Quantification of Chlorophyll in Reservoirs of
the Little Washita River Watershed
Using Airborne Video
M.M. Avard, F.R. Schiebe, and J.H. Everitt
Abstract
-.
Airborne video cameras equipped with narrow-band filters
were used to assess chlorophyll-a concentration in flood
control reservoirs of the Little Washitn River Watershed in
central Oklahoma. This study utilizes airborne video camems
equipped with narrow 10-nm band filters centered at the
critical wavelengths to assess chlorophyll-a concentration.
The video cameras were calibrated using a series of panels
and a hand-held spectroradiometer to convert digital numbers
into radiance values (fiW/cm2/sr). This was then processed
into reflectance values by the incorpomtion of solar irradiance
data. Results indicate that the relationship between emergent
radiance and chlorophyll concentration is best described by
the model y = ao(l - e-X'C),and that the ability to estimate
chlorophyll-a concentration in reservoirs using airborne video
imagery has a great deal of potential.
Introduction
To effectively utilize surface water resources, methods to assess
the quality of water in impoundments in relation to their intended use need to become more rapid and economical. Traditionally, surface water quality has been assessed using various
limnological methods and laboratory analyses. This is time
consuming, requires field sampling by trained personnel, and
is expensive.
The main objective of this study is to develop convenient,
rapid, economic methodology based on aerial video remote
sensing to assess water quality and productivity in surface
impoundments. Useful indicators include any suspended constituents visible to the human eye or to optical instruments sensitive to electromagnetic regions outside the human visible
range. The most likely candidates include suspended inorganic
particles, phytoplankton, organic detritus, and dyes. This study
concentrates on remote sensing of phytoplankton and, more
specifically,the associated chlorophyll.
Many studies have been performed examining the nature
of the relationships between reflectance, suspended solid concentration (SSC),and chlorophyll-a concentration (Dekker et al.,
1991; Gitelson, 1992; Han et al., 1994; Schalles et al., 1997).
Whether performed in the laboratory,in the field, or from remote
sensing platforms, the same general concIusions have been
reached. Exoatmospheric reflectance of water bodies in the near
infrared wavelength range depends on the amount of suspended solids present: as SSC increases, reflectance increases.
This relationship may be detected even by broad band [loo-nm
width) sensors of the various satellites. Using these broad bands
sensors, the detection of chlorophyll-a is complicated by the
presence of ssc. However, detection and quantification may be
possible using narrow band filters (10-nm width) at lower altitudes (Ritchie et al., 1994). Because chlorophyll-a absorbs light
strongly near 675 nm and strongly scatters light near 700 nm
(Schalles et al., 1997), the behavior of reflectance at these two
wavebands is utilized to develop a practical and economical
procedure to determine chlorophyll-a concentrations. This
study utilizes airborne video cameras equipped with narrow
10-nm band filters centered at the critical wavelengths to assess
chlorophyll-aconcentrationin flood control reservoirs ofthe Little Washita River Watershed located in central Oklahoma. The
primary goal is to determine the relationship between narrowband reflectance in these two bands with the chlorophyll-a concentration. This was accomplished using a single sampling site
within each reservoir. It is not meant to imply that chlorophyll
concentration is absolutely constant across the surface of each
reservoir; however, the epilimnion of water bodies of this size
are generally well-mixed over broad areas of their surfaces.
Methods
flw Information
In August of 1994,three flights were made over reservoirs of the
Little Washita River Watershed: the morning and afternoon of
the 19th and the morning of the 23rd to obtain reflectance
images from the surface waters of each of the flood control reservoirs within the watershed. The majority of the data for this
study was obtained under overcast sky conditions.
A twin engine airplane was equipped with three panchromatic (black-and-white) video cameras with fixed 12.5-mm
focal length lenses. It was flown at an altitude of 1980 m at an
average plane speed of 260 km/hr (125 knots). The video cameras were controlled by f-stop only with the automatic gain control (AGC) disabled. Each camera was fitted with a narrow-band
filter (Table 1). Images obtained by the video cameras were
recorded on video tape using video cassette recorders (VCR).
fleld Sampllng
Samples were collected by personnel of the Agricultural
Research Service (ARS),Oklahoma Conservation Commission,
M.M. Avard is with Southeastern Oklahoma State University,
Station A, Box 4200, Durant, OK 74701 (mavard @sosu.edu).
F.R. Schiebe is with SST Development Group, Inc., 824 North
Country Club Road, Stillwater, OK 74075 ([email protected]).
J.H. Everitt is with the USDAIARS, Remote Sensing Unit, 2413
East Hwy. 83, Weslaco, TX 78596 ([email protected]).
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Photogrammetric Engineering & Remote Sensing
Vol. 66, No. 2, February 2000, pp. 213-218.
0099-1112/00/6602-213$3.00/0
O 2000 American Society for Photogrammetry
and Remote Sensing
February 2000
213
phyll-a concentration from 1.18 to 206.28 mg/m3 and suspended sediment concentration from 0.5 to 21 mg1L.
Filter center
Aerial video images of each reservoir were "frameBand width
Effective wavelength
wavelength
grabbed"
from each of the three video tapes, providing an
(=I
F-Stop
(nm)
band (nml
image of each reservoir at each of the designated wavelengths:
670 nm, 700 nm, and 800 nm. These were imported into the
geographic information system (GIS) IDRISI for analysis. Water
quality data were assembled by the ARS Durant Laboratory and
the solar radiation (irradiance) data were obtained from the
Oklahoma Micronet system.
The cameras had a 16.9 mm (213 in.) cCD (charge coupled
detector) array with 640 by 480 active sensors in the horizontal
and Oklahoma State University from each reservoir on the same and vertical directions, respectively. The field of view was 1402
m (38.5") in the horizontal direction and 1052 m (29.5') in the
dates that the airborne data were collected. Surface water in
vertical direction. However, because images were recorded on
the reservoirs was sampled in open water near the dams. Data
a VCR with 400 lines resolution, the basic pixel size was reset
collected in situ included depth, Secchi depth, pH, and conductivity. Subsequent laboratory analyses performed at the ARS to 1052 m/400 = 2.63 m. Thus, each pixel was 2.63 m by 2.63 m
at the ground.
Water Quality Laboratory in Durant, Oklahoma included turbidity, total solids, total dissolved solids, total suspended solTo determine an appropriate pixel array size, a 10 by 10
window of pixels was selected and evaluated for the 700 nm
ids, total nitrogen, ammonia, nitrate, total phosphorus, dissolved phosphorus, and chlorophyll concentration (pheophywaveband from four reservoirs to determine a representative
tin corrected, trichromatic equations).
average value of radiance based on methods described by Ritchie and Cooper (1987). Pixel array sizes from 1by 1 to 10 by
Solar Radiation
10 were evaluated. The 1by 1 array was located in the upper
left-hand corner of the window. Successive pixel arrays were
Solar radiation from the Oklahoma MesonetIMicronet network
(Oklahoma Climatological Survey, 1994, personal cornmunichosen down and to the right of the previous array. The average
cation) was used to determine incoming solar radiation. The
radiance approaches a fairly constant value by a 7 by 7 pixel
Mesonet is a comprehensive network of 111weather stations in array size for all of the reservoirs (Figure 2) and would therefore
be the minimum acceptable size. This study utilized a pixel
the state of Oklahoma. The Micronet is a system of 42 stations
located on a five-kilometer grid over the Little Washita River
array size of 10 by 10, which corresponds to an area of approxiWatershed.
mately 692 m2. Array sizes larger than this may become problematic, especially in the smaller reservoirs, because values
Experimental Design
may vary as a result of changing surface water conditions (nearThere are 45 flood retention reservoirs in the Little Washita
shore sediment influx or plant growth) or unintentional incorRiver Watershed (Figure 1). A complete set of data (video,
poration of land surfaces into average water values.
water sample, solar radiation) necessary for this study was
The locations of the arrays were selected to encompass
available for 33 of the reservoirs. The reservoirs varied in chloro- the location where the field samples were collected. For each
AND FILTERlNFORMATlON
TABLE1. VIDEOCAMERASETTINGS
I
I
R-1-W
Figure 1 . Reservoirs of the Little Washita River Watershed, central Oklahoma (after Allen
and Naney, 1991).
U4
February 2000
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
g
L(670) = 4.31 + 0.069 (DN) DN < 172 (Figure 3a)
L(700) = -3.37 + 0.154 (DN) DN < 160 (Figure 3b)
8
PE
-
(2)
% 7 ' 5 u mRes41
t + - a
6.5 -
P
i
6
"
1x1
'
'
3x3
'
5x5
~
7x7
'
~
'
'
9x9
Pixel Array Size
Figure 2. Pixel array size versus average radiance for four representative reservoirs.
zoo
$80
-
*
T
where the multiplicative coefficients have units of pW/cm2/sr.
The video camera settings and filter configurations are summarized in Table 1.
Radiance images are converted into reflectance (R) images
by examining the irradiance (I), or incoming solar radiation.
The irradiance data for each reservoir was estimated using data
from the nearest Micronet site(s).Data were corrected for both
spatial and temporal differences. This irradiance data, however, is representative of the entire visible and near-infrared
spectrum. Only a small fraction of this in the appropriate wavelength band is required to normalize the measured radiance.
Utilizing spectra (Figure4) collected on-site by Harrington
(1996, personal communication), the total area under the curve
and the area of therelevant wavebands (i.e., 665 to 675 nm and
695 to 705 nm) was determined and ratioed. This resulted in
the fraction of incoming radiation in these wavebands with
respect to incoming total solar radiation measured at the
Micronet stations: i.e.,
b = spectral area^,,,,^,^^ /spectral area^,^,
This resulted in b(665-675) = 0.029844 and b(695-705) =
0.02433. Taking this into account, reflectance was calculated
using
0.00
0.11
0.87
0
0.74
0.02
(E-0
Radiance (mW/cmz/sr)
(a)
100
la0
-
0
0.00
where Zis the irradiance interpolated in space and time from
the Micronet data.
Results
*
0.11
0.61
0.87
0.74
Radiance (mW/cm2/sr)
0.02
(6-41
(b)
Figure 3. Video camera calibration for (a) the
670-nm waveband and (b) the 700-nm waveband. Sensor saturation occurs at approximately 170 digital counts.
of the 33 reservoirs, an average value for the 10by 10 digital
number pixel array was determined for each wavelength (670
nrn, 700 nm, 800 nm).
The cameras were calibrated by flying over a series of five
panels: white, gray-white, gray, gray-black, and black. As the
airborne video cameras recorded the shades of gray, radiometers on the ground were hand-held directly above the panels
to determine the reflected radiance values. Three points on
each panel were evaluated to establish the relationship
between digital numbers recorded by the video camera and
radiance measured using the spectroradiometer. Using regression analysis, the panel data was used to generate a relationship
between radiance (L) and digital number (DN) for each camera:
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Previously reported results indicate that the difference between
emergent radiance in the 670-nm and 700-nm wavebands is
related to chlorophyll-a concentration. This simple difference
method between images taken in each of these wavebands is
related to the first derivative method described by Han et al.
(1994) and Rundquist et al. (1996). The present study hypothesizes that the relationship between remotely sensed data and
chlorophyll-a concentration (chl-a) should improve as the raw
video data is processed into physically meaningful radiance values and improve even further with the normalization by solar
irradiance data. Mathematical models were applied to the data
60
-
200
400
600
800
lo00
1200
Wavelength (nm)
Figure 4. Representative spectrum of incoming solar radiation used in the Little Washita River Watershed study (after
Harrington, 1996, personal communication).
February 2000
215
and analyzed to determine which model most reasonably and
accurately described the relationship between emergent radiance and chlorophyll-a concentration.
In all models, the independent variable xrepresents chlorophyll-a concentration, while the dependent variable y is representative of digital number (DN),radiance (L), or reflectance (R).
This dependent variable is calculated by taking the difference
between data at 700 nm and the data at 670 nm. Linear and logarithmic (log) models were examined as well as other models,
including a simple transformation equation ( l l y = a, + a,lx), a
variation of the Normalized Difference Vegetation Index (NDVI),
and y = a, (1 - e-"Ic) of Schiebe et al. (1987).
Models were evaluated using several goodness-of-fitmeasures: r2,d-statistic (d-stat),and root-mean-square error (RMSE).
The d-statistic (d-stat) is a goodness-of-fitmethod proposed by
Willmott (19821: i.e.,
Model
Y
log Y
NRF.1
Statistical
Schiebe Model
computer fit
manual fit
DN
Radiance
Reflectance
0.500
0.227
0.014
0.516
0.758
0.502
0.440
0.706
0.501
0.004
0.222
0.517
0.490
0.426
0.790
0.815
0.686
0.731
Comparing linear NREI to chlorophyll concentration (Tables 2
and 3), NREI predicted an improvement in processing DN data
into physically based radiance values but no improvement as a
result of incorporating solar irradiance data.
The Schiebe model (1987)
x
where yi is the actual value, is the value predicted by the
equation, and Mis the mean of the data. It is interpreted in the
same manner as r".
The RMSE values were inconsistent and highly variable. It
is hypothesized that the RMSE is not an appropriate means of
evaluating the data in this study because the scale of the three
data sets varies by a factor of at least 10 and the RMSE does not
adjust for such variations.
Linear regression performed on the data and the log of the
data indicated significant improvement in predicting chl-a
concentration with radiance values. This would be expected
because the raw data were not calibrated, and converting to
radiance served to do so. Further improvement was expected
after incorporation of irradiance data but did not occur.
Reflectance data did not describe chlorophyll-a concentration
significantly better than the unprocessed DN data.
For a simple transformation of the data in the form of
is based on the physics of light scattering .from particles suspended in water. It describes a saturating exponential relationship with a, being the asymptotic value of y and c representing
a chlorophyll constant. The equation was evaluated (1) by
allowing the computer to choose the best values of a, and c
using an iterative, least-squares algorithm, and (2)by manually
inserting various values of c and having the computer minimize
a,. These results were compared to visual estimates of a, and c
determined by inspection of scatterplots of the data: i.e.,
(Figure 5a)
Rad y = 3.25(1 - e-x'25)
(Figure 5b)
Refl y = 0.014(1 - epXl7)
(Figure 5c)
Other than the chlorophyll constant, c, of radiance, the
computer analysis provided results similar to that of the visual
l l y = a, + a,lx,
estimate (Table 4). Because, however, c should theoretically be
a constant, the manual method mentioned above was
prediction potential was greater both after initial processing
performed.
into units of radiance and after incorporation of the irradiance
a, was minimized for various values of c (Table5). The best
data in determining reflectance values. This equation yielded
value of c was between 40 and 50 because both radiance and
the best fit (r2)between reflectance and chlorophyll concentrareflectance have relatively high r2values in that range. For a
tion (Table 2), but it is purely statistical in nature and so is diffi- value of c = 40, there is a slight improvement when processing
cult to determine the physical meaning of a, and a,.
DN into radiance and no real improvement with the incorporaNDVI is commonly used in remote sensing applications
tion of solar irradiance data in determining reflectance.
because it compensates for differences in illumination and surComparisons of the various models are illustrated in Figface slope. NDVI is normally calculated using the signals, s, at
ures 6a, 6b, and 6c. In summarizing r2and d-stat values, regard-~ ~ ~ ~ n
+ m less
) / of(the~ mathematical
~ ~ ~ ~ model
~
wavelengths of 850 nm and 630 nm ((sssonm
chosen, an improvement was
s,~,,,)) rather than the 700-nm and 670-nm bands used in this
noticed when processing raw video data into physically meanstudy. It is proposed that a similar equation utilizing the 700ingful radiance values (Tables 2 and 3). The final processing
nm and 670-nm wavebands be used and referred to as the Norstep of incorporating solar irradiance data did not significantly
malized Red-Edge Index (NREI). NREI would, therefore, be calimprove results. It is believed that the noise, or scatter, introculated as
duced by the diffuse light of the overcast sky prohibited a
meaningful improvement by the normalization process. Under
clear sky conditions, a much cleaner data set would have been
expected.
Overall, the model which best described the data was a
variation of the Schiebe model. For each processing step, the
equations are
Model
Y
1% Y
NRF.1
Statistical
Schiebe Model
computer fit
manual fit
216
February 2000
DN
Radiance
Reflectance
0.157
0.197
0.002
0.219
0.424
0.429
0.326
0.391
0.171
0.361
0.326
0.618
0.184
0.527
0.364
0.638
0.349
0.533
Y
DN
Radiance
Reflectance
equation
rZ
0.184
0.364
0.349
y
= 51.7(1 - e-XI1.')
y = 2.5(1 - e - ~ / 6 . "
y = 0.0108(1
1
-
e-x17.6)
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
90
90
80
80
+
.
data
- - - - - linear
70
L
2
60
5
Z
50
manual
. . . . . . statlstlcal
-
i
i 40
5
D 30
.
d
- computer
20
10
0
100
200
300
0
0
100
200
Chlorophyll (mglm.3)
300
(a)
Chlorophyll (mglm~3)
(a)
c
6
g
5
Y'
E
a
-
2
-
5
4
z
3
-5
2
a
1
8
+
data
5
-----
linear
4
3
manual
m
...... statistical
2
5
m
t
1
- computer
~e
0
'0
0
100
200
300
Chlorophyll (mglm~3)
0
0
100
200
(b)
300
0.30
Chlorophyll (mglm~3)
(b)
-
8 ui
0.24
E
m
OE
0
0.18
B
a
0.18
manual
0.12
., statistic
0.06
- computer
'
m k!
data
- - - - - linear
0.24
8
0.30
5:
+
0.00
0
0.12
100
200
300
Chlorophyll (mglm~3)
0.06
(c)
0.00
0
100
Chl;phyl
200
y . 3 )
300
,
Figure 5. Plots of the Schiebe model using the manual fit
through (a) the digital number data, (b) the radiance data,
and (c) the reflectance data.
DN y = 74.5111
- e-x140)
Rad y = 3.56(1 - e-x140)
Refly = 0.0148(1 - e-X'40)
Summary
Evaluating the r2 and d statistics, the physics-based Schiebe
model, y = a. (l-e-xlc), describes the relationship between
chlorophyll and aerial video data most accurately. Further
improvement is evidenced when manually setting c equal to
40. Thus, the relationship between chlorophyll-a concentration
and remotely sensed data in this study is best described by a
saturating exponential model with a chlorophyll constant of 40.
Figure 6. Comparison of the linear, statistical, and Schiebe
models (manual and computer-generated)through (a) the
digital number data, (b) the radiance data, and (c) the
reflectance data.
That the model predicts chlorophyll-a concentration better
for radiance than for reflectance was not the anticipated result,
but may easily be explained. The amount of radiance emanating
from a water body depends, in part, upon the amount of radiation striking the surface (irradiance). These were the data gathered from the Micronet stations nearest each reservoir, which
required interpolation in both time and space. Further, it required
an approximation of the irradiance available in the bands of
interest. The flights occurred on overcast days, so these irradiance interpolations probably Iack the accuracy required for the
analysis and may serve more appropriately as gross irradiance
approximations. As a result, irradiance values interpolated from
h4icronet stations cannot effectively be used to determine
reflectance values for identification of chlorophylla concentration for these experiments. It is expected that under clear sky
conditions the incoming solar irradiance would be sufficiently
well behaved and the method could be applied more effectively.
TABLE
5.
DN
Rad
Refl
0.528
0.299
0.255
0.53
0.434
0.384
0.53
0.456
0.404
0.529
0.576
0.502
1' FOR
VARIOUS VALUESOF C
0.529
0.603
0.519
In summary, this study provides a valuable first step in
describing the relationship between remotely sensed data and
chlorophyll-a concentration using aerial video cameras. Radiance determined from airborne video cameras is a fairly good
indicator of chlorophyll-a concentration in resemoirs, even on
cloudy days having highly variable irradiance. Currently, airborne video imaging of reservoirs may be used as a first approximation in quantifying the chlorophyll concentration in surface
waters of inland water bodies. The proposed technique has a
great deal of potential, and studies on this topic should continue
to improve the process and obtain more accurate results. Subsequent studies might focus on video imaging under cloudless
conditions, incorporating irradiance sensors capable of
determining incoming solar radiation spatially and contemporaneously with the video imaging process, and adapting the
technology so it may effectively be utilized under variable sky
conditions. This study provides a solid foundation for airborne
video imaging of water reservoirs and should be encouraging
to those interested in using airborne videography as a reservoir
management tool.
Acknowledgments
Thanks to James Everitt, David Escobar, Gerald Coleman, and
Michael Renee Davis for their assistance in obtaining the airborne video data; to William Troeger and Dale Pardue for the
water analyses; to John Ross and Gary Heathman for their technical assistance; and to Dr. Sherwood Mcintyre for organizing
and conducting the ground truth data collection.
Allen, P.B., and J.W. Naney, 1991.Hydrology of theLittle Washita River
Watershed, Oklahoma, U.S. Department of Agriculture, Agricultural Research Service, ARS-90.
218
February 2000
0.528
0.619
0.527
0.528
0.629
0.532
0.528
0.634
0.533
0.528
0.638
0.533
0.527
0.639
0.532
0.527
0.639
0.531
Dekker, A.G., T.J. Malthus, and E. Seyhan, 1991,Quantitative modeling
of inland water quality for high resolution MSS systems, IEEE
Transactions on Geoscience and Remote Sensing, 29(1):89-95.
Gitelson, A.A., 1992.The peak near 700 nm on reflectance spectra of
algae and water: relationships of its magnitude and position with
chlorophyll concentration, Int. J. Remote Sens., 13:3367-3373.
Han, L., D.C. Rundquist, L.L. Liu, R.N. Fraser, and J.F. Schalles, 1994.
The spectral responses of algal chlorophyll in water with varying
levels of suspended sediment, Int. J. Remote Sens., 15(18):
3707-3718.
Ritchie, J.C., and C.M. Cooper, 1987.Comparison of Landsat MSS pixel
array size for estimating water quality, Photogrammetric Engineering b Remote Sensing, 53:1549-1553.
Ritchie, J.C., F.R. Schiebe, C.M. Cooper, and J.A. Harrington, 1994.
Chlorophyll measurements in the presence of suspended sediment using broadband spectral sensors aboard satellites, lour.
Freshwater Ecology, 9~197-206.
Rundquist, D.C., L. Han, J.F. Schalles, and J.S. Peake, 1996. Remote
measurement of algal chlorophyll in surface waters: The case for
the first derivative of reflectance near 690 nm,Photogrammetric
Engineering B Remote Sensing, 62(2):195-200.
Schalles, J.E, F.R. Schiebe, P.J. Starks, and W.W. Troeger, 1997.Estimation of algal and suspended sediment loads (singly and combined)
using hyperspectral sensors and integrated mesocosm experiments, Proceedings, Fourth International Conference on Remote
Sensing for Marine and Coastal Environments, Orlando, Florida,
pp. 111-120.
Schiebe, F.R., J.A. Harrington, and J.C. Ritchie, 1988.Remote sensing
of suspended sediments of Lake Chicot, Arkansas, Proceedings
of the 6th Corps of Eng. Remote Sens. Symposium, Galveston,
Texas, pp. 77-85.
Willmott, C.J., 1982.Some comments on the evaluation of model performance, Bull. Am. Meteor. Soc., 63(11]:1309-1313.
(Received 06 April 1998;revised and accepted 16 November 1998;
revised 26 February 1999)
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING