as the following
(2)
need to deal with
. The constant C is
ated on the wrong
the shape of the
rectly affects the
ibove linear SVM
eparable cases. For
tied into a higher
ir mapping 0. The
lewly transformed
but only the kernel
(3)
(4)
n is used (Vapnik,
(5)
des the distance of
/ing a rule image,
ition result,
y classifier. Hence
solve multi-class
\A) approach, a set
iss from the others,
etc. The maximum
¡termines the final
(OAO) strategy, a
one for each pair-
urban. As done in
assifier provides a
ich, each pixel is
•er of votes.
SAR data multispectral image
7^7 X 1 X
fmal
land cover classification
Figure 1. Schematic diagram of the of the fusion strategy
3. DATA SET AND PREPROCESSING
The multisource data for the experiments contains a set of
multitemporal SAR data and a Landsat 5 TM image from May
2005. The SAR data includes Envisat ASAR alternating
polarization and ERS-2 imagery from different swaths and
polarizations (Table 1 and Figure 2).
Sensor
Image characteristics SAR data
Date
Swath
Polariz.
Orbit
ASAR
12-Apr-05
IS 6
HH/HV
Asc
ERS-2
21 -Apr-05
-
VV
Des
ERS-2
26-May-05
-
W
Des
ERS-2
30-Jun-05
-
VV
Des
ASAR
13-M-05
IS3
HH/HV
Asc
ASAR
22-Jul-05
IS 7
HH/HV
Asc
ERS-2
4-Aug-05
-
VV
Des
ASAR
14-Aug-05
IS 2
HH/HV
Asc
Table 1. Multitemporal SAR data
4. METHODS
In the experiment eight land cover classes were considered:
Arable crops, Cereals, Forest, Grassland, Orchard, Rape seed,
Root crops and Urban. One hundred samples per class were
selected from the ground mapping by equalized random
sampling, guaranteeing that all 8 classes are included in the
sample set. These training set was employed for both stages of
the classification strategy, the pre-classification and the fmal
fusion approach. Again using equalized random sampling an
independent validation set with 500 samples per class was
generated.
The SVMs were applied separately on the SAR data and the
multispectral image, following the OAO strategy. The OAO
concept was preferred, because first tests showed differences in
terms of accuracy after the data fusion. The training of the SVM
with a Gaussian kernel and the generation of the rule images
were performed using imageSVM (Janz et al., 2007), which is a
freely available IDL/ENVI extension that employs LIBSVM by
Chen and Lin (Chen and Lin, 2001) for the training of the SVM.
The best parameters for y and C were selected from a user
defined range of possible parameters based on 10-fold cross
validation.
The SVM fusion was performed applying imageSVM with OAO
strategy and majority vote to the first sets of rule images. In
addition to the SVM classifier, three different methods were
applied to the imagery. A maximum likelihood classifier (MLC),
a decision tree (DT) and a boosted decision tree (DT boost).
The DT is based on the algorithm C4.5 (Quinlan, 1993).
Boosting is a well-known method for generating a classifier
ensemble, which is based on adaptively changing the
distribution of the training samples during the classifier training.
In the initial phase, all samples are equally weighted.
Afterwards the weights of the samples are modified and
misclassified samples assigned a higher weight than those
classified accurately. The next DT in the classifier system is
trained on the newly distributed samples. In our experiments,
AdaBoost.Ml (Freund and Schapire, 1996) is used.
t sensor sources is
; a set of SVMs is
rule images are
mages is fused by
containing all rule
Figure 2. Subsets of SAR (Apr-2 l/May-26/Jun-30) and
Landsat 5 TM data
The nearly flat study site is located near Bonn, Germany. The
region is dominantly used for agriculture and characterized by
typical spatial patterns caused by differences in the phenology
of planted crops. An extensive ground truth campaign was
conducted in summer 2005. The ground truth information is
used as reference for the generation of training and validation
data sets.
5. RESULTS
The accuracy assessment (Table 2) shows that the DT achieves
the lowest classification accuracies, irrespective of the image
type. But a boosted DT outperforms the MLC in terms of the
overall accuracy. The SVM applied on both data sources further
improves the single-source results, and outperforms the other
classifier algorithms. Nevertheless the boosted DT performs
better in terms of accuracy when classifying the multisensor
imagery.