jidable difference in
lumination or even
1c differences in the
ngly, in the corres-
compensated as part
ghtness and contrast
with the heights re-
ymetric observation
N(CDN+b) (4)
own parameters: the
quired contrast and
additional equations
on the Xoffset and
ned by equation (3).
oth (3) and (4); the
‘) impact on the re-
d Intensity is based
operly, the relation
> and after the ad-
ntensity observation
hts for the intensity
oretically, needs to
tation. However, it
g with the overall
to level largely dif-
ding on terrain and
ors between 1 and
regions where the
exclusively deter-
Analysis input, the
etric matching are
of all, the roles of
ween reference and
ont offset computa-
The average offset
indicators are mea-
of point/plane pairs
ons as well as the
| limits can depend
errain and imaging
t general settings —
S, based on at least
ber of input image
rrect results; but as
d in a fairly dense
he Shear Analysis.
sitives.
DATA SETS
cloud matching ap-
data — heights and
; — and artificially
id Shear Analysis
f ADS blocks with
for the replacement
the verification of
automatically derived offsets against those human measure-
ments. The comparative analysis of the geometric and the com-
bined matching approaches is documented below.
3.1 Data Sets
The ADS test data used here have been captured and processed
by North West’s production. In the context of Shear Analysis
verification, strip offsets were automatically derived from a
very dense, practically continuous pattern of info cloud patches
of 512 x 512 image pixels in size (Table 1 and Figure 2); a re-
presentative number of manual QC measurements is available.
3.1.1 Georgian Bay: Located on the coast of Lake Huron's
Georgian Bay in Ontario, Canada, this block is dominated by
forest. It is in parts dense but generally includes clearings and
aisles, and features different tree species of various heights. The
imagery has been captured in 2009 for the Ontario Ministry of
Natural Resources (OMNR) for forest inventory; it is a typical
forest data set.
3.1.2 Lansing: This block, captured in fall 2011, shows the
City of Lansing, Michigan, approximately in its center. Accor-
dingly, the data contains predominantly urban and suburban
areas as well as some forest, fields and smaller lakes.
This block is analyzed in more detail by Gehrke et al. (2012),
comparing different georeferencing and also demonstrating the
possibilities of Shear Analysis. See also section 5.
3.1.3 New Mexico: This 2011 data set is part of the National
Agriculture Imagery Program (NAIP). From a very large ADS
block in South-Eastern New Mexico, a single strip overlap was
selected for this investigation. It features mountains and flat
desert areas with only little vegetation (Figure 2).
Data Set CS Terrain Strips | Patches
Georgian Bay 0.30 | Forest, Water 4 113
Lansing 0.30 | (Sub-)Urban 14 808
New Mexico 1.00 | Mountains 2 378
Table 1. ADS data sets used for verification of the point cloud
matching.
3.2 Accuracy in Comparison with Manual Measurements
One way of verifying the automatically derived ADS strip off-
sets is their comparison with manual QC measurements, which
are available for all ADS blocks in North West production. This
comparison was carried out for all patch locations that feature
corresponding measurements, provided that successful and
reliable offsets from both solely geometric and combined geo-
metric/radiometric point cloud matching exist. (See below for
the success rates of both methods.) Resulting averages and
standard deviations of the X, Y and Z differences between
manually measured and automatically derived strip offsets are
shown in Table 2. Note that, even though offset locations on
ground are practically identical, the orientation parameters used
in their computations differ: Human measurements are naturally
based on a stereo pair — ADS backward and forward bands in
this case —, but the SGM for the automatic method utilizes all
three ADS views to increase redundancy. Especially after aerial
triangulation, the impact of remaining orientation errors is ex-
pected to be very small but, nevertheless, can act systematically
for this comparison.
105
Combined Matching
Data /
Axis Average Std. Dev. Average Std. Dev.
[GSD] [GSD] [GSD] [GSD]
Geometric Matching
Georgian Bay, 23 Points/Patches
X 0:37 + 0.09 0.43 0.34 + 0.08 0.36
Y 0.37 + 0.08 0.38 0.14 + 0.08 0.37
Z 0.07 3: 0.12 0.59 0.06 +£0.12 0.59
Lansing, 33 Points/Patches
0.01 £0.08 0.46
0.05 + 0.04 0.23
0.05 + 0.06 0:32
X 0.17 + 0.09 0.53
Y 0.05 + 0.05 0.30
Z 0.05 + 0.06 0.33
New Mexico, 28 Points/Patches
X 0.43 + 0.16 0.85
Y 0.22 + 0.08 0.42
0.29 + 0.09 0.45
0.20 + 0.06 0.34
Z 0.30 + 0.07 0.39 0.30 + 0.07 0.39
Table 2. Averages and standard deviations of differences be-
tween manually measured and automatically computed ADS
strip offsets.
Looking at the offset differences and standard deviations in
Table 2, it can be seen that the combined geometric/radiometric
point cloud matching agrees better with manual measurements
than the geometric matching. As expected, the consideration of
intensity improves planimetric offset components for all data
sets. The majority of the average differences is not significant;
however, the highest significance level occurs in the forest data
set for geometric matching. This can be assigned, at least partly,
to the limitations of the approach, but the above-mentioned ori-
entation differences might add to that. The largest planimetric
discrepancies tend to occur across flight direction. This corres-
ponds with the general observation of North West’s production
that, after aerial triangulation, remaining orientation inaccura-
cies and, therefore, local strip offsets are largest in this direction
(but well within customer specifications); see also Gehrke et al.
(2012):
Manual | Geometric | Combined
Data Set / Axis Results Matching | Matching
[GSD] [GSD] [GSD]
Georgian Bay
X Flight Dir. 0.26 0.34 0.18
Y 0.48 0.44 0.33
Z 0.53 0.22 0.22
Lansing
X 0.45 0.40
Y Flight Dir. n/a 0.26 0.13
Z 0.15 0.15
New Mexico
X 0.29 0.72 0.29
Y Flight Dir. 0.32 0.34 0.26
Z 0.54 0.33 0.33
Table 3. Offset standard deviations throughout strip overlaps,
for Lansing and Georgian Bay RMS values based on all over-
laps. The number of manual measurements per overlap in Lan-
sing varies between 3 and 5, which is not representative for the
derivation of reliable statistics.