Monitoring turbidity in Tampa Bay using MODIS/Aqua 250-m imagery

Transcription

1 Remote Sensing of Environment 109 (2007) Monitoring turbidity in Tampa Bay using MODIS/Aqua 250-m imagery Zhiqiang Chen, Chuanmin Hu, Frank Muller-Karger Institute for Marine Remote Sensing (IMaRS), College of Marine Science, University of South Florida 140 Seventh Ave. S., St. Petersburg, FL 33701, USA Received 22 July 2006; received in revised form 27 December 2006; accepted 29 December 2006 Abstract We developed an approach to map turbidity in estuaries using a time series (May 2003 to April 2006) of 250-m resolution images from the Moderate Resolution Imaging Spectroradiometer (MODIS) on NASA's Aqua satellite, using Tampa Bay as a case study. Cross-calibration of the MODIS 250-m data (originally designed for land use) with the well-calibrated MODIS 1-km ocean data showed that the pre-launch radiometric calibration of the 250-m bands was adequate. A simple single scattering atmospheric correction provided reliable retrievals of remote sensing reflectance at 645 nm (0.002bR rs (645) b0.015 sr 1, median bias= 7%, slope=0.95, intercept =0.00, r 2 =0.97, n=15). A more rigorous approach, using a multiple scattering atmospheric correction of the cross-calibrated at-sensor radiances, retrieved similar R rs (645). R rs (645) estimates, after stringent data quality control, showed a close correlation with in situ turbidity (turbidity= R rs (645) 1.087, 0.9bturbidityb8.0 NTU, r 2 =0.73, n=43). MODIS turbidity imagery derived using the developed approach showed that turbidity in Hillsborough Bay (HB) was consistently higher than that in other sub-regions except in August and September, when higher concentrations of colored dissolved organic matter seem to have caused underestimates of turbidity. In comparison, turbidity in Middle Tampa Bay (MTB) was generally lowest among the Bay throughout the year. Both Old Tampa Bay (OTB) and Low Tampa Bay (LTB) showed marked seasonal variations with higher turbidity in LTB during the dry season and in OTB during the wet season, respectively. This seasonality is linked to wind-driven bottom resuspension events in lower portion of the Bay and river inputs of sediments in the upper portion of the Bay. The Bay also experiences significant interannual variation in turbidity, which was attributed primarily to changes in wind forcing. Compared with the once-per-month, nonsynoptic in situ surveys, synoptic and frequent sampling facilitated by satellite remote sensing provides improved assessments of turbidity patterns and thus a valuable tool for operational monitoring of water quality of estuarine and coastal waters such as in Tampa Bay Elsevier Inc. All rights reserved. Keywords: MODIS; Aqua; Ocean color; Remote sensing reflectance; Atmospheric correction; Bio-optical inversion; Turbidity; River runoff; Sediment resuspension; Tampa Bay estuary 1. Introduction Turbidity is a fundamental index used to assess coastal and estuarine water quality conditions that affect light attenuation and therefore the productivity of planktonic and benthic algae (Cloern, 1987; Cole & Cloern, 1987; Fisher et al., 1999; Pennock & Sharp, 1994), seagrass and coral reefs (Anthony et al., 2004; Moore et al., 1997). Variations in turbidity also help understand the distribution of total suspended solids or sediments (TSS), and therefore processes like coastal erosion and mobilization of chemicals or pollutants (Heyes et al., 2004). Turbidity shows a wide range in spatial and temporal variability in coastal and estuarine waters. Rivers deliver terrestrial Corresponding author. Tel.: address: zchen@marine.usf.edu (Z. Chen). materials to estuaries typically following seasonal patterns, but also in events that may trigger marked interannual variability (Doxaran et al., 2006; Pribble et al., 2001). Currents and waves lead to suspension of bottom sediments, changing turbidity in response to storms and other wind events at tidal and subtidal frequencies (Cloern et al., 1989; Uncles et al., 2002; Wolanski & Spagnol, 2003). Human activities such as transportation and dredging also influence the magnitude and distribution of turbidity (Schoellhamer, 1996). As a result, conventional sampling methods often fail to characterize turbidity dynamics because of the limitations in temporal and spatial sampling (Chen et al., in revision). Satellite-borne sensors could be an ideal tool to assess turbidity because they provide synoptic and frequent mapping capabilities. However, satellite platforms also have limitations such as in spatial resolution, revisit time, accessibility to data, /$ - see front matter 2007 Elsevier Inc. All rights reserved. doi: /j.rse

2 208 Z. Chen et al. / Remote Sensing of Environment 109 (2007) sensor calibration and image processing softwares (Doxaran et al., 2002; Miller et al., 2005; Ruddick et al., 2003; Stumpf & Pennock, 1989). For example, although Landsat ETM+ provides high resolution (30-m) imagery, the revisit time is 16 days. This long revisit time, together with possible cloud cover, makes it inadequate to resolve turbidity dynamics in coastal and estuarine waters. In contrast, some sensors designed to observe marine waters do have shorter revisit time, e.g., neardaily overpasses, but the spatial resolutions are typically too coarse ( 1 km/pixel) to characterize fine-features in turbidity (e.g., the Advanced Very High Resolution Radiometer AVHRR, the Sea-viewing Wide Field-of-View Sensor SeaWiFS, and ocean color bands of the Moderate Resolution Imaging Spectroradiometer MODIS). Although the European Medium Resolution Imaging Spectrometer (MERIS) has full-resolution of 300-m bands with near-daily coverage, the cost of acquisition of these data has limited its wide application. The MODIS flown aboard the Aqua spacecraft, launched in May, 2002, provides near-daily coverage of the subtropical ocean and has two bands that observe the Earth at 250-m resolution (band 1: 645 nm, from nm; band 2: 859 nm, from nm). These bands have sufficient sensitivity to detect a wide range of changes in color of estuarine waters (Hu et al., 2004). Several studies have demonstrated the potential of these bands to monitor water quality in coastal and estuarine waters (Hu et al., 2004; Miller & Mckee, 2004). Similar coverage is possible from the MODIS on the Terra satellite, but at present both the MODIS/Terra 1-km data and 250-m are not well-calibrated and contain more noise compared to the MODIS/ Aqua data (Hu et al., 2004). Thus we focused this study on using MODIS/Aqua data, but similar approach developed in this study should be applicable to MODIS/Terra data as well. Three key issues must be resolved before MODIS 250-m data can be routinely applied in coastal studies as proposed by Hu et al. (2004). First, the reliability of at-sensor radiances at the top of atmosphere (TOA) has to be assessed. The MODIS 250-m bands were originally designed to serve as sharpening bands to detect land, aerosol, and cloud features. Therefore, there has been no effort to apply an oceanic vicarious calibration such as done for the ocean color 1-km bands. However, remote sensing of aquatic environments demands rigorous and accurate sensor calibration because a 5% error in at-sensor radiances may result in a 50% remote sensing reflectance (R rs ) error. In addition, like most sensors Aqua signals may degrade with time, thus correction for potential degradation effects is prerequisite to use the MODIS 250-m data for a consistent time series analysis. Second, a method has to be available to estimate R rs of aquatic environments from at-sensor radiances. Accurate retrieval of R rs is critical as all other satellite products are derived from R rs estimates. Although several atmospheric correction schemes were developed for the 250-m bands (Hu et al., 2004; Miller & Mckee, 2004), the accuracy of these methods has remained untested and requires a full validation using in situ R rs measurements. Finally, a robust bio-optical algorithm (an algorithm to convert R rs to water quality parameters such as turbidity or TSS) needs to be developed and validated as well for a study region. The objective of the study was to develop an approach for routine application of MODIS 250-m data in monitoring turbidity in Tampa Bay, Florida (USA), for better characterizing its temporal and spatial variability, and thereby also provide a pathfinder approach to better study other estuarine environments. To this end, we systematically evaluated the integrity of the methods developed for applying MODIS 250-m data. We first cross-calibrated MODIS 250-m bands with the wellcalibrated 1-km ocean color bands. Second, we validated an atmospheric correction algorithm using in situ R rs measurements. Third, we developed and validated an empirical algorithm by applying rigorous data quality control process to convert the atmospherically corrected R rs to turbidity. Finally, to evaluate the algorithm and potential for operational application of MODIS 250-m data, a time series of MODIS/Aqua 250-m images collected between May 2003 and April 2006 were processed. Then, composites of MODIS turbidity images were created to represent calendar monthly means for each year and long-term cross-year monthly means (i.e. averaging turbidity over the same months across all years). The variation patterns observed in the MODIS turbidity composites were compared with in situ observations from an independent monitoring program. 2. Materials and methods 2.1. Tampa Bay estuary Tampa Bay is Florida's largest open water estuary, featuring a surface area of 1000 km 2. It is conventionally subdivided into 4 mainstem sub-regions, namely Old Tampa Bay (OTB), Hillsborough Bay (HB), Middle Tampa Bay (MTB), and Lower Tampa Bay (LTM) (Fig. 1). The average water depth is 4.0 m with a navigation channel N10 m running along the central portion of the Bay (Fig. 2). This channel plays an important role in the circulation patterns and water quality distribution across Tampa Bay (Bendis, 1999; Weisberg & Zheng, 2006). In the decades prior to the 1980's, Tampa Bay was heavily polluted by nutrient loading from sources like sewage and wastewater, which caused significant eutrophication problems. The high phytoplankton concentrations decreased water clarity in the Bay and thus led to substantial losses of seagrass cover (Janicki et al., 2001; Johansson, 2000). Since then, ecosystem restoration efforts have gradually improved the water quality of Tampa Bay and some of the seagrass has recovered (Lewis et al., 1998; Tomasko et al., 2005). At present variation in water quality in Tampa Bay is suggested generally due to seasonal and inter-annual variability in precipitation or river runoff to the Bay (Janicki et al., 2001; Lipp et al., 2001; Schmidt & Luther, 2002) Field data Turbidity data (reported in nephelometric turbidity units or NTU) were obtained from the Environmental Protection Commission of Hillsborough County's (EPCHC) Tampa Bay water quality monitoring program (Boler et al., 1991). The EPCHC conducts monthly surveys that span 3 weeks with each week covering approximately one segment of the Bay (Fig. 1). Water

3 Z. Chen et al. / Remote Sensing of Environment 109 (2007) Fig. 2. NOAA/USGS merged bathymetric/topographic digital elevation model (DEM) of Tampa Bay at 30 m spatial resolution (Gesch & Wilson, 2001). Data were binned to 250-m to match the MODIS 250-m resolution. Grey and white represent land and missing data, respectively. Fig. 1. MODIS 250-m image of Tampa Bay showing the four sub-regions of the Bay and major tributaries. The sub-regions are Old Tampa Bay (OTB), Hillsborough Bay (HB), Middle Tampa Bay (MTB), and Lower Tampa Bay (LTB). The rivers are the Hillsborough River (HB), the Alafia River (AR), the Little Manatee River (LMR), and the Manatee River (MR). The Environmental Protection Commission of Hillsborough County's (EPCHC) water quality monitoring stations are overlaid with various symbols to indicate different sampling times in each month: diamond (OTB) generally in the first week, triangle (HB) in the second week, and square (MTB and LTB) in the last week. Five stations marked with white crosses and labeled with numbers are also overlaid (stations 92, 23, 14, 40, 55 from the EPCHC program). The inset shows the location of the Tampa Bay estuary in the state of Florida, USA. samples were collected at mid-depth of a station where depth was greater than 3 m, otherwise only surface samples were collected. In the lab, the turbidity of water samples was measured with a Hach Model 2100N Turbidimeter. The delay between sample collection and analysis was usually b 24 h. Turbidity readings are expected stable within this timeframe based on protocols sanctioned by the US Environmental Protection Agency (EPA 180.1) (Joe Barron/EPCHC, personal communication). The device measures light intensity (peak spectral response is 570 nm) at 3 different angles (90 scattered, forward scattered, and transmitted light). The light scattered at 90 normalized to the total scattered light at 3 angles was calibrated with four standard solutions ranging from 2 to 2000 NTU (Hach Model 2100N Turbidimeter manual, 1999). Thus, turbidity measures a bulk scattering by particles in water samples, and has been frequently used as an indicator of sediment concentrations. However, the relationship between turbidity and sediment concentration varies depending on sediment properties (e.g., size distribution, shape, composition, and refractive index) (Twardowski et al., 2001; Ulloa et al., 1994) and measurement uncertainties. For example, when Table 1 Symbols, definitions and units Symbols Definitions Units t 0 Diffuse transmittance from the sun Dimensionless to the ground due to Rayleigh and aerosol scattering t v Diffuse transmittance from the Dimensionless ground to a sensor due to Rayleigh and aerosol scattering t oz Diffuse transmittance from the sun to the Dimensionless ground and from the ground to a sensor due to absorption by ozone L a Radiance from aerosol scattering and mw cm 2 μm 1 sr 1 aerosol-rayleigh interaction L r Radiance from Rayleigh scattering mw cm 2 μm 1 sr 1 in the absence of aerosols L t Total at-sensor radiance mw cm 2 μm 1 sr 1 L w Water-leaving radiance at the sea surface mw cm 2 μm 1 sr 1 L gn Normalized sun glint radiance if there were Sr 1 no atmosphere and solar irradiance F 0 =1 θ 0 Solar zenith angle Degree θ v Sensor zenith angle Degree F 0 Adjusted extraterrestrial solar irradiance mw cm 2 μm 1 For simplicity, the wavelength dependency of all terms (except solar and sensor angles) is suppressed.

4 210 Z. Chen et al. / Remote Sensing of Environment 109 (2007) sediment concentrations are low, relative uncertainty in measurements of sediment concentrations using the traditional dry method can be larger (Christian & Sheng, 2003). Remote sensing reflectance (R rs,sr 1 ) data were collected during several field surveys on October 2003, 1 3 June 2004, October 2004, 13 December 2005, and 27 February 2006, respectively, using a handheld PR650 (Photo Research Inc.) or Analytical Spectral Device (ASD Inc.) spectroradiometer, following the method described in Hu et al. (2004). The in situ hyperspectral R rs data were integrated over the relative spectral response (RSR) function of MODIS band 1 to obtain R rs (645) for MODIS validations Cross-calibration of the MODIS 250-m with 1-km bands Vicarious calibration of satellite at-sensor radiance is the procedure used to adjust a measured radiance to a predicted radiance, based on well-calibrated measurements and a radiative Fig. 3. (a) MODIS at-sensor radiance at 667 nm (1-km) on 13 December 2004, 18:43 GMT over the study area; (b) MODIS at-sensor radiance at 645 nm (250-m) from the same satellite pass. The dashed lines on the 250-m image are due to sensor scan errors. Data in the rectangular box offshore (10 10 and pixels for the 1-km and 250-m images, respectively) were used for cross-calibration in this and other ten images collected between 2003 and 2006.

5 Z. Chen et al. / Remote Sensing of Environment 109 (2007) After correction for atmospheric light absorption due to ozone (Hu et al., 2004), at-sensor radiances, L t, can be predicted as (for brevity the wavelength dependency is suppressed here): L t ¼ L r þ L a þ t v L w ; ð1þ (Definitions of terms are provided in Table 1.) In this equation the effects of sun glint and whitecaps are omitted because during the quality control step data associated with these artifacts can be discarded. The effects of water vapor, oxygen absorption, or light polarization are negligible for the 250-m bands (e.g., Meister et al., 2005). The terms on the right-hand side of Eq. (1) were estimated from the MODIS/ Aqua 1-km ocean bands in the following way: Fig. 4. Scatter plots of the at-sensor radiance (mw cm 2 μm 1 sr 1 ) in the two 250-m bands (645 nm in (a) and 859 nm in (b)) as measured by the sensor and predicted by the calibrated 1-km ocean color bands. The dashed lines are 1:1 lines. (See Table 2 to for statistics). transfer model. This procedure is critical because the pre-launch laboratory calibration of ocean color sensors typically has uncertainties of 2 5%, which can translate to 20 50% relative errors in the retrieved R rs after atmospheric correction. The vicarious calibration and the subsequent atmospheric correction using identical radiative transfer codes can be considered to be a self-tuning process. Vicarious calibration of ocean color sensors is typically performed at a well-characterized open ocean site (e.g., spatially homogeneous water away from land), and with well-calibrated instrument (e.g., the Marine Optical Buoy or MOBY at Hawaii, USA). When this option is not available, an alternative is to calibrate one sensor against another well-calibrated sensor, for example calibrating the Moderate Optoelectrical Scanner (MOS) on the Indian Remote Sensing Satellite (IRS-P3) using SeaWiFS (Wang & Franz, 2000) or calibrating Landsat/ETM+ using SeaWiFS (Hu et al., 2001b). We hereby refer to the calibration using observations from a satellite sensor as a crosscalibration to differentiate a calibration using in situ measured radiances. We took an approach similar to that of Hu et al. (2001b) to calibrate the 250-m bands using MODIS/Aqua 1-km ocean bands. Because all MODIS bands have identical solar/ viewing geometry, the procedure is simpler than those used to cross-calibrate two sensors on separate satellites as in Hu et al. (2001b). In short, the 250-m data are adjusted according to the at-sensor radiances predicted by the 1-km data. For clarity we briefly describe this cross-calibration procedure below. 1. A group of MODIS images from May 2003 to April 2006 were chosen for the cross-calibration based on following criteria: 1) cloud free in the entire study area (central west Florida shelf); 2) large dynamic range in L t to apply the cross-calibration over a wide range of radiances; 3) the scenes are as evenly distributed as possible in time series. Ultimately, eleven images were selected randomly from the group for this study. 2. For each image, a clear-water ocean area adjacent to Tampa Bay was chosen as a calibration site (Fig. 3). The choice of a clear-water site ensures a better performance of standard atmospheric correction schemes and avoids possible saturation issue at 667 nm because L t (667) typically ranges from mw cm 2 μm 1 sr 1, as shown in Fig. 4, which is below the saturation radiance specified for this band (4.2 mw cm 2 μm 1 sr 1, Franz et al., 2006). 3. MODIS 1-km Level 1 data were processed using SeaDAS 4.8 to estimate L r, L a, t v, and L w for each of the 1-km wavelengths (λ=412, 443, 488, 531, 551, 667, 678, 748, 869 nm, respectively), using the similar procedure described by Hu et al. (2001b) for SeaWiFS, including adjustment to the center wavelength. 4. These multi-spectral data were used to construct an artificial hyperspectral dataset by interpolation, covering 400 to 900 nm. L w (859) for the clear-water site was assumed to be equal to zero, and L w (645) was approximated as 1.30 L w (667) according to extensive in situ hyperspectral measurement collected from the West Florida shelf Table 2 Comparison of the 250-m at-sensor radiances (mw cm 2 μm 1 sr 1 ) between the measurements and those predicted from the 1-km ocean color bands Band Ratio a % difference b Slope Intercept r 2 RMSE c n L t (645) L t (859) a Median ratio between the measured and predicted radiances, indicating overall biases. b Median absolute percentage difference (MPD) between the measured and predicted at-sensor radiance, indicating typical uncertainties (Bailey & Werdell, 2006). c RMSE represents root mean square errors of the linear regression fitting, in units of mw cm 2 μm 1 sr 1.

6 212 Z. Chen et al. / Remote Sensing of Environment 109 (2007) Fig. 5. Remote sensing reflectance in Tampa Bay at band 1 (R rs (645), sr 1 ) from in situ measurements and MODIS estimates derived using 1) multiple scattering atmospheric correction of the cross-calibrated at-sensor radiance (filled circles); and 2) single scattering atmospheric correction of the original (i.e., pre-launch calibration) at-sensor radiance (open circles). The dashed and dotted line represents the 1:1 ratio (Comparison results are listed in Table 3). (Cannizzaro et al., in press). Sensitivity analyses indicated that results were similar for factors in the range , because L w (645) typically accounts for b5% of L t (645). A 40% increase in L w (645) would increase L t (645) by b2%. At the clear-water location the contribution of L w (645) to L t (645)isexpectedtobeevensmaller.Lee et al. (2005) used a threshold of 1.2 to approximate a R rs relationship between at 640 nm and at 667 nm, which is close to our value. 5. The simulated hyperspectral data (except the L w term) were integrated over the bandpass of the 250-m bands, modulated by the relative spectral response (RSR) functions R Lx ðλþsðλþdλ L x ðbandþ ¼ R ; ð2þ SðλÞdλ where S(λ) is the RSR function of a 250-m band, x can be r or a from (1), and band is 645 or 859 nm, i.e. the center wavelengths of the 250-m resolution bands Satellite image processing MODIS L1B direct broadcast data were captured in real-time by an X-band antenna located at the University of South Florida (USF) in Saint Petersburg, Florida. For regions without a local antenna, historical data can be obtained from the NASA Table 3 Comparison between in situ versus MODIS remote sensing reflectance at 645 nm (R rs (645), sr 1 ), with the latter derived using 1) single scattering approximation and the original at-sensor radiance (i.e., pre-launch calibration) and 2) multiple scattering method and the cross-calibrated at-sensor radiance Calibration Ratio % difference Slope Intercept r 2 RMSE n Pre-launch Cross-calibration The ratio, % difference, and RMSE are defined in the same way as in Table 2. Fig. 6. Relationship between MODIS remote sensing reflectance at band 1 (R rs (645), sr 1 ) and in situ turbidity (NTU) in a log log scale. The power function instead of a linear function is used here to take into account of the non-linear interaction between R rs (645) and turbidity when turbidity is high. Goddard Space Flight Center's Level 1 and atmosphere archive and distribution system (LAADS) ( gov/data). Clouds were masked according to the TOA reflectance (sr 1 ) at 859 nm, R t ¼ L t F 0 cosh 0 ; with a threshold of sr 1. This filters out most clouds and severe sun glint contaminations. Atmospheric correction was then carried out using the method of Hu et al. (2004) summarized as the following equation: R rs ð645þ ¼ ð Þ L rð859þ F 0ð645Þ F 0 ð859þ L rð645þ : t 0 ð645þt v ð645þf 0 ð645þcosh 0 L t ð645þ t oz ð645þ L tð859þ t oz 859 Both single scattering approximation and multiple scattering exact calculations (used for cross-calibrated at-sensor radiances) were used to estimate L r Satellite in situ comparison In coastal and estuarine waters, it is difficult to make direct comparisons between satellite and in situ observations due to the high spatial and temporal variability in water and atmosphere properties. To accomplish this, satellite data must be collected within certain hours of in situ measurements (Bailey & Werdell, 2006) to be considered as concurrent. Here the number of hours was chosen to be 2 for turbidity comparison but 4 for R rs validation to allow for a larger number of matching observations. Second, to minimize sensor noise (Hu et al., 2001a), a 3 3 pixel box centered on an in situ measurement site was sampled. Within this box, there had to be at least 4 valid (see below) pixels, and the coefficient of variance of the valid pixels had to be b0.4, following the value used by Harding et al. ð3þ ð4þ

7 Z. Chen et al. / Remote Sensing of Environment 109 (2007) (2005). The median value of the valid pixels was then chosen in a comparison to in situ data. Besides above general quality controls, some more rigorous quality controls were applied during MODIS-in situ comparison to remove those suspicious MODIS observations. By trial and error, pixels that meet one of the following criteria were considered invalid: 1) Water depth b 2.8 m (large bottom reflectance contamination). This value was chosen as such that it yielded a close relationship between turbidity and R rs (645) and further increasing this depth threshold did not significantly improve the relationship. Interestingly, this value is very close to the predicted light penetration depth at 645 nm (1 / K d (645), where K d (645) is the diffuse attenuation coefficient of the Fig. 7. Long-term cross-year (May 2003 to April 2006) monthly means of turbidity derived from MODIS 250-m data. White color within Tampa Bay represents the mask of shallow water (bottom depth b2.8 m). Grey color represents land.

8 214 Z. Chen et al. / Remote Sensing of Environment 109 (2007) Fig. 7 (continued ). water column that can be approximated as absorption coefficient of water molecules for medium to low turbid and CDOM waters: 0.33 m 1 )(Lee et al. 2005; Pope & Fry, 1997); 2) θ v N50 (large scan edge); 3) (R t (859) R r (859)) N0.019 sr 1 (large aerosol conditions); 4) Normalized sun glint radiance (L gn )N sr 1, following the method of Wang and Bailey (2001) (large sun glint contamination). 3. Results 3.1. Cross-calibration and validation Fig. 4 shows the comparison between the ozone-corrected atsensor radiances in the 250-m bands and those predicted from the 1-km bands. Comparison statistics are presented in Table 2. The comparison clearly shows that both MODIS 250-m bands have excellent linear relationship with the estimates derived

9 Z. Chen et al. / Remote Sensing of Environment 109 (2007) from the 1-km bands, with median ratios (overall biases) of 0.96 and 1.01 for the 645 nm and 859 nm bands, respectively. The median absolute percentage differences (MPD, overall uncertainties, Bailey & Werdell, 2006) were 3.61% and 1.30%, respectively. These uncertainties correspond to 2 counts for the 645 nm band (L t (645) is normally N60 counts with TSS ranging between 2 15 mg L 1 in Tampa Bay) and b1 count for the 859-nm band. From the slopes and intercepts of the linear regressions, the vicarious gains and offsets for the two 250-m bands are [1.0904, 0.067] and [1.0303, ], respectively. The at-sensor radiances were adjusted using these gains and offsets and then the rigorous atmospheric correction (using multiple scattering for L r ) was applied to process the MODIS imagery available for the match-up of R rs (645). Fig. 5 shows the comparison of R rs (645) between MODIS and in situ observations and Table 3 presents the statistics of the comparison. To reveal what improvement can be achieved using more rigorous methods (i.e. cross-calibration and multiple scattering), the comparison of R rs (645) derived from a single scattering atmospheric correction and the original at-sensor radiance (i.e. without cross-calibration) is also shown in Fig. 5 and Table 3. The improvement in terms of bias (5% versus 7%) or other measures is not significant (paired Student's t-test, p b 0.001), indicating that the pre-launched radiometric calibration for MODIS 250-m bands and the single scattering atmospheric correction scheme are adequate for this time period. Therefore, the pre-launch calibrations (i.e., original at-sensor radiances) and the single scattering atmospheric correction scheme were used to derive the entire time series of MODIS R rs (645) imagery between May 2003 and April Turbidity from satellite R rs After stringent quality control of the satellite data (see Materials and methods Section), a total of 43 satellite and in situ matching pairs were obtained. These matching stations are distributed across the four bay sub-regions and throughout four seasons, thus representing an overall relationship between MODIS and in situ measurements. Fig. 6 shows that MODIS R rs (645) ranging from to sr 1 is closely correlated with in situ turbidity values from 0.9 to 8.0 NTU (turbidity= R rs (645) 1.087, r 2 =0.73, n=43). More importantly, this relationship appeared to be stable over 2004 and 2005 (Fig. 6), indicating that the regression is timeindependent. Indeed, the relationship derived for 2005 alone can accurately predict 2004 turbidity (median predicted versus in situ ratio of 0.98, with median absolute percentage difference/mpd of 10%). Thus the regression relationship showed in Fig. 6 and the rigorous data quality control criteria (see Materials and methods Section) were applied to the MODIS R rs (645) imagery to obtain the time series of turbidity maps Image series of turbidity The long-term cross-year monthly mean MODIS maps (May 2003 to April 2006) help characterize the spatial and temporal patterns of turbidity in Tampa Bay (Fig. 7). Turbidity in Hillsborough Bay (HB) was consistently higher than in other sub-regions except in August and September. At this time of year, however, turbidity in HB may be underestimated due to high input of colored dissolved organic matter (CDOM) from major rivers (see Fig. 1 and discussion below). In comparison, turbidity in the upper Middle Tampa Bay (MTB, around Station 14) was generally lowest as compared to estimates in other bay sub-regions throughout the year. Both Old Tampa Bay (OTB) and Low Tampa Bay (LTB) showed marked seasonal variations. LTB, particularly near the bay mouth, showed high turbidity between November March and lower turbidity between May October, while opposite seasonality was observed in OTB. A bay-wide higher turbidity was observed in April. This seasonality across the Bay becomes more apparent in extracted time series of turbidity at stations from the sub-regions (Fig. 8). For example, the upper bay (Sta. 55 in HB and Sta. 40 in OTB; see Fig. 1 for locations) typically showed higher turbidity than MTB and LTB, particularly in the wet season (from May to October, turbidity N3.0 NTU). MTB (Sta. 14) showed the lowest turbidity and seasonality (typically 2.0 NTU). LTB showed moderate turbidity but the largest seasonal variations. The MODIS turbidity images also show that the Bay experiences significant interannual variations in turbidity. Fig. 9 shows the turbidity variations for April of 2004, 2005, and 2006 with highest turbidity in April, 2005 and lowest turbidity in April, The month of April was chosen because it typically is the month with highest wind speed and turbidity, and therefore turbidity in this month can be used to highlight the potential interannual variations. The observed interannual variation is consistent with changes in wind forcing (Fig. 10). Based on the observations of Chen et al. (in revision), the threshold of daily averaged wind speed, above which sediment resuspension occurs, is about 6.0 m s 1. Thus there existed 7, 8, and 2 sediment resuspension events for April of 2004, 2005, and 2006, respectively. Indeed, previous studies showed that wind is the most important factor controlling sediment Fig. 8. Long-term cross-year monthly means of turbidity (May 2003 and April 2006) derived from MODIS at several stations in Tampa Bay (see Fig. 1 for station locations). Station 55 was affected by high CDOM and cloud cover during late summer (i.e. no valid data in August, and possible underestimate of turbidity in September).

10 216 Z. Chen et al. / Remote Sensing of Environment 109 (2007) Fig. 9. Monthly turbidity images from MODIS 250-m data (top panels) and from interpolated in situ measurements (lower panels) for April of 2004, 2005, and 2006 (no in situ measurement was available for April 2006). Crosses in lower panels represent in situ sample stations with bottom depth N2.8 m, and the values shown are in situ turbidity measurements. The in situ monthly map is a graphical composite of single observations taken at different times in different parts of Tampa Bay and is not a true monthly mean. The number of satellite images used to create turbidity composites for April 2004, 2005, and 2006 are 4, 5, and 3, respectively, for most bay areas. The days when images were used are superimposed on Fig. 10 using arrows. The variation in the number of observations across the Bay is shown in Fig. 11.No value is shown for April resuspension in Tampa Bay relative to other factors (e.g., tides) (Chen et al., in revision; Schoellhamer, 1995) and river inputs of sediments seem to influence on turbidity only in the upper portion of the Bay, but have minimal or negligible impact in lower portion of the Bay (Chen et al., in press). In April, river inputs of sediments are usually at minimum level, thus wind forcing becomes a dominant factor determining turbidity variations. Fig. 9 also shows the contrast of turbidity between satellite observations and in situ measurements. Over larger scales, the monthly MODIS turbidity estimates were generally consistent with those determined from in situ observations. For example, both in situ and MODIS data show turbidity in April, 2005 was higher than in April, However, there were some important differences at finer scales. The differences are partly due to artifacts introduced by interpolation (Barnes, 1994) of the

11 Z. Chen et al. / Remote Sensing of Environment 109 (2007) Fig. 10. Daily averaged wind speed measured at Port of Manatee (27.64 N, W) in April 2004, 2005, and 2006 (data courtesy of NOAA). The dashed line represents 6.0 m s 1 above which a sediment resuspension is assumed to occur. Superimposed arrows mark the days when MODIS had valid observations for 2004 (blue, 4 instances), 2005 (red, 5 instances), and 2006 (black, 3 instances). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) in situ data, but in great part due to the mismatch in sampling frequencies between the two data sets. Unlike non-synoptic and single snapshot in situ measurements, MODIS composites were created using multiple and synoptic observations (Fig. 10). Similarly, while the single monthly in situ turbidity observations are frequently within 1 standard deviation of the corresponding satellite mean estimates (Fig. 11), MODIS mean turbidity showed less short-term temporal variability but more pronounced seasonal and interannual variability than in situ observations. For example, high in situ turbidity values (N10.0 NTU) were observed at Sta. 55 in October and December 2003, respectively, but the monthly MODIS means showed lower turbidity comparable to those seen in other months. These high turbidity events sampled by in situ measurements mask the broader and more persistent relative increases in turbidity observed during the wet season. Therefore the present in situ sampling protocol is unable to effectively detect seasonal changes (Fig. 8). The high in situ turbidity observations at Sta. 23 in late 2003 and at Sta. 92 in early 2004 also obscure the overall increases in turbidity for late 2004 and early 2005, which are, however, pronounced from MODIS data (Fig. 11). These improvements are clearly related to increased number of high-quality MODIS observations. MODIS typically has N 4 high-quality observations per month (about once per week), and N10 observations per month were collected frequently between October and March (Fig. 11). These multiple observations help minimize aliasing, thus construct more realistic monthly means than using the single in situ monthly measurements. In addition to this increased temporal resolution, the satellite observations also provide higher detail of the spatial heterogeneity in turbidity patterns and help reduce introduced artifacts during spatially extrapolating a limited number of in situ measurements using an interpolation method (e.g., Barnes, 1994). However, MODIS observations have limitations because valid satellite observations require cloudless skies and minimal sun glint interference. These requirements reduce the number of MODIS observation, thus MODIS did not have valid observations in some months of the wet season (Figs. 8 and 11) Fig. 11. Example time series of monthly turbidity estimates derived from in situ and MODIS measurements at selected stations shown in Fig. 1. The period covered is May 2003 to December 2005 (no in situ data available for 2006). The number of MODIS observations during each month is shown on the right-hand side. Where number of samples was N1, the standard deviation of the monthly mean is also shown on the MODIS data. Please note that there is only one in situ observation in each month. Circles highlight the relatively high turbidity observed from late 2004 to early 2005 in Lower Tampa Bay, which are not observed with the in situ measurements. The y-axis scale for station 55 is different from that in the other panels.

12 218 Z. Chen et al. / Remote Sensing of Environment 109 (2007) or missed some larger sediment resuspension occurrences (Fig. 10). 4. Discussion 4.1. Calibration and algorithm issues The cross-calibration and multiple scattering atmospheric correction schemes produced similar results to those obtained with the simpler atmospheric correction of data calibrated with pre-launch calibration coefficients. The agreement indicates that the radiometric calibration at 250-m bands is stable throughout our 3-year study period and the pre-launch calibration is adequate for a consistent time series analysis during this period. In the long run, however, periodic cross-calibration of the 250-m bands using the 1-km bands is preferred because the latter are corrected for sensor degradation effects using periodic lunar or solar calibrations. Also, at present atmospheric correction over coastal waters is still a challenge due to non-zero R rs in the near-infrared and blue-absorbing aerosols. Although atmospheric parameters in the 250-m bands may be estimated from the 1-km ocean color bands (e.g., Franz et al., 2006), the latter are known to have problems over coastal waters (Hu et al. 2000; Siegel et al., 2000). While the validation result showed good agreement between in situ and satellite R rs using our simple schemes, a potential error source may be the white aerosol assumption, which may yield errors of several counts (e.g., Hu et al., 2001b) or in the order of 5 10% in the retrieved R rs (645) in case of non-white aerosol. The black-pixel assumption (L w (859) =0) may also be problematic in very turbid waters. In such cases, an alternative correction may be effected by transferring nearby aerosol reflectance from clear waters to the turbid pixels (e.g., Hu et al., 2000; Miller & Mckee, 2004). Recently, Wang and Shi (2005) proposed to use longer wavelengths in the short-wave infrared (SWIR) to circumvent this problem, but the increased distance between the wavelength of interest (645 nm) and the reference wavelength (1640 nm) may yield some errors in the aerosol reflectance extrapolation. Future sensors may include another SWIR band to facilitate a more accurate atmospheric correction. 5. Effects of colored dissolved organic matter (CDOM) The low turbidity observed in Hillsborough Bay (HB) with MODIS in the wet season (e.g, August and September) is likely an artifact due to the interference by CDOM. The CDOM absorption coefficient at 645 nm (a cdom (645)) in the wet season can be 0.40 m 1 (Chen et al., in press), i.e. greater than the absorption due to pure water (a w (645) 0.33 m 1, Pope & Fry, 1997). In theory, when bottom contribution is negligible, R rs is proportional to b b /(b b +a) where b b is the total backscattering coefficient (dominated by particles in the red wavelengths), equivalent to turbidity, and a is the total absorption coefficient. Therefore, high CDOM absorption will reduce R rs (645) and yield low, unrealistic turbidity retrievals. Similar effects have also been observed elsewhere (e.g., Woodruff et al., 1999). This problem was, however, limited to HB in the wet season only and is generally negligible in other bay segments and other seasons, since a cdom (645) is typically b0.03 m 1 in those regions and periods (Chen et al., in press). The similarity in large-scale patterns between MODIS and in situ turbidity measurements (Fig. 11) also supports this conclusion. 6. Applicability to other regions The validations and derived MODIS images show that our approach successfully mapped turbidity in Tampa Bay, can the same approach be applied to other estuaries? The approach was based on the satellite-derived, validated R rs, which is proportional to b b (equivalent to turbidity) for medium to low turbid waters where both bottom and CDOM contributions at the 645 nm wavelength are negligible. Under these circumstances, R rs is proportional to turbidity /a w where a w is a constant (0.33 m 1 ). Therefore, the algorithm developed here should be site-independent and applicable to other estuaries as long as a same turbidity protocol (e.g., the same turbidity calibration standards) is adopted, unless in extremely turbid waters where R rs becomes a non-linear function of b b so the algorithm needs to be tuned to account for the non-linearity. On the other hand, because the mass-specific absorption and scattering coefficients of sediments can vary substantially from a region to another region (Babin et al, 2003; Chami et al., 2005), estimating sediment concentration from remote sensing does require a regional bio-optical algorithm. Similarly, when applied to other regions, those data quality control criteria empirically developed for the Tampa Bay may also need to be modified for other regions. 7. Conclusions We addressed a range of issues that advance the transition of research towards operational application of MODIS 250-m data for monitoring the turbidity of estuarine waters, using Tampa Bay, Florida, as a case study. This included cross-calibration of the MODIS 250-m bands with the 1-km ocean color bands and a rigorous, multiple scattering atmospheric correction. The combination of the two led to no substantial improvements over using the pre-launch calibration and a simple, single scattering atmospheric correction approach. We also found that turbidity can be estimated with a simple algorithm that is relatively stable from year to year, and possibly site-independent under certain circumstances. However, there still remain several issues to be addressed, specifically the interference of high concentrations of colored dissolved organic matter (CDOM) on turbidity estimates, and the contamination of the signal by shallow (b2.8 m) bright bottom reflectance. The synoptic MODIS 250-m resolution images provided multiple observations of turbidity of Tampa Bay per month. This time series showed some features that were similar to those inferred from the once per month in situ turbidity observations. However, the MODIS 250-m maps of turbidity helped differentiate between short-time variability and seasonal and interannual changes. MODIS data provided an improved

13 Z. Chen et al. / Remote Sensing of Environment 109 (2007) representation of the monthly mean state of turbidity across Tampa Bay, as well as of seasonal and inter-annual variability. These results show that turbidity in Hillsborough Bay (HB) was consistently higher than in other sub-regions except in August and September, when higher concentrations of CDOM seemed to cause underestimates of turbidity. Turbidity in Middle Tampa Bay was generally lowest across the Bay throughout the year. Both Old Tampa Bay (OTB) and Low Tampa Bay (LTB) showed marked seasonal variations with higher turbidity in LTB during the dry season and in OTB in wet season respectively. This seasonality is related to wind-driven bottom resuspension events in the lower portion of the Bay and river inputs of sediments in upper portion of the Bay. The Bay also experiences large interannual variation, primarily caused by the changes in wind forcing over the region. Because of the repeated, frequent, and synoptic coverage of estuarine regions provided by MODIS satellite data across the globe, and because these data are robust, low cost, and simple to use, satellite data from sensors such MODIS are important complements of traditional in situ water quality surveys. We recommend incorporating MODIS observations in operational monitoring of water quality of moderate size estuaries and of coastal regions. Acknowledgements This work was funded and logistically supported by the Tampa Bay Integrated Science Project of the U.S. Geological Survey (USGS), Coastal and Marine Geology Program. Additional financial support was provided by the USF- USGS Cooperative Graduate Assistantship Program, and by the National Aeronautics and Space Administration (NASA NAG ). Logistical support was provided by the operational section of the USGS Center for Coastal and Watershed Studies. MODIS data collection and processing were made possible by the data distribution and software development efforts of the NASA GSFC MODIS Project and affiliated Science Team members. We also particularly thank the Environmental Protection Commission of Hillsborough County (EPCHC) for their willing to share Tampa Bay water quality monitoring data. We also thank two anonymous reviewers for their comments and suggestions, which significantly improve this manuscript. IMaRS contribution # 109. References Anthony, K. R. N., Ridd, P. V., Orpin, A. R., Larcombe, P., & Lough, J. (2004). Temporal variation of light availability in coastal benthic habitats: Effects of clouds, turbidity, and tides. Limnology and Oceanography, 49(6), Babin, M., Stramski, D., Ferrari, G. M., Claustre, H., Bricaud, A., Obolensky, G., & Hoepffner, N. (2003). Variations in the light absorption coefficients of phytoplankton, nonalgal particles, and dissolved organic matter in coastal waters around Europe.Journal of Geophysical.Research, 108(C7), doi: /2001jc000882, Bailey, S. W., & Werdell, P. J. (2006). A multi-sensor approach for the on-orbit validation of ocean color satellite data products. Remote Sensing of Environment, 102(1 2), Barnes, S. L. (1994). Applications of the Barnes objective analysis scheme. Part III: Tuning for minimum error. Journal of Atmospheric and Oceanic Technology, 11(6), Bendis, B. (1999). Water quality trends in Tampa Bay, Florida. University of South Florida, Master Thesis, 74p. Boler, R. N., Molloy, R. C., & Lesnett, E. M. (1991). Surface water quality monitoring by the Environmental Protection Commission of Hillsborough Country. In S. F. Treat & P.A. Clark (Eds.), Proceedings from BASIS II Tampa, FL., 528 pp. Cannizzaro, J. P., Carder, K. L., Chen, F. R., Heil, C. A., Vargo, G. A. (in press). A novel technique for detection of the toxic dinoflagellate Karenia brevis in the Gulf of Mexico from remotely sensed ocean color data. Continental Shelf Research. Chami, M., Shybanov, E. B., Churilova, T. Y., Khomenko, G. A., Lee, M. E. -G., Martynov, O. V., et al. (2005). Optical properties of the particles in the Crimea coastal waters (Black Sea). Journal of Geophysical Research, 110, C doi: /2005jc Chen, Z., Hu, C., Conmy, R. N., Swarzenski, P., & Mullerr-Karger, F. (in press). Colored dissolved organic matter in Tampa Bay, Florida. Marine Chemistry. doi: /j.marchem Chen, Z., Hu, C., Mullerr-Karger, F., & Luther, M. (in revision). Physical forcing of short bio-optical variability in Tampa Bay: Observations from a coastal tower. Limnology and Oceanography. Christian, D., & Sheng, Y. P. (2003). Relative influence of various water quality parameters on light attenuation in Indian River Lagoon. Estuarine Coastal and Shelf Science, 57(5 6), Cloern, J. E. (1987). Turbidity as a control on phytoplankton biomass and productivity in estuaries. Continental Shelf Research, 7(11), Cloern, J. E., Powell, T. M., & Huzzey, L. M. (1989). Spatial and temporal variability in South San Francisco Bay. II. Temporal changes in salinity, suspended sediments, and phytoplankton biomass and productivity over tidal time scales. Estuarine, Coastal and Shelf Science, 28, Cole, B. E., & Cloern, J. E. (1987). An empirical model for estimating phytoplankton productivity in estuaries. Marine Ecology- Progress Series, 36, Doxaran, D., Castaing, P., & Lavender, S. J. (2006). Monitoring the maximum turbidity zone and detecting fine-scale turbidity features in the Gironde estuary using high spatial resolution satellite sensor (SPOT HRV, Landsat ETM+) data. International Journal of Remote Sensing, 27, Doxaran, D., Froidefond, J. M., Lavender, S., & Castaing, P. (2002). Spectral signature of highly turbid waters. Application with SPOT data to quantify suspended particulate matter concentrations. Remote Sensing of Environment, 81, Fisher, T. R., Gustafson, A. B., Sellner, K., Lacouture, R., Haas, L. W., Wetzel, R. L., et al. (1999). Spatial and temporal variation of resource limitation in Chesapeake Bay. Marine Biology, 133(4), Franz, B. A., Werdell, P. J., Meister, G., Kwiatkowska, E. J., Bailey, S. W., Ahmad, Z., et al. (2006). MODIS land bands for ocean remote sensing applications. Proc. Ocean Optics XVIII, Montreal, Canada, 9 13 October Gesch, D., & Wilson, R. (2001). Development of a seamless multisource topographic/bathymetric elevation model of Tampa Bay. Marine Technology Society Journal, 35(4), Hach Model 2100N laboratory turbidimeter instruction manual for use with software version 1 (pp. 1 80). (199). Hach Company. Harding, L. W., Magnusona, A., & Mallonee, M. E. (2005). SeaWiFS retrievals of chlorophyll in Chesapeake Bay and the mid-atlantic bight. Estuarine, Coastal and Shelf Science, 62, Heyes, A., Miller, C., & Mason, R. P. (2004). Mercury and methylmercury in Hudson River sediment: Impact of tidal resuspension on partitioning and methylation. Marine Chemistry, 90(1 4), Hu, C., Carder, K. L., & Muller-Karger, F. E. (2000). Atmospheric correction of SeaWiFS imagery over turbid coastal waters: A practical method. Remote Sensing of Environment, 74, Hu, C., Carder, K. L., & Muller-Karger, F. E. (2001). How precise are SeaWiFS ocean color estimates? Implication of digitization-noise errors. Remote Sensing of Environment, 76, Hu, C., Chen, Z., Clayton, T., Swarnzenski, P., Brock, J., & Muller-Karger, F. (2004). Assessment of estuarine water-quality indicators using MODIS