Remote sensing of high resolution urban impervious surfaces - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 238

Our analysis also demonstrates that adding the two extra
feature variables of ''Ratio of Band 4'' and ''Ratio of Band 1'' did
affect the classification accuracy. While the Kappa value increases
1%when adding the ''Ratio of Band 1'' only, the overall accuracy
decreases 3% by adding the ''Ratio of Band 4''. This is probably
because the ''Ratio of Band 4'' shows similar values for some
forests and non-forested vegetation, which lead to decreased
accuracy for forest and non-forested areas. Nonetheless, when
both ratios of band 1 and band 4 were added, the classification
results increased about 2% (Table 17.4).
The land cover map was classified based on the optimal
shape and scale factor values and with the two extra feature
variables included (Fig. 17.8a). Its accuracy has been assessed by
creating error matrix using the same test area as reference data
(Table 17.5). An overall accuracy of 92.5% was achieved for the

17.4.3 Results

The results of the object-oriented classification change over
the user-defined values of shape and scale factors (Fig. 17.7).
Different land cover classes show different change trends. For
example, the accuracy of water is consistent while the impervious
class shows steadily decreasing accuracy with increases in the
shape factor value. The overall classification accuracy reaches the
highest point when the shape factor approximately equals to 0.3
and then it trends down to the lowest when the value is 0.9. On
the other hand, the highest accuracy can be found for majority
of the classes when scale parameter reaches about 20. Therefore,
the optimal shape factor and scale parameter are set to 0.3 and
20, respectively for our classification.

Impervious Surface
Forested Area

Water

Non-forested Area
Shadow

(a)

(b)

FIGURE 17.8 Object-oriented classification maps based on the QuickBird multispectral data. (a) General land covers
classified by the object-oriented classifier, (b) Final impervious surface map after unmixing shadow class into impervious
surface or other classes.