Environmental Engineering Reference
In-Depth Information
spectrum as a combination of spectra for pure land cover
types, called endmembers. SMA has been successfully applied
for estimating the fractions of impervious surface for a pixel of
medium resolution remote sensing imagery. For example, Phinn
et al
. (2002) successfully estimated impervious surface distribu-
tion using a constrained spectral mixture analysis method with
endmembers chosen from aerial photos. Wu and Murray (2003)
implemented a constrained linear SMA to generate impervious
surface distribution in Columbus OH, and found that impervi-
ous surface fraction can be estimated by a linear model of low
and high albedo endmembers. Further, Wu (2004) proposed a
normalized spectral mixture analysis (NSMA) to achieve a better
estimation accuracy of impervious surface distribution. For high
resolution impervious surface estimation, Lu and Weng (2009)
developed linear spectral mixture analysis (LSMA) and deci-
sion tree classifier (DTC) to quantify urban imperviousness.
In addition, considering the spectral variations and boundary
effects of IKONOS imagery, Wu (2009) developed a modified
NSMA model.
Pu
et al
., 2008) and high resolution (Mohapatra and Wu, 2007)
remote sensing imagery.
17.2.2
Object-basedmodels
Besides pixel-based models, object-based classification has also
been developed for classifying impervious surface areas from
high resolution remote sensing imagery (e.g., QuickBird). Com-
pared to traditional per-pixel classificationmethods, object-based
approaches provide unique capabilities to incorporate large-scale
textural and contextual information, and numerous object-based
features in the classification process. The unique characteristic of
object-oriented classification is creating image objects by image
segmentation and performing classification on image objects
rather than image pixels. The purpose of image segmentation is
to provide optimal information that simultaneously represents
objects in different spatial resolutions for further classification
(Gitas, Mitri and Ventura, 2004). The complexity of urban land-
scape inhigh resolution remote sensing imagerymakes traditional
digital image classification difficult. Object-oriented approach is
considered more appropriate since it differentiates land cover
classes based on both spectral and spatial information of the
image data. Some studies demonstrated significantly higher accu-
racy for the object-oriented approach (Benz
et al
., 2004; Wang,
Sousa, and Gong, 2004), while other investigations reported
object-based method produced similar results with comparable
accuracy as other methods (Willhauck, 2000; Sun, 2003).
17.2.1.3
Regression treemodel
In addition to SMA, regression tree model is another popu-
lar method for generating impervious surface data in an urban
area. Like regression analysis, a regression tree constructs the
relationships between a dependent variable (e.g., impervious
surface fraction) and independent variables (e.g., reflectance
for a particular band). However, the regression tree model is
more complicated than regression analysis. It grows a (inverted)
categorical tree by repeatedly splitting the data according to
specific rules depending on how the dependent variable and
the independent variables interact with each other. The goal of
the algorithm is to categorize the data into more homogeneous
groups by uncovering the predictive structure of the problem
under consideration (Breiman
et al
., 1984). The performance
of the regression tree model was reported to be more accurate
than ordinary regression models in many applications (Huang
and Townshend, 2003). Yang
et al
. (2003) reported the results
of applying regression tree models in quantifying impervious
surface fraction applied to Landsat ETM
17.3
Pixel-based models for
estimating high-resolution
impervious surface
17.3.1
Introduction
images. This method
has been adopted by United States Geological Survey (USGS)
to produce 30
+
The objective of this section is to develop a pixel-based method
for generating accurate and high-resolution impervious surface
data from IKONOS imagery. In particular, this section explores
whether the existing impervious surface generationmethods (e.g.,
SMA and RT models) can be applied to high resolution remote
sensing imagery. Moreover, an integrated SMA and RT model
has been developed to further improve the estimation accuracy.
The rest of this section is organized as follows. Section 17.3.2
describes the study area and remotely sensed data, including
IKONOS imagery and aerial photographs. Two popular pixel-
based impervious surface estimation methods, SMA and RT
models, and the integrated SMA and RT model, are detailed
in Section 17.3.3. Section 17.3.4 reports the results of accuracy
assessment.
×
30 m national land cover dataset (NLCD). For
high resolution impervious surface estimation, Goetz
et al
. (2003)
estimated impervious surface using a classification tree approach
for Montgomery County, MD, in which 11 IKONOS image tiles
were acquired in six swaths to provide complete coverage of
the study site. Lu and Weng (2007) used a hybrid approach
based on the combination of decision tree classifier and unsu-
pervised classification to extract impervious surface areas from
IKONOS data.
17.2.1.4
Artificial neural network
In addition to the above pixel-based models, ANN is another
alternative to traditional classification models and it can address
the problems of nonlinearity among variables (Ji, 2000). Fur-
ther, when compared with other classifiers (e.g., maximum
likelihood), ANN models generated better classification results
(Flanagan and Civco, 2001; Kavzoglu and Mather, 2003). ANN
has been applied to estimate urban impervious surface informa-
tion from both moderate resolution (Flanagan and Civco, 2001,
17.3.2
Study area and data
Grafton (including the village and town) in Ozaukee County,
WI (see Fig. 17.1) was chosen as the study area. Being about
Search WWH ::
Custom Search