An object-oriented pattern recognition approach for urban classification - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 127

(a) (b) (c) (d) (e)

FIGURE 9.4 A subset image and segmented images at different object scales using shape ( S sh )0 . 1 and compactness
( S cm )0 . 5. (a) Original subset; (b) level 1 (scale parameter 10); (c) level 2 (scale parameter 25), (d) level 3 (scale parameters
50), (e) level 4 (scale parameter 100).

(a) (b) (c)

FIGURE 9.5 (a) Mean value of the original band 4 feature (Dark tone in pool areas). (b) Mean value of PCA 2 feature (Dark
tone in pool areas), (c) Mean value of brightness feature (Dark tone in pool areas).

FIGURE 9.6 Mean value of the original band 4 feature (a
threshold value of 10 000 to extract pools - blue areas).
Note: Some other land-cover areas were mistakenly identi-
fied as pools.

FIGURE 9.7 Mean value of PCA 2 feature (a threshold value
of 22 000 to extract pools - blue areas). Note: No mistakenly
identified areas.

outputs has the highest accuracy since delineated features are
similar among the four outputs. Thus, an accuracy assessment
was conducted for all four classified outputs. We provided the
rule sets developed to extract pools in Table 9.2.
We produced error matrices in order to analyze and evaluate
each approach. These error matrices show the contingency of the
class to which each pixel truly belongs (columns) on the map
unit to which it is allocated by the selected analysis (rows). From
the error matrix, overall accuracy, producer's accuracy, user's
accuracy, and kappa coefficient were generated.
To evaluate different rule sets applied to extract swimming
pool more effectively we digitized the swimming pool in the
first subset using a visual interpretation approach via heads-up
digitizing to produce a reference map that contains two classes
(i.e., pool and non-pool class). We used each pixel in the whole

been mistakenly identified as pools, when generated by the mean
values of the original band 4 and brightness features.
Figures 9.10 and 9.11 show the output maps after masking the
shadows. It was found that the rule set with PCA 2 feature did
not produce shadows in the area and hence, the above procedure
was not necessary. However, we provided the output of Rule set 4
(PCA 2 feature value less than 22 000 at scale level 1) in Fig 9.12.
To demonstrate the effect of using different image object levels
we also identified the pool using the same approach with PCA
2 feature at level 1. By visual inspection, the output at level 1
provided more realistic representations of pool edges than the
one produced at level 4. However, we found three small objects
identified as pools adjacent to the actual pools, though they
are not very noticeable. It is uncertain which one of these four