Geoscience Reference
In-Depth Information
and from aerial color 35-mm reference photography acquired under the same conditions. With the
addition of Global Positioning System (GPS) coordinate data encoded directly onto the videotape
for georeferencing, sample points can be rapidly located for interpretation during playback.
9.2.3
Reporting Accuracy Assessment Results
The current standard for reporting results of classification accuracy assessment focuses on the
error or confusion matrix, which summarizes the comparison of map class labels with reference
data labels. Some easily computed summary statistics for the error matrix include overall map
accuracy, proportion correct by classes (user and producer accuracy), and errors of omission and
commission. Additional summary statistics usually include a Kappa (Khat) coefficient that adjusts
the overall proportion correct for the possibility of chance agreement (Congalton et al., 1983;
Rosenfield and Fitzpatrick-Lins, 1986; Congalton and Green, 1999). Although Kappa is widely
used, some authors have criticized its characterization of actual map accuracy (Foody, 1992). Ma
and Redmond (1995) proposed some alternatives to the Kappa coefficient, including a Tau statistic
that is more readily computed and easier to interpret than Kappa. Stehman (1997) reviewed a variety
of summary statistics and concluded that overall map accuracy and user and producer accuracies
have direct probabilistic interpretations for a given map, whereas other summary statistics must be
used with caution. The error matrix itself is recognized as the most important accuracy assessment
result when accompanied by descriptions of classification protocols, accuracy assessment design,
source of reference data, and confidence in reference sample labels (Stehman and Czaplewski,
1998; Congalton and Green, 1999; Foody, 2002).
9.3 METHODS
Four LC maps for the upper San Pedro River Watershed (Plate 9.1) were generated using 1973,
1986, and 1992 North American Landscape Characterization (NALC) project MSS data (Lunetta
et al., 1993) and the 1997 TM data. All images were coregistered and georeferenced to a 60-
60-
m Universal Transverse Mercator (UTM) ground coordinate grid with a nominal geometric precision
of 1 to 1.5 pixels (60 to 90 m).
¥
9.3.1
Image Classification
= 10) were used to develop all four maps (Table 9.1). Vegetation cover
classes represented very broad biome-level categories of biological organization, similar to the
ecological formation levels as described in the classification system for biotic communities of North
America (Brown et al., 1979). The classes included forest, oak woodland, mesquite woodland,
grassland, desertscrub, riparian, agriculture, urban, water, and barren and were selected after direct
consultation with the major land managers and stakeholder groups within the San Pedro watershed
in Arizona and Mexico (Kepner et al., 2000). Most of the watershed was covered by grassland,
desert scrub, and mesquite and oak woodland (Table 9.2).
The classification process for each data set began with an unsupervised classification using the
green, red, and near-infrared spectral bands to produce a map with 60 spectrally distinct classes.
The choice of 60 classes was based on previous experience with NALC data that usually gave a
satisfactory trade-off between the total number of classes and the number of mixed classes. In this
context, it proved helpful to define a set of 21 intermediate classes, which were easier to relate to
the spectral information. For example, the barren class contained bare rock, chalk deposits, mines,
tailing ponds, etc., that had unique spectral signatures. Each class was then displayed over the false-
color image and assigned to one of the LC categories or to a mixed class.
The same LC classes (
n
