EM 1110-2-2907
1 October 2003
which the disk is visible. Remote sensors generally detect water clarity to 2050% of the Sd
measurement. Sampling sites were chosen evenly across the reservoir and adjacent tribu-
taries. A global positioning unit was used to locate 109 sample sites. Drift during sampling
occurred but was compensated for with the use of a 33 kernel during image classification.
Samples and data were collected during two periods--summer and fall of 1991.
c. Sensor System. Two Landsat TM data sets separated in time were used to develop a
was employed with a stratified sampling technique. Using a variety of band ratio, the work-
ers were able to accurately develop a blended classification scheme, which is detailed be-
low.
d. Study Results. The authors adapted multivariate density estimation with the use of an
algorithm k-NN density estimator. This was used to group spectrally similar pixels. The
spectral classes and class structures (or groupings that separated the spectral classes) were
developed using an unsupervised classification. Within each of the two scenes, 16 unique
classes were determined. These classes were combined with the empirical data, leaving four
logarithmic algorithms. Applying a 33 kernel to the data compensated for the drift that oc-
curred during data collection. This placed the positional accuracy to within 30 m.
(1) The average spectral value was determined by a log estimation of the band ratio
for the given pixel within the kernel (Equation 6-2). Combinations of band ratios were
tested. A middle infrared ratio against the visible red showed the largest correlation with Sd
and Ca. Visible green versus near infrared also provided a good separation of the spectral
response for estimating Sd and Ca.
1 3
3
(mid IR)
IR = ∑
∑ ln
6-2
9 x =1
(visible red)x,y
y =1
estimate. Focused sampling and spectral blending led to the development of an accurate un-
supervised classification with a 95% confidence interval. Sampling positions near tributar-
ies were overestimated at only five sampling sites (relative to 109 sampling sites).
e. Conclusions. A strong correlation was made between the Landsat TM middle IR and
the empirical Secchi Disk and chlorophyll concentrations. Chlorophyll was shown to have
increased from 12.64 to 17.03 mg/m3, contributing to a decline in water clarity. The appli-
cation of this log estimate now eliminates the need to collect empirical water quality data,
likely reducing the cost in a water quality survey.
Point of Contact: Robert Bolus, Phone: (603) 646-4307
6-13
Previous Page
