194
Recent remote sensing literature has shown that SVM methods
generally outperform traditional statistical and neural methods
in classification.
2.3 The Processing of Object-oriented Classification based
on MRF and SVM
In this section, a new procedure of object-oriented image
classification based on MRF and SVM is presented. It
synthesizes image preprocessing (e.g., information fusion,
image segmentation), geographical information system (vector-
based feature selection) and data mining (intelligent SVM
classification) technology to interpret image from pixels to
segments and then to thematic information, and it is an
integrative, iterative classification system. The flowchart of the
object-oriented classification algorithm is shown in figure 2,
and there are three main processes: the MRF segmentation
progress, the object selection process, and the object-oriented
SVM classification progress.
(3) Object-oriented Classitkation j
VCH
Feature extraction ~|
HR MS
3Z
HR PAN
HR fused image
j
[ (1) segmentation ^4-
Build
Fig.2. The flowchart of object-oriented classification based on
MRF and SVM.
(1) In the segmentation progress, the multi-spectral (MS) image
and panchromatic (PAN) are co-registered and fused before
segmentation. The fused image is segmented iteratively by the
MRF segmentation method until getting the best segmentation
results. Through the segmentation process, the image is
subdivided into a huge number of homogeneous elementary
objects.
(2) In the object selection progress, the features (e.g., spectral,
spatial, and contextual) of sample objects from different classes
are calculated. Typical features of some objects act as the
training samples of SVM, and the whole objects act as the
testing samples.
(3) In the object-oriented SVM classification progress, firstly,
the training samples and the testing samples are scaled. The
main purpose of this step is to avoid the domination of greater
numeric ranges on smaller numeric ranges. Another purpose is
to avoid numerical difficulties during the calculation. Secondly,
the parameters are set. In this step, the training samples are
mapped to the higher dimensional feature space to find the
optimum space of various classes using SVM, thus the support
vectors and VC are gained reliably, and the decision function
and training model are formed. The testing samples acting as
the input of decision function are tested to get the classification
map. At last, the misclassified objects detected by the proposed
classification method are classified again. This step is iterative
until getting the best classification result. For the
implementation of SVM, the software package LIBSVM by
Chang and Lin (2001) was adopted.
3. CLASSIFICATION EXPERIMENT
3.1 Study Area
The experimental data are panchromatic and multi-spectral
QuickBird data at 0.61-m spatial resolution, which was
acquired in May 2005 in HeFei Province of China. The data
were geo-referenced to UTM projection, and then fused by the
Smoothing Filter-based Intensity Modulation (SFIM) method
using Imagelnfo remote sensing imagery processing software
(www.image-info.com) developed by Chinese Academy of
Surveying and Mapping. Figure 4 shows the panchromatic
QuickBird imagery, and Figure 5 shows the multi-spectral
QuickBird composite imagery of bands 4, 2, and 1. Figure 6
shows the fused image compositing from the same bands as
figure 5. Six land cover classes were identified: 1 water bodies,
2 road, 3 trail, 4 shrub and grassland, 5 agriculture, and 6
building.
3.2 Classification Process
Before segmentation, the textural features (variance, contrast)
of the fused imagery were calculated using Gray Level Co
occurrence Matrix (GLCM) filtering, and then the segmentation
was achieved by the MRF segmentation method, which uses a
finite number of parameters to characterize spatial interactions
of pixels to describe an image region. The image textures can
be viewed as realizations of samples from a parametric
probability in the image space. In our case, we chose ICM
method with beta=0.9, t=0.05, T0=4, and c=0.98, where beta is
the strength of second order clique potential, t is the stop
threshold, TO is initial temperature , and c is temperature
scheduler's factor. Figure 5 shows the segmented fusion
imagery.
After the segmentation, we selected several objects from each
class as the training samples, and defined the whole objects as
the testing samples.
During the classification processing, first, we scaled the training
samples and the testing samples before applying SVM.
Secondly, we chose RBF kernel, which is more suitable to land
cover classification. Thirdly, we used cross-validation and grid-
search method to get the best parameters of RBF. The highest
overall accuracy on the multi-source image information was
obtained with g=0.125 and C=1024, where g is the width of the
kernel function, and C is the regularization parameter. At last,
the training samples were used to train the support vector
machine, and the resulted model was used to classify the entire
image with the OAO method to get the preliminary
classification result which is perhaps not the best one. So we
selected samples and classified sequentially for various classes
by analyzing the classification errors of omission and
commission until getting the best result. The final classification
map is shown in figure 6. For the purpose of getting the
preferable classification result, it is better to consider small
misclassified objects.
