MODELING AND MAPPING WATER QUALITY IN BARNEGAT
BAY, NJ, USING LANDSAT-TM

Jack Schooley

Remote Sensing for Marine and Coastal Environments Conference
Seattle, Washington, USA, 18-20 September 1995.

ABSTRACT

Landsat-5 Thematic Mapper imagery combined with ground-truth data were used to model and map water quality in Barnegat Bay, NJ. Water depth, secchi depth and surface temperature were measured at sample sites in the bay near the time of satellite overpass on 9 October 1993. Bottled samples were analyzed in a laboratory for salinity, turbidity, total suspended solids, chlorophyll-a, and dissolved organic carbon. Satellite data and ground-truth data for the sample sites were used to develop regression models. Models and maps for salinity, chlorophyll-a, and temperature were produced. The maps provide an unprecedented view of the distribution of these variables in Barnegat Bay. Water was relatively clear on the image date, so bottom reflectance interfered with modeling and mapping for a large portion of the bay. Also, variation in reflectance in the study area was low compared with satellite sensor electronic noise. Additional research is recommended using mid-summer satellite imagery, when biological productivity, turbidity, and water quality variability are at annual maxima.

INTRODUCTION

This project evaluated the effectiveness of Landsat Thematic Mapper (TM) data for providing a synoptic and quantitative overview of water quality in Barnegat Bay, New Jersey (Figure 1). Barnegat Bay is a shallow, lagoon-type estuary with residential development pressure and intensive recreational use. Water depth in the bay ranges to 4.0 meters in the central portion, shoaling rapidly toward shore. The mean tide range is 0.15 meters, and total volume of the bay is approximately 2.4x108 m 3 (Durand, 1984). Long-term mean total freshwater discharge to the bay is approximately 65 m3 s-1, with the majority from groundwater discharge. The turnover rate for bay water is about 45 days (Kennish, 1978). Salinity for the sample date ranges from a minimum of 15 ppt near tributary stream discharges, to a maximum near Barnegat Inlet of 31 ppt.

The 1200 km2 coastal plain watershed consists of sandy soils and extensive water-table aquifers supporting a vast pine and scrub-oak forest, interspersed with cedar bogs. Bay tributaries contribute considerable organic matter and humic acids, which produce the characteristic tea color of much of the bay water. Fringing salt marsh has been filled, and most of the shoreline is bulkheaded. The remaining marsh has been drained by ditching for mosquito control. A combination of loss of buffering marshes and an increase in impermeable surface area has increased direct discharge of unfiltered overland flow to the bay, boosting suspended solid, nutrient, and contaminant levels.

METHODS

GROUND-TRUTH DATA

Thirty samples were collected from a boat by 3 technicians between 10:18 am and 12:36 pm on 9 October 1993 (Figure 1, below). Satellite overpass was at 10:56 am. Water temperature (TMP), secchi depth (SD), water depth (DEP), and sample location coordinates were measured at each sample site as bottled samples were collected. Temperature was measured by barely submerging the length of the probe of a digital thermometer until the temperature display stabilized. Secchi depth was measured using a secchi rod specially designed for this project. It consists of a secchi disk attached to the side of an incremented sounding rod. This simple instrument allowed quick measurement of water depth

hfgh

and secchi depth in one motion. Global Positioning System (GPS) coordinates were collected using a waterproof marine GPS receiver with differential beacon receiver attached. This system allowed a real-time differentially corrected latitude/longitude coordinate, accurate time, and sample number to be stored with the touch of 1 button on the receiver. A bottle cage allowed 3 bottles to be filled simultaneously at each site at a bottleneck depth of 0.4 meters. Amber glass bottles were used and samples were kept on ice. On-site sampling time was approximately 75 seconds, and travel time between sites was 3 to 4 minutes.

The sampling date was preceded by an extended dry and calm period throughout the basin. At satellite overpass time the sky was exceptionally clear, and the SE wind had been blowing for 45 minutes at a speed of 6.2 m s-1. There were several boats and wakes in the study area, and approximately 2 wind-generated whitecaps per image pixel (30 x 30 meters). At satellite overpass time, the tide had been flooding at the inlet for more than 2 hours, but the tide flow was ebbing southward throughout most of the bay.

The bottled samples were analyzed in a laboratory for salinity, turbidity, total suspended solids, chlorophyll-a, and dissolved organic carbon. American Public Health Association (APHA) methods were used throughout the laboratory procedures (APHA, 1989). Salinity (SAL) was quantified using the electrical conductivity method (methods 2510-a and 2520-b). The procedure used both a standard seawater solution and a 15 ppt solution for calibration. Turbidity (TRB) was measured using a nephelometric turbidity meter (method 2130-b). Total suspended solids (TSS) concentration was determined using the gravimetric method (method 2540-d). Chlorophyll-a (CHa) analyses used the trichromatic spectrophotometric method (method 10200-h). Dissolved organic carbon (DOC) was calculated using the wet-oxidation method (method 5310-d). Acid blanks, filter blanks, and calibration standard samples were used to minimize error.

SATELLITE DATA

The coordinates for the samples sites were transformed into the satellite image pixel path and row reference system using 25 control points well distributed around the shoreline of the bay (RMS error = 14 meters). Digital numbers (DN) for the 3 x 3 pixel matrix centered on each sample site were extracted from the TM image. A matrix was extracted to account for any remaining positional error in the field sample coordinates or the satellite image, and to improve the signal-to-noise ratio (Lathrop, 1989). For the 7 sites where boats and their wakes may have corrupted the image data, the closest valid pixels were chosen. A mean DN was calculated for each band for the 9 pixels in the matrix at each site. For band 6, a DN for only one pixel at each site was extracted since these pixels cover a large area (120 x 120 meters). Bands 5 and 7 (mid-infrared) were omitted from further consideration because DNs approached zero, due to strong absorption of mid-infrared light by water. DNs were converted to radiance values (mW m-2 sr -1) for quantitative work (EOSAT, 1993).

MODEL DEVELOPMENT

The ground-truth data and radiance data for each sample site were entered into a statistical software package. Logarithmic transformations, band combinations, and other derivative variables were generated. Temperature and TM band 6 radiance values were not transformed since there is a well documented linear relationship between surface temperature and the satellite data (Wukelic, 1986). Data relationships were investigated using scatter plots and correlation matrices. The natural log (Ln) of ground-truth and satellite data, with no other transformations or band combinations, yielded the strongest scientifically defensible relationships.

The data were subset based on region of the bay, depth vs. secchi depth, and time between satellite overpass and sample collection. Setting and time subsets yielded no significant improvement in correlation coefficients (except for temperature and band 6: see below). This was expected since previous sampling has confirmed that regions of the bay are not significantly different, and because the sampling time interval was short and flow rate in the bay was low. There was a strong increase in r values when only samples were considered where the depth was 1.7 times or greater than the secchi depth. No additional improvement was noted when only samples were considered where this ratio was 2.0 or greater. The 2.0 ratio has been suggested for avoidance of bottom reflectance (Lathrop, 1986; Thiruvengadachari, 1983). The 20 sample sites where the ratio was 1.7 times or greater were used for model development.

Linear regression models were attempted for all the dependent ground-truth variables, using TM bands 1, 2, 3, 4 and 6 as independent variables. R2adj, F/Fcr, Cp/p, and S.E.y for the equation, and the t-ratio for each coefficient, were calculated and evaluated against criteria suggested by Whitlock (1982). The resulting models are in the form of linear regression equations which predict values for the water quality variables based on the satellite data. Successful salinity, chlorophyll-a, and temperature models were produced. The temperature model used samples collected within 30 minutes of satellite overpass without consideration of depth. Improved r values for this subset of 14 samples indicates there was a significant change in temperature at the sampling sites over the full sampling time interval.

MAPPING

Before mapping, the areas of the bay where bottom reflectance is a potential problem needed to be omitted. These areas were determined by considering the depth to secchi depth ratio (DEP/SD) at nearby sample sites. For most of the bay, the 1.83 m ( 6 ft.) bathymetric isoline can be used as a limit. Therefore, models were applied to TM pixels which occurred in water greater than 1.83 meters deep. The extent of this area is indicted by the area modeled in Figure 2, and includes the central axis of the main bay and most of the Toms and Metedeconk Rivers. The temperature model was applied to the entire bay.

The resulting one band images were converted to raster GIS files and filtered to omit single pixels, or small groups of adjacent pixels with one value which would clutter the final maps. The salt-and-pepper appearance in unfiltered files is due largely to TM sensor noise. The unfiltered files were archived for future study since some of the extreme or localized values may be due to real variation in the water column which can be determined through field investigations. Five data classes for each variable were created for mapping purposes. The classification scheme is a combination of ‘equal steps based on the data range’, and ‘equal area per class’. This scheme was necessary to show the intricacies of the spatial distribution of water quality in the bay.

RESULTS

DATA

Although sampling was restricted to the deepest regions of the bay, the secchi depth approached the water depth near the inlet (Table 1, samples 1-5). This is due to a high turnover rate in this area and inflow of relatively clear ocean water with the flooding tide. The data range for temperature, turbidity, and dissolved organic carbon were very low on this date. This may be caused by the low stream discharge at the time, and the period of calm, dry weather preceding sampling. The error indicated at the bottom of Table 1 is estimated from duplicate sampling and from the error inherent in the field and laboratory procedures (APHA, 1989). It is expressed here as the error divided by the range, multiplied by 100. Not shown is the error in time and GPS coordinates. Time is accurate to within 1 second, while the coordinates are accurate to within 10 meters at the 0.95 confidence level (MSC, 1992).

bsrdfh


The relatively low range and variance in the ground-truth data contributed to the low range and variance in the satellite data (Table 2). Band 3 traditionally has shown a strong relationship
with several water quality variables, such as total suspended solids (Khorram, 1985; Tassan, 1987). However, the relatively great range for TSS in the ground-truth data is not reflected in the range in band 3. The somewhat greater range in DNs for band 4 (near-IR) may indicate that most variation in water column reflectance is at the surface, since light sensed in this band emanates from a depth of less than 20 cm. Visible bands provide data over a greater depth range (Lepley, 1975).

khk

The small range in the satellite data required careful attention to the sensor electronic noise contribution. The noise value in digital numbers for each band was calculated by analyzing the ladder-like pattern and magnitude of noise apparent in the open ocean. The semi-systematic pattern is random enough to elude a systematic correction algorithm. Given the noise pattern and magnitude of approximately 1.0 DN, 5 of 9 pixels in any 3 x 3 matrix may be affected, and the total influence for a site is (5/9) x 1.0 DN, or 0.56 DN. This value agrees well with data from Wrigley (1985). Since only one pixel is extracted for each site for band 6, the noise component remains at 1.0 DN. The error displayed at the bottom of Table 2 is calculated by dividing the noise component (0.56 for bands 1-4, and 1.0 for band 6) by the range of DN values for the 30 sample sites, then multiplying by 100.

Table 3 shows that the correlations (r) are better than might be expected considering the estimated error in the satellite data. The upper portion of the matrix indicates the strong inverse relationship between turbidity and salinity, and good positive relationship between turbidity and chlorophyll-a. On the lower right side of the matrix, the r values for adjacent bands are relatively low. This indicates unique

etyhj

information is sensed in each band, or randomness due to the noise component. The lower central portion of the matrix shows the low r values for band 3 and the ground-truth variables, which is expected given the small range and high percent error for that band. Bands 1, 2, and 4 show good relationships with salinity, turbidity, and chlorophyll-a. The relatively strong relationships between some bands and salinity, turbidity, and chlorophyll-a indicates that regression models might be successful. Actually, the standard error for the turbidity model was greater than the data range. Therefore, regression models were only produced for salinity, chlorophyll-a, and temperature.

MODELS

Variation in dissolved salts have a negligible effect on upwelling radiance. However, salinity co-varies with secchi depth, turbidity, chlorophyll-a, and dissolved organic carbon, and other unmeasured variables in the water column that do influence upwelling radiance. In addition, salinity is one of the most characterizing and important variables in an estuary. The best model used LnB1, LnB2, and LnB4 (Table 4).

djhd


For the CHa model, although the coefficients are all significant at the 95% confidence level, and the standard error is good (1.3 mg l-1 ), the model has only a fair R2adj (Table 5). However, the relationships implied by this model are substantiated in the literature. An inverse relationship between CHa concentration and band 1 radiance is expected, since there is a CHa absorption peak in band 1 (APHA, 1989). The positive relationship between band 4 radiance and CHa is due to an increase in turbidity associated with phytoplankton population increase (Stumph, 1988).

ghd


The relatively low R2adj for the temperature model reflects the small range in temperature in the samples and across the study area on this date relative to the noise component in band 6 (Table 6). If all samples are used in the model the R2adj is 42%, emphasizing the influence of time.

wret


MAPS

Elevated salinity at the northernmost end of the bay is produced by tidal flow through an Intracoastal Waterway navigational canal (Figure 2, below, North Section). The discharge of the Metedeconk River limits the influence of this saline water. In the Toms River, eddies of more saline, less turbid water hug the northern shoreline. In the South Section, the channel which connects the bay with the inlet discharges across from Oyster Creek. The plume of more saline water meanders northward in the deeper portion of the bay. Stratification contributes to the patchy pattern of this plume. To the south, another channel introduces more saline water near the town of Barnegat.

ryu


Elevated chlorophyll-a occupies the central reach of the Metedeconk River, and several areas in the Toms River (Figure 3, below, North Section). The pattern is associated with nutrients provided by tributary streams in these areas, along with non-point residential sources. A localized area of elevated chlorophyll-a shown at the bottom of the North Section may be produced by nutrient rich discharge from lagoon developments near this area on the mainland. The low concentrations in the South Section are due to lack of nutrients at this time of year. If the map classes were subdivided, the map would more strongly indicate slightly elevated productivity toward the shallower eastern side of the bay, where nutrients may be available through resuspension of bottom sediment.

ghhhh

Variation in temperature toward the eastern shore of the bay in the North Section is mainly due to variation in water depth, with shallower areas being warmer (Figure 4, below). Discharge from small tributaries and marsh drainage ditches produces localized elevated temperatures in other areas. The discharge plume from the power plant is most obvious in the Southern Section. The plume is starting to displace northward with the flooding tide. The elevated temperatures along the west shore north and south of the plume are due to the migration of the plume with flood and ebb of the tide. The deeper axis of the bay is represented by the coolest temperature class. Along the eastern shore, storm overwash fans produce small areas of slightly shallower and warmer water.

df

CONCLUSIONS

Low spatial variation in water quality resulted in satellite data which have correspondingly low variance. Sensor electronic noise was a large component of satellite data variance. Bottom reflectance reduced the number of samples useful for modeling and restricted the area which could be mapped. Since the error component in the satellite data approached limits of acceptability, the contribution of field and lab error to total error also became significant in producing marginal models.

The salinity model is the most statistically robust, with high R2adj, low standard error and low bias. The model presented for chlorophyll-a has a low standard error and makes use of defensible relationships between satellite data bands and ground-truth data, but has only a fair R2adj value compared to results of similar studies (Lathrop, 1986, 1989). Temperature has very low field measurement error, so the unremarkable R2adj must be explained by sensor noise coupled with the low variation in temperature across the study area. In other studies, the relationship between band 6 digital numbers and water surface temperature has been shown to be strong and linear (Lathrop, 1986; Wukelic, 1986).

The maps provide a unique and near-instantaneous view of the distribution of salinity, chlorophyll-a, and temperature for a large contiguous area of Barnegat Bay. The maps show the distribution of these variables in relation to potential sources, causes, or influences. Circulation patterns and areas of upwelling are also obvious. The maps can be used to help focus future point sampling on problem areas. The maps must be interpreted in light of the shortcomings in the data and models discussed above. However, much of the influence of measurement error, satellite data noise, and potential bottom reflectance are removed in the data smoothing (generalization) procedure used in map production.

REFERENCES

APHA (American Public Health Association), Standard Methods for Examination of Water and Wastewater. Port City Press, Baltimore, 1989.

Durand, J.B., “Nitrogen Distribution in New Jersey Coastal Bays.” In Ecology of Barnegat Bay, N.J.: Lecture Notes on Coastal and Estuarine Studies, M.J. Kennish and R.A. Lutz, Springer-Verlag, New York, 1984.

EOSAT (Earth Observation Satellite Company), Fast Format Document for TM Data Products, Version B, Revised December 1, 1993.

Kennish, M.J., “Effects of Thermal Discharges on Mortality of Mercenaria Mercenaria in Barnegat Bay, New Jersey.” Environmental Geology, Vol. 2 pp. 223-254, 1978.

Khorram, S., “Development of Water Quality Models Applicable Throughout the Entire San Francisco Bay and Delta.” PE&RS, Vol. 51, No. 1, pp 53-62, January 1985.

Lathrop, R. G. and T. M. Lillesand, “Use of Thematic Mapper Data to Assess Water Quality in Green Bay and Central Lake Michigan.” PE&RS, Vol. 52, No. 5, pp 671-680, May 1986.

Lathrop, R. G. and T. M. Lillesand, “Monitoring Water Quality and River Plume Transport in Green Bay, Lake Michigan with SPOT-1 Imagery.” PE&RS, Vol. 55, No. 3,pp 349-354, March 1989.

Lepley, L.K., K.E. Foster, and L.G. Everett, “Water Quality Monitoring of Reservoirs on the Colorado River Utilizing ERTS-1 Imagery.” In Remote Sensing and Resources Management, Proceedings American Water Resources Association, No. 17, Urbana, Illinois, pp 105-111, 1975.

MSC (Magellan Systems Corporation), Magellan GPS NAV 5000D Users Guide, Magellan Systems Corporation, San Dimas, CA, 1992.

Stumph, R.P., M.A. Tyler, “ Satellite Detection of Bloom and Pigment Distributions in Estuaries. “ Remote Sensing of Environment, Vol. 24, pp 385-404, 1988.

Tassan, S., “Evaluation of the Potential of the Thematic Mapper for Marine Application.” PE&RS, Vol. 8, No. 10, pp 1455-1478, 1987.

Thiruvengadachari, S., N. G. Inamdar, H.G. Tambralli, S.V. Kotbagi, and B.L. Deekshatulu, “Some Ground-Truth Considerations in Inland Water Quality Surveys.” International Journal of Remote Sensing, Vol. 4, No. 3, pp 537-544, February 1983.

Wrigley, R.C., C.A. Hlavka, D.H. Card, “Evaluation of Thematic Mapper Interband Registration and Noise Characteristics.” PE&RS, Vol. 51, No. 9, pp 1417-1425, September 1985.

Wukelic, J.E., J.C. Barnard, G.M. Petrie, and H.P. Foote, “ Opportunities and Difficulties Associated With Landsat TM Data for Determining Surface Water Temperature.” Pacific Northwest Laboratory, Richland WA, 1986.