WSP11837

• File Number: 8200.760
• Description: Yampa River General
• State: CO
• Basin: Yampa/White
• Water Division: 6
• Date: 10/1/1978
• Author: USGS
• Title: Preliminary Applications of Landsat Images and Aerial Photography for Determining Land-Use - Geologic - and Hydrologic Characteristics - Yampa River Basin - Colorado and Wyoming - October 1978
• Water Supply Pro - Doc Type: Report/Study

(fig. 1), was selected for initial development of training sets for various
land-use classes. Four areas on figure 9 were chosen for possible use in
selecting training sets. These four areas were selected because they
contained the majority of land-use and vegetative-cover types present in the
eastern one-half of the Yampa River basin.

Unsupervised classification was applied to area ,. The system was
instructed to divide all of the pixels within area 1 into 20 spectral
classes. Although the spectral separation for all of the 20 classes was
good, the resulting spectral classes did not correlate well with land-use or
vegetative-cover classes. For example, a specific land-use class would
include several spectral classes. Water was the single exception, because
its spectral characteristics were unique. Limited correlation between
spectral classes selected by clustering and actual land-use classes
undoubtedly is caused by a combination of the spectral variability within a
particular land-use class and division of the area into too many spectral
classes (20 groups). Both of these factors could result in more than one
spectral class within a particular land-use class. Clustering may have given
better results if fewer spectral classes had been selected, but time
iimitations precluded further investigation of this approach.

Several training sets were selected from the four areas for each land-
use or vegetative-cover class that could be identified on the imagery. In
addition, because clouds and cloud shadows were present in other parts of the
imagery, training sets were selected for these features as well. This was
done because the maximum-likelihood classifer used by IDIMS segregates all
pixels in the image into one of the preselected land-use classes. Most of
the training-set selection was done within the four areas delineated on
figure 9. However, training sets for conifers, clouds, and cloud shadows
were selected from other, more appropriate areas on the imagery. Maximum
detail within each of the four areas was obtained by expanding each of the
areas to fill the display monitor. This allowed individual pixels to be
identified. Training sets for each land-use class were located on the
imagery primarily by reference to the appropriate black-and-white aerial
photographs, topographic maps, and field-reconnaissance data.

Several training sets outlined by polygons in area 2 are illustrated on
figure 10. Any number of training sets may be selected for each land-use
class that is identified. When the selection of training sets is completed,
the computer records the distribution of spectral values for each land-use
class and calculates the necessary statistics to define the spectral
signature of that land-use class.

A series of two-dimensional plots of the distribution of pixel values
for land-use classes can be used to evaluate the spectral separation between
classes. These plots consist of ellipses that show the distribution of data
between two spectral bands for each of the selected land-use classes. The
area inside an ell ipse represents the data (plotted in two-dimensional space)
that are within one standard deviation from the mean for that class.

20