4
On the other hand it could be assumed that a SVM performs
better, because it achieves the highest results when classifying
single-sensor data. This assumption is confirmed by the
experimental results: The separate training of the SVM and the
fusion by another SVM further increases the classification
accuracy up to 76.5%.
Classifier
algorithm
SAR
TM
SAR+TM
MLC
55.5
66.0
70.1
DT
46.0
63.5
64.0
DT boost
61.5
67.7
75.6
SVM
62.7
68.2
74.2
SVM Fusion
-
-
76.5
Table 2. Overall accuracies [%], using different classifier
algorithms and data sets
The achieved overall accuracies clearly illustrate the value of
using multisensor imagery for classification. Comparing the
multisensor classification by SVM, the accuracy is increased by
up to 6% compared to the results of the Landsat image and up
to 11.5% compared to the classification accuracy achieved with
the SAR imagery. The positive impact of multisensor imagery
on the classifier performance is also confirmed by the class-
specific accuracies (Table 3 and Table 4).
Land cover
Class
Input data
SAR
TM
Arable crops
51.2
56.2
Cereals
73.4
75.0
Forest
74.6
89.4
Grassland
71.0
56.8
Orchards
54.8
47.6
Rapseed
58.2
73.4
Root crops
58.2
66.0
Urban
60.2
81.4
Table 3. Class-specific accuracies, using single-source imagery,
bold numbers indicates best results
Land cover
Class
Classifier algorithm
DT boost. SVM Fusion
Arable crops
70.6
70.8
Cereals
77.8
76.6
Forest
93.2
91.2
Grassland
74.2
72
Orchards
63.6
68.6
Rapseed
76.0
75.2
Root crops
66.4
73.8
Urban
83.0
84
Table 4. Class-specific accuracies, using a boosted DT and
SVM Fusion, bold numbers indicates best results
On one hand the single-source results show that the two data
sources are differently reliable. Whereas grassland and
orchards are more accurately classified using the SAR data, the
other classes are better described by the Landsat image.
Moreover all class-specific accuracies are improved by the use
of multisensor imagery. The accuracies of classes that achieve
relatively low values on both image sources (e.g., arable crops,
orchards) are significantly increased by the multisensor
approach.
The comparison of the class-specific accuracies by a boosted
DT and the proposed fusion strategy demonstrate that the
results from the SVM Fusion are more balanced (Table 4).
Whereas the accuracies vary between 63.6% and 93.2% for the
boosted DT, the accuracies achieved by the SVM Fusion vary
between 68.6% and 91.2%.
The visual assessment of the classification results underlines the
positive effect of multisensor fusion and the complementary
character of the multitemporal SAR and the multispectral
optical data (Figure 2): The resulting maps seem more
homogeneous in some regions and errors for critical classes are
decreased. Contrary to the visibly good performance of the
multisensor fusion, the disappearance of the highway in the
middle of the map subset, which was recognized in the TM data,
shows possible drawbacks of the multisensor approach.
Figure 2. Classification results, using different data sets
6. CONCLUSION
The challenge of classifying multisensor data sets, containing
multitemporal SAR and multispectral imagery has been
addressed. In experiments, the separate training and further
fusion by a SVM outperformed all other parametric and
nonparametric classification techniques including a
conventional SVM.
We assume the
concept to be
different imagei
equally reliable
more adequate
source individi
function on the
by the first S
transformation,
space that cons
rule images. T!
original imager
to describe the i
Moreover it h<
classifiers in tei
data sets. Henc
proposed classi
classifier perfc
segmentation a
would be wort!
increased multii
Generally, the r
multitemporal
worthwhile. He
increased by si
applied classify
approaches for
ongoing researc
reference:
Benediktsson, J
of Multisource
Fusion. IEEE I
pp. 1367-1377.
Benediktsson, J
Network Apj
Classification i
Trans. Geoscie)
Blaes, X., Vaa
crop identificat
Remote Sensing
Briem, G.J., E
Multiple Class
Data. IEEE Trc
2291-2299
Brown de Cc
Commisso, K.,
vegetation map
decision tree cl
pp. 316-327.
Camps-Vails,
Frances, J., Q
hyperspectral
Remote Sensing
Chen, C.-C. £
support vec
http://www.csii
