Potential of hyperspectral remote sensing for analyzing the urban environment - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 50

several functions, such as ratios and vegetation indices (Bannari
et al ., 1995).
Figure 4.1 depicts exemplary feature functions for numeri-
cal description of spectral features based on spectra of different
man-made materials. The functions mean and standard devia-
tion (Fig. 4.1a) characterize the albedo and are mostly used for
identification of materials with typical brightness and flat curve
progression and weak or no spectral absorption features, such as
concrete and a number of dark materials. Another feature func-
tion is the ratio describing increase and decrease of reflectance
occurring in a specific wavelength range. The example shown
in Fig. 4.1b depicts the use of this function for describing the
absorption feature of polyethylene near 1700 nm. Broad absorp-
tion bands which are typical for aluminum and zinc are well
expressed by the feature function of the area that is enclosed
by the reflectance curve and the hull function between two
wavelength positions (Fig. 4.1c). The spectrum of red loose chip-
pings contains the above mentioned kaolinite absorption band
between 2134 and 2272 nmwith the maximum absorption depth
at 2200 nm (Fig. 4.1d). The applied feature function calculates
the maximum of the difference between the reflectance curve and
the hull function within a defined wavelength range.
So far, most of the approaches for hyperspectral image analysis
only consider one spectral feature type for a certain analytical
task. Due to the wide range of surface materials and their
spectral variations occurring in the urban environment there
is the need for techniques which are capable of simultaneously
analyzing different spectral feature types in order to derive the
most characteristic combination for a certain material. For this
purpose the authors have developed an interactive approach
for identification of robust spectral features based on image
spectra since they contain the most comprehensive variability
of spectral characteristics for the analyzed surface materials.
The approach is described in detail in Heiden et al . (2007).
It has led to the identification of distinct spectral features for
21 manmade surface materials. Figure 4.2 shows the results for
roofing polyethylene (a), and the open surface materials red loose
chippings (b) and asphalt (c). For these examples representative
image spectra obtained from different flight lines are plotted
against a single field spectrum. Additionally, for each material
the number of analyzed image spectra, the specific feature types,
the used feature functions and the wavelength ranges where
the features occur are specified, whereas the wavelength values
correspond to the center wavelength of the HyMap bands. In
case of these three materials the number of identified spectral
features varies between nine for roofing polyethylene and three
for asphalt reflecting the overall finding that the biggest number
of spectral features can be identified for bright materials, such
as roofing polyethylene and roofing tiles. In contrast, some of
the dark materials, such as asphalt and roofing tar paper do not
show distinct absorption features. Their separation from other
materials had only been possible based on brightness represented
by the feature functions mean and standard deviation of all
reflectance values between two wavelength positions.
All of the interactively identified spectral features have been
validated with regard to their robustness determining the sep-
arability. The developed approach and the obtained results are
discussed in detail in Heiden et al . (2007). They show that the
identified spectral features have resulted in overall good separa-
tion between materials. Best separability has been achieved for
bright materials with distinct spectral features (e.g., polyethy-
lene). Problems have occurred for dark materials with low

reflectance and weak spectral characteristics. These results con-
firm the capability of hyperspectral imagery to resolve robust
spectral features for most of the investigated urban surface mate-
rials. However, the investigations have also revealed that the
interactive approach is very time consuming and thus can only
be applied in test areas with a limited number of surfacematerials.
Hence, there is the goal to develop methods which are capable of
automatically determining such robust spectral features as part
of an automated material mapping system.

4.3 Automated
identification of urban
surfacematerials

Automatedmapping approaches for the urban environment have
to be capable of fully exploiting the spectral information content
of the hyperspectral imagery and related expert knowledge in
form of spectral libraries. Furthermore, such approaches need to
be able to handle a large number of possible surface materials
including their spectral variations and a high percentage of mixed
pixels in the image data due to the dominance of heterogeneous
and small-sized structures. In the following, standard approaches
of automated hyperspectral image analysis are briefly discussed
and evaluated in regard to their suitability for automatedmapping
of urban surface materials (Section 4.3.1). The obtained findings
have formed the basis for the development of a multi-step
hyperspectral processing system for urban material mapping
which is presented in Section 4.3.2.

4.3.1 State of the art of automated
hyperspectral image analysis

Most of the classification and spectral unmixing approaches
follow a three-step processing scheme. In the first step, the
spectral dimensionality of the reflectance imagery is reduced
(Section 4.3.1.1). In the second step, key endmembers are
identified that are represented by spectrally pure image pixels
(Section 4.3.1.2). In the third step, these endmembers are used to
perform automated classification and spectral unmixing of sur-
face materials for the whole image data set. This inversion process
results in classification maps or in case of spectral unmixing in
surface abundance maps for each material. In the following, the
steps of this overall approach are discussed in more detail focus-
ing on the suitability of existing methods for analyzing the urban
environment.

4.3.1.1 Dimensionality reduction

Dimensionality reduction aims at eliminating redundant infor-
mation from a dataset because many algorithms perform better
in a low-dimensional and uncorrelated subspace. In case of
hyperspectral data, dimensionality reduction is a requirement
for many endmember detection methods that rely on convex
geometry concepts (Section 4.3.1.2). In this context, methods,
such as principal components analysis (PCA), the minimum