using the ArcGIS Calculate Geometry function. Each road
segment in the Michigan Framework roads data layer has a
unique identifier known as the Linear Reference System Link
file (
LRS_LINK
). Using the
LRS_LINK
column in the attribute
table, the total length of each eCognition classified unpaved
road segment is calculated using the Summarize tool. The
result is a sum of the length of unpaved road segments along
each
LRS_LINK
segment. To calculate the percent coverage
value, the sum of the length of the unpaved segments is di-
vided by the length of the road in the corresponding
LRS_LINK
segment. The result is the percentage of the
LRS_LINK
segment
that is unpaved. The entire
LRS_LINK
road segment is classified
unpaved if a certain percentage of the total road length (called
“percent coverage”) is classified as unpaved. The paved/
unpaved classification is captured as a conditional statement
that results in a “Y” or “N” value in the “paved” column in
the attribute table. The percent unpaved rule was required to
account for tree or building shadows that can be cast on the
road as well as variability in spectral response/brightness that
can be caused by wet roads.
Preprocessing
The initial classification effort followed the lead of Nobrega
(2008) where a subset of an Ikonos scene was classified to find
the road network in an area where no maps of the roads ex-
isted. Where Nobrega’s work was limited to a relatively small
area with the goal of mapping the road network in the favelas
of Sao Paolo, Brazil, our study area included the unincorpo-
rated areas of six counties in southeastern Michigan, where
the extent of the total road network is already well known.
SEMCOG
supplied the project team with four-band (red/
green/blue/near-infrared) 30 cm
GSD
aerial imagery flown in
spring (leaf-off) 2010. Each image had a 1,524 m × 1,524 m
footprint. In order to reduce the total amount of aerial image
tiles that would need to be processed and therefore the total
amount of processing time, the aerial images were mosaicked
3,048 m × 3,048 m tiles.
Initially, the team followed Nobrega’s methodology and
processed several 1,524 m × 1,524 m tiles to locate linear
features that were spectrally similar to unpaved roads. While
this approach works well in a small heterogeneous study
area, in this case, image processing times were long and it
proved difficult to separate unpaved roads from other spec-
trally similar features which often were farm fields. To reduce
data processing time and improve classification accuracy, a
9.1 meter buffer around the road network was used to limit
the area to be processed. A visual inspection of Michigan’s
Framework roads layer (2011 version) superimposed over the
aerial photographs revealed significant location variability
between the roads represented in the Michigan Framework
Roads shapefile and the location of the same road in the aerial
imagery. Buffering the road network minimized the effect of
this misalignment and improved classification accuracy.
The road centerline polygons that resulted from the buffer-
ing process were dissolved into a single, large, county-wide
road centerline polygon feature. Processing only the buffered
road network not only significantly reduced image processing
time, it also excluded features that were spectrally similar to
unpaved roads but were not adjacent to any known road. The
buffer around the road network enabled better definition of
spectral responses specific to the different surface materials
used for roads in the study area.
Classification
The heterogeneous nature of the landscape, the spectral simi-
larity of unpaved roads to tilled farm fields particularly in
the spring when crops have not yet been planted, the similar-
ity of concrete roads to unpaved roads made with crushed
limestone and the frequent presence of tree canopy over the
roads to be evaluated, all presented significant challenges to
the classification process. Classification processes often use
spectral characteristics to assign pixels to a class. However, as
Dezso
et al
. (2012) notes: “Traditional-pixel-based classifica-
tion methods completely disregard spatial relations” which
are important to object-based classification.
To address these issues, eCognition was used in this
analysis to develop a multi-step image classification process,
shown in Figure 3 as a flowchart. Chessboard segmentation,
which creates sets of spectrally similar, contiguous pixels,
was run within the polygons that were created using the road
centerline shapefiles using pre-defined eCognition values for
input parameters. Quadtree segmentation segments the poly-
gon that contains the road into a grid based on spectral differ-
ences within the object and was assigned a scale value of 60
with image layer weights assigned to the red, green, blue, and
infrared bands. These processes run recursively until there
are no further significant changes in any resulting square. A
multi-resolution segmentation region grow process is then ap-
plied to combine spectrally similar areas into objects. Spectral
difference segmentation then merges objects according to a
user defined mean layer intensity value.
Figure 3. The eCognition unpaved versus paved road clas-
sification process.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2017
379
