advantage of lower cost when compared to higher spatial
resolution datasets.
In order to investigate the principles of using
HSR
multi-
spectral digital aerial photography to infer
ODR
, we performed
various linear regression models (Models 3 to 7) by using only
one visible band and by using only spectral features (mean
values) or texture features (standard deviation values). The
results were summarized in Table 5. The results revealed that
the visible red band best predicts
ODR
at all spatial resolutions
(e.g., 6-inch dataset R
2
>93% and
RMSE
<28), while visible
blue band predicts
ODR
at the lowest certainty. Table 5 also
revealed that when compared to texture features, spectral
features predict
ODR
at a higher certainty (e.g., 6-inch dataset
R
2
>92% and
RMSE
<29).
Results revealed that the regression model that uses all six
principal components exhibited the best capability to predict
ODR
. Therefore, this model was selected for validation. Table 6
shows the results of the 6-inch, 12-inch, and 24-inch regression
models when using only half of the study sites for calibra-
tion (sample size =25). It shows that the R-squared value is
decreased while the
RMSE
is increased for all three models, but
not substantially. All three models are still valid at a 95 percent
confidence interval (joint P-value (Prob >F) is less than 0.001).
The other 25 study sites were used to independently
validate predicted (model-generated)
ODR
values versus actual
(ground reference)
ODR
values, the
RMSE
, mean absolute er-
ror, and standard error of which are shown in Table 7. Not
surprisingly, the
RMSE
for each model is higher when vali-
dated using holdout samples and predicted using the smaller
sample size of 25 to develop the model, but not substantially.
In addition, the mean absolute error and standard error are in-
creased when the resolution becomes coarser, but all are less
than an error of 84 that manual evaluation can exhibit
.
T
able
7. E
rror
S
ummary
for
P
redicted
ODR
Dataset (Size: 25) RMSE Mean Absolute Error Standard Error
6-inch
42.8826
35.0000
5.0577
12-inch 63.1958
43.7600
9.3070
24-inch 72.5551
66.2400
12.0770
Validation results, consistent with model fits, show that
the 6-inch aerial photography results in the lowest error when
compared to manual evaluation results, whether measured by
RMSE
, mean absolute error, or standard error. Therefore, we
conclude that
ODR
can be most effectively predicted by the
6-inch aerial photography. While none of the models can be
used to detect detailed distress (e.g., cracks) or vertical dis-
tress (e.g., rutting), all models indicate potential for the direct
estimation of
ODR
with less error than manual approaches
.
One limitation of the proposed method is that it cannot
be used for high traffic volume sections. This is because
vehicles are considered as unwanted features on the pave-
ment. Too many vehicles present in the images could reduce
the area of pavement observed to such a degree that distress
cannot be accurately evaluated. This proposed method also
must use reference pavement surface distress rates (collected
either through manual evaluation or automatic evaluation) to
develop initial model calibrations.
Conclusions
Routine evaluation of pavement surface condition is a chal-
lenge to all transportation agencies. In the real world, it is
impossible to get exhaustive condition data for all pavement
surfaces. Current methods for pavement surface distress
T
able
6. M
odel
V
alidation
for
P
rediction
of
ODR V
alues
Dataset (Size: 25)
Variables
Coefficient
Standard Error
t
P>|t|
R
2
Adjusted R
2
RMSE Prob > F
6-inch
PC
A1
46.03
3.22
14.29 <0.001*
0.9232 0.8976
37.206 <0.001*
PC
A2
-13.17
11.74
-1.12 0.277
PC
A3
-102.89
61.00
-1.69 0.109
PC
A4
264.22
129.98
2.03
0.057
PC
A5
-43.32
152.78
-0.28 0.780
PC
A6
-539.59
272.10
-1.98 0.063
Intercept
137.28
7.44
18.45 <0.001*
12-inch
PC
A1
45.33
5.12
8.85 <0.001*
0.8178 0.7571
57.309 <0.001*
PC
A2
-4.55
13.65
-0.33 0.743
PC
A3
42.88
59.73
0.72
0.482
PC
A4
-175.53
143.74
-1.22 0.238
PC
A5
50.69
241.64
0.21
0.836
PC
A6
187.02
335.35
0.56
0.584
Intercept
137.28
11.46
11.98 <0.001*
24-inch
PC
A1
33.06
7.73
4.28 <0.001*
0.7166 0.6222
71.476 <0.001*
PC
A2
-42.78
9.45
-4.53 <0.001*
PC
A3
125.42
76.79
1.63
0.120
PC
A4
-231.38
118.32
-1.96 0.066
PC
A5
-88.60
300.07
-0.30 0.771
PC
A6
194.58
486.38
0.40
0.694
Intercept
137.28
14.30
9.60 <0.001*
Note:
PC
A1
and
PC
A6
indicate the six principal components extracted from the mean and standard deviation values of each of the three visible
bands; RMSE indicates root mean squared error; and * indicates the independent variable is significant at p = 0.05 level.
718
September 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
