Ground-Water Hydrology of the Upper Klamath Basin, Oregon and ...

Appendix A. 71
Appendix A. Landsat Analysis
Methods
The imagery classification was based on two levels of
resolution: (1) a generalized nonirrigated level that consisted
of four Anderson Level I/II classes: ice and snow, evergreen
forest, water, and shrub lands that represented a combination
of bare soil, rock, sagebrush and other nonirrigated vegetation
(Anderson and others, 1976); and (2) irrigated lands clustered
into five vegetative classes based on their spectral similarity
and potential crop water needs. These classes were labeled
alfalfa and irrigated grasses, small grains, onions-garlic,
potatoes-corn, and strawberries. Although it would have been
desirable, identification of individual crop types proved to be
impossible because of the lack of unique spectral signatures
(Paul Seevers, EROS Data Center, written commun., 2000).
The use of three satellite images from different times during
the growing season increased the probability that certain croptypes
could be distinguished based on the development of their
spectral signatures. To aid in the classification of the imagery,
field work was conducted to map directly crop types in areas
totaling about 17,000 acres. A variety of representative landcover
types were observed, including native trees, pasture
and most of the agricultural crops. Most data were collected
in mid-July. Data on crop type and height, percentage of crop
cover, sprinkler type, and other parameters also were collected.
The three images were processed in succession. The May
21, 2000, scene (from Landsat 5) showed full canopy cover of
the perennial crops such as alfalfa, irrigated grasses, and any
winter wheat that may have been planted the previous fall.
The timing of the scene put it before any annual row crops
had enough growth to show a vegetative signature. Areas
with vigorous growth in the May scene served to mask over
the same areas in the August scene. This step reduced the
amount of data in the August scene that required analysis. The
irrigation district boundaries were used to segregate irrigated
areas outside of the Klamath Project. The August 1 scene
(from Landsat 7) revealed the full canopy of the annual row
crops planted during the spring. The September 18 scene (also
from Landsat 7) was used to show any crops that might have a
vigorous vegetative signature beyond the harvest dates of the
small grains.
A review was performed on the results of the final landcover
classification by creating an error matrix to evaluate
how well the classification of the imagery matched what was
actually mapped on the ground. The review looked at how
accurately the classification identified specific crop-types
and how accurately the classification did with respect to all
irrigated crop-classes. To do this, the new land-cover map
created from the classified Landsat imagery was converted
to a polygon dataset, with each polygon being 30 meters, the
resolution of the original imagery. Each polygon contained a
code for crop class determined by the classification process.
The ground reference boundaries were then used to clip out
the same areas in the classified land cover. The clipped landcover
polygons were then evaluated by crop class against the
ground reference. For example, if a ground reference area
was identified as alfalfa, then the same area was compared in
the land cover. The results of the review showed that largeacreage
crops such as alfalfa had a correct classification ratio
of about 64 percent. However, if all irrigated crop classes were
included then the accuracy of identifying irrigated lands within
that area increased to about 73 percent. For crops grown on
smaller fields, the accuracy of the classification to identify
specific crop class decreased to a range of 20 to 30 percent. If
all irrigated lands were included then the accuracy increased to
the 50 to 70 percent range. Sugar beet fields for example, were
correctly classified 34 percent of the time, but if all irrigated
lands for the same area were included then the accuracy
increased to 62 percent. All mint and strawberry fields were
located during the ground truth and were included in the final
classification.