values are just above the cutoff of six that is used to classify a
road as unpaved. Typically, paved roads have mean IR-Green
values that are negative or slightly positive.
Occasionally, paved roads in developed areas will be clas-
sified as unpaved as a result of IR-Green values in excess of
the threshold of six. Errors of omission (where the algorithm
classified roads as paved and the
PASER
data did not) occurred
most frequently as a result of road centerline misalignment
or unpaved roads where the IR-Green value was negative,
which is more typical of a paved road. The phenomenon of an
unpaved road having a strong spectral resemblance to a paved
road may be a result of local road commissions using crushed
limestone, a major component of both concrete and macadam
pavement, for the road.
Available Oakland County
PASER
data is not as comprehen-
sive as Monroe County data and could not be used directly
as a complete ground reference data set. The Michigan
Framework roads layer for Oakland County contains a total
of 12,331 kilometers of roads, although not all are the respon-
sibility of the Road Commission for Oakland County (
RCOC
).
A statement on the
RCOC
website says that “More than 750
of the 2,700-plus miles of the Road Commission for Oakland
County’s (
RCOC
) county roads are not paved…” (
RCOC
, 2013).
Error matrices as described by Story and Congalton (1986)
and Lunetta
et al
. (1991) provide a well understood statistical
methodology to assess the performance of the classification
algorithm. Error matrices for each county were calculated as
part of the accuracy assessment process. Matrices were calcu-
lated for different coverage percentage values, and generally,
the coverage value that had the best overall accuracy was
chosen to represent the roads in that particular county. Cover-
age values varied from one county to the next as a result of
differences in geography, i.e., some areas had significant tree
cover over the roads, limiting the view of the roads and mak-
ing classification less accurate; others were more open, which
generally improved classification accuracy.
Processing challenges have primarily been the variable
road network centerline accuracy which is most noticeable
when displayed over high spatial resolution aerial imagery.
Some road centerlines align very closely to their associated
feature in the high resolution aerial imagery while for fairly
extensive distances, others within the same roads dataset can
be misregistered by as much as 5 meters. This may be a func-
tion of the scale at which the roads are digitized; however
the state-led Framework roads effort is working to improve
centerline accuracy.
Spectral similarities in the four band aerial imagery
between roads and non-road features are another source of
potential error in the classification process. Roads built with
concrete, old macadam pavement, and unpaved roads made
with crushed limestone (which is a component of both) ap-
pear to be spectrally very similar, which can lead to misclas-
sification in both directions. Another potential source of error
is bare soil and natural aggregate (such as locally sourced
river sand and gravel), which also are very spectrally similar.
This becomes less of a problem when the classification is con-
strained to a known road network and a small buffered area
around the roads, as was done for this effort.
Shadows that obscure the road, particularly where there is
extensive forest cover, make it difficult to classify roads that
pass under the canopy. This is a well-known challenge for
remote sensing processes where forest cover limits surface
visibility. One solution to this issue is to follow Nobrega
et
al
. (2006) and mask out the shadows to increase the ability
to only detect those areas of bare soil and therefore eliminate
shadows as a “land cover” class and its impact on overall
accuracy. However, the project team used the “percent cover-
age” rule to address this problem, whereby only a certain
percentage of a road segments needed to be called unpaved
for the entire segment to be labeled as such. An additional so-
lution could potentially include examining the other defined
classes along the road to determine if trees or shadows are
mixed with defined unpaved road segments, therefore making
the road underneath the shadow more likely to be unpaved
(or vice versa).
Conclusions
Knowing the location, length, and condition of unpaved roads
in a regional road network is important to transportation
agencies that need to cost-effectively manage their roads. Lim-
ited budgets and resources add to the maintenance challenges
faced by regional, county and local road commissions. This
paper outlines a methodology to identify unpaved roads in a
local road network using high spatial resolution (30 cm pixel)
four band imagery. The tiled four band imagery is loaded into
eCognition along with a 9.1 meter buffer polygon derived
from county road centerline data.
The imagery is segmented and classified to identify
unpaved roads in the dataset. It was found that the value
resulting from subtracting mean IR values from the mean
Green values provided a useful method for separating paved
roads from unpaved roads. A
ROC
curve was applied to find
the optimal Infra-red-Green (IR-Green) threshold value for
unpaved road detection. An IR - Green value of six provided
the best compromise for maximizing the true positive clas-
sification results while minimizing false positives. The result
of the classification process is a shapefile containing unpaved
road polygons.
The shapefile output from the eCognition classification
process form the basis of identifying unpaved roads in the
road network. The unpaved shapefiles are imported into Arc-
GIS and intersected with the road network, creating a shape-
file that is the linear segments of the road network that are
considered to be unpaved. Each road segment in this shape-
file is compared to the overall length of the original segment;
if more a preset percentage of the segment was classified as
unpaved, then the entire segment is classified as unpaved.
The comparison process is run on all the roads in a county
and the results compared to a
PASER
ground truth dataset
shared by the project partners at
SEMCOG
. The project team’s
classification at 25 percent coverage found 620.6 kilometers
of unpaved roads in the Monroe County network, compared
to the
PASER
data which reported 629.2 kilometers of unpaved
roads. Oakland County had significantly more road centerline
length than Monroe County but a less complete
PASER
dataset.
When run using the same methodology as Monroe County,
the classification found 1,116.8 kilometers of unpaved road
in Oakland County, which estimates there are approximately
1,207 kilometers of unpaved roads. The classification results
were used as mission planning input for a spring and sum-
mer 2013 field campaign in Southeastern Michigan to assess
unpaved road condition from an Unmanned Aerial Vehicle
and manned fixed-wing aircraft. The classified data are used
as input into the larger “Characterization of Unpaved Road
Conditions through the Use of Remote Sensing” project that
needs to know where the unpaved roads are located before
data collection missions can be flown. The unpaved versus
paved mapping results have been shared with
SEMCOG
and
other project partners as well providing additional access to
information pertaining to where unpaved roads are located
in their regions of interest. Using an automated classification
methodology allows
SEMCOG
(or other similar groups) to now
have access to automated road type classification informa-
tion updated regularly upon collection of new aerial imagery,
aiding in the determination of which roads have changed type
(i.e. unpaved to paved or vice versa).
382
May 2017
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
