detection and delineation methods were tested. Pixel-based
supervised methods were used to classify images into either
two, three, four, five, or six classes. These classifications were
based on training samples that depicted known crack damage
features present in the scene, as well as non-damage surfaces
such as vegetation, road surface, concrete, shadow, and bare
ground. Between three and six training samples of each class
were digitized and combined into a signature file used for
classification. Tests were conducted on samples created both
by region growing, as well as manual selection. Unsupervised
methods were tested using both K-Means and
ISODATA
clus-
tering routines that created 5, 10, 25, 50, or 100 classes and
required between 85 and 95 percent convergence to be classi-
fied. The maximum number of iterations was set to 10.
After classification, like-class outputs were converted into
pseudo-objects through pixel clumping (i.e., aggregation of
like-class, contiguous pixels) and sieving (i.e., removal of
pixel clumps smaller than a specified size) processes, and
those clumps meeting a 20-pixel minimum size criterion
were retained. The two classes created by the unsupervised
classification that comprise the highest area of actual crack
features present in the scene were retained and combined
into the final crack classification product tested. The clump
and sieve processes were used to standardize the outputs for
testing based on minimum size criteria also used in the bi-
temporal approaches. The minimum size for retention was set
at 20 pixels based on the size of the smallest target object (2 ×
10 pixel crack).
To test the relative effectiveness of techniques that perform
segmentation and classify true image objects, eCognition
OBIA
software was utilized to develop rulesets for the detection
and delineation of cracks. The workflow shown in Figure 2
depicts the two stages of segmentation and classification that
were performed. The first stage identified roads surfaces on
which cracks could be present. The classified road surface
was then re-segmented at a finer scale, and crack features
present are detected and delineated. Variables tested using
object-based methods related to segmentation type and scale,
spectral characteristics, location, and orientation of detected
objects in relation to the road surface. Due to the considerable
variation of surfaces present in the four test scenes, differ-
ent rulesets needed to be developed for each location. While
containing unique parameters and thresholds, each ruleset
followed the same two stage segmentation and classification
approach.
Bi-temporal Approaches
Evaluating the effectiveness of bi-temporal, change detec-
tion approaches to delineate new crack damage involved the
creation of raster models using
ERDAS
Imagine software, post-
classification comparison of
OBIA
classifications created in
eCognition; and both supervised and unsupervised classifica-
tions of multitemporal layer stack images. Multiple outputs
were generated using each method to test the effectiveness of
various combinations of techniques and thresholds.
With the
ERDAS
Imagine 2016 software, a crack feature
detection model was created using a pixel-based approach to
compare the bi-temporal test imagery. Figure 3 illustrates the
workflow utilized in this approach. At each stage of the mod-
el, various parameters and thresholds were tested, and the
most effective values retained. Testing then began on the next
parameter. The final accuracy tests were performed on the
fully tuned model with either the inclusion or exclusion of
key tools to gauge their effectiveness. The model workflow be-
gins by detecting and masking edge features
using a Sobel filter. This filter detects edges
in the time 1 image that are not of interest
in the change detection application, as they
cannot represent new damage. The mask
was applied to both image dates to avoid
false detections in these areas. Iterative test-
ing of the edge intensity magnitude used to
generate the mask, as well as the size of the
edges in the time 1 image that may repre-
sent preexisting damage were performed to
determine optimal values. Edge intensity
values between .05 and 2.0 standard devia-
tions from the mean and kernel sizes of 3 ×
3, 5 × 5, and 7 × 7 pixels were tested.
To compensate for small misregistration
effects, a kernel based focal minimum filter
is applied to the time 1 image. This filtering
technique minimizes the effects of slight
misregistration by comparing the BVs in the
time 2 image to the minimum value of BVs
contained in the kernel window of time 1.
When the
BV
of the time 2 pixel is less than
the minimum value contained in that focal
window, it is considered to represent a true
change. The focal minimum filter effectively
expands the footprint of features in time 1
so they can be compared accurately to the
time 2 image, even if slightly misregistered.
In an example where a narrow feature such
as a crack is misregistered by two pixels,
the change detection routine would display
the misregistered after position in the time
2 image as a change, when in fact it is an
artifact of misregistration and not an actual
change. The effectiveness of focal filtering at
various kernel sizes between 3 × 3 and 13 ×
13 was evaluated.
Figure 2. Object-based crack detection workflow. The first stage of
segmentation and classification identifies the road surface on which cracks
will be present. A scale factor of 50 is used for the primary segmentation.
Secondary segmentation operates on a finer scale factor of 15 and is only
applied to the road surface. After secondary segmentation, cracks are
delineated based on spatial and spectral information.
78
February 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
