Image Processing Reference
In-Depth Information
35°35'35" N
N
Phoenix
33°5'35" N
0
200 km
113°41'55" W111°11'55" W108°41'55" W
Fig. 12.2
Location of the Phoenix, AZ study region and recoded expert land cover classification
results
minimize potential confusion in the reference dataset used for accuracy assessment.
The reference dataset was constructed from 3 m/pixel digital aerial orthophotos for
the Phoenix metropolitan region collected in 1999. The 2000 ASTER VNIR data
were also examined during the assessment process to minimize reference dataset
error due to temporal change in classes such as Agricultural Vegetation. Assessment
points were selected using a stratified random approach to ensure that each class had
at least 50 associated reference points. Points were selected using a 3 × 3 moving
window and a 9-pixel majority rule; in other words all 9 pixels in the moving window
were required to be the class of interest for the central point to be selected. Each
accuracy assessment point was then examined to ensure that it did not fall within the
associated class of interest's training regions; any point that did was removed from
the reference dataset. If the number of reference points for a given class fell below 40,
new points were selected to achieve this minimum number. Table
12.2
presents
producer, user, and conditional Kappa accuracy assessment values.
12.2.4
Landscape Metric Calculation and Grid Construction
The land cover classification data were further aggregated to six superclasses
(Asphalt, Soil and Bedrock, Agriculture, Undifferentiated Vegetation, Built, and Water)
to reduce the computational load for landscape metric calculation. The Water class
was not included in further analysis due to its minimal representation in the land
