gravel roads are easily distinguishable from asphalt roads. It is not
easy to differentiate between some roads and roof types such as
dark new roads and dark tile roofs. However, there are specific
absorption features for each urban object in the short wave
infrared which help in differentiating them. Segl et al. (2003)
confirm that it is a challenge to use these methods in identifying
urban surface materials due to the variation of these materials in
relatively small regions. A hyperspectral pixel in an urban scene is
generally a mixture of different material components which makes
it difficult to discriminate successfully between certain classes of
materials (Bhaskaran and Datt 2000).
Most of recent methods for analysis of hyperspectral data do not
directly determine the materials and just define how similar the
unknown material is to the known material. There are other
techniques which are able to directly identify the material using
spectral features. A lot of research has been done for applying
these techniques to image spectrometry to have better
classification results. Freek van der Meer (2004) develops a
method for the analysis of hyperspectral images using absorption
band depth and position over mainly mineral material. This
method consists of a simple linear regression formula to estimate
absorption-band parameters from hyperspectral image data which
is easily implemented in ENVI tools for those users that are not
familiar with programming languages. The sensitivity analysis
shows that more reliable results would be achieved by more
accurate absorption band parameters (shoulders, absorption
points). Heiden et al. (2007) develop an approach for urban
feature identification using specific robust characteristics. These
spectral characteristics include absorption bands (depth and
position) or sharp increase and decrease of reflectance.
Separability analysis is used to evaluate the robustness of spectral
features. It is concluded that urban materials need to be described
by more than one type of feature. A similar approach is followed
in this research. It could be helpful in identifying road surface
materials since the focus is on special features such as absorption
depth which may be unique for a particular material. For a
comprehensive review of related work please refer to Mohammadi
(2011). So far there is no standard approach to produce reasonable
results. The analysis techniques for hyperspectral data can be
unsupervised or supervised. Supervised techniques include
similarity measures such as spectral angle mapper (SAM), spectral
information divergence (SID) and spectral correlation mapper
(SCM). Spectral similarity measures are used to identify unknown
spectra based on a comparison with reference spectra. In addition,
unique spectral characteristics within certain wavelength ranges of
some spectra can be used directly to identify materials.
In the following chapters the data of the study area are described.
For the classification of roads using hyperspectral data various
methods have been investigated but only those which led to most
successful experimental results are discussed. The analysis on the
road surface condition is focused on asphalt roads with good
success in identifying roads with good, intermediate and bad
surface condition.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
2. STUDY AREA; HYPERSPECTRAL IMAGE DATA
AND PREPROCESSING
The analysis of hyperspectral data to support road surface material
classification is done for a case study in the city of Ludwigsburg.
This city is situated in Baden-Württemberg, Germany, in
particular in the north part of Stuttgart region, near to river
Neckar. The imagery data covers the urban area extending
approximately 11 km x 16 km of this city. The hyperspectral data
was acquired during the HyMap campaign on 20th August, 2010
by German Aerospace Center (DLR) and consists of six strips.
The data consists of 125 bands (ranging from 0.4 um to 2.5 um)
and has a ground sample distance of 4 m. ALK vector data for
roads is provided by Fachbereich Stadtplanung und Vermessung
der Stadt Ludwigsburg. The vector data is used for limiting the
analysis to roads. Information on road condition based on field
visits is available from the municipality of Ludwigsburg and the
Vermessungsbüro Praxl und Partner GmbH. The preprocessing of
the hyperspectral data has been done by DLR. The data is
corrected for radiometric, geometric and atmospheric effects. A
high resolution LiDAR surface model of Ludwigsburg with 2 m
raster size was made available for this purpose.
? à Bayem T.
Orien Wuertemberg* ri
; i d
í (
{ Lam i fe b
NAM m Por
et
Figure 1 . Research area, Ludwigsburg.
In order to determine how well the HyMap and vector data fit, the
two data sets are overlaid. It is observed that there is a shift
between the two data sets in the order of 10 m. The vector layer
for roads is selected as a source of ground control points. The
GCPs are used to georeference or correct the geographic location
of the HyMap data. After performing an affine transformation, it
is observed that an overall RMS error of about 0.7 pixels is
obtained. A road mask is created from the road vector data.
However, it is observed that the road mask also covers some areas
with vegetation and thus these areas have to be eliminated.
Vegetation spectra for different areas in the HyMap data are
collected and used to create a spectral library. The HyMap data is
classified using the spectral library in order to identify vegetation
areas. A mask is created from the output classification map. This
mask is subtracted from the road mask in order to obtain a mask
limiting subsequent analyses to road surfaces only.
In
cla
pai
TO
on
3.1
As
sui
ari
