Remote sensing of high resolution urban impervious surfaces - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 234

impervious surface estimation accuracy (Song, 2005). For the
RT model, the selection of independent variables may influence
the final estimates. In this paper, four reflectance bands and four
Tasseled Cap components have been chosen following the studies
of Yang et al . (2003). Although this selection seems reasonable,
other variables may be more closely related to impervious surface
distribution. As an approach to partially address these problems,
an integrated approach of the SMA and RT model is proposed.
In this approach, the SMA provides an 'initial' estimate, and the
RT method is utilized to further calibrate the initial estimate. In
particular, the estimates of vegetation, impervious surface, and
soil fractions, generated by the SMA model, were inputted to
the RT model as independent variables, and the dependent vari-
able is the 'true' impervious surface fraction digitized from the
aerial photograph. The same sets of training and testing samples
were applied, and the Cubist program was utilized to construct
the RT model. The resulting regression tree rules are shown in
Table 17.2.

TABLE 17.2 Rule definition for the integrated spectral mixture
analysis and regression tree model.

Rule Condition Impervious surface fraction

1 imp ≤ 15 . 36628 − 0.05834 + 0.78 imp

2 imp > 15.36628 26.59405 + 0.69 imp − 0.51 veg

17.3.3.4 Accuracy assessment

In order to compare the performances of these three models, an
accuracy assessment has been carried out. In detail, 150 testing
samples with a sampling unit of 5 5 pixels were generated

using a stratified random methodology for evaluating the esti-
mation accuracy. For each sample, a 20

×

20msamplingarea
was identified, and the impervious surfaces were digitized using
ERDAS IMAGINE

×

™ Area of Interest (AOI) tools. The fraction

(a) (b)

(c)

FIGURE 17.5 Impervious surface fraction imagery generated from (a) spectral mixture analysis model, (b)regression tree
method, and (c) integrated model of spectral mixture analysis and regression tree.