The overall accuracy of the general
LULC
classification was 74
percent, with a Kappa statistic of 0.66 (Table 1). Water, urban,
and bare soil were classified with relatively higher accuracies
due to their unique spectral and spatial attributes. However,
other vegetation class had the lowest user’s accuracy of 63 per-
cent, whereas cropland had the lowest producer’s accuracy at
57 percent, due to the similarity of these two classes. Other veg-
etation class was also somewhat confused with forest. This may
explain the relatively low overall accuracy of the classification.
The overall accuracy of the impervious surface classifica-
tion was 95 percent, with a Kappa statistic of 0.79 (Table 2).
The user’s accuracy of 90 percent and producer’s accuracy of
76 percent for impervious surface indicate that the omission
error was 14 percentage points higher than the commission
error. Accuracy of impervious surface estimation has seasonal
variation (Wu and Yuan, 2007). The
NAIP
images were flown
during the peak growing season. The tree canopies could
cover part of impervious features in the summer.
Compared with Two Landsat-based Classifications in TCMA
A previous
LULC
classification in the
TCMA
was conducted
by Yuan (2010) using 2006 Landsat imagery, which includes
six classes: urban, forest, cropland, wetland, grassland, and
water. An overall accuracy of 85.7 percent was obtained based
on 300 randomly sampled pixels. This study tried to match
the classification scheme and accuracy assessment method
used by Yuan (2010) as closely as possible. In addition, for
the purpose of comparison, the most recent 2011 edition of
2006
NLCD
was downloaded
). It was based on a
decision tree classification of Landsat imagery with an overall
accuracy of 78 percent for the 16 Level II classes (Wickham
et al
., 2013). To compare it with the other two classifications,
some of the
NLCD
Level II classes were combined.
Plate 1b provides a comparison of the three clas-
sifications at the representative scale of 1 to 100 000.
The 1 m
NAIP
-based classification offers much more
spatially detailed land feature information than the
two Landsat-based classifications (Plate 1b). There are
also big discrepancies in the area statistics between the
NAIP
-based classification and the other two classifica-
tions, excluding water class (Table 3). The classified urban area
of 867 km
2
in this study was much smaller than the urban esti-
mates of 2006
NLCD
(2,462 km
2
) and the 2006 Landsat classifica-
tion (2,798 km
2
). Due to differences in urban class definition,
some features that were not considered as urban such as yards,
lawns, urban trees, public parks, and other developed open
space in the
NAIP
-based classification were included in the ur-
ban/developed classes of the other two classifications (Plate 1b).
The
TCMA
has a high amount of urban features, many of which
are smaller than a 30 m Landsat pixel. Hence, mixed-pixel
problem has a higher impact in Landsat images than in the
NAIP
images. This study also estimated much higher other vegeta-
tion and forest covers, while the opposite was true for crop-
land. Mixed pixel problem along with actual differences to the
definitions of the classes could again contribute to the discrep-
ancies. For example, shrub was included in the forest class in
the 2006 classification by Yuan (2010) whereas it was combined
with other vegetation in this study. Misclassifications caused
by confusions among the other vegetation, cropland, and forest
classes could be another reason. Further, large contiguous areas
of vegetation, such as cropland, can be better captured by Land-
sat imagery because of the repeat coverage and the additional
spectral information the imagery provides whereas some small
patches of forest and other vegetation in urban areas may be
underestimated in the Landsat classifications.
Discussion
Accuracies of the Classifications and Comparison of the Two Classifiers
The accuracies for cropland and other vegetation classes
in the
LULC
classification map were relatively low due to
the spectral confusion between the two classes. This result
T
able
1. E
rror
M
atrix
of
G
eneral
LULC C
lassification
Classified Data
Reference Data
Water Forest
Crop-
land
Other
Veg. Urban
Bare
Soil
Grand
Total
User’s
Accuracy
Water
16 0
0
0
0 0 16 100%
Forest
1 56 4 11 2 0 74 76%
Cropland 0
1 40 8
2 0 51 78%
Other veg.
0 15 26 72 1 1 115 63%
Urban
1
1
0
2 33 0 37 89%
Bare Soil
0
0
0
0
2 5 7
71%
Grand
Total
18 73 70 93 40 6 300
Producer’s
Accuracy
89% 77% 57% 77% 83% 83%
Overall Accuracy = 74%; Kappa statistic = 0.66
T
able
2. E
rror
M
atrix
of
I
mpervious
S
urface
C
lassification
Classified data
Reference Data
Pervious Impervious Total
User’s
Accuracy
Pervious
250
11
261 96%
Impervious
4
35
39 90%
Total
254
46
300
Producer’s
Accuracy
98% 76%
Overall Accuracy = 95%; Kappa statistic = 0.79
T
able
3. A
rea
S
tatistics
of
the
T
hree
LULC C
lassifications
2010 NAIP Classification
2006 NLCD
2006 Landsat Classification (Yuan, 2010)
Class
km
2
Class
km
2
Class
km
2
Water
439
Water
452
Water
426
Urban
867 Developed low, medium, and high intensity + developed open space 2461
Urban
2798
Forest
1999
Deciduous, Evergreen, and Mixed Forest
932
Forest
980
Agriculture 1366
Pasture/Hay and Cultivated Crops
3011 Agriculture
2965
Other Vegetation 2943
Shrub/Scrub and Grassland/Herbaceous
313
Grass
57
Wetland
N/A
Woody Wetlands and Emergent Herbaceous Wetlands
520
Wetland
479
Barren Land 90
Barren Land
13
Barren Land
N/A
68
January 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
