Geoscience Reference
In-Depth Information
clump function classifies the isolated pixel into the class which has the highest
occurrence of its surrounding pixels. Some water bodies such as a river and some
forest which are represented by a pixel or line of pixels were likely to be removed
by the sieve-clump process, therefore the original water and forest classes were
reserved.
Finally, an accuracy assessment was conducted for the classified image by using
the reference points, which were separated from the training sample at the begin-
ning. The reference points were compared with the classified map to check to see if
the reference field was classified correctly or not in ArcMap with the cardinal
directions of the reference point visualized in a different color. This information
was typed into Excel, which has columns of reference point numbers and classified
classes. For the forest and water class, random points were created by an accuracy
assessment function, and they were visually assessed by using the satellite image
with the color composition used for the creation of the training set described earlier
in this section.
An error matrix with columns of reference classes and rows of classified classes
was created. By using the matrix, overall accuracy was estimated by dividing the
total amount of diagonal pixels by the total amount of all of the pixels used for the
accuracy assessment. An accuracy of 85% or more overall accuracy is considered
acceptable. The accuracy of each class was estimated. A producer's accuracy or an
omission was calculated by dividing the total number of the correctly classified
pixels in a certain class by the total number of pixels of that class derived from the
reference data. A user's accuracy or commission errors is calculated by dividing
the total number of correctly classified pixels in a certain class by the total number
of pixels of that class derived from the classified data and tells how much the
classified pixels match with the actual validation points. A further assessment was
performed by conducting a kappa analysis, which indicates the accuracy between
the classified map and the reference data and if accuracy was derived from an actual
agreement between the two data or by chance. Actual agreement would be strong
with a kappa value of more than 0.80, fair with a value between 0.80 and 0.40, and
poor with a value below 0.40.
An example of an annual land cover/land use and crop type classification (from
2005) is shown in Fig.
8.6
, with farmland with planted crops the most common land
cover/land use. Soybean (47%) and corn (18%) crops are the most common rural
land use/land cover types. Forest cover was found to represent 17% of the land area.
The reference points of hay and wheat were small due to the limited amount of
the ground truth points. Overall accuracy of the classification was 87.96% and the
Kappa value was 0.82.
Tillage classification was conducted in the same way as the crop types were
classified. For example the tillage classification in 2006 used a Landsat TM image
of path 20/row 31 acquired on May 23, 2006 and obtained from the OhioView
website. The classes created for the map were traditional tillage (
30 %), mulch-till
<
(30-90 %), no-till (
90 %), forest, and water. Tillage systems within the Maumee
watershed were documented at 8,927 farmfield data points that were checked by
USDA personnel at the county level. Approximately ½ of the farm field points were
<
