intervals, between −3 and +3, for a total of 19 different EVs
captured (see Table 1).
WB
variable images were captured
utilizing the camera presets listed in Table 1 at the EVs of 0,
−1 and +1, for a total of 27 images. Variable
light metering
tests were conducted using the “center weighted” light meter
setting, at the EVs listed in Table 1, for a total of seven images.
The variable
ISO
tests were carried out at the EVs of −1, 0, and
+1, using the
ISO
values listed in Table 1, for a total of 18 im-
ages. The variable aperture tests were conducted using 19 dif-
ferent apertures ranging from
f/
2 to
f/
16, utilizing a different
camera system from all the other tests, with
ISO
= 50,
f
= 105
mm, four different degrees of onboard vignetting control, and
EVs of −
⅓
,
⅔
, 0, and +
⅔
, for a total of 228 images.
Two
DSLR
camera systems were used in this study, a Nikon
D800E full frame camera paired with a Nikon DC-Nikkor 105
mm
f/
2 to
f/
16 relative aperture lens, which was used for the
variable aperture study, and a Sony Alpha 65 APS-C camera
(Bockaert, 2006) paired with a Sony
f
= 18-55 mm
f/
3.5 to 5.6
relative aperture lens, which was used for the variable
EV
,
WB
,
light meter, and
ISO
studies. The Sony a65, shown setup for
our experimental imaging in Figure 3, would have been the
only camera utilized for this study, however the widest aper-
ture usable to capture the 55 mm images,
f/
5.6, is an aperture
at which vignetting effects are negligible. For this reason,
the Nikon D800E with a wider
f/
2 aperture lens with more
pronounced vignetting effects was utilized for the aperture
testing (DXOMARK, 2012).
Table 1. Experimental camera exposure variables and their
settings.
EVs
(19)
White
Balance (9)
Light Metering
EVs (7)
ISO
(6)
Relative
aperture (19)
-3
AWB
-3
100
f/
2
-2
⅔
2500K
-2
200
f/
2.2
-2
⅓
3500K
-1
400
f/
2.5
-2
4500K
0
800
f/
2.8
-1
5500K
+1
1600
f/
3.2
0
6500K
+2
3200
f/
3.5
+1
7500K
+3
f/
9
+2
8500K
f/
10
+2
⅓
9500K
f/
11
+2
⅔
f/
13
+3
f/
14
f/
16
The Nikon D800E was used to test across-image variability
during a single collection, utilizing a reflectance calibration
panel as a surface of uniform brightness. The reflectance
panel was positioned perpendicular to the sun, with the
camera positioned as close to perpendicular with the sun and
reflectance panel as possible, without having the shadow of
the camera in the resulting image. Images were captured in
the mid-afternoon, to lessen the chances of accidental overex-
posure when capturing images with a high
f
-stop.
The Sony camera captured four image sets for two different
scenes in support of dynamic range and spatial acuity tests
with variable
EV
,
WB
, light metering, and
ISO
. Two collections
were carried out in the afternoon with a solar elevation of
45°, and two collections were carried out in the late afternoon
with a solar elevation of 30°, to achieve variations in illumi-
nation conditions and degrees of shadowing. Images were
captured on clear and hazy days, and yielding image data sets
influenced by two differing atmospheric optical conditions.
The sites chosen were selected to capture differing urban
scene compositions, and to achieve an image viewing azimuth
angle parallel to the solar azimuth at time of collection.
To mimic airborne imagery collection from a terrestrial
location, we attempted to approximate an equivalent path
radiance to match a range of altitudes (750 to 1500 m) above
ground level (
AGL
), that we commonly use as platform alti-
tudes for our simulated damage assessment research. Optical
depth increases horizontally along with the zenith angle, from
a factor of one at zenith angle 0°, to around 40° at zenith angle
90° (Allen, 1973), and vertically, with a zenith angle of 0°, the
magnitude of an object at the top of the atmosphere decreases
by 0.28 at sea level, 0.24 at 500 m, and 0.21 at 1,000 m above
sea level, to 0 at around 100 km (Green, 1992). With 25
percent of atmospheric scattering due to atmosphere located
between sea level and 1,000 m, this would mean a horizontal
image at 1,000 m would have 25 percent less path radiance
than the same image captured at sea level. Given both the
non-linear and temporally varying nature of optical depth, a
25 percent reduction in path radiance was an estimate used
for this study. We simulated vertical imaging altitudes of 750
m and 1,500 m
AGL
, such that the 565 m to 1,125 m in the
horizontal plane, above ground level.
In addition to images datasets generated for the change
detection noise minimization tests (described above), an-
other set was created that simulated infrastructure damage.
Simulated cracks were used to test the sensitivity of different
capture settings in detecting cracks in a wall or road as prox-
ies for infrastructure damage. Image pairs taken of simulated
cracks simulated were analyzed to determine the settings
for which the smallest crack can be resolved from difference
images and to examine the influence of
EV
settings on crack
detection. To assess the ability to detect damage signals, simu-
lated cracks made out of black painter’s tape were placed on
a concrete wall with widths approximately equal to one, two,
three, and four pixels (
GSD
= 4.5 cm). Vertical cracks were cre-
ated with varying widths, equal to one, two, three, and four
times the estimated image ground sampling distance (
GSD
),
and diagonal cracks were created equal to one, two, and four
times the estimated
GSD
, using matte black tape. An oblique
imagery collection was performed with the Sony camera
mounted on a tripod from the top of building 42 m vertically,
960 m horizontally distant from the concrete wall. A second
collection (repetition of
EV
settings) made after the tape had
been removed. Image pairs for each
EV
setting were regis-
tered and then subset to the extent covering the simulated
cracks and their surrounding background (concrete wall). The
images for this test were collected with only a single
WB
of
AWB
, compared to the nine
WB
settings captured for the noise
minimization test imagery, all other settings utilizing the con-
trolled constants from the noise minimization image set. The
ideal
EV
level is that which enables detection of the smallest
simulated cracks, with the greatest temporal brightness differ-
ence across the simulated crack features.
Data Analysis
Automated change detection products are typically based on
image differencing of one or more wavebands or transform im-
ages for a
RSI
pair (in our case), creating an output a raster with
band values representing the magnitude and sign of
DN
value
changes. The most appropriate (if not optimal) imagery collec-
tion settings were those combination of settings which produce
images with the lowest root mean square difference (
RMSD
) val-
ues for DNs in the difference images, for test images or subsets
where scenes contained few actual land surface changes.
Prior to image analyses, test images were downloaded,
sorted, and assessed for quality, for both the noise minimiza-
tion and simulated crack signal maximization assessments.
Test image pairs used for dynamic range and
light metering,
vignetting, and simulated crack signal assessments, required
152
March 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
