m b e r 2 0 1 2 1173

12-023.qxd

10/19/12

8:15 PM

Page 1174

Raqueno, R.V., 2003. Hyperspectral Analysis Tools for Multiparameter Inversion of Water Quality Factors in the Lake Ontario Rochester Embayment, State University of New York, 2004 College of Environmental Science & Forestry, Ph.D. Environmental Resource Engineering Dissertation, 165 p. Schott, J.R., 2007. Remote Sensing - The Image Chain Approach, New York, Oxford Unversity Press.

1174 N o v e m b e r 2 0 1 2

Schott, J.R., J.A. Barsi, B.L. Nordgren, N.G. Raqueño, and D. de Alwis, 2001. Calibration of Landsat thermal data and application to water resource studies, Remote Sensing of Environment, 78(1-2):108–117. Wang, J., 2001. A Nowcast/Forecast System for coastal ocean circulation using simple nudging data assimilation, Journal of Atmospheric and Oceanic Technology, 18(6):1037–1047.
