Research Approach
The research approach used to establish the best technique to
detect and delineate post-hazard crack damage to critical infra-
structure was based on the iterative testing of multiple param-
eters related to image analysis. Pixel-based image processing
routines were developed using
ERDAS
Imagine
®
,
OBIA
rule sets
using eCognition, and accuracy layers created and evaluated
using ArcGIS
®
. Through the systematic investigation of chang-
ing variables related to the use of spatial filters, thresholding,
segmentation, edge detection, and
vector orientation, the optimal order
and combination of processes result-
ing in the most accurate classifica-
tion that simultaneously minimized
false detections of crack features was
determined. Multiple outputs were
generated with each methodological
approach using different combina-
tions of parameters and thresholds.
These outputs were tested for accura-
cy by comparing them to vector layers
created in ArcGIS showing the known
location of crack features placed in
the scenes by the research team.
To standardize the multi-temporal
images for analysis, several necessary
preprocessing steps were conducted
prior to change detection to minimize
radiometric and geometric differ-
ences between the images in each
set (Stow
et al.
, 2015). Images were
first automatically registered to one
another using the
SARA
point match-
ing software system. This software
generates tie points in each image and
applies a second-order polynomial
warping function to directly align the
images in the multitemporal dataset
to one other. Only as a result of the
RSI
capture technique minimizing the
differential parallax in the images are
these computationally simple meth-
ods of image to image registration
successful (Coulter and Stow, 2008;
Coulter
et al.
, 2012; Stow
et al.
, 2016).
Second-order warping functions cali-
brated with a small number (e.g., <10)
match points are more than capable
of achieving accurate registration (<2
pixels) when applied to
RSI
pairs, as
the view geometries of the two images
are nearly replicated (Stow
et al.
,
2016).
In this study, the time 2 (after)
image was used as the master im-
age, with the time 1 (before) image
registered to the master. Histogram
normalization was performed to align
image brightness values (BVs) be-
tween the different times of capture.
This normalization was based on the
detection of pixels having similar
shadow or non-shadow conditions
at both collection times. The bright-
ness values of one image were shifted
based on the mean and standard
deviation of the other image prior
to change detection. In addition to
histogram normalization for similarly illuminated pixels, a
novel shadow detection and re-brightening technique devel-
oped by Storey
et al.
(2017) was applied to pixels having tran-
sient shadows (i.e., shadows in one image but not the other)
to minimize the effect of shadowing on change detection.
Single-Date Approaches
To evaluate the effectiveness of methods using a single (post-
event) image as input, both supervised and unsupervised
Figure 1. Example image from each scene used in this study: Bridge Blvd.
SW
(top
left), Lake Murray 1 (top right), Lake Murray 2 (bottom left) and Camp Roberts
(bottom right).
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
February 2018
77
