crack objects and the minimization of false detections. As the
target cracks in the scene were created through the addition
of black tape to the road surface by members of the research
team, the positions of target features are known. Using Arc-
Map GIS software, an accuracy template was created through
heads-up digitization of crack features added to the scene.
Cracks from all images were first digitized as linear objects;
then a three-pixel buffer was created forming an epsilon enve-
lope to compensate for misregistration and slight differences
in the hand digitized layers (Chrisman, 1983). These polygons
were used as the accuracy template for single-date detection
methods. To isolate new cracks in multitemporal image pairs,
a symmetrical difference operation was performed between
the time 1 and time 2 accuracy layers, and the result of this
operation was compared using intersect with those polygons
present in the time 2 accuracy template. The final accuracy
template product for bi-temporal image pairs depicts only
new cracks, i.e., those present in time 2 and not in time 1.
Agreement and disagreement for accuracy assessment was
based on visual interpretation by determining the presence or
absence of detected cracks delineated through semi-automat-
ed image processing. If any part of a process-identified crack
present on a road or bridge surface intersected a known crack
feature in the digitized accuracy template, it was considered
correctly identified. Detected features on the road or bridge
surface that did not overlap with a known crack feature were
considered false detections. The ratio of successful detected
cracks to the total number of actual cracks in the scene (pro-
ducer’s accuracy) was calculated using Equation 2:
A
C
C
P
detected
actual
=
(2)
where
A
P
= producer’s accuracy,
C
detected
= number of cracks
successfully detected, and
C
actual
= number of cracks present in
the scene.
The ratio of the total number of detected cracks to the
number of correctly detected cracks (user’s accuracy) was
calculated using Equation 3:
A
CD
CD
U
actual
total
=
(3)
where
A
U
= user’s accuracy,
CD
actual
= number of actual cracks
detected (correct detections only), and
CD
total
= total number of
cracks detected (both correct and false detections).
The composite accuracy statistic is used to judge the most
successful overall method by considering both the producer’s
and user’s accuracy. This metric was calculated using Equation 4:
A
C
=
A
P
×
A
U
(4)
where
A
C
= composite accuracy,
A
P
= producer’s accuracy, and
A
U
= user’s accuracy.
Results
Accuracy of Single-Date Products
Single-date methods employed one post-event image as input
to the various crack delineation approaches, using both pixel-
and object-based methods. For the supervised classifications,
the two-class designation consists of: crack and all other
surfaces; three-class: crack, road and bare ground; four-class:
crack, road, bare ground and vegetation; five-class: crack,
road, bare ground, vegetation and concrete; and six-class:
crack, road, bare ground, vegetation, concrete and shadow.
Regardless of the number of classes used, the effectiveness of
the pixel-based single-date methods is poor. In many classifi-
cation products, nearly the entire road surface was classified
as crack, along with many medium to dark grey, grey brown
or dull green features in the scene. Such a high level of false
detection made reporting the number of falsely detected
objects an inappropriate measure of accuracy. While there are
few falsely detected crack objects, they occupy a large part of
the scene because of their large size. The pixel-based super-
vised classification method is generally able to detect cracks
in the scene, but unable to isolate them. This level of false
detection makes the products unusable. Figure 4 shows an
example of two crack classification outputs using the pixel-
based supervised classification that are typical of the level of
false detection observed.
The unsupervised single-date approach is more success-
ful than the supervised methods in avoiding false detections
in the scene, but still did not produce a useful product. With
a 38 percent producer’s accuracy and 8 percent user’s accu-
racy, the effectiveness of the unsupervised classification for
detecting cracks is low. Coupled with the high level of false
detection, the products generated by the unsupervised clas-
sification are not effective in the detection and delineation of
fine-scale cracks. The unsupervised approach was also not
easily automated, as the classes that most accurately repre-
sented actual cracks needed to be manually identified and
combined into the final crack classification product used for
testing. In this study, pixels in the two classes that comprised
the highest area of actual cracks were identified, spatially
aggregated, and clumps meeting the 20 pixel minimum size
criteria retained for accuracy testing. Testing revealed no
substantial difference between the classifications performed
using K-Means versus
ISODATA
clustering methods, nor based
on the difference in convergence values in the range tested.
While pixel-based supervised and unsupervised methods are
computationally simple and easy to implement, the limited
spectral information present in an 8-bit,
RGB
image does not
enable accurate classification. Without the inclusion of spatial
or contextual information to reduce the number of false detec-
tions, detection and delineation of cracks using pixel-based
supervised and unsupervised methods is unsuccessful.
Leveraging the power of geographic object-based image
analysis (
GEOBIA
) techniques to incorporate spatial, contextual
and multiple levels of location information resulted in a much
more accurate classification than was generated by the pixel
based, single-date methods. Based on the size of the targets to
be classified, a scale factor of 50 was used in the first stage seg-
mentation, and a scale factor of 15 is used in second stage seg-
mentation. The multi-resolution segmentation in eCognition
worked effectively and is the segmentation algorithm used for
all object-based approaches. Table 1 shows a total producer’s
accuracy of 86.2 percent and user’s accuracy of 47.6 percent
when summarized for all scenes. This indicates the
GEOBIA
method was quite successful in detecting crack features, yet
was limited by a substantial number of false detections.
As image texture was a significant metric used in the clas-
sification rulesets, several high texture, linear features such as
road lines and center dividers were falsely classified as cracks.
One of the most powerful tools created that limited the number
of false detections present in the classification is an exclusion
criterion based on the angular orientation of a detected feature.
Having the ability to exclude detections with an angular
orientation similar to that of the road surface as a whole is
extremely effective at reducing false detections caused by road
markings and dividers. With the ability to calculate the main
direction of the road or bridge surface and exclude objects
with a main direction that was similar (+/- 6°), many of these
false detections were eliminated. A shortcoming of this
GEOBIA
80
February 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
