An object-oriented pattern recognition approach for urban classification - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 128

FIGURE 9.8 Mean value of brightness feature (a threshold
value of 22 000 to extract pools - blue areas). Note: Some
other land-cover areas were mistakenly identified as pools.

FIGURE 9.10 Output generated by rule set 1 (Mean band
4feature < 10 000 and not mean PCA 1 < 19 000 at scale
level 4). Note: A few other land-cover areas were mistakenly
identified as pools.

FIGURE 9.9 Mean value of the original band 5 feature
(a threshold value of 19 000 to extract shadows - yellow
areas).

FIGURE 9.11 Output generated by rule set 2 (Mean bright-
ness feature < 22 000 and not mean PCA 1 < 19 000 at
scale level 4). Note: A few other land-cover areas were
mistakenly identified as pools.

TABLE 9.2 Selected rule sets to identify swimming pools.

(Congalton, 1991). To be consistent and for precise comparison
purposes, we applied the same sample points generated for the
output generated by the first data set (mean values of the original
bands and PCA bands 1, 2, and 3) to the output produced by
the second dataset (mean values of the original bands, brightness
band, and maximum difference band).

Decision Rule
Rule 1 Means band 4 feature < 10,000 and not mean PCA
1 < 19,000 at scale level 4
Rule 2 Means brightness feature < 22,000 and not mean
PCA < 19,000 at scale level 4
Rule 3 PCA 2 feature value less than 22,000 at scale level 4
Rule 4 PCA 2 feature value less than 22,000 at scale level 1

9.4.2 Nearest neighbor classifier to
extract urban land covers

subset to perform the accuracy assessment instead of randomly
sampling pixels from the image. The reference map can be
assumed to contain a negligible amount of error. There are total
of 175 062 pixels in the image, and only 1973 pixels belong to the
pool area.
We selected 200 samples points that led to approximately 30
points per class (seven total classes) for the accuracy assessment.
A minimum of 15 points per class was selected to generate
200 test points using a stratified random sampling approach

We attempted to identify seven urban classes including building,
grass, impervious, swimming pool, shadow, soil, and tree. There
are two features or data sets used for the analysis: (1) mean values
of the original bands and PCA bands 1, 2, and 3; and (2) mean
values of the original bands, brightness band, and maximum
difference band. We observed that tree land-cover class stands
out as bright objects and unmanaged soil cover turned out to be
one of the darkest object types for the maximum difference band