ATA
naterial
gsburg.
ny, in
O river
tending
ral data
t, 2010
Strips.
.9 um)
lata for
lessung
ing the
n field
and the
sing of
data is
ects. À
th 2 m
fit, the
a shift
or layer
ts. The
ocation
ation, it
ixels is
yr data.
1e areas
jinated.
ata are
data is
zetation
ip. This
a mask
3. METHODS AND RESULTS
In order to identify road material types, different supervised
classification approaches are tested. The investigation is in two
parts. The first part discusses the methods and results obtained for
roads surface material identification while the second part focuses
on conditions determination.
3.1 Road surface material identification
Asphalt or more specifically bituminous asphalt is the most used
surface material of road surfaces (pavements) in Ludwigsburg. It
is a composite material of construction aggregate (e.g. gravel,
crushed stone, stand, etc.) and asphalt which serves as a binder.
The mixing is formed in various ways which leads to a certain
“asphalt variety” within classification. Other surfaces in the city
are concrete surfaces, in particular the concrete pavers, and gravel
surfaces. These three main classes, simply called ‘asphalt’,
‘concrete’ and ‘gravel’ are taken into account for road material
identification. In order to map these materials, the SAM criterion
and the brightness spectral feature are used. Training regions are
selected over the calibration sites for the three materials. Each
training region defines a spectrum as an ROI average. Using the
ROI spectra, the roads within the investigation area are classified
based on SAM with the default threshold setting (ENVI) of 0.1
radian. Angles between two spectra larger than this threshold lead
to unclassified pixels. In addition to 0.1 two more spectral angle
thresholds are tested: 0.08 and 0.15 (radians). Best classification
results have been achieved with the lowest threshold of 0.08 but at
the cost of a high percentage of unclassified pixels. The output
classification map based on the threshold of 0.08 is shown in
figure 2. Four different subclasses are defined for asphalt which
are combined to form one asphalt class.
MW Unclassified
Masphaiti
U Gravel
BlConerste
asked Pixels
Figure 2. SAM classification using threshold of 0.08 (radians).
In the classification results (excluding non-road pixels), about
29% of the pixels are identified as asphalt, 2% as concrete, 3% as
gravel and 66% of the pixels are unclassified. Figure 3 is an
example of an area that shows a road segment of concrete pavers
which is correctly classified. The corresponding spectra (Figure 3)
indicate that the SAM similarity measure is relatively insensitive
to illumination and albedo effects. Besides, field visits confirm
that the material is concrete.
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
OrthoPhoto M Class Image 2 2 AT ^ Le
Figure 3. An example of a correctly classified area of
concrete pavers.
materials involves the use of the brightness spectral feature.
Brightness is one of the spectral features which is more distinct for
materials with relatively flat low reflectance curves (Figure 4)
such as asphalt (Heiden et al., 2005).
Road surafce materials
T T T
ascol. GORTERIA een z
E conos d T 5 7
s 2000L | NO el
5 LÍ id
5 | ^&
B [| | Yu
gen à de ae
= 1500 7 asphalt Le Y AN ^&]
x e qM i
Lu I
7 Jj
L1 i rd 1 LL ier 5 1 zi
0.5 1.0 À =.
Wavelength
Figure 4. Asphalt, concrete and gravel spectra.
For evaluating the brightness of surface materials the mean and
standard deviation functions are used. In order to distinguish
asphalt, concrete and gravel, the mean function is used over the
complete wavelength range of the HyMap sensor (445nm-
2448nm). It is observed that asphalt has the lowest mean in this
range and can be easily distinguished from the other two
materials. Concrete and gravel have similar mean values over the
specified wavelength range which makes it difficult to
differentiate the two materials. By looking at the signatures of
concrete and gravel it can be expected that the standard deviations
distinguish significantly if the standard deviation function is
applied over the wavelength range of 619.9nm-1323.7nm. This is
experimentally confirmed; in particular, concrete has a low
standard deviation in this wavelength range. Based on these
findings a simple procedure for the identification of asphalt,
concrete and gravel is shown in figure 5. The sequential process of
applying mean and standard deviation features has a desired side
effect which is the detection of vegetation pixels (high standard
deviation) which might have remained after the first attempt of
vegetation removal.
