The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
190
characteristics of these road network structures, roads in USA
tend to be parallel and cross each other orthogonally forming T-
junctions or crossroads, whereas European roads tend to wiggle
and meet or cross each other at roundabouts. Thus it seems
natural that Hp < 2 bits are necessary to encode information
about road segments at junctions for road networks in the USA,
whereas for road networks in Europe, Hp> 2 bits are necessary.
The same measure can also be used to distinguish between
Mountains and Fields, while the ‘density’ features distinguish
rural networks from urban networks.
A ‘distribution’ measure of edges at a vertex provides us with
information as to how the edges at a vertex are distributed in the
network. Let E D i be the proportion of junction points with i
edges at them. We use mean(£^ fi ) and var(£ D l ) as features. The
variance of the edge distribution is lower in the case of
networks in urban areas as opposed to rural, and it is lower also
in the case of urban networks in the USA as opposed to in
Europe.
Notation
Description
N,
Junction density
L
Network length
L
Length density
Ä
Network area density
Pe
Ratio of length
var(p)
Ratio of lengths variance
mean(p)
Ratio of lengths mean
K
Average curvature
var(k)
Average curvature variance
mean(k)
Average curvature mean
Ep,
Number of junction with m v =i
var(E Dli )
Edge distribution variance
mean(E D ,i)
Edge distribution mean
Ej
Number of junction edges
Ej
Junction edges density
Mj, a
Density of junction edges per quadrant
var( Mj)
Junction edges density variance
mean( Mj)
Junction edges density mean
K,
Local junction density
var(N ri )
Variance of the local junction densities
mean(N r ,)
Mean of the local junction densities
ßj
Vectors of angles between segments at junction
j
Jk
Entropy of road segment orientation
Table 1: Summary of the features computed from road networks
2.3 Urban Region Features
We focus on the last four features in Table 2. These features
enable us to distinguish between rural classes (Villages and
Fields) and urban class (Europe), which otherwise were
misclassified due to the lack of extracted network information
from the small compact urban regions in the images, shown in
Figure 3(a) and Figure 3(c). Let £2 and Q R be the area of the
image and the area of the extracted regions respectively and L\|/
and be the network length in ¥ = £2 - Q R and perimeter of
the extracted regions respectively.
We define two descriptors, R A , the extracted region density and
Cf A = fV'r/ the extracted region compactness factor. These
two features help us to distinguish the Villages class from the
rest of the classes: for example, R A ~ 1 for Urban classes and R A
~ 0 for Mountains and Fields classes.
The number of urban regions in an image, the feature R v , is used
to distinguish between complete Urban, Villages, Fields and
Mountains. A complete Urban (USA and Europe) will have R v
= 1, whereas, a Villages will have R v >1, and Fields and
Mountains will have R v = 0. Another feature A n = Cl R / L\|/, the
inverse fractional length density, is also computed to separate
the Village class from Urban and Mountains and Fields. For
complete Urban classes (USA and Europe), L\j/ = 0, and for
Mountains and Field classes = £2. Hence for Mountains and
Fields classes, A n = 0, while for complete Urban classes, A n = oo,
and for the Village class 0 < A n < oo.We augment these urban
region features with the features computed from the graph
representation of the road network as described earlier to
improve the classification of the geographical environments
which otherwise were misclassified due to the loss of
information from small dense urban regions.
Notation
Description
£2
Area of image
Q r
Area of extracted regions
Ly
Network length in 'F = £2 - £2 Ä
r*
Perimeter of extracted regions
Ra
Region area density : £2 fi /£2
Cf A
Region compactness factor T R 2 / £2*
R v
Number of regions : # R
Inverse fractional length density : £2 R /Ly
Table 2: Summary of features computed for urban areas.
3. CLASSIFICATION
The 32 features (16 features for each network extraction
method) described in section 2.2 were computed for a database
of 497 SPOT5, 5m resolution images. To provide ground truth,
these images were manually classified into the 7 classes
described in section 1 representing various kinds of urban and
rural environments. Machine classification was done with a
five-fold cross validation on the data set, with 80% of data for
training and the remaining 20% for testing in each fold. We
performed feature selection using a Fisher Linear Discriminant
(FLD) analysis (Duda et al., 2000), followed by a SVM linear
kernel classification on the selected feature set. The result of the
classification is shown in Table 3. The SVM linear kernel
classification on the 30-dimensional feature space selected by
the FLD shows a mean error of 24.5% with a standard deviation
of 2.92%. As can be clearly seen in the confusion matrix Table
3, the Villages class is confused with the Fields class and also
there is a slight confusion between the Urban USA and Urban
Europe classes. These confusions arise because, as stated above,
the road extraction methods fail to detect the fine and densely
structured roads present in some images. Table 4 shows the
results of classification of the same set of images with 20
selected feature out of 36 features (32 road network features
plus the 4 features computed from the segmented urban areas).
As can be seen, there is an improvement in the confusion matrix.
The Villages class is less confused with the Fields class than
before. The SVM linear kernel classification error is drastically
reduced from 24.5%, with only road network features to 12.9%,
with the combined feature set with a standard deviation of