registration (or co-alignment), which was performed
vis a vis
the “image translation, scale, and rotation”, alignment tool
in Adobe Photoshop
®
CC. Difference images were generated
from these registered pairs by subtracting the per band
DN
and
intensity (of the hue-saturation-intensity transform) values of
the pre-damage image from those of the post-damage image.
Dynamic Range Assessment
Dynamic range was assessed for each waveband, for all 19
EV
levels captured per imagery collection, and for two different
light metering
zonal calculation methods. The optimal
EV
and light metering method, were assessed as the level which
produced the largest
DN
value range, the lowest
RMSD
value,
and the lowest pixel count of truncated 0 or 255 value DNs
for each band. The “lowest” and “highest”
DN
value bins had
to contain at least 10 pixels to avoid counting random pixel
noise that can artificially expand the dynamic range. The
dynamic range of the different
EV
0 images was compared to
determine if the center, or overall weighted
light metering
method consistently produces images with greater dynamic
range. This test was performed with both the 8-bit and 12-bit
RAW
images, to determine if the optimal
EV
level for 8-bit im-
ages, is the same
EV
level for 12-bit images.
Intra Frame Brightness Variation Assessment
The effects of white balance were assessed with the
AIM
of de-
termining the optimal camera setting or processing method, to
achieve temporal consistency between bi-temporal image pairs.
Temporal inconsistency across dates, as a result of uninten-
tional variations in
WB
color temperature, could result in a bias
or offset of
DN
values in image differencing products. The dif-
ference between the
AWB
color temperature for each collection,
was used to establish the typical degree of variation, between
the
WB
on different dates. The minimum and maximum color
temperature values from all images collected, to determine if
fixing the color temperature offers substantial benefit.
The influence of relative aperture settings on vignett-
ing was assessed. This involved stacking a large number of
images of a uniform reflectance target captured for multiple
relative aperture settings, and extracting diagonal pixel
DN
profiles to estimate radial brightness trends that represent the
vignetting effect. The brightness trend images can be used
to create an anti-vignetting correction mask, as an alterna-
tive to the correction masks provided by the camera and lens
manufacturers (Stow
et al
., 1996). Radial trends were com-
pared between pixel profiles generated with the four different
internal camera vignetting compensation settings (none, low,
medium, and high).
Image Visual Acuity Assessment
The influence of
ISO
setting on image acuity was visually
evaluated, in order to determine at what value threshold im-
age acuity begins to decline. This was accomplished by taking
images with different
ISO
values, randomizing them, and then
trying to determine which images are visually noisier than the
image captured with
ISO
100.
Temporal Consistency
Magnitudes of noise in bi-temporal pairs were assessed for
each band, for all 19
EV
levels captured, per imagery collec-
tion, and for two different
light metering
zonal calculation
methods. For both the comparison of the 19
EV
levels and
light meter weighting methods, red, green, blue, and intensity
RMSD
values were derived. The image settings which yielded
the lowest calculated
RMSD
for each band represented the
optimal camera settings.
For seven of the 19 EVs evaluated,
light metering perfor-
mance was assessed by comparing how consistently the
APEX
value recorded in the image
EXIF
metadata as “Brightness Value”
and
EV
offsets, maintain parity with each other for the center
weighted metering, compared with the seven images captured
using overall weighted metering. The two zonal calculation
methods are referred to by the camera’s manufacturer (Sony), as
“multi-segment” and “center weighted”, both methods utilize
different proprietary formulas, and as such, the computation
differences between them is unknown. Whichever metering
method had
APEX
values that more consistently aligned with
the
EV
values was deemed the more consistent
light metering
method, and the preferable camera settings choice.
Simulated Damage Assessment
Crack detection was assessed by analyzing pixel transects/
profiles across the simulated crack features and quantitatively
by generating a type of signal-to-noise ratio (
SNR
) metric. The
signal was quantified as the spatial gradient (forward difference
operator) of bi-temporal image difference values calculated for
the pixel transects crossing the crack features. Noise was quan-
tified as the
RMSD
of bi-temporal image difference values for a
subset of the background surface surrounding but not including
the cracks. The
SNR
value is the quotient of the signal divided
by the noise metric for bi-temporal subset pairs for each
EV
trial.
Results
Dynamic Range Assessment
Dynamic range was assessed for 133 pairs of images in 8-bit
and 12-bit radiometric resolution, all captured on the same
time of day, spanning two months and three locations, and
for 19EV settings in
⅓
increments between -3EV and +3EV.
Dynamic range values for the 12-bit
RAW
images were re-
corded as 16-bit containers, meaning that the values should
have filled histogram bins contiguously, taking up 4,096 of
the 65,536 total bins. However, the
RAW
file format used by
the camera (Sony ARW format) automatically applied lossy
compression to all of the images, which had the effect of non-
contiguously spreading
DN
values from the 12-bit source data,
across the full 16-bit container value range, leaving empty
DN
bins between ones containing data.
All 133 images contain pixels with the maximum (max)
value (65535), while 49 images also have pixels with the min
value (0). Images that have pixels in both the max and min
bins, do not necessarily represent the optimal dynamic range,
as many of those images were highly underexposed and had
the majority of pixels spread across only a handful of bins.
For this reason, the standard deviation of both the 8-bit and
12-bit images were calculated in order to contextualize the
dynamic range differences.
A plot of
DN
values for 12-bit images, shown Figure 4a, illus-
trates that the highest standard deviation of intensity values for
the entire image, occurs around
EV
0, with the images between
EVs −3 and −2/3, having maximum ranges. Visually, images
between
EV
−1 to
EV
0 exhibit the highest degree of acuity, with
this range of values achieving a good balance between maxi-
mized dynamic range and a wide distribution of
DN
values.
The 8-bit images exhibit a trend similar to the 16-bit image
tests, and are visually indistinguishable from their 12-bit coun-
terparts. A 2
nd
degree polynomial trend line was used, because
the trend was parabolic, with the
r
2
included to highlight the
difference between the 8-bit and 12-bit trends. As seen in Fig-
ure 4b,
EV
−1 to
EV
0, as with the 12-bit images, seem to be the
ideal range for achieving both a maximized dynamic range, a
wide distribution of
DN
values, and the highest degree of acuity.
Intra-frame Brightness Variation Assessment - White Balance Influence
Across 20 collections on 10 dates, at three sites, over three
months, the color temperature range was 5150K to 5300K
utilizing the onboard camera
AWB
. This variation of 150K
will cause minor hue, saturation, and intensity fluctuations
between dates, if a set value is not fixed onboard the camera,
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
March 2018
153
