conforms to the finding of Nagel
et al
. (2014) who mapped
LULC
and wetland across Minnesota and Wisconsin using
NAIP
images. The accuracies could be improved if the two
classes are combined (Yang and Zhou, 2011). Additionally,
using multi-temporal data may produce higher classification
accuracy (Coppin and Bauer, 1994; Wolter
et al
., 1995; Reese
et al
., 2002; Yuan
et al
., 2005). However, high-resolution
multi-seasonal or multi-source images over large area are still
impractical due to the high cost of data collection.
Comparatively, the impervious surface classification has a
high overall accuracy. However, it only included two classes,
while the
LULC
classification extracted six classes. More classes
in a classification model results in more chances for confu-
sion between classes and therefore a higher degree of misclas-
sification. In addition, while Feature Analyst appears to make
use of advanced techniques, much of its processing happens
behind the scenes. In contrast, methods used in See5 are well
described in the literature and the analyst has more influence
over the software. The original idea of this study was to classify
the general
LULC
map using the object-based classifier and then
use the
LULC
map, along with the texture layer and the lidar
DEM
, as additional attributes in the decision tree analysis for
impervious surface extraction. However, after the winnowing
and pruning analyses, only bands 1 (B), 3 (R), and 4 (
NIR
) from
the
NAIP
image were chosen in the final decision tree. Winnow-
ing and pruning removes insignificant attributes and branches
from the decision tree model while improving processing
efficiency and accuracy. A reason that the texture layers were
not selected by the decision tree classifier was because, unlike
other land covers with unique textures such as water, forest, or
cropland, the textures of urban impervious surfaces are diverse,
which might not contribute to the classification significantly,
but could make the decision tree more complex. The study site
is flat without much elevation change. Consequently, the
DEM
layer was not chosen by the decision tree classifier either. A
DEM
coupled with a digital surface model (
DSM
) might be more
useful in classifying urban impervious and other
LULC
features.
The final decision tree for impervious surface extraction did
not use the general
LULC
map, which might be attributed to
the fact that the decision tree analysis was applied to pixels
whereas the
LULC
map was based on image objects. In addition,
the structure of the final decision tree in this study was quite
simple, which may explains why the boosting function did
not increase much accuracy of the results. Furthermore, in a
previous study (Yuan, 2008), it was found that high-resolution
texture derivatives were critical for classifying general
LULC
from historical black and white aerial images. In the future, a
sensitivity analysis can be performed to explore how the inclu-
sion of texture and elevation will affect the accuracy of general
LULC
classification from three-band or four-band
NAIP
imagery.
The results showed that the classified impervious surfaces
are similar to those of urban and bare soil from the
LULC
clas-
sification. Likewise, the amounts of areas mapped by the two
approaches are close to each other. The user’s and producer’s
accuracies of the impervious surface class are also comparable
to those of the urban and bare soil classes from the
LULC
clas-
sification. Hence, these results indicate both techniques can be
used to extract impervious surfaces from the
NAIP
imagery with
relatively high accuracy. The impervious surface map can be de-
rived alternatively by combining the urban and bare soil classes
from the
LULC
map generated by the object-based classifier.
Implications for Future Classification Studies
Any study that relies on remotely sensed data has inherent
uncertainties regarding the accuracy of the final product,
dependent on the characteristics and quality of the data used
as well as the specific approaches implemented. A limita-
tion of this study was that neither the training data nor the
testing samples could be collected in the field. While the
highest possible diligence was used in correctly identifying
these samples, some amount of uncertainty could still be
introduced. In addition, due to the limitation of resources,
we only employed the 1 m lidar
DEM
product in the clas-
sifications. Nevertheless, this study demonstrated that even
though handling the large amounts of
NAIP
and lidar
DEM
data
was challenging, high-resolution classification maps could be
extracted successfully using the proposed techniques.
In fact, high-resolution
NAIP
imagery with a
NIR
band cover-
ing large areas has only recently become available. When
the focus of a study is a relatively small area, or urban and
suburban areas, the use of this type of data is beneficial for
measuring impervious surfaces and patches of specific forest
assemblages, such as deciduous and hardwoods. However,
additional spectral bands beyond the
NIR
region are required
to differentiate various vegetation classes, such as cultivated
cropland from grassland and forest, accurately. The medium-
resolution Landsat imagery has more spectral bands and
higher temporal resolution than the
NAIP
imagery. Therefore,
Landsat images are suitable for regional studies that involve
multi-temporal
LULC
classification or wetland mapping (Lu-
netta and Balogh, 1999; Ozesmi and Bauer, 2002; Yuan
et al
.,
2005). Despite its cost, high-resolution satellite images, espe-
cially those that have spectral bands beyond the
NIR
region,
such as red-edge and middle infrared bands, can be used as
an alternative to produce high-resolution
LULC
maps.
Conclusions
In this study, two sets of high-resolution classification maps
in
TCMA
were produced using an object-based classifier and
a decision tree method. The
LULC
map had lower overall ac-
curacy compared to the impervious surface map due to the
higher number of classes, likelihood of confusion between
cropland and other vegetation classes, and in part might be
attributed to the methodology used. While the object-based
Feature Analyst appears to make use of advanced techniques,
the decision tree classifier See5 is more transparent. However,
both techniques can be used to extract urban impervious
surfaces over large areas from the
NAIP
imagery with relatively
high accuracy. Comparatively, the decision tree method is
more efficient in processing large data sets and more flexible
for handling multi-source input data.
In conclusion, extracting
LULC
and impervious surface
information over large areas from high-resolution remote
sensing data is challenging but feasible. From an application
standpoint, this allows for more accurate environmental man-
agement at a localized scale, with storm water runoff and is-
sues pertaining to habitat fragmentation or patches of specific
vegetative assemblages being chief examples. Methodologi-
cally, this study provides an example of how freely available
high-resolution
NAIP
imagery can be used to produce high-res-
olution land cover data. In the future, it would be beneficial to
assess how well the decision tree technique would be able to
classify the general
LULC
classes from high-resolution airborne
or satellite images over large areas. In addition, generating ad-
ditional derivative from the lidar data such as a three-dimen-
sion canopy layer or a
DSM
may be helpful for classifying both
high-resolution
LULC
and impervious surfaces, even though it
will take an enormous amount of data processing resources.
Acknowledgments
This study was funded by a Faculty Research Grant of the
Minnesota State University, Mankato. The authors would
like to extend our sincere appreciation to the editor and three
anonymous reviewers for their constructive suggestions and
comments.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
January 2016
69
