and the denoised image, which obtained metric Q values of
0.0020 and 0.0023, respectively.
Enhanced Discriminative Feature Description near Missing Regions
To validate the function of the
VIS
image gap inpainting for
the classification, five single spatial features and the multiple-
feature
VS
approach were utilized, respectively. Table 7 re-
cords the classification accuracy results, and the quantitative
assessments for each class are shown in Figure 8.
T
able
7. C
lassification
A
ccuracy
on
the
E
dge
S
ub
-P
art
of
the
S
econd
T
est
S
et
for
E
ach
S
patial
F
eature
R
elated
A
pproach
for
the
R
emaining
C
lass
C
lassification
of
the
S
tudy
A
rea
, B
efore
and
A
fter
I
mage
I
npainting
Index Area SD DB MI
GLCM
VS
With
inpainting
OA 0.9096 0.9092 0.9103 0.8960 0.9127
0.9350
AA 0.8819 0.8628 0.8856 0.8430 0.9322
0.9448
κ
0.8790 0.8778 0.8800 0.8604 0.8843
0.9133
Without
inpainting
OA 0.8935 0.8388 0.8950 0.8951 0.9080 0.9274
AA 0.8721 0.8138 0.8807 0.8321 0.9293 0.9365
κ
0.8572 0.7881 0.8596 0.8590 0.8781 0.9031
In Table 7, it can be seen that in all the comparison pairs, es-
pecially the
SD
pair, the image gap inpainting can improve the
classification accuracy. In Figure 8, the classification omis-
sion, agreement, and commission (Li
et al.
, 2014) are shown
as the sub-bars for each category in a group, and the detailed
number in each sub-bar denotes the associated proportion.
Overall, the classification performance of each category has
been improved by the image gap inpainting, as shown in Fig-
ure 8. Specifically, comparing
VIS-VS
with In-
VIS-VS
, it can be
seen that the omission decreases for the vegetation, concrete
roof, and red roof classes, and the agreement for tree pixels
increases. For the commission, similar observations can be
made, which further confirm the effectiveness of the simple
inpainting step.
Discriminative Ability Analysis for each Class
To illustrate the discriminability between the land-use and
land-cover types, the classification omission, agreement, and
commission for the global test set are shown as the sub-bars
for each class in a group in Figure 9. Here, it can be first seen
that the road discriminability of the
TI-HSI
data is superior to
that of the
VIS
data, although the spatial features are repre-
sented in the
VIS
data. It should be mentioned that the pro-
posed
OOIFA
approach can achieve the goal of road extraction,
and can also alleviate the omission error of the other classes.
The remaining classes consist of bare soil, vegetation, trees,
and buildings (i.e., red roof, concrete roof, and gray roof).
For these classes, it is believed that their spatial description
is more useful. For the first three class types in Figure 9 (i.e.,
bare soil, vegetation, and concrete roof), it is suggested that
the contextual feature is superior, as each of these classes has
a specific contextual pattern, which can be seen in Figure 1a.
It can also be observed that the
VS
approach can maintain the
superiority of the contextual description, to conduct the clas-
sification task. For the three building (roof) classes in Figure
9, it can be seen that a single feature cannot describe the dis-
criminative characteristic satisfactorily, and the
VS
approach
utilizes the complementary spatial descriptions to improve
the classification accuracy.
Conclusions
The study of the newly released multi-sensor, multi-spectral-
spatial resolution, and multi-swath width
TI-HSI
and
VIS
datasets is a challenging topic, and this paper presents a
multi-level fusion approach for discriminative information
mining to achieve urban land-use and land-cover classifica-
tion. The specific superiorities of the
TI-HSI
and
VIS
datasets
are integrated to improve the classification accuracy, which
was confirmed by a quantitative assessment. In particular, a
novel image gap inpainting method for the
VIS
data with the
guidance of the
TI-HSI
data is applied to deal with the swath
width inconsistency and facilitate accurate spatial feature
extraction, thereby improving the overall classification ac-
curacy. In summary, it is suggested that utilizing the
TI-HSI
data together with the
VIS
data for urban surface exploitation
is both promising and meaningful.
Acknowledgments
The authors would like to thank Telops, Inc. (Canada) for
providing the data used in this study, and the
IEEE GRSS
Image
Analysis and Data Fusion Technical Committee for organizing
2014 Data Fusion Contest. Thanks are also due to the handling
editor and the anonymous reviewers for their insightful and
constructive comments. The authors would also like to thank
the supporting from the National Basic Research Program of
China (973 Program) under Grant 2011CB707105, the National
Natural Science Foundation of China under Grants 41571362
and 41431175, and the Key Laboratory of Agri-informatics,
Ministry of Agriculture, P.R.China..
References
Boser, B.E., I.M. Guyon, and V.N. Vapnik, 1992. A training algorithm
for optimal margin classifiers,
Proceedings of the Fifth Annual
Workshop on Computational Learning Theory
, 1992, pp.
144–152.
Comaniciu, D., and P. Meer, 2002. Mean shift: A robust approach
toward feature space analysis,
IEEE Transactions on Pattern
Analysis and Machine Intelligence
, 24(5):603–619.
Congalton, R., 1991. A review of assessing the accuracy of
classifications of remotely sensed data,
Remote Sensing of
Environment
, 37(1):35-46.
Dalla Mura, M., J.A. Benediktsson, B. Waske, and L. Bruzzone, 2010.
Morphological attribute profiles for the analysis of very high
resolution images,
IEEE Transactions on Geoscience and Remote
Sensing
, 48(10):3747–3762.
Fan, R.-E., K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin, 2008.
LIBLINEAR: A library for large linear classification,
The Journal
of Machine Learning Research
, 9:1871–1874.
Foody, G.M., 2004. Thematic map comparison: Evaluating the
statistical significance of differences in classification accuracy,
Photogrammetric Engineering & Remote Sensing
, 70(5):627–634.
Ghamisi, P., M. Dalla Mura, and J.A. Benediktsson, 2015. A survey
on spectral-spatial classification techniques based on attribute
profiles,
IEEE Transactions on Geoscience and Remote Sensing
,
53(5):2335–2353.
Guo, M., H. Zhang, J. Li, L. Zhang, and H. Shen, 2014. An online
coupled dictionary learning approach for remote sensing
image fusion,
IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing
, 7(4):1284–1294.
Hay, G.J., and T. Blaschke, 2010. Special issue: Geographic object-
based image analysis (GEOBIA),
Photogrammetric Engineering &
Remote Sensing
, 76(2):121–122.
Huang, X., and L. Zhang, 2011. A multilevel decision fusion
approach for urban mapping using very high-resolution multi/
hyperspectral imagery,
International Journal of Remote Sensing
,
33(11):3354–3372.
Huang, X., and L. Zhang, 2013. An SVM ensemble approach
combining spectral, structural, and semantic features for the
classification of high-resolution remotely sensed imagery,
IEEE
Transactions on Geoscience and Remote Sensing
, 51(1):257–272.
910
December 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
